Condition-Specific Readmission Risk Stratification in a Predominantly Black Statewide Cohort Using Machine Learning: Development of Subtype-Specific Models for Heart Failure, Acute Myocardial Infarction, Atrial Fibrillation/Flutter, and Hypertensive Heart Disease

Condition-Specific Readmission Risk Stratification in a Predominantly Black Statewide Cohort Using Machine Learning: Development of Subtype-Specific Models for Heart Failure, Acute Myocardial Infarction, Atrial Fibrillation/Flutter, and Hypertensive Heart Disease | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Condition-Specific Readmission Risk Stratification in a Predominantly Black Statewide Cohort Using Machine Learning: Development of Subtype-Specific Models for Heart Failure, Acute Myocardial Infarction, Atrial Fibrillation/Flutter, and Hypertensive Heart Disease Ismail El Moudden, Michael Bittner, Sunita Dodani This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9098008/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 8 You are reading this latest preprint version Abstract Cardiovascular disease (CVD) readmissions impose substantial clinical and economic burden. Machine learning (ML) may improve risk stratification, yet most predictive models aggregate CVD subtypes into a single outcome and underrepresent Black populations. Using Virginia Health Information database records (2010 to 2020), we analyzed 157,791 discharge records from 123,272 unique patients (96.6% Black) to develop condition-specific 30-day readmission models for heart failure (HF; n = 91,752), acute myocardial infarction (AMI; n = 34,497), atrial fibrillation/flutter (AF/AFL; n = 18,424), and hypertensive heart disease (HHD; n = 13,118). Four algorithms (XGBoost, LightGBM, Random Forest, Elastic Net) plus a Super Learner ensemble were trained on patient-grouped 70/30 splits with and without Synthetic Minority Oversampling Technique balancing. Models incorporated validated clinical indices (LACE, Charlson, Elixhauser) and administrative social determinants of health proxies. The overall 30-day readmission rate was 18.9%. Best area under the receiver operating characteristic curve (AUC) values by condition were HF 0.708 (95% CI, 0.701 to 0.716), AMI 0.706 (95% CI, 0.691 to 0.721), AF/AFL 0.732 (95% CI, 0.715 to 0.750), and HHD 0.758 (95% CI, 0.735 to 0.777). XGBoost was the top-performing algorithm for three of four subtypes. The LACE Index, Charlson Comorbidity Index, and insurance type were consistently the strongest predictors. Algorithm-native, aggregated, and SHAP-based importance measures converged on these key features. In this largest-to-date, predominantly Black statewide cohort, condition-specific ML models achieved moderate-to-high discrimination for HF, AMI, AF/AFL, and HHD. Key clinical indices and administrative social determinants proxies emerged as dominant predictors, highlighting modifiable targets and high-risk subgroups. These findings support the development of precision, equity-informed readmission interventions and provide a scalable framework for deploying ML-driven decision support in safety-net and minority-serving healthcare systems. Health sciences/Cardiology Health sciences/Diseases Health sciences/Health care Health sciences/Medical research Health sciences/Risk factors cardiovascular diseases heart failure machine learning patient readmission risk assessment healthcare disparities African Americans social determinants of health Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Cardiovascular disease (CVD) remains the leading cause of mortality in the United States, accounting for approximately 931,578 deaths annually and $ 252 billion in direct medical costs. 1,2 Heart failure (HF), acute myocardial infarction (AMI), atrial fibrillation and flutter (AF/AFL), and hypertensive heart disease (HHD) are the principal drivers of cardiovascular hospitalization and 30-day readmission. 1–3 Reported 30-day readmission rates range from 18% to 25% for HF, 3–5 12% to 17% for AMI, 6,7 and 10% to 18% for AF/AFL, 8,9 while HHD drives acute care utilization with disproportionately elevated rates in minority populations. 2,10 Centers for Medicare & Medicaid Services (CMS) value-based penalty programs, including the Hospital Readmissions Reduction Program (HRRP) and the Hospital-Acquired Condition Reduction Program (HACRP), have produced modest improvements in quality metrics, yet safety-net hospitals serving predominantly Black and low-income patients bear a disproportionate share of financial penalties. 11–13 Racial differences in AMI readmission persisted after HRRP and were attributable to patient-level rather than hospital-level factors. 12 Black patients hospitalized for HF have 3.9% to 6.8% higher composite readmission and mortality rates than White patients across deprivation strata, even after covariate adjustment. 14 Historical structural factors such as residential redlining compound these effects at the population level rather than at individual hospitals, pointing to the need for upstream interventions. 15,16 Machine learning (ML) handles high-dimensional data and non-linear relationships more flexibly than logistic regression; AUC values from 0.51 to 0.93 have been reported across heart failure readmission studies. 17–20 Most of these studies lacked external validation, rarely assessed calibration, and relied on racially heterogeneous populations unlikely to generalize to racially concentrated cohorts. 20 Integrating social determinants of health (SDOH) into ML models has improved prediction in other cardiovascular contexts (AUC 0.694 to 0.823 for stroke), but application to readmission in predominantly minority populations has not been studied. 21,22 Despite this growing body of work, ML models for CVD readmission have rarely been developed in predominantly minority populations, 14,16,19,20 head-to-head algorithm comparisons across distinct cardiovascular conditions using standardized feature sets remain sparse, 5,19,20 and no prior study has integrated validated clinical indices with administrative SDOH proxy data into condition-specific models within racially concentrated cohorts. 14,21,22 We compared four ML algorithms (XGBoost, LightGBM, Random Forest, Elastic Net) and a Super Learner stacked ensemble across four conditions (HF, AMI, AF/AFL, HHD) in 157,791 discharge records from 123,272 unique patients (96.6% Black/African American) using an algorithm benchmarking design (4 base algorithms × 4 conditions × 2 dataset configurations plus 8 ensemble models, 40 total; TRIPOD Type 1a: development) 23 with validated clinical indices and consensus-selected features. CVD conditions were classified using AHA-aligned ICD-9/ICD-10 hierarchies. The study design follows TRIPOD guidelines for model development studies (Type 1a), 23 with the crossed design extending the framework to support simultaneous algorithm comparison within each condition. Similar crossed designs have been used in cardiovascular ML benchmarking. 5,18–20 This study was conducted to evaluate the performance of machine learning models across different cardiovascular disease (CVD) subtypes in a minority population and to identify the relative importance of validated clinical indices and administrative social determinants of health (SDOH) proxies, particularly insurance status and comorbidity burden, on condition-specific risk prediction. We hypothesize that (1) ML model performance will differ across CVD subtypes, with condition-specific risk architectures requiring tailored modeling, and (2) validated clinical indices and administrative SDOH proxies particularly insurance status and comorbidity burden, will be significant predictors, with their relative importance varying by CVD subtype in this minority population. HHD was included as a standalone category because it disproportionately affects Black populations, who constitute 96.6% of our cohort, and is a leading contributor to heart failure progression, yet it has been largely absent from the readmission prediction literature as a separately modeled condition. Methods Study Design and Data Source. We conducted a retrospective cohort study using de-identified inpatient discharge records from the Virginia Health Information (VHI) statewide all-payer database (January 2010 through December 2020). 24 VHI captures demographic, clinical, financial, and administrative data from all acute-care hospitalizations in Virginia (approximately 9.24 million total discharges). The study objective was to develop and compare ML models for 30-day all-cause readmission across four CVD subtypes in a predominantly minority population. Cohort flow is illustrated in Fig. 1 . CVD Identification. CVD patients were identified from the primary diagnosis field using AHA guideline-aligned hierarchical classification. Four categories were defined: HF, AMI, AF/AFL, and HHD. Dual-era diagnostic codes (ICD-9 and ICD-10) were mapped to each category; complete code lists are available in Supplementary Table S1 . Exclusion Criteria. We excluded records with non-CVD primary diagnoses, race/ethnicity other than Black/African American or Hispanic/Latino (per the grant’s minority health focus), age 89 years, planned readmissions, in-hospital death, or missing readmission outcome. After exclusions, 157,791 discharge records from 123,272 unique patients comprised the analytic cohort; no records were excluded for missing patient identifiers. Baseline characteristics are presented in Table 1 . The analytic cohort was 96.6% Black/African American and 3.4% Hispanic/Latino, consistent with the grant’s minority health focus. Given the small Hispanic subgroup, no separate models were estimated for this population; however, model performance was confirmed to be comparable when restricted to Black patients alone (results available upon request). Table 1 Baseline Demographics and Clinical Characteristics by Readmission Status (N = 157,791 Discharge Records, 123,272 Unique patients, 1.28 Discharge-to-patient ratio, 18.9% Readmission rate) Characteristic Overall (N = 157,791) No Readmission (n = 127,913) Readmission (n = 29,878) SMD Demographics and Encounter Characteristics Age, years, mean (SD) 62.75 (14.19) 62.55 (14.17) 63.60 (14.23) 0.074 Female sex, % 49.0 49.0 51.0 0.036 Length of stay, days, mean (SD) 4.99 (6.01) 4.81 (5.98) 5.73 (6.08) 0.151 Total diagnoses, mean (SD) 14.41 (4.14) 14.13 (4.24) 15.60 (3.44) 0.381 Total procedures, mean (SD) 1.72 (1.96) 1.76 (1.99) 1.56 (1.85) 0.100 Total charges ( $ ), mean (SD) 47,910 (88,915) 47,727 (88,549) 48,692 (90,459) 0.011 Admission Order, mean (SD) 1.49 (1.59) 1.31 (1.04) 2.28 (2.82) 0.456 N Admissions per Patient, mean (SD) 1.99 (2.63) 1.60 (1.68) 3.65 (4.59) 0.592 Validated Clinical Indices LACE Index (0–19), mean (SD) 9.67 (3.02) 9.34 (2.96) 11.09 (2.86) 0.604 Charlson Comorbidity Index, mean (SD) 3.15 (2.05) 2.99 (2.04) 3.81 (1.99) 0.405 Age-Adjusted Charlson, mean (SD) 5.04 (2.65) 4.87 (2.65) 5.78 (2.48) 0.357 van Walraven Elixhauser Index, mean (SD) 11.99 (8.74) 11.34 (8.70) 14.74 (8.37) 0.398 Elixhauser Comorbidity Count, mean (SD) 4.63 (2.00) 4.49 (2.00) 5.27 (1.85) 0.410 CVD Severity Score, mean (SD) 3.49 (0.81) 3.45 (0.83) 3.67 (0.65) 0.294 CVD Mortality Risk, mean (SD) 2.65 (0.90) 2.62 (0.92) 2.79 (0.83) 0.199 Total CVD Conditions, mean (SD) 4.57 (1.94) 4.50 (1.94) 4.88 (1.91) 0.195 LACE Risk Category, n (%) 0.502 Low 6,188 (3.9) 5,827 (4.6) 361 (1.2) Moderate 67,567 (42.8) 59,606 (46.6) 7,961 (26.6) High 84,036 (53.3) 62,480 (48.8) 21,556 (72.1) Race/Ethnicity, n (%) 0.066 Black 152,492 (96.6) 123,344 (96.4) 29,148 (97.6) Hispanic 5,299 (3.4) 4,569 (3.6) 730 (2.4) Insurance Type, n (%) 0.342 Commercial 33,341 (21.1) 29,266 (22.9) 4,075 (13.6) Medicaid 16,868 (10.7) 12,881 (10.1) 3,987 (13.3) Medicare 92,250 (58.5) 72,024 (56.3) 20,226 (67.7) Self-Pay/Uninsured 15,332 (9.7) 13,742 (10.7) 1,590 (5.3) Primary CVD Category, n (%) 0.302 Acute Myocardial Infarction 30,126 (19.1) 26,207 (20.5) 3,919 (13.1) Atrial Fibrillation/Flutter 18,782 (11.9) 15,902 (12.4) 2,880 (9.6) Heart Failure 94,562 (59.9) 73,253 (57.3) 21,309 (71.3) Hypertensive Heart Disease 14,321 (9.1) 12,551 (9.8) 1,770 (5.9) Teaching Hospital Status, n (%) 0.083 ACGME 22,062 (14.0) 18,105 (14.2) 3,957 (13.3) Council for GME 13,626 (8.6) 11,029 (8.6) 2,597 (8.7) Council of Teaching 9,082 (5.8) 7,006 (5.5) 2,076 (7.0) Council of Teaching Hospitals 19,534 (12.4) 15,447 (12.1) 4,087 (13.7) None 93,325 (59.2) 76,190 (59.6) 17,135 (57.4) Hospital Size (beds), n (%) 0.069 Small 17,088 (10.8) 13,823 (10.8) 3,265 (10.9) Medium 60,458 (38.3) 49,587 (38.8) 10,871 (36.4) Large 47,702 (30.2) 38,757 (30.3) 8,945 (29.9) Major 32,543 (20.6) 25,746 (20.1) 6,797 (22.7) Hospital Ownership, n (%) 0.028 Not-for-Profit 125,919 (79.8) 102,345 (80.0) 23,574 (78.9) Proprietary 31,871 (20.2) 25,567 (20.0) 6,304 (21.1) Elixhauser Comorbidities (prevalence), % Congestive Heart Failure 68 66 78 0.268 Renal Failure 45 42 59 0.325 Cardiac Arrhythmia 37 36 42 0.132 Chronic Pulmonary Disease 35 33 43 0.199 Fluid & Electrolyte Disorder 32 30 36 0.120 Hypertension, Complicated 31 29 35 0.124 Obesity 28 29 25 0.085 Hypertension, Uncomplicated 25 27 18 0.201 Diabetes, Uncomplicated 25 25 26 0.018 Diabetes with Complications 25 24 30 0.142 Pulmonary Circulation Disorder 19 18 23 0.135 Valvular Disease 18 18 20 0.067 Peripheral Vascular Disease 11 11 13 0.079 Depression 8 7 10 0.079 Hypothyroidism 8 8 9 0.050 Other Neurological Disorder 6 6 8 0.090 Drug Abuse 6 6 7 0.055 Deficiency Anemia 6 6 7 0.047 Alcohol Abuse 5 5 5 0.008 Coagulopathy 5 5 7 0.059 Weight Loss 4 4 6 0.098 Liver Disease 4 4 6 0.087 Rheumatoid Arthritis/Collagen 3 3 4 0.049 Psychoses 2 2 2 0.057 Solid Tumor w/o Metastasis 2 2 3 0.063 Metastatic Cancer 1 1 1 0.061 Lymphoma 1 1 1 0.044 Blood Loss Anemia 1 1 1 0.032 HIV/AIDS 1 0 1 0.033 Peptic Ulcer Disease 0 0 1 0.013 Paralysis 1 1 1 0.013 Abbreviations: ACGME, Accreditation Council for Graduate Medical Education; CVD, cardiovascular disease; GME, Graduate Medical Education; LACE, Length of stay + Acuity + Comorbidities + Emergency visits; SD, standard deviation; SMD, standardized mean difference . Total discharge records: N = 157,791 from 123,272 unique patients (discharge−to−patient ratio = 1.28) . Values are mean (SD) for continuous variables and n (%) or prevalence (%) for categorical variables. SMD ≥ 0.20 (bolded) indicates a meaningful difference between groups. For categorical variables with >2 levels, a single overall SMD is reported on the section header row . Validated indices computed from discharge data: LACE Index (van Walraven et al., 2010), Charlson Comorbidity Index (Charlson et al., 1987; Quan et al., 2011), van Walraven Elixhauser Index (van Walraven et al., 2009). All 31 Elixhauser comorbidity flags shown with individual SMDs . Table 2. Best-Performing Models by CVD Category Abbreviations: AUC, area under the receiver operating characteristic curve; AUPRC, area under the precision-recall curve; CI, confidence interval; CVD, cardiovascular disease; Sens, sensitivity; Spec, specificity. Best model selected per category based on highest AUC across balanced (SMOTE) and unbalanced datasets. Both top-performing configurations shown per subtype. Brier score range: 0 (perfect) to 1 (worst). Lower values indicate better calibration. AUC interpretation: 0.70–0.80 = moderate; >0.80 = high discrimination. Super Learner is a stacked ensemble combining XGBoost, LightGBM, Ranger, and Elastic Net via non-negative least squares meta-learning. All models validated using patient-grouped 70/30 splits with clustered bootstrap 95% confidence intervals (1,000 replicates). CVD Category Dataset Algorithm AUC (95% CI) AUPRC Sens Spec Brier AMI Unbalanced XGBoost 0.706 (0.691–0.721) 0.275 0.571 0.728 0.213 Balanced Elastic Net 0.697 (0.679–0.713) 0.266 0.679 0.607 0.203 AF/AFL Balanced XGBoost 0.732 (0.715–0.750) 0.339 0.708 0.648 0.120 Unbalanced XGBoost 0.731 (0.714–0.749) 0.339 0.658 0.684 0.212 HF Balanced XGBoost 0.708 (0.701–0.716) 0.418 0.632 0.680 0.158 Unbalanced XGBoost 0.707 (0.700–0.714) 0.418 0.678 0.630 0.215 HHD Balanced XGBoost 0.758 (0.735–0.777) 0.338 0.734 0.652 0.116 Unbalanced Super Learner 0.754 (0.731–0.776) 0.338 0.680 0.698 0.096 Unit of Analysis and Patient Grouping. The unit of analysis was the discharge record, matching the HRRP penalty structure. 11,12 Because patients could contribute multiple admissions, all data partitions were grouped at the patient level to prevent information leakage. A 70/30 train-test split assigned all discharges from a given patient exclusively to one partition. Hyperparameter tuning used inner 5-fold patient-grouped cross-validation. Zero patient overlap was programmatically verified. Prior large-scale administrative readmission studies have used the same grouping strategy. 4,6,18,20 To evaluate whether discrimination differed by admission history, test-set performance was stratified by admission order (first admission vs. subsequent admissions) for each best-performing model ( Supplementary Table S6 ). Outcome Definition. The primary outcome was 30-day all-cause readmission to any Virginia acute-care hospital, identified through longitudinal patient linkage using Readmissions and Transfers (RATs) supplemental files. Feature Engineering. Raw administrative data were supplemented with validated clinical indices pre-computed during data extraction: the LACE Index (van Walraven et al., 2010), 32 Charlson Comorbidity Index (Charlson 1987; Quan 2005), 33,34 Age-Adjusted Charlson Index, van Walraven Elixhauser Index (van Walraven et al., 2009), 35 and Elixhauser Comorbidity Count. All 31 AHRQ Elixhauser comorbidity flags were pre-computed from secondary diagnosis fields (DX2–DX18) using published ICD-9/ICD-10 mappings. 25 Additional features captured hospital characteristics (teaching status, ownership, bed size), admission characteristics (length of stay, total diagnoses, total procedures), and patient history (admission order [sequential index of each discharge], admissions per patient [cumulative count]). Complete feature specifications appear in Supplementary Table S1 . The study period (January 2010 through December 2020) spans the October 1, 2015 transition from ICD-9-CM to ICD-10-CM coding. Primary CVD diagnosis classification used published dual-era diagnostic mappings ( Supplementary Table S1 ). However, ICD-10’s substantially greater coding granularity may have inflated comorbidity counts and altered Elixhauser flag prevalence in the post-transition period, as previously documented in administrative database studies. Admission Year was included as a temporal feature (selected by 3 of 4 feature selection methods), which may partially absorb coding era effects. A descriptive comparison of key comorbidity features across coding eras is provided in Supplementary Table S1 0 . Feature Selection. From 60 candidate features after technical filtering (removal of zero-variance, near-zero-variance features with > 95% single value, and one member of pairs with |r| > 0.90), four independent selection methods were applied to the training set: (a) Boruta all-relevant selection, 36 (b) SHAP-based importance from a reference XGBoost model, (c) recursive feature elimination with patient-grouped cross-validation, and (d) minimum redundancy maximum relevance (mRMR). 37 Features endorsed by at least two of four methods were retained, producing 45 consensus-selected features applied uniformly across all CVD categories and configurations. This data-driven consensus approach replaced the univariate correlation ranking used in earlier analyses, removing the arbitrary top-k cap so that retained features received converging evidence from multiple methodologically distinct selection criteria (Supplementary Table S1 , Figure S1 ). Machine Learning Model Development. Four base algorithms were evaluated: XGBoost, LightGBM, Random Forest (ranger), Elastic Net (glmnet). Each was trained independently for each CVD category under two configurations (unbalanced and SMOTE-balanced), for a total of 32 base models. A Super Learner stacked ensemble combined out-of-fold predictions from all base learners through a non-negatively constrained logistic regression meta-learner,40 adding 8 ensemble models (one per CVD category per dataset configuration) for a total of 40 models (4 base × 4 categories × 2 configurations + 8 ensemble; Fig. 1 ). All metrics were evaluated on the held-out test set, held out from training, tuning, and SMOTE augmentation. Balanced training sets used SMOTE with K = 5 nearest neighbors applied exclusively to the training partition. For unbalanced datasets, class weights were assigned through algorithm-native methods (scale_pos_weight, inverse-class-frequency weighting, or native logistic loss). All algorithms underwent grid search with inner 5-fold patient-grouped cross-validation maximizing AUC. XGBoost and LightGBM each searched 27 configurations with early stopping at 15 rounds; Random Forest searched 18 configurations selected by out-of-bag error; Elastic Net optimized 11 alpha values with lambda selected by the one-standard-error rule. The term “ensemble” in the title refers to the Super Learner stacked ensemble approach, which combines out-of-fold predictions from all base learners through a non-negatively constrained logistic regression meta-learner. Performance Assessment and Validation. Model performance followed TRIPOD guidelines 23 on the 30% held-out test set. Discrimination was assessed by AUC with 95% bootstrap CIs (1,000 iterations, percentile method) and area under the precision-recall curve (AUPRC), which is more informative than AUC-ROC for imbalanced outcomes. The optimal threshold was determined by maximizing the Youden index, at which sensitivity, specificity, positive and negative predictive values, F1 score, balanced accuracy, and Matthews correlation coefficient were computed. Clinical utility was classified as Excellent (≥ 0.80), High (≥ 0.75), Moderate (≥ 0.70), or Limited (< 0.70). Calibration was assessed using Brier scores, calibration slopes, calibration-in-the-large (CITL), and expected-to-observed (E/O) ratio per TRIPOD guidelines (Fig. 4 ). ROC curves are presented in Fig. 2 and the AUC performance heatmap in Fig. 3 . Pairwise algorithm comparisons used DeLong tests within the best-performing dataset configuration per CVD category ( Supplementary Table S3 ). To assess residual within-patient correlation, we fitted GEE models with exchangeable working correlation and sandwich standard errors on multi-admission patient subsets within each test partition. Patient-resampled clustered bootstrap CIs (1,000 iterations) were also computed. Results appear in Supplementary Tables S4–S5 . Feature Importance Analysis. Feature importance was quantified using algorithm-native methods (gain-based for XGBoost, LightGBM; impurity-based for Random Forest; absolute standardized coefficients for Elastic Net), normalized to [0, 1] within each model. Three complementary perspectives are reported: condition-specific importance from each best-performing model (Fig. 5 ), aggregated importance averaged across all 32 base models requiring appearance in at least three ( Supplementary Tables S7 and S8 ), and SHapley Additive exPlanations (SHAP)-based importance providing directional, additive decompositions grounded in cooperative game theory (Fig. 6 , Supplementary Figures S3–S4 ). Algorithm-native importance served as the primary measure; SHAP analysis provides supplementary mechanistic insight. Post-hoc SHAP analysis was performed on each condition’s best-performing model ( Table 2 ) using shapviz in R. Exact TreeSHAP values were computed for tree-based models (XGBoost, LightGBM) via native contribution output. SHAP values were computed on a random subsample of 500 test observations per condition (set.seed(42)). Three visualizations were generated: beeswarm summary plots (Fig. 6 ), waterfall plots for representative true positive and true negative cases ( Supplementary Figure S4 ), and dependence plots for continuous predictors ( Supplementary Figure S3 ). SHAP analysis was performed post-hoc and did not influence model selection or hyperparameter tuning. SHAP values were computed on a random subsample of 500 test observations per CVD category due to computational cost. The subsample was verified for representativeness by comparing event rates and distributions of the top 5 predictors (LACE Index, Charlson Comorbidity Index, Length of Stay, Elixhauser Comorbidity Count, and Admission Order) between the subsample and the full test set, confirming no meaningful divergence (all standardized mean differences < 0.10). Model Specification and Reproducibility. Tree ensemble models (XGBoost, LightGBM, Random Forest) and the Super Learner cannot be expressed as closed-form equations; their prediction logic is distributed across learned parameters. Serialized model objects containing all trained models, preprocessing parameters, hyperparameter configurations, selected features, and optimal thresholds were archived and are available from the corresponding author. A standalone prediction function accepting raw patient-level input and returning readmission probabilities was also archived for prospective deployment. For Elastic Net, which yields interpretable coefficient vectors, Supplementary Table S8 reports the full coefficient matrix (intercept plus 45 features) for each CVD category under both dataset configurations. These coefficients enable exact reproduction of predicted log-odds without access to the serialized model objects: log-odds(readmission) = β₀ + Σβⱼxⱼ, where β values are provided in Supplementary Table S8 and λ was selected by the one-standard-error rule. Algorithmic Bias Assessment. Bias assessment included demographic parity across insurance types and age groups, equalized odds evaluation, calibration consistency analysis, and post-hoc auditing of differential prediction by insurance type. Statistical Analysis. All analyses were conducted in R version 4.4.2 using xgboost, lightgbm, ranger, glmnet, caret, smotefamily, pROC, geepack, Boruta, mRMRe, shapviz, and tableone. Baseline characteristics were compared using Welch’s t-tests and Pearson’s chi-squared tests. Effect sizes were reported as standardized mean differences (SMDs) computed via the tableone package, with SMD ≥ 0.10 considered indicative of meaningful imbalance. All 31 Elixhauser comorbidity flags were pre-computed during data extraction using published AHRQ ICD-9/ICD-10 mappings applied to secondary diagnosis fields. 25 Two variables had missing values: Teaching Status (0.1%) and Hospital Ownership (< 0.01%), handled by mode imputation. Four candidate features (APR-DRG Severity of Illness, APR-DRG Risk of Mortality, Admission Type, Discharge Disposition) were unavailable in the analytic dataset and excluded. A fixed random seed (set.seed(42)) was applied prior to all stochastic procedures. 26 With a sample of 157,791 discharges, conventional significance tests (Welch’s t and Pearson’s chi-squared) have near-certainty of rejecting the null hypothesis for even trivially small differences. Therefore, while p-values are reported in Table 1 for completeness, clinical meaningfulness of between-group differences was assessed using standardized mean differences (SMDs), with SMD > 0.10 indicating meaningful imbalance per established thresholds (Austin, 2009). No formal multiplicity adjustment was applied to model selection, consistent with standard practice in machine learning benchmarking studies. For each CVD category, the best-performing model was selected from 10 candidates (5 algorithms × 2 configurations); the optimistic bias inherent in selecting the maximum of multiple correlated estimates (“winner’s curse”) is acknowledged as a limitation. Ethical Considerations. The study was approved by the Eastern Virginia Medical School IRB (IRB #23-09-NH-0248) with waiver of informed consent for de-identified data analysis. All analyses used de-identified data with no attempt to re-identify patients. Results Cohort Characteristics. The analytic cohort comprised 157,791 discharge records from 123,272 unique patients identified through the VHI statewide database (discharge-to-patient ratio: 1.28), with 83.3% of patients contributing a single admission and 16.7% contributing two or more admissions. No records were excluded for missing outcome or patient identifiers. All ML model training and evaluation used this full cohort; cohort flow is detailed in Fig. 1 . Baseline demographics and clinical characteristics are presented in Table 1 . The population was 96.6% Black/African American and 49.0% female, with a mean age of 62.75 ± 14.19 years. HF was the most common primary CVD category (59.9% of discharges), followed by AMI (19.1%), AF/AFL (11.9%), and HHD (9.1%). Readmitted patients had higher LACE Index scores (11.09 ± 2.86 vs 9.34 ± 2.96; SMD = 0.604), more admissions per patient (3.65 ± 4.59 vs 1.60 ± 1.68; SMD = 0.592), higher admission order (2.28 ± 2.82 vs 1.31 ± 1.04; SMD = 0.456), greater Elixhauser comorbidity count (5.27 ± 1.85 vs 4.49 ± 2.00; SMD = 0.410), higher Charlson Comorbidity Index (3.81 ± 1.99 vs 2.99 ± 2.04; SMD = 0.405), higher van Walraven Elixhauser Index (14.74 ± 8.37 vs 11.34 ± 8.70; SMD = 0.398), more total diagnoses (15.60 ± 3.44 vs 14.13 ± 4.24; SMD = 0.381), and higher prevalence of renal failure (59% vs 42%; SMD = 0.325). Insurance type showed substantial separation (SMD = 0.342), with Medicare overrepresented among readmissions (67.7% vs 56.3%) and commercial insurance underrepresented (13.6% vs 22.9%). 30-Day Readmission Outcome. The overall 30-day all-cause readmission rate was 18.9% (n = 29,878). Rates varied across CVD categories: HF 22.5%, AF/AFL 15.3%, AMI 13.0%, and HHD 12.4%. HF was disproportionately represented among readmissions (71.3% of readmitted discharges vs 57.3% of non-readmitted). Medicare beneficiaries accounted for 67.7% of readmissions compared with 56.3% of non-readmitted discharges (SMD = 0.342), while self-pay/uninsured patients were underrepresented among readmissions (5.3% vs 10.7%). The LACE Risk Category showed the strongest categorical separation: 72.1% of readmitted patients were classified as high-risk compared with 48.8% of non-readmitted patients (SMD = 0.502). Model Discrimination. From 60 candidate features after technical filtering, four independent selection methods yielded 45 consensus-selected features, applied uniformly across all CVD categories and dataset configurations. Forty prediction models were then trained by crossing four base algorithms plus a Super Learner ensemble with four CVD categories and two dataset configurations. All models used patient-grouped 70/30 splitting with zero patient leakage between partitions. Complete results for all 40 models appear in Supplementary Table S2 ; the best-performing models per category are summarized in Table 2 . Discrimination varied across CVD subtypes, with all four categories achieving moderate or high discrimination (AUC 0.706–0.758), though the highest-performing HHD model (AUC 0.758) exhibited calibration limitations (slope 1.479–1.646) requiring recalibration before absolute risk estimation. HHD yielded the highest AUC: XGBoost on the balanced dataset achieved 0.758 (95% CI, 0.735 to 0.777), classified as high utility. The Super Learner ensemble was competitive for HHD on the unbalanced dataset (0.754; 95% CI, 0.731 to 0.776). AF/AFL followed at 0.732 (95% CI, 0.715 to 0.750; XGBoost, balanced), representing moderate utility. HF achieved 0.708 (95% CI, 0.701 to 0.716; XGBoost, balanced), with the narrow confidence interval driven by the large test sample (28,380 records from 21,149 patients). AMI achieved 0.706 (95% CI, 0.691 to 0.721; XGBoost, unbalanced). XGBoost was the best-performing base algorithm across all four categories. ROC curves are shown in Fig. 2 ; the cross-category algorithm comparison appears in Fig. 3 . AUPRC values further characterized discrimination in the context of varying event rates: HF 0.418 (event rate 22.5%), AF/AFL 0.339 (15.3%), HHD 0.338 (12.4%), and AMI 0.275 (13.0%). The relatively higher AUPRC for HF reflects the higher base rate, while AMI’s lower AUPRC despite comparable AUC suggests that positive prediction remains challenging for the least prevalent readmission category. Effect of SMOTE Balancing. Balanced XGBoost achieved the best AUC for HHD (0.758) and HF (0.708), while unbalanced XGBoost was best for AMI (0.706). AF/AFL showed virtually no difference between balanced and unbalanced XGBoost (0.732 vs 0.731). Across categories, SMOTE generally improved sensitivity at the expense of specificity, a tradeoff that was most beneficial in categories with lower event rates (HHD 12.4%, AMI 13.0%; Supplementary Table S9 ). Model Calibration. Calibration was assessed using Brier scores, calibration slopes, and E/O ratios (Fig. 4 ). The AMI XGBoost model demonstrated excellent relative calibration (slope = 0.993), indicating well-preserved rank ordering of predicted probabilities. However, the expected-to-observed (E/O) ratio of 3.47 indicates substantial overestimation of absolute risk, meaning the model predicted 3.47 times more readmission events than were observed. While the model reliably identifies higher-risk patients relative to lower-risk patients, absolute probability estimates for AMI require recalibration (e.g., Platt scaling or isotonic regression) before clinical deployment. HF models showed excellent calibration (balanced slope 1.014, unbalanced slope 1.025), with Brier scores of 0.158 (balanced) and 0.215 (unbalanced). AF/AFL exhibited slight underconfidence (slopes 1.037–1.065). HHD showed the greatest miscalibration (slopes 1.479–1.646), indicating systematic overconfidence at the extremes of predicted risk; Platt recalibration is recommended before deployment for this subtype. Raw Brier scores are reported in Table 2 ; for all categories, rank-ordering of risk was preserved even when probability estimates required rescaling. Pairwise Algorithm Comparisons. DeLong tests for correlated ROC curves compared all algorithm pairs within the best-performing dataset configuration per CVD category ( Supplementary Table S3 ). XGBoost significantly outperformed Random Forest and Elastic Net across most conditions. Within each condition, XGBoost and LightGBM were generally statistically indistinguishable, though XGBoost held a consistent numerical edge. The Super Learner was competitive with or slightly below the best base learner across all categories, with the closest match in HHD. Patient-Grouped Validation. GEE models with exchangeable working correlation and sandwich standard errors were fitted to the multi-admission subset of each category’s test partition ( Supplementary Table S4 ). The GEE coefficient for predicted risk was significant in HF (α = 0.120, Wald P < 0.001), AMI (α = 0.041, Wald P < 0.001), and HHD (α = 0.141, Wald P = 0.017), confirming that model-predicted risk remained a significant readmission predictor after accounting for within-patient correlation. AF/AFL showed a non-significant correlation (α = 0.170, P = 0.229), likely due to the smaller multi-admission sample (1,067 records from 461 patients). Clustered bootstrap resampling (1,000 iterations resampling patients with replacement) produced confidence intervals comparable to standard record-level bootstrap intervals ( Supplementary Table S5 ). The mean CI width ratio across categories was 1.13; patient clustering widened intervals by approximately 13% on average. No model’s point estimate fell outside its clustered interval, and all utility classifications remained unchanged. Taken together, the patient-grouped splitting strategy adequately controlled within-patient correlation. The clustered bootstrap confidence interval for HHD (0.733–0.784) extends below the 0.75 threshold defining “high utility” in our classification schema, despite the point estimate (0.758) exceeding this threshold. The 21.2% CI widening from patient-level clustering for HHD was the largest among all four categories, reflecting greater within-patient correlation in this subgroup. The “high utility” classification for HHD should therefore be interpreted with this uncertainty in mind, particularly when considering prospective deployment. Performance by Admission Order. To assess whether prior hospitalization history influenced discrimination, best-model AUC was stratified by admission order within each test partition ( Supplementary Table S6 ). Discrimination varied by admission history, with greater heterogeneity among repeat admissions in categories with lower event rates. The Super Learner provided the most stable performance across admission strata. LACE Index Sensitivity Analysis. To assess the impact of construct overlap between the LACE Index composite score and its constituent components (LACE-L, LACE-A, LACE-C, LACE-E), we conducted a sensitivity analysis comparing three feature configurations using XGBoost: (1) the full consensus feature set including both the composite and components, (2) LACE composite index only (individual components removed), and (3) LACE components only (composite removed). AUC differences were negligible across configurations ( Supplementary Table S11). For AMI, HF, and AF/AFL, all three configurations produced identical AUCs to the third decimal place (all DeLong p > 0.20). For HHD, the composite-only configuration yielded marginally higher AUC (0.755) than both the full set (0.747, DeLong p = 0.053) and components-only (0.747, p = 0.015), suggesting the LACE Index composite alone is sufficient and may even reduce noise from redundant component features. These results support using the composite LACE Index alone for parsimony and clinical interpretability, and confirm that construct overlap did not artificially inflate model discrimination. Feature Importance. Aggregated importance across all 32 base models produced a stable predictor ranking (Fig. 5 , Supplementary Table S7 ). Validated clinical indices dominated: the LACE Index ranked among the top predictors in every category, in line with its strong separation between readmitted and non-readmitted patients (SMD = 0.604). The Charlson Comorbidity Index, van Walraven Elixhauser Index, and Elixhauser Comorbidity Count also ranked highly, a pattern that follows the shift from ad-hoc composites to published indices with established reference ranges. Category-specific patterns supported the case for condition-specific modeling (Fig. 5 ). In HF, the LACE Index and renal failure dominated, with admission order contributing strongly. In AMI, total diagnoses and the LACE Index led. In AF/AFL, the importance distribution was broader, with insurance type contributing more prominently. In HHD, the LACE Index, insurance type, and renal failure led, with obesity and depression contributing more than in other categories. Renal failure showed consistent importance across all categories (SMD = 0.325 for readmission vs non-readmission), in keeping with the cardiorenal syndrome’s known role in CVD readmission. Elastic Net coefficient vectors ( Supplementary Table 7 ) offered a view of predictor direction and regularization-driven sparsity. The LACE Index and renal failure carried the largest positive coefficients across conditions, while obesity consistently received negative coefficients, suggesting a protective association after adjustment for comorbidity burden. SHAP analysis replicated the global importance rankings at the individual prediction level (Fig. 6 ). The LACE Index, admission order, renal failure, and insurance type produced the largest absolute SHAP values across predictions. Waterfall decompositions of representative cases ( Supplementary Figure S4 ) illustrated how these top features combined to produce individual risk estimates. Dependence plots ( Supplementary Figure S3 ) revealed non-linear relationships, including threshold effects for the LACE Index and differential effects of insurance type across predicted risk levels. Discussion This study is, to our knowledge, the largest ML-based cardiovascular readmission prediction analysis in a predominantly Black cohort and the first to compare five algorithms plus a Super Learner ensemble across four CVD subtypes using standardized validated features in a racially concentrated population. Condition-Specific Discrimination. All four CVD categories reached moderate or high predictive utility, exceeding most prior administrative data-based benchmarks. The HF AUC of 0.708 exceeded both Sharma and colleagues’ benchmark of 0.65 using XGBoost on Canadian administrative data 27 and Mortazavi and colleagues’ Random Forest AUC of 0.628 in the Tele-HF clinical trial cohort, 39 suggesting that validated indices and consensus feature selection narrow the discrimination gap previously attributed to administrative data limitations. Fine and colleagues achieved 0.74 for a composite HF endpoint using CatBoost with deep feature synthesis, though their outcome included ED revisits and mortality. 28 Awan and colleagues similarly reported a best AUC of 0.62 for HF readmission or death using administrative data with explicit attention to class imbalance. 18 The HHD model reached high utility (AUC 0.758). Because most studies subsume HHD within broader HF or hypertension categories, 1,2 direct comparison with published benchmarks is limited. The narrow AUC range across algorithms for HHD likely reflects a more uniform pathophysiology, making this subtype a practical candidate for early deployment. AMI models achieved moderate utility (0.706) with the best calibration in the study (slope 0.993), within the range of published administrative data AMI models (AUC 0.65–0.75 without clinical variables). 6,7 Pandey and colleagues found that racial differences in AMI readmission were attributable to patient-level rather than hospital-level factors, 12 which aligns with the strong insurance and comorbidity contributions in our AMI models. AF/AFL reached moderate utility (0.732), likely because validated indices better captured the comorbidity burden driving arrhythmia-related readmissions. Algorithmic Hierarchy. XGBoost outperformed all alternatives across four categories, extending results from mixed-race populations 5,17,19 to a predominantly Black cohort with identical features. The Super Learner ensemble matched the best base learner most closely for HHD (unbalanced AUC 0.754 vs XGBoost 0.753), where LightGBM and Ranger received the largest meta-learner weights ( Supplementary Figure S2 ), suggesting that ensemble gains require base learners with complementary decision boundaries. Balanced datasets produced the best AUC in two of four categories (HHD, HF), with unbalanced best for AMI and no meaningful difference for AF/AFL. Insurance, Social Determinants, and Health Equity. The LACE Index was the strongest discriminator between readmitted and non-readmitted patients (SMD = 0.604) and, unlike ad-hoc composite scores, has established reference ranges and external validity across settings. 32 Insurance type ranked among the top predictors (SMD = 0.342). Bahiru and colleagues similarly showed that hospitals serving higher proportions of dual-eligible patients had significantly higher 30-day readmission rates. 30 The low readmission representation among self-pay and uninsured discharges (5.3% vs 10.7%) more likely reflects access barriers to rehospitalization than better outcomes, 11,14,22 meaning ML models trained on these data may underestimate true risk for uninsured patients. Chaiyachati and colleagues showed that racial disparities in readmission widened within safety-net hospitals after HRRP implementation, especially for conditions not targeted by HRRP, 31 adding weight to the case that HRRP risk adjustment models should incorporate social determinant measures. Condition-specific feature importance patterns, such as the prominent role of obesity and depression in HHD, would have been missed in pooled modeling. Predictive contribution is distinct from clinical actionability. The LACE Index, renal failure, and obesity improve risk stratification but are not directly modifiable in the acute window. Insurance type is partly modifiable through policy yet also proxies for age, disability, and post-discharge access. Care process features (admission order, length of stay, hospital characteristics) identify where targeted transition-of-care interventions may reduce readmission probability. Model outputs should be interpreted accordingly: non-modifiable features refine risk classification, while care process features indicate intervention opportunities. Feature Importance Concordance. Algorithm-native, aggregated, and SHAP-based importance rankings all placed the LACE Index, renal failure, and insurance type among the top predictors, paralleling the emphasis on appropriate model and metric selection for administrative HF readmission data reported by Awan and colleagues. 18 The LACE Index, Charlson, and van Walraven Elixhauser indices carry published reference ranges and established predictive validity, making them more reproducible and interpretable than ad-hoc composites. However, the LACE Index requires careful interpretation as a predictor because its components (length of stay, acuity of admission, comorbidity, and emergency department visits) overlap with readmission risk factors, creating potential construct overlap. While this does not invalidate its predictive utility, it means that LACE-based risk stratification identifies high-utilization patients through a partially tautological construct. SHAP dependence plots ( Supplementary Figure S3 ) revealed non-linear LACE–readmission relationships, including diminishing marginal effects at the highest scores. Our sensitivity analysis ( Supplementary Table S11 ) confirmed that the simultaneous inclusion of the LACE Index and its constituent components did not inflate model discrimination: AUC was unchanged across all three feature configurations for AMI, HF, and AF/AFL (all DeLong p > 0.20), and the composite-only configuration actually performed marginally better for HHD (DeLong p = 0.015 vs. components-only). The composite LACE Index alone is therefore sufficient for prediction, and its high feature importance ranking reflects genuine predictive value rather than redundant information from overlapping components. Practical Implications. All four CVD models demonstrated moderate to high utility for risk stratification, supporting their integration into clinical workflows. The HHD model (AUC 0.758, high utility) is the most immediate candidate for pilot clinical decision support in safety-net settings. HF and AF/AFL models, now at moderate utility, can support tiered risk stratification within the planned Readmission Analytics and Interactive Platform (RAIP) dashboard. Low PPVs across all models, as seen in comparable studies, 27,28 indicate that integration within a tiered risk stratification system rather than standalone screening remains appropriate. The Super Learner’s competitive performance for HHD provides an ensemble option for deployment contexts where model stability is valued over marginal discrimination gains. The HHD model shows high discrimination but suboptimal calibration (slopes 1.479–1.646), suggesting that predicted probabilities at extremes are overstated. Thus, it is suitable for ranking patients by risk but requires recalibration, via Platt scaling or isotonic regression, before absolute risk communication. Key strengths include a large, predominantly Black cohort (157,791 discharges, 96.6% Black), a fully crossed 40-model design enabling unbiased algorithm comparison, data-driven consensus feature selection, and validated clinical indices for reproducibility. 14,16,20 Patient-grouped splitting with GEE post-hoc validation and clustered bootstrap resampling ensured adequate control for within-patient correlation. Identifying HHD as a standalone prediction target with the highest discrimination offers an actionable result for a condition routinely subsumed within broader categories. Several limitations warrant discussion. The 96.6% Black cohort limits generalizability to diverse populations and requires external validation. 16,20 Lack of laboratory values, vital signs, ejection fraction, medications, and four key administrative features (APR-DRG Severity, Risk of Mortality, Admission Type, Discharge Disposition) may have constrained model performance. ICD-based classification cannot distinguish HFpEF from HFrEF. Low readmission among self-pay/uninsured patients introduces potential bias, as observed outcomes may underestimate true clinical risk. SMOTE-generated synthetic observations, particularly for binary comorbidity flags, may not fully reflect the true feature space, and alternative resampling strategies (e.g., SMOTE-NC or conditional generative models) may better preserve data fidelity. Model selection from multiple candidates likely overestimates performance (optimistic bias), emphasizing the need for independent external validation (Supplementary Table S3). The LACE Index, as both a predictor and a construct overlapping with readmission determinants, introduces potential circularity that warrants evaluation in external validation. The ICD-9 to ICD-10 transition may have introduced systematic differences in comorbidity ascertainment over the study period, affecting feature distributions despite adjustment for admission year (Supplementary Table S10). Future studies should consider multi-site validation and stratification by coding era to mitigate these temporal effects. Finally, another limitation is the absence of independent external validation. Although models were evaluated using a patient-grouped 70/30 train–test split with additional robustness checks including clustered bootstrap resampling and GEE analyses to account for within-patient correlation, these procedures represent internal validation within a single statewide dataset. Consequently, model performance estimates may remain optimistic relative to deployment in new clinical environments. Administrative coding practices, case mix, health system structure, and social determinants vary across regions and institutions, and these factors may influence both predictor distributions and readmission risk relationships. Therefore, the present models should be interpreted as development-stage prediction tools rather than finalized clinical decision support systems. Independent validation in geographically distinct populations and healthcare systems will be necessary to assess generalizability and confirm model calibration and discrimination Conclusions Condition-specific ML models achieved moderate-to-high discrimination for 30-day readmission across four CVD subtypes in a large, predominantly Black cohort. Model performance was driven more by CVD subtype than algorithm choice, with XGBoost performing consistently and the Super Learner competitive for HHD. Key predictors, including the LACE Index, renal failure, and insurance status, highlight the combined influence of clinical and structural factors. These results support deploying subtype-specific, equity-informed readmission models using validated clinical indices rather than ad-hoc composites. External validation in diverse populations, incorporation of area-level social determinants, and prospective evaluation of model-guided interventions, particularly for HHD and HF, are needed to confirm generalizability and assess impact on readmission disparities. Declarations Acknowledgements This research was supported by American Heart Association Second Century Implementation Science Award #23SCISA1145640. The opinions expressed in this article are solely those of the authors and do not necessarily represent those of the American Heart Association. This project was initiated at the EVMS-Sentara Healthcare Analytics and Delivery Science Institute (HADSI), Eastern Virginia Medical School, and continued at the Macon & Joan Brock Virginia Health Sciences at Old Dominion University following the institutional integration of EVMS and Old Dominion University. The authors acknowledge collaborative research support provided through a subaward to the University of Illinois at Chicago. Data access was facilitated by the M. Foscue Brock Institute for Community and Global Health and the Research and Infrastructure Service Enterprise (RISE) at the Macon & Joan Brock Virginia Health Sciences, Old Dominion University. Virginia Health Information (VHI) has provided non-confidential patient-level information used in this study, which it has compiled in accordance with Virginia law but which it has no authority to independently verify. By using this data, the authors agree to assume all risks that may be associated with or arise from the use of inaccurate data. VHI cannot and does not represent that the use of VHI’s data was appropriate for this study or endorse or support any conclusions or inferences that may be drawn from the use of VHI’s data. This study was approved by the Eastern Virginia Medical School Institutional Review Board (IRB #23-09-NH-0248) with a waiver of informed consent for de-identified data analysis. Author Contributions I.E.M. conceived and designed the study, developed the machine learning pipeline and feature engineering framework, conducted all model training and validation analyses, curated the data, created all figures and tables, wrote the original manuscript draft, supervised the project, administered all aspects of the research as Principal Investigator, and acquired funding. M.B. investigated readmission patterns across demographic, clinical, and administrative dimensions and contributed to data cleaning and preparation of the VHI Patient Level Database and Readmissions and Transfers Supplemental Data Set under the supervision of I.E.M. S.D. provided scientific oversight, cardiovascular epidemiology expertise, and mentorship throughout the project, contributed to study conceptualization and to the interpretation of health equity findings and clinical contextualization of results, and facilitated the institutional research infrastructure. All authors reviewed and approved the final manuscript. Competing Interests Dr. El Moudden reports research funding from the American Heart Association (Award #23SCISA1145640) during the conduct of this study; no other relationships or activities that could appear to have influenced the submitted work. Dr. Dodani reports the following relationships: Vice-Chair, American Heart Association Health Equity Research Network (HERN) Oversight Advisory Committee; subaward through the University of Illinois at Chicago for collaborative research support on this project (Award #23SCISA1145640). She holds appointments at the University of Illinois College of Medicine–Peoria (Founding Director, Center 4 Health Research; Professor of Clinical Medicine, Department of Medicine) and at the Macon & Joan Brock Virginia Health Sciences, Eastern Virginia Medical School at Old Dominion University (Professor of Medicine, Community Non-Tenure Track). No other relationships or activities that could appear to have influenced the submitted work. Mr. Bittner reports no disclosures. Data Availability The de-identified datasets analyzed in this study were derived from the Virginia Health Information (VHI) Patient Level Database (PLD) and the Readmissions and Transfers Supplemental Data Set (RATs), licensed inpatient discharge files encompassing all verified acute-care hospital discharges in Virginia from 2010 through 2020 (~9.24 million total discharges). Data access was provided through the M. Foscue Brock Institute for Community and Global Health and the Research and Infrastructure Service Enterprise (RISE) at the Macon & Joan Brock Virginia Health Sciences, Old Dominion University, under a site license agreement with VHI. Due to VHI licensing restrictions and the Virginia Patient Level Database System Act of 1993, the data cannot be publicly shared or deposited in an open repository. Access may be requested from the corresponding author upon reasonable request and is contingent upon execution of a VHI data license agreement and approval from the Brock Institute. Researchers seeking comparable data must apply directly to VHI, complete the licensing process, and cover applicable fees (www.vhi.org/pld). Correspondence regarding data extraction or PLD-to-RATs linkage methodology should be directed to Dr. Ismail El Moudden ( [email protected] ). Code Availability All R analysis scripts for data preprocessing, feature engineering, model training, hyperparameter tuning, performance evaluation, and figure generation are available at https://github.com/isamil/ML-CVD-Readmission-in-Black-Cohort and will be made publicly available upon publication. Access for peer review is granted upon request to the corresponding author. The archived model package includes: (1) serialized trained model objects for all 32 base learner models and 8 Super Learner ensembles, (2) a standalone prediction function accepting raw patient-level input, (3) preprocessing parameters and hyperparameter configurations, and (4) Elastic Net coefficient vectors enabling exact replication without model objects (Supplementary Table S8). The archived model package is available from the corresponding author upon reasonable request. References Martin, S. S. et al. 2025 Heart Disease and Stroke Statistics: A Report of US and Global Data From the American Heart Association. Circulation 151 , e41-e660 (2025). https://doi.org/10.1161/CIR.0000000000001303 Martin, S. S. et al. 2024 Heart Disease and Stroke Statistics: A Report of US and Global Data From the American Heart Association. Circulation 149 , e347-e913 (2024). https://doi.org/10.1161/CIR.0000000000001209 Heidenreich, P. A. et al. 2022 AHA/ACC/HFSA Guideline for the Management of Heart Failure. Circulation 145 , e895-e1032 (2022). https://doi.org/10.1161/CIR.0000000000001063 Dharmarajan, K. et al. Diagnoses and timing of 30-day readmissions after hospitalization for heart failure, acute myocardial infarction, or pneumonia. JAMA 309 , 355–363 (2013). https://doi.org/10.1001/jama.2012.216476 Shin, S. et al. Machine learning vs. conventional statistical models for predicting heart failure readmission and mortality. ESC Heart Fail 8 , 106-115 (2021). https://doi.org/10.1002/ehf2.13073 Krumholz, H. M. et al. Relationship between hospital readmission and mortality rates for patients hospitalized with acute myocardial infarction, heart failure, or pneumonia. JAMA 309 , 587-593 (2013). https://doi.org/10.1001/jama.2013.333 Khera, R. et al. Comparison of Readmission Rates After Acute Myocardial Infarction in 3 Patient Age Groups. Am J Cardiol 120 , 1761-1767 (2017). https://doi.org/10.1016/j.amjcard.2017.07.081 Freeman, J. V., Wang, Y., Akar, J. G., Desai, N. R. & Krumholz, H. M. National Trends in Atrial Fibrillation Hospitalization, Readmission, and Mortality for Medicare Beneficiaries, 1999-2013. Circulation 135 , 1227-1239 (2017). https://doi.org/10.1161/CIRCULATIONAHA.116.022388 Tripathi, B. et al. Outcomes and Resource Utilization Associated With Readmissions After Atrial Fibrillation Hospitalizations. J Am Heart Assoc 8 , e013026 (2019). https://doi.org/10.1161/JAHA.119.013026 Muntner, P. et al. Trends in Blood Pressure Control Among US Adults With Hypertension, 1999-2000 to 2017-2018. JAMA 324 , 1190-1200 (2020). https://doi.org/10.1001/jama.2020.14545 Shashikumar, S. A., Waken, R. J., Luke, A. A., Nerenz, D. R. & Joynt Maddox, K. E. Association of Stratification by Proportion of Patients Dually Enrolled in Medicare and Medicaid With Financial Penalties in the Hospital-Acquired Condition Reduction Program. JAMA Intern Med 181 , 330-338 (2021). https://doi.org/10.1001/jamainternmed.2020.7386 Pandey, A. et al. Temporal Trends in Racial Differences in 30-Day Readmission and Mortality Rates After Acute Myocardial Infarction Among Medicare Beneficiaries. JAMA Cardiol 5 , 136-145 (2020). https://doi.org/10.1001/jamacardio.2019.4845 Wadhera, R. K., Yeh, R. W. & Joynt Maddox, K. E. The Hospital Readmissions Reduction Program, Time for a Reboot. N Engl J Med 380 , 2289-2291 (2019). https://doi.org/10.1056/NEJMp1901225 Patel, S. A. et al. Excess 30-Day Heart Failure Readmissions and Mortality in Black Patients Increases With Neighborhood Deprivation. Circ Heart Fail 13 , e007947 (2020). https://doi.org/10.1161/CIRCHEARTFAILURE.120.007947 Fields, N. D. et al. Historical Redlining and Heart Failure Outcomes Following Hospitalization in the Southeastern United States. J Am Heart Assoc 13 , e032019 (2024). https://doi.org/10.1161/JAHA.123.032019 Downing, N. S. et al. Association of Racial and Socioeconomic Disparities With Outcomes Among Patients Hospitalized With Acute Myocardial Infarction, Heart Failure, and Pneumonia. JAMA Netw Open 1 , e182044 (2018). https://doi.org/10.1001/jamanetworkopen.2018.2044 Krittanawong, C., Zhang, H., Wang, Z., Aydar, M. & Kitai, T. Artificial Intelligence in Precision Cardiovascular Medicine. J Am Coll Cardiol 69 , 2657-2664 (2017). https://doi.org/10.1016/j.jacc.2017.03.571 Awan, S. E., Bennamoun, M., Sohel, F., Sanfilippo, F. M. & Dwivedi, G. Machine learning-based prediction of heart failure readmission or death: implications of choosing the right model and the right metrics. ESC Heart Fail 6 , 428-435 (2019). https://doi.org/10.1002/ehf2.12419 Rajkomar, A. et al. Scalable and accurate deep learning with electronic health records. NPJ Digit Med 1 , 18 (2018). https://doi.org/10.1038/s41746-018-0029-1 Yu, M. Y. & Son, Y. J. Machine learning-based 30-day readmission prediction models for patients with heart failure: a systematic review. Eur J Cardiovasc Nurs 23 , 711-719 (2024). https://doi.org/10.1093/eurjcn/zvae031 Chen, M., Tan, X. & Padman, R. A Machine Learning Approach to Support Urgent Stroke Triage Using Administrative Data and Social Determinants of Health at Hospital Presentation. J Med Internet Res 25 , e36477 (2023). https://doi.org/10.2196/36477 White-Williams, C. et al. Addressing Social Determinants of Health in the Care of Patients With Heart Failure: A Scientific Statement From the American Heart Association. Circulation 141 , e841-e863 (2020). https://doi.org/10.1161/CIR.0000000000000767 Collins, G. S., Reitsma, J. B., Altman, D. G. & Moons, K. G. M. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. BMJ 350 , g7594 (2015). https://doi.org/10.1136/bmj.g7594 Virginia Health Information. Patient Level Database (PLD) and Readmissions and Transfers Supplemental Data Set (RATs). Richmond, VA: Virginia Health Information; 2021. Accessed January 15, 2023. https://www.vhi.org/pld Elixhauser, A., Steiner, C., Harris, D. R. & Coffey, R. M. Comorbidity measures for use with administrative data. Med Care 36 , 8-27 (1998). https://doi.org/10.1097/00005650-199801000-00004 Carpenter, J. & Bithell, J. Bootstrap confidence intervals: when, which, what? A practical guide for medical statisticians. Stat Med 19 , 1141-1164 (2000). https://doi.org/10.1002/(sici)1097-0258(20000515)19:93.0.co;2-f Sharma, V. et al. Predicting 30-day readmissions in patients with heart failure using administrative data: a machine learning approach. J Card Fail 28 , 710–722 (2022). https://doi.org/10.1016/j.cardfail.2021.12.004 Fine, N. M. et al. Machine learning for risk prediction after heart failure emergency department visit or hospital admission using administrative health data. PLOS Digit Health 3 , e0000636 (2024). https://doi.org/10.1371/journal.pdig.0000636 He, H. & Garcia, E. A. Learning from imbalanced data. IEEE Trans Knowl Data Eng 21 , 1263–1284 (2009). https://doi.org/10.1109/TKDE.2008.239 Bahiru, E. et al. Association of Dual Eligibility for Medicare and Medicaid With Heart Failure Quality and Outcomes Among Get With The Guidelines–Heart Failure Hospitals. JAMA Cardiol 6 , 791-800 (2021). https://doi.org/10.1001/jamacardio.2021.0611 Chaiyachati, K. H., Qi, M. & Werner, R. M. Changes to racial disparities in readmission rates after Medicare’s Hospital Readmissions Reduction Program within safety-net and non-safety-net hospitals. JAMA Netw Open 1 , e184154 (2018). https://doi.org/10.1001/jamanetworkopen.2018.4154 van Walraven, C. et al. Derivation and validation of an index to predict early death or unplanned readmission after discharge from hospital to the community. CMAJ 182 , 551-557 (2010). https://doi.org/10.1503/cmaj.091117 Charlson, M. E., Pompei, P., Ales, K. L. & MacKenzie, C. R. A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. J Chronic Dis 40 , 373-383 (1987). https://doi.org/10.1016/0021-9681(87)90171-8 Quan, H. et al. Coding algorithms for defining comorbidities in ICD-9-CM and ICD-10 administrative data. Med Care 43 , 1130-1139 (2005). https://doi.org/10.1097/01.mlr.0000182534.19832.83 van Walraven, C., Austin, P. C., Jennings, A., Quan, H. & Forster, A. J. A modification of the Elixhauser comorbidity measures into a point system for hospital death using administrative data. Med Care 47 , 626-633 (2009). https://doi.org/10.1097/MLR.0b013e31819432e5 Kursa, M. B. & Rudnicki, W. R. Feature Selection with the Boruta Package. J Stat Softw 36 , 1-13 (2010). https://doi.org/10.18637/jss.v036.i11 Ding, C. & Peng, H. Minimum Redundancy Feature Selection from Microarray Gene Expression Data. J Bioinform Comput Biol 3 , 185-205 (2005). https://doi.org/10.1142/S0219720005001004 van der Laan, M. J., Polley, E. C. & Hubbard, A. E. Super Learner. Stat Appl Genet Mol Biol 6 , Article 25 (2007). https://doi.org/10.2202/1544-6115.1309 Mortazavi, B. J. et al. Analysis of Machine Learning Techniques for Heart Failure Readmissions. Circ Cardiovasc Qual Outcomes 9 , 629-640 (2016). https://doi.org/10.1161/CIRCOUTCOMES.116.003039 Additional Declarations Competing interest reported. Competing Interests Dr. El Moudden reports research funding from the American Heart Association (Award #23SCISA1145640) during the conduct of this study; no other relationships or activities that could appear to have influenced the submitted work. Dr. Dodani reports the following relationships: Vice-Chair, American Heart Association Health Equity Research Network (HERN) Oversight Advisory Committee; subaward through the University of Illinois at Chicago for collaborative research support on this project (Award #23SCISA1145640). She holds appointments at the University of Illinois College of Medicine–Peoria (Founding Director, Center 4 Health Research; Professor of Clinical Medicine, Department of Medicine) and at the Macon & Joan Brock Virginia Health Sciences, Eastern Virginia Medical School at Old Dominion University (Professor of Medicine, Community Non-Tenure Track). No other relationships or activities that could appear to have influenced the submitted work. Mr. Bittner reports no disclosures. Supplementary Files npjDMSupplementaryInformationIEM3112026.docx Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 16 May, 2026 Reviewers agreed at journal 15 May, 2026 Reviews received at journal 07 Apr, 2026 Reviewers agreed at journal 20 Mar, 2026 Reviewers invited by journal 17 Mar, 2026 Editor assigned by journal 14 Mar, 2026 Submission checks completed at journal 14 Mar, 2026 First submitted to journal 11 Mar, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9098008","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":607811486,"identity":"603b4868-7a6c-4671-9380-6da497aaf985","order_by":0,"name":"Ismail El Moudden","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAArklEQVRIiWNgGAWjYDACdhDBJiHHwJAAYhCjhRmixZhkLQyJDURr4WfmTnzMU2aR3s+eY8DwoewwYS2SzbybjXnOSeTO7HljwDjjHBFaDA7zbpPmbZPI3XAjx4CZt40ILfZQLekGIC1/idECNBmsJQGshZEYLRKHeTcbzjknYTiz51nBwZ5z6YS18Lf3bnzwpqxOnp89eeODH2XWhLWggAMkqh8Fo2AUjIJRgAsAAHGQM94rNKgWAAAAAElFTkSuQmCC","orcid":"","institution":"Macon \u0026 Joan Brock Virginia Health Sciences at Old Dominion University","correspondingAuthor":true,"prefix":"","firstName":"Ismail","middleName":"El","lastName":"Moudden","suffix":""},{"id":607811490,"identity":"43a9421b-7de9-4c4c-bd0a-7a32f3a48901","order_by":1,"name":"Michael Bittner","email":"","orcid":"","institution":"Macon \u0026 Joan Brock Virginia Health Sciences at Old Dominion University","correspondingAuthor":false,"prefix":"","firstName":"Michael","middleName":"","lastName":"Bittner","suffix":""},{"id":607811491,"identity":"bd98ebab-e072-4976-8c8f-7192a9300341","order_by":2,"name":"Sunita Dodani","email":"","orcid":"","institution":"University of Illinois College of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Sunita","middleName":"","lastName":"Dodani","suffix":""}],"badges":[],"createdAt":"2026-03-11 20:54:06","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9098008/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9098008/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":104999458,"identity":"b17f1fcb-82a3-439d-b200-0bc701cb15ab","added_by":"auto","created_at":"2026-03-19 16:33:55","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":730653,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStudy Design and Cohort Flow\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003csup\u003eCONSORT-style diagram showing cohort derivation from the VHI statewide database (157,791 discharges from 123,272 patients), patient-grouped 70/30 train-test split, four CVD category subsets, balanced and unbalanced configurations, 4 base algorithms plus Super Learner ensemble, totaling 40 models. Abbreviations: AF/AFL, atrial fibrillation/flutter; AMI, acute myocardial infarction; CVD, cardiovascular disease; HF, heart failure; HHD, hypertensive heart disease; ML, machine learning; SMOTE, Synthetic Minority Oversampling Technique; VHI, Virginia Health Information.\u003c/sup\u003e\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-9098008/v1/e0391c1ea4354f4868108a69.png"},{"id":104999459,"identity":"038fe007-e4e3-452c-8457-a91ccee991f5","added_by":"auto","created_at":"2026-03-19 16:33:55","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1066427,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eReceiver Operating Characteristic Curves by CVD Category\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003csup\u003eROC curves for all algorithms across four CVD categories (HF, AMI, AF/AFL, HHD) under balanced and unbalanced configurations. AUC values reported in legends. Abbreviations: AF/AFL, atrial fibrillation/flutter; AMI, acute myocardial infarction; AUC, area under the receiver operating characteristic curve; CI, confidence interval; CVD, cardiovascular disease; FPR, false positive rate; HF, heart failure; HHD, hypertensive heart disease; ROC, receiver operating characteristic; TPR, true positive rate.\u0026nbsp;\u003c/sup\u003e\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-9098008/v1/e36309cdab36dda96da5c0a2.png"},{"id":104999463,"identity":"939affa1-ab4f-4735-8626-2a2760c38505","added_by":"auto","created_at":"2026-03-19 16:33:55","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":294520,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAUC Performance Heatmap\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003csup\u003eHeatmap displaying AUC values for all 40 models (5 algorithms × 4 CVD categories × 2 configurations). Color scale indicates utility classification. Abbreviations: AF/AFL, atrial fibrillation/flutter; AMI, acute myocardial infarction; AUC, area under the receiver operating characteristic curve; CI, confidence interval; CVD, cardiovascular disease; HF, heart failure; HHD, hypertensive heart disease.\u003c/sup\u003e\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-9098008/v1/b70453c2556b156c6929f048.png"},{"id":105035239,"identity":"f2617e28-d278-405b-adee-16b0f7337dfc","added_by":"auto","created_at":"2026-03-20 07:25:43","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":478611,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCalibration Plots for Best-Performing Models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003csup\u003eCalibration curves for the best model per CVD category with calibration slope, intercept, CITL, and E/O ratio annotated. Abbreviations: AF/AFL, atrial fibrillation/flutter; AMI, acute myocardial infarction; CI, confidence interval; CVD, cardiovascular disease; E/O, expected-to-observed ratio; HF, heart failure; HHD, hypertensive heart disease.\u003c/sup\u003e\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-9098008/v1/8b8361d1bb3ea47e2a5ed625.png"},{"id":104999466,"identity":"e237ac44-f396-45f4-bb5a-d28bde2f1223","added_by":"auto","created_at":"2026-03-19 16:33:56","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":597938,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFeature Importance: Top 15 Predictors by Mean Gain\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003csup\u003eHorizontal bar chart showing top 15 features by mean gain-based importance across best-performing models, colored by feature domain. Abbreviations: AF/AFL, atrial fibrillation/flutter; AMI, acute myocardial infarction; AUC, area under the receiver operating characteristic curve; CVD, cardiovascular disease; HF, heart failure; HHD, hypertensive heart disease.\u003c/sup\u003e\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-9098008/v1/abe33fe0db087d1088e6eaf7.png"},{"id":104999461,"identity":"9be32e80-9a8b-44c2-83a4-fcb0c26855d0","added_by":"auto","created_at":"2026-03-19 16:33:55","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":637666,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSHAP Beeswarm Summary Plots\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003csup\u003eSHAP summary plots for the best-performing model per CVD category showing top 20 features with direction and magnitude of effects on predicted readmission risk. Abbreviations: AF/AFL, atrial fibrillation/flutter; AMI, acute myocardial infarction; AUC, area under the receiver operating characteristic curve; CVD, cardiovascular disease; HF, heart failure; HHD, hypertensive heart disease; SHAP, SHapley Additive exPlanations.\u003c/sup\u003e\u003c/p\u003e","description":"","filename":"image6.png","url":"https://assets-eu.researchsquare.com/files/rs-9098008/v1/43a950e6a124494299f58457.png"},{"id":105562556,"identity":"e8d8833d-2982-42c5-8f5f-2e3332de7752","added_by":"auto","created_at":"2026-03-27 12:42:48","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":5854783,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9098008/v1/bcd11540-a2f3-48d5-a49f-742e7dd00f30.pdf"},{"id":104999464,"identity":"c2968ef1-5722-4d6b-9b49-6c53272c7478","added_by":"auto","created_at":"2026-03-19 16:33:55","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":1009675,"visible":true,"origin":"","legend":"","description":"","filename":"npjDMSupplementaryInformationIEM3112026.docx","url":"https://assets-eu.researchsquare.com/files/rs-9098008/v1/6c198facc1095985901dfddb.docx"}],"financialInterests":"Competing interest reported. Competing Interests\nDr. El Moudden reports research funding from the American Heart Association (Award #23SCISA1145640) during the conduct of this study; no other relationships or activities that could appear to have influenced the submitted work. Dr. Dodani reports the following relationships: Vice-Chair, American Heart Association Health Equity Research Network (HERN) Oversight Advisory Committee; subaward through the University of Illinois at Chicago for collaborative research support on this project (Award #23SCISA1145640). She holds appointments at the University of Illinois College of Medicine–Peoria (Founding Director, Center 4 Health Research; Professor of Clinical Medicine, Department of Medicine) and at the Macon \u0026 Joan Brock Virginia Health Sciences, Eastern Virginia Medical School at Old Dominion University (Professor of Medicine, Community Non-Tenure Track). No other relationships or activities that could appear to have influenced the submitted work. Mr. Bittner reports no disclosures.","formattedTitle":"Condition-Specific Readmission Risk Stratification in a Predominantly Black Statewide Cohort Using Machine Learning: Development of Subtype-Specific Models for Heart Failure, Acute Myocardial Infarction, Atrial Fibrillation/Flutter, and Hypertensive Heart Disease","fulltext":[{"header":"Introduction","content":"\u003cp\u003eCardiovascular disease (CVD) remains the leading cause of mortality in the United States, accounting for approximately 931,578 deaths annually and \u003cspan\u003e$\u003c/span\u003e252\u0026nbsp;billion in direct medical costs.\u003csup\u003e1,2\u003c/sup\u003e Heart failure (HF), acute myocardial infarction (AMI), atrial fibrillation and flutter (AF/AFL), and hypertensive heart disease (HHD) are the principal drivers of cardiovascular hospitalization and 30-day readmission.\u003csup\u003e1\u0026ndash;3\u003c/sup\u003e Reported 30-day readmission rates range from 18% to 25% for HF,\u003csup\u003e3\u0026ndash;5\u003c/sup\u003e 12% to 17% for AMI,\u003csup\u003e6,7\u003c/sup\u003e and 10% to 18% for AF/AFL,\u003csup\u003e8,9\u003c/sup\u003e while HHD drives acute care utilization with disproportionately elevated rates in minority populations.\u003csup\u003e2,10\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eCenters for Medicare \u0026amp; Medicaid Services (CMS) value-based penalty programs, including the Hospital Readmissions Reduction Program (HRRP) and the Hospital-Acquired Condition Reduction Program (HACRP), have produced modest improvements in quality metrics, yet safety-net hospitals serving predominantly Black and low-income patients bear a disproportionate share of financial penalties.\u003csup\u003e11\u0026ndash;13\u003c/sup\u003e Racial differences in AMI readmission persisted after HRRP and were attributable to patient-level rather than hospital-level factors.\u003csup\u003e12\u003c/sup\u003e Black patients hospitalized for HF have 3.9% to 6.8% higher composite readmission and mortality rates than White patients across deprivation strata, even after covariate adjustment.\u003csup\u003e14\u003c/sup\u003e Historical structural factors such as residential redlining compound these effects at the population level rather than at individual hospitals, pointing to the need for upstream interventions.\u003csup\u003e15,16\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eMachine learning (ML) handles high-dimensional data and non-linear relationships more flexibly than logistic regression; AUC values from 0.51 to 0.93 have been reported across heart failure readmission studies.\u003csup\u003e17\u0026ndash;20\u003c/sup\u003e Most of these studies lacked external validation, rarely assessed calibration, and relied on racially heterogeneous populations unlikely to generalize to racially concentrated cohorts.\u003csup\u003e20\u003c/sup\u003e Integrating social determinants of health (SDOH) into ML models has improved prediction in other cardiovascular contexts (AUC 0.694 to 0.823 for stroke), but application to readmission in predominantly minority populations has not been studied.\u003csup\u003e21,22\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eDespite this growing body of work, ML models for CVD readmission have rarely been developed in predominantly minority populations,\u003csup\u003e14,16,19,20\u003c/sup\u003e head-to-head algorithm comparisons across distinct cardiovascular conditions using standardized feature sets remain sparse,\u003csup\u003e5,19,20\u003c/sup\u003e and no prior study has integrated validated clinical indices with administrative SDOH proxy data into condition-specific models within racially concentrated cohorts.\u003csup\u003e14,21,22\u003c/sup\u003e We compared four ML algorithms (XGBoost, LightGBM, Random Forest, Elastic Net) and a Super Learner stacked ensemble across four conditions (HF, AMI, AF/AFL, HHD) in 157,791 discharge records from 123,272 unique patients (96.6% Black/African American) using an algorithm benchmarking design (4 base algorithms \u0026times; 4 conditions \u0026times; 2 dataset configurations plus 8 ensemble models, 40 total; TRIPOD Type 1a: development)\u003csup\u003e23\u003c/sup\u003e with validated clinical indices and consensus-selected features. CVD conditions were classified using AHA-aligned ICD-9/ICD-10 hierarchies. The study design follows TRIPOD guidelines for model development studies (Type 1a),\u003csup\u003e23\u003c/sup\u003e with the crossed design extending the framework to support simultaneous algorithm comparison within each condition. Similar crossed designs have been used in cardiovascular ML benchmarking.\u003csup\u003e5,18\u0026ndash;20\u003c/sup\u003e This study was conducted to evaluate the performance of machine learning models across different cardiovascular disease (CVD) subtypes in a minority population and to identify the relative importance of validated clinical indices and administrative social determinants of health (SDOH) proxies, particularly insurance status and comorbidity burden, on condition-specific risk prediction. We hypothesize that (1) ML model performance will differ across CVD subtypes, with condition-specific risk architectures requiring tailored modeling, and (2) validated clinical indices and administrative SDOH proxies particularly insurance status and comorbidity burden, will be significant predictors, with their relative importance varying by CVD subtype in this minority population. HHD was included as a standalone category because it disproportionately affects Black populations, who constitute 96.6% of our cohort, and is a leading contributor to heart failure progression, yet it has been largely absent from the readmission prediction literature as a separately modeled condition.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e \u003cb\u003eStudy Design and Data Source.\u003c/b\u003e We conducted a retrospective cohort study using de-identified inpatient discharge records from the Virginia Health Information (VHI) statewide all-payer database (January 2010 through December 2020).\u003csup\u003e24\u003c/sup\u003e VHI captures demographic, clinical, financial, and administrative data from all acute-care hospitalizations in Virginia (approximately 9.24\u0026nbsp;million total discharges). The study objective was to develop and compare ML models for 30-day all-cause readmission across four CVD subtypes in a predominantly minority population. Cohort flow is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eCVD Identification.\u003c/b\u003e CVD patients were identified from the primary diagnosis field using AHA guideline-aligned hierarchical classification. Four categories were defined: HF, AMI, AF/AFL, and HHD. Dual-era diagnostic codes (ICD-9 and ICD-10) were mapped to each category; complete code lists are available in Supplementary \u003cb\u003eTable S1\u003c/b\u003e.\u003c/p\u003e \u003cp\u003e \u003cb\u003eExclusion Criteria.\u003c/b\u003e We excluded records with non-CVD primary diagnoses, race/ethnicity other than Black/African American or Hispanic/Latino (per the grant\u0026rsquo;s minority health focus), age\u0026thinsp;\u0026lt;\u0026thinsp;18 or \u0026gt;\u0026thinsp;89 years, planned readmissions, in-hospital death, or missing readmission outcome. After exclusions, 157,791 discharge records from 123,272 unique patients comprised the analytic cohort; no records were excluded for missing patient identifiers. Baseline characteristics are presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The analytic cohort was 96.6% Black/African American and 3.4% Hispanic/Latino, consistent with the grant\u0026rsquo;s minority health focus. Given the small Hispanic subgroup, no separate models were estimated for this population; however, model performance was confirmed to be comparable when restricted to Black patients alone (results available upon request).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003eBaseline Demographics and Clinical Characteristics by Readmission Status\u003c/b\u003e\u003c/p\u003e \u003cp\u003e(N\u0026thinsp;=\u0026thinsp;157,791 Discharge Records, 123,272 Unique patients, 1.28 Discharge-to-patient ratio, 18.9% Readmission rate)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCharacteristic\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOverall \u003c/p\u003e \u003cp\u003e(N\u0026thinsp;=\u0026thinsp;157,791)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo Readmission \u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;127,913)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eReadmission \u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;29,878)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSMD\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"5\" nameend=\"c5\" namest=\"c1\"\u003e \u003cp\u003eDemographics and Encounter Characteristics\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge, years, mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e62.75 (14.19)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e62.55 (14.17)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e63.60 (14.23)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.074\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale sex, %\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e49.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e49.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e51.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.036\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLength of stay, days, mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.99 (6.01)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.81 (5.98)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5.73 (6.08)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.151\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal diagnoses, mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14.41 (4.14)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e14.13 (4.24)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e15.60 (3.44)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.381\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal procedures, mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.72 (1.96)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.76 (1.99)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.56 (1.85)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.100\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal charges (\u003cspan\u003e$\u003c/span\u003e), mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e47,910 (88,915)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e47,727 (88,549)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e48,692 (90,459)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.011\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAdmission Order, mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.49 (1.59)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.31 (1.04)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.28 (2.82)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.456\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eN Admissions per Patient, mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.99 (2.63)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.60 (1.68)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3.65 (4.59)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.592\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"5\" nameend=\"c5\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eValidated Clinical Indices\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLACE Index (0\u0026ndash;19), mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9.67 (3.02)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e9.34 (2.96)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e11.09 (2.86)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.604\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCharlson Comorbidity Index, mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.15 (2.05)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.99 (2.04)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3.81 (1.99)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.405\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge-Adjusted Charlson, mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5.04 (2.65)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.87 (2.65)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5.78 (2.48)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.357\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003evan Walraven Elixhauser Index, mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11.99 (8.74)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e11.34 (8.70)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e14.74 (8.37)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.398\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eElixhauser Comorbidity Count, mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.63 (2.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.49 (2.00)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5.27 (1.85)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.410\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCVD Severity Score, mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.49 (0.81)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.45 (0.83)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3.67 (0.65)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.294\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCVD Mortality Risk, mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.65 (0.90)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.62 (0.92)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.79 (0.83)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.199\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal CVD Conditions, mean (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.57 (1.94)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.50 (1.94)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4.88 (1.91)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.195\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eLACE Risk Category, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.502\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLow\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6,188 (3.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5,827 (4.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e361 (1.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModerate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e67,567 (42.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e59,606 (46.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7,961 (26.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHigh\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e84,036 (53.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e62,480 (48.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e21,556 (72.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eRace/Ethnicity, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.066\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBlack\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e152,492 (96.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e123,344 (96.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e29,148 (97.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHispanic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5,299 (3.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4,569 (3.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e730 (2.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eInsurance Type, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.342\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCommercial\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e33,341 (21.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e29,266 (22.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4,075 (13.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMedicaid\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e16,868 (10.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12,881 (10.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3,987 (13.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMedicare\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e92,250 (58.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e72,024 (56.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e20,226 (67.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSelf-Pay/Uninsured\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e15,332 (9.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e13,742 (10.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1,590 (5.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePrimary CVD Category, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.302\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAcute Myocardial Infarction\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e30,126 (19.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e26,207 (20.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3,919 (13.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAtrial Fibrillation/Flutter\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e18,782 (11.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e15,902 (12.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2,880 (9.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHeart Failure\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e94,562 (59.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e73,253 (57.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e21,309 (71.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHypertensive Heart Disease\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14,321 (9.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12,551 (9.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1,770 (5.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTeaching Hospital Status, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.083\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eACGME\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e22,062 (14.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e18,105 (14.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3,957 (13.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCouncil for GME\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e13,626 (8.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e11,029 (8.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2,597 (8.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCouncil of Teaching\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9,082 (5.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7,006 (5.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2,076 (7.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCouncil of Teaching Hospitals\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e19,534 (12.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e15,447 (12.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4,087 (13.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNone\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e93,325 (59.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e76,190 (59.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e17,135 (57.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eHospital Size (beds), n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.069\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSmall\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e17,088 (10.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e13,823 (10.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3,265 (10.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMedium\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e60,458 (38.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e49,587 (38.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10,871 (36.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLarge\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e47,702 (30.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e38,757 (30.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e8,945 (29.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMajor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e32,543 (20.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e25,746 (20.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6,797 (22.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eHospital Ownership, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.028\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNot-for-Profit\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e125,919 (79.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e102,345 (80.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e23,574 (78.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eProprietary\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e31,871 (20.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e25,567 (20.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6,304 (21.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"5\" nameend=\"c5\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eElixhauser Comorbidities (prevalence), %\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCongestive Heart Failure\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.268\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRenal Failure\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e59\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.325\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCardiac Arrhythmia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e37\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.132\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChronic Pulmonary Disease\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e33\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.199\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFluid \u0026amp; Electrolyte Disorder\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.120\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHypertension, Complicated\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e29\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.124\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eObesity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e29\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.085\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHypertension, Uncomplicated\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.201\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiabetes, Uncomplicated\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.018\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiabetes with Complications\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.142\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePulmonary Circulation Disorder\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.135\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eValvular Disease\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.067\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePeripheral Vascular Disease\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.079\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDepression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.079\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHypothyroidism\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.050\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOther Neurological Disorder\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.090\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDrug Abuse\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.055\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDeficiency Anemia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.047\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAlcohol Abuse\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.008\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCoagulopathy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.059\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWeight Loss\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.098\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLiver Disease\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.087\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRheumatoid Arthritis/Collagen\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.049\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePsychoses\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.057\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSolid Tumor w/o Metastasis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.063\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMetastatic Cancer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.061\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLymphoma\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.044\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBlood Loss Anemia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.032\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHIV/AIDS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.033\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePeptic Ulcer Disease\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.013\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eParalysis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.013\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003e\u003csup\u003eAbbreviations: ACGME, Accreditation Council for Graduate Medical Education; CVD, cardiovascular disease; GME, Graduate Medical Education; LACE, Length of stay + Acuity + Comorbidities + Emergency visits; SD, standard deviation; SMD, standardized mean difference\u003c/sup\u003e.\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003e\u003csup\u003eTotal discharge records: N = 157,791 from 123,272 unique patients (discharge\u0026minus;to\u0026minus;patient ratio = 1.28)\u003c/sup\u003e.\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003e\u003csup\u003eValues are mean (SD) for continuous variables and n (%) or prevalence (%) for categorical variables. SMD \u0026ge; 0.20 (bolded) indicates a meaningful difference between groups. For categorical variables with \u0026gt;2 levels, a single overall SMD is reported on the section header row\u003c/sup\u003e.\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003e\u003csup\u003eValidated indices computed from discharge data: LACE Index (van Walraven et al., 2010), Charlson Comorbidity Index (Charlson et al., 1987; Quan et al., 2011), van Walraven Elixhauser Index (van Walraven et al., 2009). All 31 Elixhauser comorbidity flags shown with individual SMDs\u003c/sup\u003e.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e\u003cp\u003e\u003cstrong\u003eTable 2. Best-Performing Models by CVD Category\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003csup\u003eAbbreviations: AUC, area under the receiver operating characteristic curve; AUPRC, area under the precision-recall curve; CI, confidence interval; CVD, cardiovascular disease; Sens, sensitivity; Spec, specificity.\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003e\u003csup\u003eBest model selected per category based on highest AUC across balanced (SMOTE) and unbalanced datasets. Both top-performing configurations shown per subtype. Brier score range: 0 (perfect) to 1 (worst). Lower values indicate better calibration. AUC interpretation: 0.70\u0026ndash;0.80 = moderate; \u0026gt;0.80 = high discrimination. Super Learner is a stacked ensemble combining XGBoost, LightGBM, Ranger, and Elastic Net via non-negative least squares meta-learning. All models validated using patient-grouped 70/30 splits with clustered bootstrap 95% confidence intervals (1,000 replicates).\u003c/sup\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.9306%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCVD Category\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14.9842%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDataset\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14.3533%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAlgorithm\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7129%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUC (95% CI)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.46372%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUPRC\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.57098%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSens\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.57098%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSpec\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.41325%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBrier\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.9306%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAMI\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14.9842%;\"\u003e\n \u003cp\u003eUnbalanced\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14.3533%;\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7129%;\"\u003e\n \u003cp\u003e0.706 (0.691\u0026ndash;0.721)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.46372%;\"\u003e\n \u003cp\u003e0.275\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.57098%;\"\u003e\n \u003cp\u003e0.571\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.57098%;\"\u003e\n \u003cp\u003e0.728\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.41325%;\"\u003e\n \u003cp\u003e0.213\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.9306%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14.9842%;\"\u003e\n \u003cp\u003eBalanced\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14.3533%;\"\u003e\n \u003cp\u003eElastic Net\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7129%;\"\u003e\n \u003cp\u003e0.697 (0.679\u0026ndash;0.713)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.46372%;\"\u003e\n \u003cp\u003e0.266\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.57098%;\"\u003e\n \u003cp\u003e0.679\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.57098%;\"\u003e\n \u003cp\u003e0.607\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.41325%;\"\u003e\n \u003cp\u003e0.203\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.9306%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAF/AFL\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14.9842%;\"\u003e\n \u003cp\u003eBalanced\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14.3533%;\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7129%;\"\u003e\n \u003cp\u003e0.732 (0.715\u0026ndash;0.750)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.46372%;\"\u003e\n \u003cp\u003e0.339\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.57098%;\"\u003e\n \u003cp\u003e0.708\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.57098%;\"\u003e\n \u003cp\u003e0.648\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.41325%;\"\u003e\n \u003cp\u003e0.120\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.9306%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14.9842%;\"\u003e\n \u003cp\u003eUnbalanced\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14.3533%;\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7129%;\"\u003e\n \u003cp\u003e0.731 (0.714\u0026ndash;0.749)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.46372%;\"\u003e\n \u003cp\u003e0.339\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.57098%;\"\u003e\n \u003cp\u003e0.658\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.57098%;\"\u003e\n \u003cp\u003e0.684\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.41325%;\"\u003e\n \u003cp\u003e0.212\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.9306%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHF\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14.9842%;\"\u003e\n \u003cp\u003eBalanced\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14.3533%;\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7129%;\"\u003e\n \u003cp\u003e0.708 (0.701\u0026ndash;0.716)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.46372%;\"\u003e\n \u003cp\u003e0.418\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.57098%;\"\u003e\n \u003cp\u003e0.632\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.57098%;\"\u003e\n \u003cp\u003e0.680\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.41325%;\"\u003e\n \u003cp\u003e0.158\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.9306%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14.9842%;\"\u003e\n \u003cp\u003eUnbalanced\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14.3533%;\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7129%;\"\u003e\n \u003cp\u003e0.707 (0.700\u0026ndash;0.714)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.46372%;\"\u003e\n \u003cp\u003e0.418\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.57098%;\"\u003e\n \u003cp\u003e0.678\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.57098%;\"\u003e\n \u003cp\u003e0.630\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.41325%;\"\u003e\n \u003cp\u003e0.215\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.9306%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHHD\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14.9842%;\"\u003e\n \u003cp\u003eBalanced\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14.3533%;\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7129%;\"\u003e\n \u003cp\u003e0.758 (0.735\u0026ndash;0.777)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.46372%;\"\u003e\n \u003cp\u003e0.338\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.57098%;\"\u003e\n \u003cp\u003e0.734\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.57098%;\"\u003e\n \u003cp\u003e0.652\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.41325%;\"\u003e\n \u003cp\u003e0.116\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.9306%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14.9842%;\"\u003e\n \u003cp\u003eUnbalanced\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 14.3533%;\"\u003e\n \u003cp\u003eSuper Learner\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 22.7129%;\"\u003e\n \u003cp\u003e0.754 (0.731\u0026ndash;0.776)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9.46372%;\"\u003e\n \u003cp\u003e0.338\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.57098%;\"\u003e\n \u003cp\u003e0.680\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.57098%;\"\u003e\n \u003cp\u003e0.698\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7.41325%;\"\u003e\n \u003cp\u003e0.096\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n \u003cp\u003e \u003cb\u003eUnit of Analysis and Patient Grouping.\u003c/b\u003e The unit of analysis was the discharge record, matching the HRRP penalty structure.\u003csup\u003e11,12\u003c/sup\u003e Because patients could contribute multiple admissions, all data partitions were grouped at the patient level to prevent information leakage. A 70/30 train-test split assigned all discharges from a given patient exclusively to one partition. Hyperparameter tuning used inner 5-fold patient-grouped cross-validation. Zero patient overlap was programmatically verified. Prior large-scale administrative readmission studies have used the same grouping strategy.\u003csup\u003e4,6,18,20\u003c/sup\u003e To evaluate whether discrimination differed by admission history, test-set performance was stratified by admission order (first admission vs. subsequent admissions) for each best-performing model (\u003cb\u003eSupplementary Table S6\u003c/b\u003e).\u003c/p\u003e \u003cp\u003e \u003cb\u003eOutcome Definition.\u003c/b\u003e The primary outcome was 30-day all-cause readmission to any Virginia acute-care hospital, identified through longitudinal patient linkage using Readmissions and Transfers (RATs) supplemental files.\u003c/p\u003e \u003cp\u003e \u003cb\u003eFeature Engineering.\u003c/b\u003e Raw administrative data were supplemented with validated clinical indices pre-computed during data extraction: the LACE Index (van Walraven et al., 2010),\u003csup\u003e32\u003c/sup\u003e Charlson Comorbidity Index (Charlson 1987; Quan 2005),\u003csup\u003e33,34\u003c/sup\u003e Age-Adjusted Charlson Index, van Walraven Elixhauser Index (van Walraven et al., 2009),\u003csup\u003e35\u003c/sup\u003e and Elixhauser Comorbidity Count. All 31 AHRQ Elixhauser comorbidity flags were pre-computed from secondary diagnosis fields (DX2\u0026ndash;DX18) using published ICD-9/ICD-10 mappings.\u003csup\u003e25\u003c/sup\u003e Additional features captured hospital characteristics (teaching status, ownership, bed size), admission characteristics (length of stay, total diagnoses, total procedures), and patient history (admission order [sequential index of each discharge], admissions per patient [cumulative count]). Complete feature specifications appear in Supplementary \u003cb\u003eTable S1\u003c/b\u003e.\u003c/p\u003e \u003cp\u003eThe study period (January 2010 through December 2020) spans the October 1, 2015 transition from ICD-9-CM to ICD-10-CM coding. Primary CVD diagnosis classification used published dual-era diagnostic mappings (\u003cb\u003eSupplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e\u003c/b\u003e). However, ICD-10\u0026rsquo;s substantially greater coding granularity may have inflated comorbidity counts and altered Elixhauser flag prevalence in the post-transition period, as previously documented in administrative database studies. Admission Year was included as a temporal feature (selected by 3 of 4 feature selection methods), which may partially absorb coding era effects. A descriptive comparison of key comorbidity features across coding eras is provided in \u003cb\u003eSupplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e0\u003c/b\u003e.\u003c/p\u003e \u003cp\u003e \u003cb\u003eFeature Selection.\u003c/b\u003e From 60 candidate features after technical filtering (removal of zero-variance, near-zero-variance features with \u0026gt;\u0026thinsp;95% single value, and one member of pairs with |r| \u0026gt; 0.90), four independent selection methods were applied to the training set: (a) Boruta all-relevant selection,\u003csup\u003e36\u003c/sup\u003e (b) SHAP-based importance from a reference XGBoost model, (c) recursive feature elimination with patient-grouped cross-validation, and (d) minimum redundancy maximum relevance (mRMR).\u003csup\u003e37\u003c/sup\u003e Features endorsed by at least two of four methods were retained, producing 45 consensus-selected features applied uniformly across all CVD categories and configurations. This data-driven consensus approach replaced the univariate correlation ranking used in earlier analyses, removing the arbitrary top-k cap so that retained features received converging evidence from multiple methodologically distinct selection criteria (Supplementary \u003cb\u003eTable S1\u003c/b\u003e, \u003cb\u003eFigure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e\u003c/b\u003e).\u003c/p\u003e \u003cp\u003e \u003cb\u003eMachine Learning Model Development.\u003c/b\u003e Four base algorithms were evaluated: XGBoost, LightGBM, Random Forest (ranger), Elastic Net (glmnet). Each was trained independently for each CVD category under two configurations (unbalanced and SMOTE-balanced), for a total of 32 base models. A Super Learner stacked ensemble combined out-of-fold predictions from all base learners through a non-negatively constrained logistic regression meta-learner,40 adding 8 ensemble models (one per CVD category per dataset configuration) for a total of 40 models (4 base \u0026times; 4 categories \u0026times; 2 configurations\u0026thinsp;+\u0026thinsp;8 ensemble; Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). All metrics were evaluated on the held-out test set, held out from training, tuning, and SMOTE augmentation. Balanced training sets used SMOTE with K\u0026thinsp;=\u0026thinsp;5 nearest neighbors applied exclusively to the training partition. For unbalanced datasets, class weights were assigned through algorithm-native methods (scale_pos_weight, inverse-class-frequency weighting, or native logistic loss). All algorithms underwent grid search with inner 5-fold patient-grouped cross-validation maximizing AUC. XGBoost and LightGBM each searched 27 configurations with early stopping at 15 rounds; Random Forest searched 18 configurations selected by out-of-bag error; Elastic Net optimized 11 alpha values with lambda selected by the one-standard-error rule. The term \u0026ldquo;ensemble\u0026rdquo; in the title refers to the Super Learner stacked ensemble approach, which combines out-of-fold predictions from all base learners through a non-negatively constrained logistic regression meta-learner.\u003c/p\u003e \u003cp\u003e \u003cb\u003ePerformance Assessment and Validation.\u003c/b\u003e Model performance followed TRIPOD guidelines\u003csup\u003e23\u003c/sup\u003e on the 30% held-out test set. Discrimination was assessed by AUC with 95% bootstrap CIs (1,000 iterations, percentile method) and area under the precision-recall curve (AUPRC), which is more informative than AUC-ROC for imbalanced outcomes. The optimal threshold was determined by maximizing the Youden index, at which sensitivity, specificity, positive and negative predictive values, F1 score, balanced accuracy, and Matthews correlation coefficient were computed. Clinical utility was classified as Excellent (\u0026ge;\u0026thinsp;0.80), High (\u0026ge;\u0026thinsp;0.75), Moderate (\u0026ge;\u0026thinsp;0.70), or Limited (\u0026lt;\u0026thinsp;0.70). Calibration was assessed using Brier scores, calibration slopes, calibration-in-the-large (CITL), and expected-to-observed (E/O) ratio per TRIPOD guidelines (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). ROC curves are presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e and the AUC performance heatmap in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Pairwise algorithm comparisons used DeLong tests within the best-performing dataset configuration per CVD category (\u003cb\u003eSupplementary Table S3\u003c/b\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo assess residual within-patient correlation, we fitted GEE models with exchangeable working correlation and sandwich standard errors on multi-admission patient subsets within each test partition. Patient-resampled clustered bootstrap CIs (1,000 iterations) were also computed. Results appear in \u003cb\u003eSupplementary Tables S4\u0026ndash;S5\u003c/b\u003e.\u003c/p\u003e \u003cp\u003e \u003cb\u003eFeature Importance Analysis.\u003c/b\u003e Feature importance was quantified using algorithm-native methods (gain-based for XGBoost, LightGBM; impurity-based for Random Forest; absolute standardized coefficients for Elastic Net), normalized to [0, 1] within each model. Three complementary perspectives are reported: condition-specific importance from each best-performing model (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e), aggregated importance averaged across all 32 base models requiring appearance in at least three (\u003cb\u003eSupplementary Tables S7 and S8\u003c/b\u003e), and SHapley Additive exPlanations (SHAP)-based importance providing directional, additive decompositions grounded in cooperative game theory (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e, \u003cb\u003eSupplementary Figures S3\u0026ndash;S4\u003c/b\u003e). Algorithm-native importance served as the primary measure; SHAP analysis provides supplementary mechanistic insight.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003ePost-hoc SHAP analysis was performed on each condition\u0026rsquo;s best-performing model (\u003cb\u003eTable\u0026nbsp;2\u003c/b\u003e) using shapviz in R. Exact TreeSHAP values were computed for tree-based models (XGBoost, LightGBM) via native contribution output. SHAP values were computed on a random subsample of 500 test observations per condition (set.seed(42)). Three visualizations were generated: beeswarm summary plots (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e), waterfall plots for representative true positive and true negative cases (\u003cb\u003eSupplementary Figure S4\u003c/b\u003e), and dependence plots for continuous predictors (\u003cb\u003eSupplementary Figure S3\u003c/b\u003e). SHAP analysis was performed post-hoc and did not influence model selection or hyperparameter tuning. SHAP values were computed on a random subsample of 500 test observations per CVD category due to computational cost. The subsample was verified for representativeness by comparing event rates and distributions of the top 5 predictors (LACE Index, Charlson Comorbidity Index, Length of Stay, Elixhauser Comorbidity Count, and Admission Order) between the subsample and the full test set, confirming no meaningful divergence (all standardized mean differences\u0026thinsp;\u0026lt;\u0026thinsp;0.10).\u003c/p\u003e \u003cp\u003e \u003cb\u003eModel Specification and Reproducibility.\u003c/b\u003e Tree ensemble models (XGBoost, LightGBM, Random Forest) and the Super Learner cannot be expressed as closed-form equations; their prediction logic is distributed across learned parameters. Serialized model objects containing all trained models, preprocessing parameters, hyperparameter configurations, selected features, and optimal thresholds were archived and are available from the corresponding author. A standalone prediction function accepting raw patient-level input and returning readmission probabilities was also archived for prospective deployment.\u003c/p\u003e \u003cp\u003eFor Elastic Net, which yields interpretable coefficient vectors, \u003cb\u003eSupplementary Table S8\u003c/b\u003e reports the full coefficient matrix (intercept plus 45 features) for each CVD category under both dataset configurations. These coefficients enable exact reproduction of predicted log-odds without access to the serialized model objects: log-odds(readmission) = β₀ + Σβⱼxⱼ, where β values are provided in Supplementary \u003cb\u003eTable S8\u003c/b\u003e and λ was selected by the one-standard-error rule.\u003c/p\u003e \u003cp\u003e \u003cb\u003eAlgorithmic Bias Assessment.\u003c/b\u003e Bias assessment included demographic parity across insurance types and age groups, equalized odds evaluation, calibration consistency analysis, and post-hoc auditing of differential prediction by insurance type.\u003c/p\u003e \u003cp\u003e \u003cb\u003eStatistical Analysis.\u003c/b\u003e All analyses were conducted in R version 4.4.2 using xgboost, lightgbm, ranger, glmnet, caret, smotefamily, pROC, geepack, Boruta, mRMRe, shapviz, and tableone. Baseline characteristics were compared using Welch\u0026rsquo;s t-tests and Pearson\u0026rsquo;s chi-squared tests. Effect sizes were reported as standardized mean differences (SMDs) computed via the tableone package, with SMD\u0026thinsp;\u0026ge;\u0026thinsp;0.10 considered indicative of meaningful imbalance. All 31 Elixhauser comorbidity flags were pre-computed during data extraction using published AHRQ ICD-9/ICD-10 mappings applied to secondary diagnosis fields.\u003csup\u003e25\u003c/sup\u003e Two variables had missing values: Teaching Status (0.1%) and Hospital Ownership (\u0026lt;\u0026thinsp;0.01%), handled by mode imputation. Four candidate features (APR-DRG Severity of Illness, APR-DRG Risk of Mortality, Admission Type, Discharge Disposition) were unavailable in the analytic dataset and excluded. A fixed random seed (set.seed(42)) was applied prior to all stochastic procedures.\u003csup\u003e26\u003c/sup\u003e With a sample of 157,791 discharges, conventional significance tests (Welch\u0026rsquo;s t and Pearson\u0026rsquo;s chi-squared) have near-certainty of rejecting the null hypothesis for even trivially small differences. Therefore, while p-values are reported in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e for completeness, clinical meaningfulness of between-group differences was assessed using standardized mean differences (SMDs), with SMD\u0026thinsp;\u0026gt;\u0026thinsp;0.10 indicating meaningful imbalance per established thresholds (Austin, 2009). No formal multiplicity adjustment was applied to model selection, consistent with standard practice in machine learning benchmarking studies. For each CVD category, the best-performing model was selected from 10 candidates (5 algorithms \u0026times; 2 configurations); the optimistic bias inherent in selecting the maximum of multiple correlated estimates (\u0026ldquo;winner\u0026rsquo;s curse\u0026rdquo;) is acknowledged as a limitation.\u003c/p\u003e \u003cp\u003e \u003cb\u003eEthical Considerations.\u003c/b\u003e The study was approved by the Eastern Virginia Medical School IRB (IRB #23-09-NH-0248) with waiver of informed consent for de-identified data analysis. All analyses used de-identified data with no attempt to re-identify patients.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e \u003cb\u003eCohort Characteristics.\u003c/b\u003e The analytic cohort comprised 157,791 discharge records from 123,272 unique patients identified through the VHI statewide database (discharge-to-patient ratio: 1.28), with 83.3% of patients contributing a single admission and 16.7% contributing two or more admissions. No records were excluded for missing outcome or patient identifiers. All ML model training and evaluation used this full cohort; cohort flow is detailed in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Baseline demographics and clinical characteristics are presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eThe population was 96.6% Black/African American and 49.0% female, with a mean age of 62.75\u0026thinsp;\u0026plusmn;\u0026thinsp;14.19 years. HF was the most common primary CVD category (59.9% of discharges), followed by AMI (19.1%), AF/AFL (11.9%), and HHD (9.1%). Readmitted patients had higher LACE Index scores (11.09\u0026thinsp;\u0026plusmn;\u0026thinsp;2.86 vs 9.34\u0026thinsp;\u0026plusmn;\u0026thinsp;2.96; SMD\u0026thinsp;=\u0026thinsp;0.604), more admissions per patient (3.65\u0026thinsp;\u0026plusmn;\u0026thinsp;4.59 vs 1.60\u0026thinsp;\u0026plusmn;\u0026thinsp;1.68; SMD\u0026thinsp;=\u0026thinsp;0.592), higher admission order (2.28\u0026thinsp;\u0026plusmn;\u0026thinsp;2.82 vs 1.31\u0026thinsp;\u0026plusmn;\u0026thinsp;1.04; SMD\u0026thinsp;=\u0026thinsp;0.456), greater Elixhauser comorbidity count (5.27\u0026thinsp;\u0026plusmn;\u0026thinsp;1.85 vs 4.49\u0026thinsp;\u0026plusmn;\u0026thinsp;2.00; SMD\u0026thinsp;=\u0026thinsp;0.410), higher Charlson Comorbidity Index (3.81\u0026thinsp;\u0026plusmn;\u0026thinsp;1.99 vs 2.99\u0026thinsp;\u0026plusmn;\u0026thinsp;2.04; SMD\u0026thinsp;=\u0026thinsp;0.405), higher van Walraven Elixhauser Index (14.74\u0026thinsp;\u0026plusmn;\u0026thinsp;8.37 vs 11.34\u0026thinsp;\u0026plusmn;\u0026thinsp;8.70; SMD\u0026thinsp;=\u0026thinsp;0.398), more total diagnoses (15.60\u0026thinsp;\u0026plusmn;\u0026thinsp;3.44 vs 14.13\u0026thinsp;\u0026plusmn;\u0026thinsp;4.24; SMD\u0026thinsp;=\u0026thinsp;0.381), and higher prevalence of renal failure (59% vs 42%; SMD\u0026thinsp;=\u0026thinsp;0.325). Insurance type showed substantial separation (SMD\u0026thinsp;=\u0026thinsp;0.342), with Medicare overrepresented among readmissions (67.7% vs 56.3%) and commercial insurance underrepresented (13.6% vs 22.9%).\u003c/p\u003e \u003cp\u003e \u003cb\u003e30-Day Readmission Outcome.\u003c/b\u003e The overall 30-day all-cause readmission rate was 18.9% (n\u0026thinsp;=\u0026thinsp;29,878). Rates varied across CVD categories: HF 22.5%, AF/AFL 15.3%, AMI 13.0%, and HHD 12.4%. HF was disproportionately represented among readmissions (71.3% of readmitted discharges vs 57.3% of non-readmitted). Medicare beneficiaries accounted for 67.7% of readmissions compared with 56.3% of non-readmitted discharges (SMD\u0026thinsp;=\u0026thinsp;0.342), while self-pay/uninsured patients were underrepresented among readmissions (5.3% vs 10.7%). The LACE Risk Category showed the strongest categorical separation: 72.1% of readmitted patients were classified as high-risk compared with 48.8% of non-readmitted patients (SMD\u0026thinsp;=\u0026thinsp;0.502).\u003c/p\u003e \u003cp\u003e \u003cb\u003eModel Discrimination.\u003c/b\u003e From 60 candidate features after technical filtering, four independent selection methods yielded 45 consensus-selected features, applied uniformly across all CVD categories and dataset configurations. Forty prediction models were then trained by crossing four base algorithms plus a Super Learner ensemble with four CVD categories and two dataset configurations. All models used patient-grouped 70/30 splitting with zero patient leakage between partitions. Complete results for all 40 models appear in \u003cb\u003eSupplementary Table S2\u003c/b\u003e; the best-performing models per category are summarized in \u003cb\u003eTable\u0026nbsp;2\u003c/b\u003e.\u003c/p\u003e \u003cp\u003eDiscrimination varied across CVD subtypes, with all four categories achieving moderate or high discrimination (AUC 0.706\u0026ndash;0.758), though the highest-performing HHD model (AUC 0.758) exhibited calibration limitations (slope 1.479\u0026ndash;1.646) requiring recalibration before absolute risk estimation. HHD yielded the highest AUC: XGBoost on the balanced dataset achieved 0.758 (95% CI, 0.735 to 0.777), classified as high utility. The Super Learner ensemble was competitive for HHD on the unbalanced dataset (0.754; 95% CI, 0.731 to 0.776). AF/AFL followed at 0.732 (95% CI, 0.715 to 0.750; XGBoost, balanced), representing moderate utility. HF achieved 0.708 (95% CI, 0.701 to 0.716; XGBoost, balanced), with the narrow confidence interval driven by the large test sample (28,380 records from 21,149 patients). AMI achieved 0.706 (95% CI, 0.691 to 0.721; XGBoost, unbalanced). XGBoost was the best-performing base algorithm across all four categories. ROC curves are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e; the cross-category algorithm comparison appears in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eAUPRC values further characterized discrimination in the context of varying event rates: HF 0.418 (event rate 22.5%), AF/AFL 0.339 (15.3%), HHD 0.338 (12.4%), and AMI 0.275 (13.0%). The relatively higher AUPRC for HF reflects the higher base rate, while AMI\u0026rsquo;s lower AUPRC despite comparable AUC suggests that positive prediction remains challenging for the least prevalent readmission category.\u003c/p\u003e \u003cp\u003e \u003cb\u003eEffect of SMOTE Balancing.\u003c/b\u003e Balanced XGBoost achieved the best AUC for HHD (0.758) and HF (0.708), while unbalanced XGBoost was best for AMI (0.706). AF/AFL showed virtually no difference between balanced and unbalanced XGBoost (0.732 vs 0.731). Across categories, SMOTE generally improved sensitivity at the expense of specificity, a tradeoff that was most beneficial in categories with lower event rates (HHD 12.4%, AMI 13.0%; \u003cb\u003eSupplementary Table S9\u003c/b\u003e).\u003c/p\u003e \u003cp\u003e \u003cb\u003eModel Calibration.\u003c/b\u003e Calibration was assessed using Brier scores, calibration slopes, and E/O ratios (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). The AMI XGBoost model demonstrated excellent relative calibration (slope\u0026thinsp;=\u0026thinsp;0.993), indicating well-preserved rank ordering of predicted probabilities. However, the expected-to-observed (E/O) ratio of 3.47 indicates substantial overestimation of absolute risk, meaning the model predicted 3.47 times more readmission events than were observed. While the model reliably identifies higher-risk patients relative to lower-risk patients, absolute probability estimates for AMI require recalibration (e.g., Platt scaling or isotonic regression) before clinical deployment. HF models showed excellent calibration (balanced slope 1.014, unbalanced slope 1.025), with Brier scores of 0.158 (balanced) and 0.215 (unbalanced). AF/AFL exhibited slight underconfidence (slopes 1.037\u0026ndash;1.065). HHD showed the greatest miscalibration (slopes 1.479\u0026ndash;1.646), indicating systematic overconfidence at the extremes of predicted risk; Platt recalibration is recommended before deployment for this subtype. Raw Brier scores are reported in \u003cb\u003eTable\u0026nbsp;2\u003c/b\u003e; for all categories, rank-ordering of risk was preserved even when probability estimates required rescaling.\u003c/p\u003e \u003cp\u003e \u003cb\u003ePairwise Algorithm Comparisons.\u003c/b\u003e DeLong tests for correlated ROC curves compared all algorithm pairs within the best-performing dataset configuration per CVD category (\u003cb\u003eSupplementary Table S3\u003c/b\u003e). XGBoost significantly outperformed Random Forest and Elastic Net across most conditions. Within each condition, XGBoost and LightGBM were generally statistically indistinguishable, though XGBoost held a consistent numerical edge. The Super Learner was competitive with or slightly below the best base learner across all categories, with the closest match in HHD.\u003c/p\u003e \u003cp\u003e \u003cb\u003ePatient-Grouped Validation.\u003c/b\u003e GEE models with exchangeable working correlation and sandwich standard errors were fitted to the multi-admission subset of each category\u0026rsquo;s test partition (\u003cb\u003eSupplementary Table S4\u003c/b\u003e). The GEE coefficient for predicted risk was significant in HF (α\u0026thinsp;=\u0026thinsp;0.120, Wald P\u0026thinsp;\u0026lt;\u0026thinsp;0.001), AMI (α\u0026thinsp;=\u0026thinsp;0.041, Wald P\u0026thinsp;\u0026lt;\u0026thinsp;0.001), and HHD (α\u0026thinsp;=\u0026thinsp;0.141, Wald P\u0026thinsp;=\u0026thinsp;0.017), confirming that model-predicted risk remained a significant readmission predictor after accounting for within-patient correlation. AF/AFL showed a non-significant correlation (α\u0026thinsp;=\u0026thinsp;0.170, P\u0026thinsp;=\u0026thinsp;0.229), likely due to the smaller multi-admission sample (1,067 records from 461 patients).\u003c/p\u003e \u003cp\u003eClustered bootstrap resampling (1,000 iterations resampling patients with replacement) produced confidence intervals comparable to standard record-level bootstrap intervals (\u003cb\u003eSupplementary Table S5\u003c/b\u003e). The mean CI width ratio across categories was 1.13; patient clustering widened intervals by approximately 13% on average. No model\u0026rsquo;s point estimate fell outside its clustered interval, and all utility classifications remained unchanged. Taken together, the patient-grouped splitting strategy adequately controlled within-patient correlation.\u003c/p\u003e \u003cp\u003eThe clustered bootstrap confidence interval for HHD (0.733\u0026ndash;0.784) extends below the 0.75 threshold defining \u0026ldquo;high utility\u0026rdquo; in our classification schema, despite the point estimate (0.758) exceeding this threshold. The 21.2% CI widening from patient-level clustering for HHD was the largest among all four categories, reflecting greater within-patient correlation in this subgroup. The \u0026ldquo;high utility\u0026rdquo; classification for HHD should therefore be interpreted with this uncertainty in mind, particularly when considering prospective deployment.\u003c/p\u003e \u003cp\u003e \u003cb\u003ePerformance by Admission Order.\u003c/b\u003e To assess whether prior hospitalization history influenced discrimination, best-model AUC was stratified by admission order within each test partition (\u003cb\u003eSupplementary Table S6\u003c/b\u003e). Discrimination varied by admission history, with greater heterogeneity among repeat admissions in categories with lower event rates. The Super Learner provided the most stable performance across admission strata.\u003c/p\u003e \u003cp\u003eLACE Index Sensitivity Analysis. To assess the impact of construct overlap between the LACE Index composite score and its constituent components (LACE-L, LACE-A, LACE-C, LACE-E), we conducted a sensitivity analysis comparing three feature configurations using XGBoost: (1) the full consensus feature set including both the composite and components, (2) LACE composite index only (individual components removed), and (3) LACE components only (composite removed). AUC differences were negligible across configurations (\u003cb\u003eSupplementary Table S11).\u003c/b\u003e For AMI, HF, and AF/AFL, all three configurations produced identical AUCs to the third decimal place (all DeLong p\u0026thinsp;\u0026gt;\u0026thinsp;0.20). For HHD, the composite-only configuration yielded marginally higher AUC (0.755) than both the full set (0.747, DeLong p\u0026thinsp;=\u0026thinsp;0.053) and components-only (0.747, p\u0026thinsp;=\u0026thinsp;0.015), suggesting the LACE Index composite alone is sufficient and may even reduce noise from redundant component features. These results support using the composite LACE Index alone for parsimony and clinical interpretability, and confirm that construct overlap did not artificially inflate model discrimination.\u003c/p\u003e \u003cp\u003e \u003cb\u003eFeature Importance.\u003c/b\u003e Aggregated importance across all 32 base models produced a stable predictor ranking (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, \u003cb\u003eSupplementary Table S7\u003c/b\u003e). Validated clinical indices dominated: the LACE Index ranked among the top predictors in every category, in line with its strong separation between readmitted and non-readmitted patients (SMD\u0026thinsp;=\u0026thinsp;0.604). The Charlson Comorbidity Index, van Walraven Elixhauser Index, and Elixhauser Comorbidity Count also ranked highly, a pattern that follows the shift from ad-hoc composites to published indices with established reference ranges.\u003c/p\u003e \u003cp\u003eCategory-specific patterns supported the case for condition-specific modeling (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). In HF, the LACE Index and renal failure dominated, with admission order contributing strongly. In AMI, total diagnoses and the LACE Index led. In AF/AFL, the importance distribution was broader, with insurance type contributing more prominently. In HHD, the LACE Index, insurance type, and renal failure led, with obesity and depression contributing more than in other categories. Renal failure showed consistent importance across all categories (SMD\u0026thinsp;=\u0026thinsp;0.325 for readmission vs non-readmission), in keeping with the cardiorenal syndrome\u0026rsquo;s known role in CVD readmission.\u003c/p\u003e \u003cp\u003eElastic Net coefficient vectors (\u003cb\u003eSupplementary Table\u0026nbsp;7\u003c/b\u003e\u003c/p\u003e \u003cp\u003e) offered a view of predictor direction and regularization-driven sparsity. The LACE Index and renal failure carried the largest positive coefficients across conditions, while obesity consistently received negative coefficients, suggesting a protective association after adjustment for comorbidity burden.\u003c/p\u003e \u003cp\u003eSHAP analysis replicated the global importance rankings at the individual prediction level (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e). The LACE Index, admission order, renal failure, and insurance type produced the largest absolute SHAP values across predictions. Waterfall decompositions of representative cases (\u003cb\u003eSupplementary Figure S4\u003c/b\u003e) illustrated how these top features combined to produce individual risk estimates. Dependence plots (\u003cb\u003eSupplementary Figure S3\u003c/b\u003e) revealed non-linear relationships, including threshold effects for the LACE Index and differential effects of insurance type across predicted risk levels.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study is, to our knowledge, the largest ML-based cardiovascular readmission prediction analysis in a predominantly Black cohort and the first to compare five algorithms plus a Super Learner ensemble across four CVD subtypes using standardized validated features in a racially concentrated population.\u003c/p\u003e \u003cp\u003e \u003cb\u003eCondition-Specific Discrimination.\u003c/b\u003e All four CVD categories reached moderate or high predictive utility, exceeding most prior administrative data-based benchmarks. The HF AUC of 0.708 exceeded both Sharma and colleagues\u0026rsquo; benchmark of 0.65 using XGBoost on Canadian administrative data\u003csup\u003e27\u003c/sup\u003e and Mortazavi and colleagues\u0026rsquo; Random Forest AUC of 0.628 in the Tele-HF clinical trial cohort,\u003csup\u003e39\u003c/sup\u003e suggesting that validated indices and consensus feature selection narrow the discrimination gap previously attributed to administrative data limitations. Fine and colleagues achieved 0.74 for a composite HF endpoint using CatBoost with deep feature synthesis, though their outcome included ED revisits and mortality.\u003csup\u003e28\u003c/sup\u003e Awan and colleagues similarly reported a best AUC of 0.62 for HF readmission or death using administrative data with explicit attention to class imbalance.\u003csup\u003e18\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eThe HHD model reached high utility (AUC 0.758). Because most studies subsume HHD within broader HF or hypertension categories,\u003csup\u003e1,2\u003c/sup\u003e direct comparison with published benchmarks is limited. The narrow AUC range across algorithms for HHD likely reflects a more uniform pathophysiology, making this subtype a practical candidate for early deployment. AMI models achieved moderate utility (0.706) with the best calibration in the study (slope 0.993), within the range of published administrative data AMI models (AUC 0.65\u0026ndash;0.75 without clinical variables).\u003csup\u003e6,7\u003c/sup\u003e Pandey and colleagues found that racial differences in AMI readmission were attributable to patient-level rather than hospital-level factors,\u003csup\u003e12\u003c/sup\u003e which aligns with the strong insurance and comorbidity contributions in our AMI models. AF/AFL reached moderate utility (0.732), likely because validated indices better captured the comorbidity burden driving arrhythmia-related readmissions.\u003c/p\u003e \u003cp\u003e \u003cb\u003eAlgorithmic Hierarchy.\u003c/b\u003e XGBoost outperformed all alternatives across four categories, extending results from mixed-race populations\u003csup\u003e5,17,19\u003c/sup\u003e to a predominantly Black cohort with identical features. The Super Learner ensemble matched the best base learner most closely for HHD (unbalanced AUC 0.754 vs XGBoost 0.753), where LightGBM and Ranger received the largest meta-learner weights (\u003cb\u003eSupplementary Figure S2\u003c/b\u003e), suggesting that ensemble gains require base learners with complementary decision boundaries. Balanced datasets produced the best AUC in two of four categories (HHD, HF), with unbalanced best for AMI and no meaningful difference for AF/AFL.\u003c/p\u003e \u003cp\u003e \u003cb\u003eInsurance, Social Determinants, and Health Equity.\u003c/b\u003e The LACE Index was the strongest discriminator between readmitted and non-readmitted patients (SMD\u0026thinsp;=\u0026thinsp;0.604) and, unlike ad-hoc composite scores, has established reference ranges and external validity across settings.\u003csup\u003e32\u003c/sup\u003e Insurance type ranked among the top predictors (SMD\u0026thinsp;=\u0026thinsp;0.342). Bahiru and colleagues similarly showed that hospitals serving higher proportions of dual-eligible patients had significantly higher 30-day readmission rates.\u003csup\u003e30\u003c/sup\u003e The low readmission representation among self-pay and uninsured discharges (5.3% vs 10.7%) more likely reflects access barriers to rehospitalization than better outcomes,\u003csup\u003e11,14,22\u003c/sup\u003e meaning ML models trained on these data may underestimate true risk for uninsured patients. Chaiyachati and colleagues showed that racial disparities in readmission widened within safety-net hospitals after HRRP implementation, especially for conditions not targeted by HRRP,\u003csup\u003e31\u003c/sup\u003e adding weight to the case that HRRP risk adjustment models should incorporate social determinant measures. Condition-specific feature importance patterns, such as the prominent role of obesity and depression in HHD, would have been missed in pooled modeling.\u003c/p\u003e \u003cp\u003ePredictive contribution is distinct from clinical actionability. The LACE Index, renal failure, and obesity improve risk stratification but are not directly modifiable in the acute window. Insurance type is partly modifiable through policy yet also proxies for age, disability, and post-discharge access. Care process features (admission order, length of stay, hospital characteristics) identify where targeted transition-of-care interventions may reduce readmission probability. Model outputs should be interpreted accordingly: non-modifiable features refine risk classification, while care process features indicate intervention opportunities.\u003c/p\u003e \u003cp\u003e \u003cb\u003eFeature Importance Concordance.\u003c/b\u003e Algorithm-native, aggregated, and SHAP-based importance rankings all placed the LACE Index, renal failure, and insurance type among the top predictors, paralleling the emphasis on appropriate model and metric selection for administrative HF readmission data reported by Awan and colleagues.\u003csup\u003e18\u003c/sup\u003e The LACE Index, Charlson, and van Walraven Elixhauser indices carry published reference ranges and established predictive validity, making them more reproducible and interpretable than ad-hoc composites.\u003c/p\u003e \u003cp\u003eHowever, the LACE Index requires careful interpretation as a predictor because its components (length of stay, acuity of admission, comorbidity, and emergency department visits) overlap with readmission risk factors, creating potential construct overlap. While this does not invalidate its predictive utility, it means that LACE-based risk stratification identifies high-utilization patients through a partially tautological construct. SHAP dependence plots (\u003cb\u003eSupplementary Figure S3\u003c/b\u003e) revealed non-linear LACE\u0026ndash;readmission relationships, including diminishing marginal effects at the highest scores.\u003c/p\u003e \u003cp\u003eOur sensitivity analysis (\u003cb\u003eSupplementary Table S11\u003c/b\u003e) confirmed that the simultaneous inclusion of the LACE Index and its constituent components did not inflate model discrimination: AUC was unchanged across all three feature configurations for AMI, HF, and AF/AFL (all DeLong p\u0026thinsp;\u0026gt;\u0026thinsp;0.20), and the composite-only configuration actually performed marginally better for HHD (DeLong p\u0026thinsp;=\u0026thinsp;0.015 vs. components-only). The composite LACE Index alone is therefore sufficient for prediction, and its high feature importance ranking reflects genuine predictive value rather than redundant information from overlapping components.\u003c/p\u003e \u003cp\u003e \u003cb\u003ePractical Implications.\u003c/b\u003e All four CVD models demonstrated moderate to high utility for risk stratification, supporting their integration into clinical workflows. The HHD model (AUC 0.758, high utility) is the most immediate candidate for pilot clinical decision support in safety-net settings. HF and AF/AFL models, now at moderate utility, can support tiered risk stratification within the planned Readmission Analytics and Interactive Platform (RAIP) dashboard. Low PPVs across all models, as seen in comparable studies,\u003csup\u003e27,28\u003c/sup\u003e indicate that integration within a tiered risk stratification system rather than standalone screening remains appropriate. The Super Learner\u0026rsquo;s competitive performance for HHD provides an ensemble option for deployment contexts where model stability is valued over marginal discrimination gains.\u003c/p\u003e \u003cp\u003eThe HHD model shows high discrimination but suboptimal calibration (slopes 1.479\u0026ndash;1.646), suggesting that predicted probabilities at extremes are overstated. Thus, it is suitable for ranking patients by risk but requires recalibration, via Platt scaling or isotonic regression, before absolute risk communication.\u003c/p\u003e \u003cp\u003eKey strengths include a large, predominantly Black cohort (157,791 discharges, 96.6% Black), a fully crossed 40-model design enabling unbiased algorithm comparison, data-driven consensus feature selection, and validated clinical indices for reproducibility.\u003csup\u003e14,16,20\u003c/sup\u003e Patient-grouped splitting with GEE post-hoc validation and clustered bootstrap resampling ensured adequate control for within-patient correlation. Identifying HHD as a standalone prediction target with the highest discrimination offers an actionable result for a condition routinely subsumed within broader categories.\u003c/p\u003e \u003cp\u003eSeveral limitations warrant discussion. The 96.6% Black cohort limits generalizability to diverse populations and requires external validation.\u003csup\u003e16,20\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eLack of laboratory values, vital signs, ejection fraction, medications, and four key administrative features (APR-DRG Severity, Risk of Mortality, Admission Type, Discharge Disposition) may have constrained model performance. ICD-based classification cannot distinguish HFpEF from HFrEF. Low readmission among self-pay/uninsured patients introduces potential bias, as observed outcomes may underestimate true clinical risk. SMOTE-generated synthetic observations, particularly for binary comorbidity flags, may not fully reflect the true feature space, and alternative resampling strategies (e.g., SMOTE-NC or conditional generative models) may better preserve data fidelity. Model selection from multiple candidates likely overestimates performance (optimistic bias), emphasizing the need for independent external validation (Supplementary Table S3). The LACE Index, as both a predictor and a construct overlapping with readmission determinants, introduces potential circularity that warrants evaluation in external validation.\u003c/p\u003e \u003cp\u003eThe ICD-9 to ICD-10 transition may have introduced systematic differences in comorbidity ascertainment over the study period, affecting feature distributions despite adjustment for admission year (Supplementary Table S10). Future studies should consider multi-site validation and stratification by coding era to mitigate these temporal effects.\u003c/p\u003e \u003cp\u003eFinally, another limitation is the absence of independent external validation. Although models were evaluated using a patient-grouped 70/30 train\u0026ndash;test split with additional robustness checks including clustered bootstrap resampling and GEE analyses to account for within-patient correlation, these procedures represent internal validation within a single statewide dataset. Consequently, model performance estimates may remain optimistic relative to deployment in new clinical environments. Administrative coding practices, case mix, health system structure, and social determinants vary across regions and institutions, and these factors may influence both predictor distributions and readmission risk relationships. Therefore, the present models should be interpreted as development-stage prediction tools rather than finalized clinical decision support systems. Independent validation in geographically distinct populations and healthcare systems will be necessary to assess generalizability and confirm model calibration and discrimination\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eCondition-specific ML models achieved moderate-to-high discrimination for 30-day readmission across four CVD subtypes in a large, predominantly Black cohort. Model performance was driven more by CVD subtype than algorithm choice, with XGBoost performing consistently and the Super Learner competitive for HHD. Key predictors, including the LACE Index, renal failure, and insurance status, highlight the combined influence of clinical and structural factors. These results support deploying subtype-specific, equity-informed readmission models using validated clinical indices rather than ad-hoc composites.\u003c/p\u003e \u003cp\u003eExternal validation in diverse populations, incorporation of area-level social determinants, and prospective evaluation of model-guided interventions, particularly for HHD and HF, are needed to confirm generalizability and assess impact on readmission disparities.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research was supported by American Heart Association Second Century Implementation Science Award #23SCISA1145640. The opinions expressed in this article are solely those of the authors and do not necessarily represent those of the American Heart Association. This project was initiated at the EVMS-Sentara Healthcare Analytics and Delivery Science Institute (HADSI), Eastern Virginia Medical School, and continued at the Macon \u0026amp; Joan Brock Virginia Health Sciences at Old Dominion University following the institutional integration of EVMS and Old Dominion University. The authors acknowledge collaborative research support provided through a subaward to the University of Illinois at Chicago. Data access was facilitated by the M. Foscue Brock Institute for Community and Global Health and the Research and Infrastructure Service Enterprise (RISE) at the Macon \u0026amp; Joan Brock Virginia Health Sciences, Old Dominion University. Virginia Health Information (VHI) has provided non-confidential patient-level information used in this study, which it has compiled in accordance with Virginia law but which it has no authority to independently verify. By using this data, the authors agree to assume all risks that may be associated with or arise from the use of inaccurate data. VHI cannot and does not represent that the use of VHI\u0026rsquo;s data was appropriate for this study or endorse or support any conclusions or inferences that may be drawn from the use of VHI\u0026rsquo;s data. This study was approved by the Eastern Virginia Medical School Institutional Review Board (IRB #23-09-NH-0248) with a waiver of informed consent for de-identified data analysis.\u003c/p\u003e\n\n\u003cp\u003e\u003cstrong\u003eAuthor Contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eI.E.M. conceived and designed the study, developed the machine learning pipeline and feature engineering framework, conducted all model training and validation analyses, curated the data, created all figures and tables, wrote the original manuscript draft, supervised the project, administered all aspects of the research as Principal Investigator, and acquired funding. M.B. investigated readmission patterns across demographic, clinical, and administrative dimensions and contributed to data cleaning and preparation of the VHI Patient Level Database and Readmissions and Transfers Supplemental Data Set under the supervision of I.E.M. S.D. provided scientific oversight, cardiovascular epidemiology expertise, and mentorship throughout the project, contributed to study conceptualization and to the interpretation of health equity findings and clinical contextualization of results, and facilitated the institutional research infrastructure. All authors reviewed and approved the final manuscript.\u003c/p\u003e\n\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDr. El Moudden reports research funding from the American Heart Association (Award #23SCISA1145640) during the conduct of this study; no other relationships or activities that could appear to have influenced the submitted work. Dr. Dodani reports the following relationships: Vice-Chair, American Heart Association Health Equity Research Network (HERN) Oversight Advisory Committee; subaward through the University of Illinois at Chicago for collaborative research support on this project (Award #23SCISA1145640). She holds appointments at the University of Illinois College of Medicine\u0026ndash;Peoria (Founding Director, Center 4 Health Research; Professor of Clinical Medicine, Department of Medicine) and at the Macon \u0026amp; Joan Brock Virginia Health Sciences, Eastern Virginia Medical School at Old Dominion University (Professor of Medicine, Community Non-Tenure Track). No other relationships or activities that could appear to have influenced the submitted work. Mr. Bittner reports no disclosures.\u003c/p\u003e\n\n\u003cp\u003e\u003cstrong\u003eData Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe de-identified datasets analyzed in this study were derived from the Virginia Health Information (VHI) Patient Level Database (PLD) and the Readmissions and Transfers Supplemental Data Set (RATs), licensed inpatient discharge files encompassing all verified acute-care hospital discharges in Virginia from 2010 through 2020 (~9.24 million total discharges). Data access was provided through the M. Foscue Brock Institute for Community and Global Health and the Research and Infrastructure Service Enterprise (RISE) at the Macon \u0026amp; Joan Brock Virginia Health Sciences, Old Dominion University, under a site license agreement with VHI. Due to VHI licensing restrictions and the Virginia Patient Level Database System Act of 1993, the data cannot be publicly shared or deposited in an open repository. Access may be requested from the corresponding author upon reasonable request and is contingent upon execution of a VHI data license agreement and approval from the Brock Institute. Researchers seeking comparable data must apply directly to VHI, complete the licensing process, and cover applicable fees (www.vhi.org/pld). Correspondence regarding data extraction or PLD-to-RATs linkage methodology should be directed to Dr. Ismail El Moudden ([email protected]).\u003c/p\u003e\n\n\u003cp\u003e\u003cstrong\u003eCode Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll R analysis scripts for data preprocessing, feature engineering, model training, hyperparameter tuning, performance evaluation, and figure generation are available at https://github.com/isamil/ML-CVD-Readmission-in-Black-Cohort and will be made publicly available upon publication. Access for peer review is granted upon request to the corresponding author. The archived model package includes: (1) serialized trained model objects for all 32 base learner models and 8 Super Learner ensembles, (2) a standalone prediction function accepting raw patient-level input, (3) preprocessing parameters and hyperparameter configurations, and (4) Elastic Net coefficient vectors enabling exact replication without model objects (Supplementary Table S8). The archived model package is available from the corresponding author upon reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eMartin, S. S. et al. 2025 Heart Disease and Stroke Statistics: A Report of US and Global Data From the American Heart Association. \u003cem\u003eCirculation \u003c/em\u003e\u003cstrong\u003e151\u003c/strong\u003e, e41-e660 (2025). https://doi.org/10.1161/CIR.0000000000001303\u003c/li\u003e\n\u003cli\u003eMartin, S. S. et al. 2024 Heart Disease and Stroke Statistics: A Report of US and Global Data From the American Heart Association. \u003cem\u003eCirculation \u003c/em\u003e\u003cstrong\u003e149\u003c/strong\u003e, e347-e913 (2024). https://doi.org/10.1161/CIR.0000000000001209\u003c/li\u003e\n\u003cli\u003eHeidenreich, P. A. et al. 2022 AHA/ACC/HFSA Guideline for the Management of Heart Failure. \u003cem\u003eCirculation \u003c/em\u003e\u003cstrong\u003e145\u003c/strong\u003e, e895-e1032 (2022). https://doi.org/10.1161/CIR.0000000000001063\u003c/li\u003e\n\u003cli\u003eDharmarajan, K. et al. Diagnoses and timing of 30-day readmissions after hospitalization for heart failure, acute myocardial infarction, or pneumonia. \u003cem\u003eJAMA \u003c/em\u003e\u003cstrong\u003e309\u003c/strong\u003e, 355\u0026ndash;363 (2013). https://doi.org/10.1001/jama.2012.216476\u003c/li\u003e\n\u003cli\u003eShin, S. et al. Machine learning vs. conventional statistical models for predicting heart failure readmission and mortality. \u003cem\u003eESC Heart Fail \u003c/em\u003e\u003cstrong\u003e8\u003c/strong\u003e, 106-115 (2021). https://doi.org/10.1002/ehf2.13073\u003c/li\u003e\n\u003cli\u003eKrumholz, H. M. et al. Relationship between hospital readmission and mortality rates for patients hospitalized with acute myocardial infarction, heart failure, or pneumonia. \u003cem\u003eJAMA \u003c/em\u003e\u003cstrong\u003e309\u003c/strong\u003e, 587-593 (2013). https://doi.org/10.1001/jama.2013.333\u003c/li\u003e\n\u003cli\u003eKhera, R. et al. Comparison of Readmission Rates After Acute Myocardial Infarction in 3 Patient Age Groups. \u003cem\u003eAm J Cardiol \u003c/em\u003e\u003cstrong\u003e120\u003c/strong\u003e, 1761-1767 (2017). https://doi.org/10.1016/j.amjcard.2017.07.081\u003c/li\u003e\n\u003cli\u003eFreeman, J. V., Wang, Y., Akar, J. G., Desai, N. R. \u0026amp; Krumholz, H. M. National Trends in Atrial Fibrillation Hospitalization, Readmission, and Mortality for Medicare Beneficiaries, 1999-2013. \u003cem\u003eCirculation \u003c/em\u003e\u003cstrong\u003e135\u003c/strong\u003e, 1227-1239 (2017). https://doi.org/10.1161/CIRCULATIONAHA.116.022388\u003c/li\u003e\n\u003cli\u003eTripathi, B. et al. Outcomes and Resource Utilization Associated With Readmissions After Atrial Fibrillation Hospitalizations. \u003cem\u003eJ Am Heart Assoc \u003c/em\u003e\u003cstrong\u003e8\u003c/strong\u003e, e013026 (2019). https://doi.org/10.1161/JAHA.119.013026\u003c/li\u003e\n\u003cli\u003eMuntner, P. et al. Trends in Blood Pressure Control Among US Adults With Hypertension, 1999-2000 to 2017-2018. \u003cem\u003eJAMA \u003c/em\u003e\u003cstrong\u003e324\u003c/strong\u003e, 1190-1200 (2020). https://doi.org/10.1001/jama.2020.14545\u003c/li\u003e\n\u003cli\u003eShashikumar, S. A., Waken, R. J., Luke, A. A., Nerenz, D. R. \u0026amp; Joynt Maddox, K. E. Association of Stratification by Proportion of Patients Dually Enrolled in Medicare and Medicaid With Financial Penalties in the Hospital-Acquired Condition Reduction Program. \u003cem\u003eJAMA Intern Med \u003c/em\u003e\u003cstrong\u003e181\u003c/strong\u003e, 330-338 (2021). https://doi.org/10.1001/jamainternmed.2020.7386\u003c/li\u003e\n\u003cli\u003ePandey, A. et al. Temporal Trends in Racial Differences in 30-Day Readmission and Mortality Rates After Acute Myocardial Infarction Among Medicare Beneficiaries. \u003cem\u003eJAMA Cardiol \u003c/em\u003e\u003cstrong\u003e5\u003c/strong\u003e, 136-145 (2020). https://doi.org/10.1001/jamacardio.2019.4845\u003c/li\u003e\n\u003cli\u003eWadhera, R. K., Yeh, R. W. \u0026amp; Joynt Maddox, K. E. The Hospital Readmissions Reduction Program, Time for a Reboot. \u003cem\u003eN Engl J Med \u003c/em\u003e\u003cstrong\u003e380\u003c/strong\u003e, 2289-2291 (2019). https://doi.org/10.1056/NEJMp1901225\u003c/li\u003e\n\u003cli\u003ePatel, S. A. et al. Excess 30-Day Heart Failure Readmissions and Mortality in Black Patients Increases With Neighborhood Deprivation. \u003cem\u003eCirc Heart Fail \u003c/em\u003e\u003cstrong\u003e13\u003c/strong\u003e, e007947 (2020). https://doi.org/10.1161/CIRCHEARTFAILURE.120.007947\u003c/li\u003e\n\u003cli\u003eFields, N. D. et al. Historical Redlining and Heart Failure Outcomes Following Hospitalization in the Southeastern United States. \u003cem\u003eJ Am Heart Assoc \u003c/em\u003e\u003cstrong\u003e13\u003c/strong\u003e, e032019 (2024). https://doi.org/10.1161/JAHA.123.032019\u003c/li\u003e\n\u003cli\u003eDowning, N. S. et al. Association of Racial and Socioeconomic Disparities With Outcomes Among Patients Hospitalized With Acute Myocardial Infarction, Heart Failure, and Pneumonia. \u003cem\u003eJAMA Netw Open \u003c/em\u003e\u003cstrong\u003e1\u003c/strong\u003e, e182044 (2018). https://doi.org/10.1001/jamanetworkopen.2018.2044\u003c/li\u003e\n\u003cli\u003eKrittanawong, C., Zhang, H., Wang, Z., Aydar, M. \u0026amp; Kitai, T. Artificial Intelligence in Precision Cardiovascular Medicine. \u003cem\u003eJ Am Coll Cardiol \u003c/em\u003e\u003cstrong\u003e69\u003c/strong\u003e, 2657-2664 (2017). https://doi.org/10.1016/j.jacc.2017.03.571\u003c/li\u003e\n\u003cli\u003eAwan, S. E., Bennamoun, M., Sohel, F., Sanfilippo, F. M. \u0026amp; Dwivedi, G. Machine learning-based prediction of heart failure readmission or death: implications of choosing the right model and the right metrics. \u003cem\u003eESC Heart Fail \u003c/em\u003e\u003cstrong\u003e6\u003c/strong\u003e, 428-435 (2019). https://doi.org/10.1002/ehf2.12419\u003c/li\u003e\n\u003cli\u003eRajkomar, A. et al. Scalable and accurate deep learning with electronic health records. \u003cem\u003eNPJ Digit Med \u003c/em\u003e\u003cstrong\u003e1\u003c/strong\u003e, 18 (2018). https://doi.org/10.1038/s41746-018-0029-1\u003c/li\u003e\n\u003cli\u003eYu, M. Y. \u0026amp; Son, Y. J. Machine learning-based 30-day readmission prediction models for patients with heart failure: a systematic review. \u003cem\u003eEur J Cardiovasc Nurs \u003c/em\u003e\u003cstrong\u003e23\u003c/strong\u003e, 711-719 (2024). https://doi.org/10.1093/eurjcn/zvae031\u003c/li\u003e\n\u003cli\u003eChen, M., Tan, X. \u0026amp; Padman, R. A Machine Learning Approach to Support Urgent Stroke Triage Using Administrative Data and Social Determinants of Health at Hospital Presentation. \u003cem\u003eJ Med Internet Res \u003c/em\u003e\u003cstrong\u003e25\u003c/strong\u003e, e36477 (2023). https://doi.org/10.2196/36477\u003c/li\u003e\n\u003cli\u003eWhite-Williams, C. et al. Addressing Social Determinants of Health in the Care of Patients With Heart Failure: A Scientific Statement From the American Heart Association. \u003cem\u003eCirculation \u003c/em\u003e\u003cstrong\u003e141\u003c/strong\u003e, e841-e863 (2020). https://doi.org/10.1161/CIR.0000000000000767\u003c/li\u003e\n\u003cli\u003eCollins, G. S., Reitsma, J. B., Altman, D. G. \u0026amp; Moons, K. G. M. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. \u003cem\u003eBMJ \u003c/em\u003e\u003cstrong\u003e350\u003c/strong\u003e, g7594 (2015). https://doi.org/10.1136/bmj.g7594\u003c/li\u003e\n\u003cli\u003eVirginia Health Information. Patient Level Database (PLD) and Readmissions and Transfers Supplemental Data Set (RATs). Richmond, VA: Virginia Health Information; 2021. Accessed January 15, 2023. https://www.vhi.org/pld\u003c/li\u003e\n\u003cli\u003eElixhauser, A., Steiner, C., Harris, D. R. \u0026amp; Coffey, R. M. Comorbidity measures for use with administrative data. \u003cem\u003eMed Care \u003c/em\u003e\u003cstrong\u003e36\u003c/strong\u003e, 8-27 (1998). https://doi.org/10.1097/00005650-199801000-00004\u003c/li\u003e\n\u003cli\u003eCarpenter, J. \u0026amp; Bithell, J. Bootstrap confidence intervals: when, which, what? A practical guide for medical statisticians. \u003cem\u003eStat Med \u003c/em\u003e\u003cstrong\u003e19\u003c/strong\u003e, 1141-1164 (2000). https://doi.org/10.1002/(sici)1097-0258(20000515)19:9\u0026lt;1141::aid-sim479\u0026gt;3.0.co;2-f\u003c/li\u003e\n\u003cli\u003eSharma, V. et al. Predicting 30-day readmissions in patients with heart failure using administrative data: a machine learning approach. \u003cem\u003eJ Card Fail \u003c/em\u003e\u003cstrong\u003e28\u003c/strong\u003e, 710\u0026ndash;722 (2022). https://doi.org/10.1016/j.cardfail.2021.12.004\u003c/li\u003e\n\u003cli\u003eFine, N. M. et al. Machine learning for risk prediction after heart failure emergency department visit or hospital admission using administrative health data. \u003cem\u003ePLOS Digit Health \u003c/em\u003e\u003cstrong\u003e3\u003c/strong\u003e, e0000636 (2024). https://doi.org/10.1371/journal.pdig.0000636\u003c/li\u003e\n\u003cli\u003eHe, H. \u0026amp; Garcia, E. A. Learning from imbalanced data. \u003cem\u003eIEEE Trans Knowl Data Eng \u003c/em\u003e\u003cstrong\u003e21\u003c/strong\u003e, 1263\u0026ndash;1284 (2009). https://doi.org/10.1109/TKDE.2008.239\u003c/li\u003e\n\u003cli\u003eBahiru, E. et al. Association of Dual Eligibility for Medicare and Medicaid With Heart Failure Quality and Outcomes Among Get With The Guidelines\u0026ndash;Heart Failure Hospitals. \u003cem\u003eJAMA Cardiol \u003c/em\u003e\u003cstrong\u003e6\u003c/strong\u003e, 791-800 (2021). https://doi.org/10.1001/jamacardio.2021.0611\u003c/li\u003e\n\u003cli\u003eChaiyachati, K. H., Qi, M. \u0026amp; Werner, R. M. Changes to racial disparities in readmission rates after Medicare\u0026rsquo;s Hospital Readmissions Reduction Program within safety-net and non-safety-net hospitals. \u003cem\u003eJAMA Netw Open \u003c/em\u003e\u003cstrong\u003e1\u003c/strong\u003e, e184154 (2018). https://doi.org/10.1001/jamanetworkopen.2018.4154\u003c/li\u003e\n\u003cli\u003evan Walraven, C. et al. Derivation and validation of an index to predict early death or unplanned readmission after discharge from hospital to the community. \u003cem\u003eCMAJ \u003c/em\u003e\u003cstrong\u003e182\u003c/strong\u003e, 551-557 (2010). https://doi.org/10.1503/cmaj.091117\u003c/li\u003e\n\u003cli\u003eCharlson, M. E., Pompei, P., Ales, K. L. \u0026amp; MacKenzie, C. R. A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. \u003cem\u003eJ Chronic Dis \u003c/em\u003e\u003cstrong\u003e40\u003c/strong\u003e, 373-383 (1987). https://doi.org/10.1016/0021-9681(87)90171-8\u003c/li\u003e\n\u003cli\u003eQuan, H. et al. Coding algorithms for defining comorbidities in ICD-9-CM and ICD-10 administrative data. \u003cem\u003eMed Care \u003c/em\u003e\u003cstrong\u003e43\u003c/strong\u003e, 1130-1139 (2005). https://doi.org/10.1097/01.mlr.0000182534.19832.83\u003c/li\u003e\n\u003cli\u003evan Walraven, C., Austin, P. C., Jennings, A., Quan, H. \u0026amp; Forster, A. J. A modification of the Elixhauser comorbidity measures into a point system for hospital death using administrative data. \u003cem\u003eMed Care \u003c/em\u003e\u003cstrong\u003e47\u003c/strong\u003e, 626-633 (2009). https://doi.org/10.1097/MLR.0b013e31819432e5\u003c/li\u003e\n\u003cli\u003eKursa, M. B. \u0026amp; Rudnicki, W. R. Feature Selection with the Boruta Package. \u003cem\u003eJ Stat Softw \u003c/em\u003e\u003cstrong\u003e36\u003c/strong\u003e, 1-13 (2010). https://doi.org/10.18637/jss.v036.i11\u003c/li\u003e\n\u003cli\u003eDing, C. \u0026amp; Peng, H. Minimum Redundancy Feature Selection from Microarray Gene Expression Data. \u003cem\u003eJ Bioinform Comput Biol \u003c/em\u003e\u003cstrong\u003e3\u003c/strong\u003e, 185-205 (2005). https://doi.org/10.1142/S0219720005001004\u003c/li\u003e\n\u003cli\u003evan der Laan, M. J., Polley, E. C. \u0026amp; Hubbard, A. E. Super Learner. \u003cem\u003eStat Appl Genet Mol Biol \u003c/em\u003e\u003cstrong\u003e6\u003c/strong\u003e, Article 25 (2007). https://doi.org/10.2202/1544-6115.1309\u003c/li\u003e\n\u003cli\u003eMortazavi, B. J. et al. Analysis of Machine Learning Techniques for Heart Failure Readmissions. \u003cem\u003eCirc Cardiovasc Qual Outcomes \u003c/em\u003e\u003cstrong\u003e9\u003c/strong\u003e, 629-640 (2016). https://doi.org/10.1161/CIRCOUTCOMES.116.003039\u003cstrong\u003e\u003cstrong\u003e\u003c/strong\u003e\u003c/strong\u003e\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"npj-digital-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"npjdigitalmed","sideBox":"Learn more about [npj Digital Medicine](http://www.nature.com/npjdigitalmed/)","snPcode":"41746","submissionUrl":"https://submission.springernature.com/new-submission/41746/3","title":"npj Digital Medicine","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"cardiovascular diseases, heart failure, machine learning, patient readmission, risk assessment, healthcare disparities, African Americans, social determinants of health","lastPublishedDoi":"10.21203/rs.3.rs-9098008/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9098008/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eCardiovascular disease (CVD) readmissions impose substantial clinical and economic burden. Machine learning (ML) may improve risk stratification, yet most predictive models aggregate CVD subtypes into a single outcome and underrepresent Black populations. Using Virginia Health Information database records (2010 to 2020), we analyzed 157,791 discharge records from 123,272 unique patients (96.6% Black) to develop condition-specific 30-day readmission models for heart failure (HF; n\u0026thinsp;=\u0026thinsp;91,752), acute myocardial infarction (AMI; n\u0026thinsp;=\u0026thinsp;34,497), atrial fibrillation/flutter (AF/AFL; n\u0026thinsp;=\u0026thinsp;18,424), and hypertensive heart disease (HHD; n\u0026thinsp;=\u0026thinsp;13,118). Four algorithms (XGBoost, LightGBM, Random Forest, Elastic Net) plus a Super Learner ensemble were trained on patient-grouped 70/30 splits with and without Synthetic Minority Oversampling Technique balancing. Models incorporated validated clinical indices (LACE, Charlson, Elixhauser) and administrative social determinants of health proxies. The overall 30-day readmission rate was 18.9%. Best area under the receiver operating characteristic curve (AUC) values by condition were HF 0.708 (95% CI, 0.701 to 0.716), AMI 0.706 (95% CI, 0.691 to 0.721), AF/AFL 0.732 (95% CI, 0.715 to 0.750), and HHD 0.758 (95% CI, 0.735 to 0.777). XGBoost was the top-performing algorithm for three of four subtypes. The LACE Index, Charlson Comorbidity Index, and insurance type were consistently the strongest predictors. Algorithm-native, aggregated, and SHAP-based importance measures converged on these key features. In this largest-to-date, predominantly Black statewide cohort, condition-specific ML models achieved moderate-to-high discrimination for HF, AMI, AF/AFL, and HHD. Key clinical indices and administrative social determinants proxies emerged as dominant predictors, highlighting modifiable targets and high-risk subgroups. These findings support the development of precision, equity-informed readmission interventions and provide a scalable framework for deploying ML-driven decision support in safety-net and minority-serving healthcare systems.\u003c/p\u003e","manuscriptTitle":"Condition-Specific Readmission Risk Stratification in a Predominantly Black Statewide Cohort Using Machine Learning: Development of Subtype-Specific Models for Heart Failure, Acute Myocardial Infarction, Atrial Fibrillation/Flutter, and Hypertensive Heart Disease","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-19 16:33:46","doi":"10.21203/rs.3.rs-9098008/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"261121231873238693915009822504707913666","date":"2026-05-16T10:58:26+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"195435175829807797375166899935483391170","date":"2026-05-15T14:22:30+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-07T14:27:24+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"69669824922263230556725075349498518007","date":"2026-03-20T11:37:08+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-03-17T14:44:53+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-03-14T12:26:36+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-03-14T08:21:11+00:00","index":"","fulltext":""},{"type":"submitted","content":"npj Digital Medicine","date":"2026-03-11T20:40:39+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"npj-digital-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"npjdigitalmed","sideBox":"Learn more about [npj Digital Medicine](http://www.nature.com/npjdigitalmed/)","snPcode":"41746","submissionUrl":"https://submission.springernature.com/new-submission/41746/3","title":"npj Digital Medicine","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"27e5668e-cb69-4373-a3fe-596916ecbcdd","owner":[],"postedDate":"March 19th, 2026","published":true,"recentEditorialEvents":[{"type":"reviewerAgreed","content":"261121231873238693915009822504707913666","date":"2026-05-16T10:58:26+00:00","index":29,"fulltext":""},{"type":"reviewerAgreed","content":"195435175829807797375166899935483391170","date":"2026-05-15T14:22:30+00:00","index":28,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":64677062,"name":"Health sciences/Cardiology"},{"id":64677063,"name":"Health sciences/Diseases"},{"id":64677064,"name":"Health sciences/Health care"},{"id":64677065,"name":"Health sciences/Medical research"},{"id":64677066,"name":"Health sciences/Risk factors"}],"tags":[],"updatedAt":"2026-03-19T16:33:47+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-19 16:33:46","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9098008","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9098008","identity":"rs-9098008","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}