Interpretable Machine Learning for Sepsis Risk Prediction in Elderly ICU Patients With Heart Failure: Integrating Genetic Evidence and Clinical Data

Interpretable Machine Learning for Sepsis Risk Prediction in Elderly ICU Patients With Heart Failure: Integrating Genetic Evidence and Clinical Data | 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 Research Article Interpretable Machine Learning for Sepsis Risk Prediction in Elderly ICU Patients With Heart Failure: Integrating Genetic Evidence and Clinical Data Lin Wang, Yiping Hu, Yu Xiong, Yang Yu, Chunxiang Zhang This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9185668/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 5 You are reading this latest preprint version Abstract Background : Heart failure (HF) and sepsis frequently coexist in critically ill patients, yet the causal direction between these conditions remains unclear. Clarifying whether genetic susceptibility to HF predisposes individuals to sepsis may improve risk stratification in high-risk populations. Methods: We performed a bidirectional two-sample Mendelian randomization (MR) analysis using FinnGen genome-wide association study summary statistics to investigate the causal relationship between HF and sepsis. Based on the MR findings, machine learning (ML) models were developed to predict sepsis among elderly intensive care unit (ICU) patients with HF using the MIMIC-IV database and externally validated in the MIMIC-III cohort. Model performance was evaluated using discrimination, calibration, and clinical utility analyses, and SHapley Additive exPlanations (SHAP) were applied to improve model interpretability. Results : Genetically predicted liability to HF was associated with an increased risk of sepsis (inverse-variance weighted odds ratio 1.29, 95% confidence interval 1.09–1.54; P = 0.003), with consistent results across sensitivity analyses and no evidence of directional pleiotropy. Reverse MR analysis showed no causal association between genetic liability to sepsis and HF. In the external validation cohort, ML models demonstrated moderate discrimination (area under the curve 0.606–0.742), with the support vector machine model achieving the best performance (AUC 0.742, 95% CI 0.721–0.763). SHAP analysis identified illness severity scores, early ICU interventions, antimicrobial therapy, and laboratory indicators as key contributors to model predictions. Conclusions : Genetic liability to HF is associated with increased susceptibility to sepsis, highlighting a direction-specific relationship between these conditions. Integrating genetic causal inference with interpretable ML models provides a potential framework for sepsis risk prediction among elderly ICU patients with HF. Heart failure Machine learning Mendelian randomization Risk prediction Sepsis Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Heart failure (HF) and sepsis are among the leading causes of morbidity and mortality in critically ill patients, particularly in the elderly population. These conditions frequently coexist in the intensive care unit (ICU), where patients with pre-existing HF often experience infectious complications, and sepsis may precipitate acute cardiac dysfunction. [ 1 ] Observational studies have consistently reported associations between HF and adverse infectious outcomes [ 2 – 5 ] ; however, the directionality of this relationship remains uncertain. Chronic HF is characterized by systemic inflammation [ 6 ] , neurohormonal activation [ 7 ] , endothelial dysfunction [ 8 ] , and multi-organ impairment [ 9 ] —factors that may plausibly increase susceptibility to infection and sepsis [ 2 ] . Conversely, sepsis is well recognized to induce transient myocardial dysfunction and long-term cardiovascular complications [ 10 , 11 ] . Shared risk factors and residual confounding further complicate causal interpretation, making it difficult to determine whether HF predisposes individuals to sepsis or whether observed associations primarily reflect reverse causation and overlapping pathophysiology [ 12 ] . Mendelian randomization (MR) offers a framework for enhancing causal inference by utilizing genetic variants as instrumental variables, thus minimizing confounding and reverse causation bias. Bidirectional two-sample MR is particularly useful in disentangling complex relationships between traits with potential reciprocal associations. Despite growing interest in the interplay between cardiovascular disease and systemic infection, no large-scale bidirectional MR study has comprehensively evaluated the directional relationship between genetic liability to HF and sepsis risk using contemporary genome-wide association data [ 13 ] . Clarifying this directionality is clinically relevant, as it determines whether HF should be regarded primarily as a downstream consequence of sepsis or as an upstream susceptibility state. Establishing causal direction at the population level, however, does not directly translate into individualized clinical decision-making. If HF represents a risk-enriched state for sepsis, patients with HF—especially elderly ICU patients—may benefit from targeted risk stratification and early surveillance strategies [ 14 ] . Existing sepsis prediction models are typically developed in heterogeneous ICU populations and rarely focus specifically on patients with HF [ 15 ] . Moreover, many models lack independent external validation and interpretable frameworks, limiting their generalizability and clinical adoption. Machine learning (ML) approaches offer the ability to model complex, nonlinear interactions among clinical variables [ 15 ] , and the incorporation of explainability methods such as SHapley Additive exPlanations (SHAP) enhances transparency and interpretability [ 16 ] . In this study, we integrated bidirectional two-sample MR with clinically applicable ML modeling to bridge genetic inference and bedside risk assessment. First, using FinnGen genome-wide association summary statistics, we investigated the directional relationship between genetic liability to HF and sepsis. Guided by the MR findings, we subsequently developed and externally validated ML models to predict Sepsis-3–defined sepsis among elderly ICU patients with HF, using MIMIC-IV for model development and MIMIC-III for validation. Through this integrated approach, we aimed to clarify the directional association between HF and sepsis and to translate genetic insights into an interpretable framework for individualized sepsis risk stratification. Methods Study Design This study utilized an integrated framework that combines genetic and clinical modeling to investigate the relationship between heart failure (HF) and sepsis. To begin, a bidirectional two-sample Mendelian randomization (MR) analysis was performed to assess potential causal connections between genetically predicted susceptibility to HF and the likelihood of developing sepsis. Second, based on the causal direction suggested by the MR findings, ML–based prediction models were developed and externally validated to estimate the occurrence of sepsis among elderly ICU patients. The overall analytical workflow is summarized in Fig. 1 . Data Sources Summary-level genome-wide association study (GWAS) data were obtained from the 2021 release of the FinnGen consortium, focusing on individuals of European descent. Genetic instruments for all-cause HF were derived from FinnGen (finn-b-I9_HEARTFAIL_ALLCAUSE; 23,397 cases and 194,811 controls), and summary statistics for sepsis were obtained from FinnGen (finn-b-AB1_SEPSIS; 6,164 cases and 197,660 controls). All GWAS datasets were based on the GRCh37 (hg19) reference genome build. Data for model training and internal validation were obtained from version 3.1 of the Medical Information Mart for Intensive Care IV (MIMIC-IV), whereas the MIMIC-III database was used for external validation. MIMIC-IV comprises de-identified health records from patients admitted to the intensive care units of Beth Israel Deaconess Medical Center (Boston, MA, USA) between 2008 and 2022. Access to the databases was granted following completion of the requisite data use agreements and approval by the institutional review board, with a waiver of informed consent (certification number 74299605). Study Population and Outcome Measurement For the machine-learning (ML) analysis, eligible patients were required to satisfy the following conditions: (1) first ICU admission during the index hospitalisation, (2) an ICU stay of at least 24 h, and (3) age between 65 and 100 years at ICU admission. HF was defined according to International Classification of Diseases (ICD) diagnostic codes documented during the hospital stay. The main outcome was the development of sepsis during the intensive care unit stay, defined according to the Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). Sepsis events were identified using the validated MIMIC code repository implementation, based on documented or suspected infection in combination with acute organ dysfunction. Predictor Variables Baseline predictive variables were extracted from the MIMIC-IV database via structured query language, with specifications that these values represented the earliest measurements obtained after ICU admission. Collected features included demographic characteristics (e.g., age, sex, and body weight), major comorbidities, medication exposures and clinical interventions, severity scores, and laboratory measurements. The detailed information of all candidate predictor variables—including specific variable names, operational definitions, data types (continuous/categorical), and measurement timelines—is summarized in Table S1 . Comorbidities encompassed chronic cardiovascular, neurologic, metabolic, respiratory, renal, and hepatic conditions, as well as malignancy. Treatment-related variables included antimicrobial exposures (glycopeptides, macrolides, β-lactam antibiotics, and quinolones), vasoactive agent use, mechanical ventilation, continuous renal replacement therapy, sedatives, albumin administration, and other commonly administered therapies. Disease severity was evaluated using the Sequential Organ Failure Assessment (SOFA) score and the Simplified Acute Physiology Score II (SAPS II). Laboratory variables consisted of routinely collected hematologic, biochemical, coagulation, and arterial blood gas parameters. Missing Data and Sample Size Considerations Predictors containing over 20% missing data were excluded from subsequent analyses. For the remaining variables, missing values were addressed through multiple imputation by chained equations.All imputation and preprocessing procedures were performed exclusively within the training dataset to minimize the risk of data leakage, and the same transformations were subsequently applied to the testing and external validation cohorts. Given the retrospective study design and the utilization of large critical care databases, no formal a priori sample size calculation was performed. Statistical Analysis MR Analysis A bidirectional two-sample MR design was employed. In the forward analysis, genetic variants linked to all-cause HF served as instrumental variables, with sepsis designated as the outcome.In the reverse analysis, genetic variants associated with sepsis were used as instrumental variables, with all-cause HF as the outcome. SNPs were selected at P < 5 × 10⁻⁶, clumped for linkage disequilibrium (r² 10. Summary statistics were standardized by aligning effect alleles across datasets, and palindromic SNPs with ambiguous strand orientation were removed. Causal effect estimates were primarily obtained using the inverse-variance weighted (IVW) method. Sensitivity analyses were carried out using MR-Egger regression, weighted median, simple mode, and weighted mode approaches. Directional horizontal pleiotropy was evaluated via the MR-Egger intercept test. Leave-one-out analyses and MR-PRESSO were utilized to assess the robustness of causal estimates and identify potential outlier instruments. ML Analysis The MIMIC-IV cohort was randomly split into a training set (70%) and an internal validation set (30%).The independent MIMIC-III cohort was used for external validation. Multiple ML algorithms were developed and compared, including logistic regression, SVM, GBM, neural network, KNN, Xgboost, AdaBoost, LightGBM, and CatBoost. Continuous variables were standardized for algorithms sensitive to feature scaling. Model hyperparameters were adjusted within the training dataset using grid-search procedures combined with cross-validation. The specific parameter grids applied to each machine-learning algorithm are detailed in Supplementary Table S3. Discriminative performance was assessed using the area under the receiver operating characteristic curve (AUC), alongside measures of sensitivity, specificity, and calibration. The identical set of metrics was employed to evaluate performance in the external validation cohort. All statistical analyses were conducted using R software (version 4.4.2). Results 1. Forward MR analysis of HF on sepsis risk In the forward Mendelian randomisation (MR) analysis, HF-related genetic variants were employed as instrumental variables to explore a potential causal relationship between HF and the subsequent risk of sepsis. In total, 24 independent single-nucleotide polymorphisms (SNPs) were selected as instruments. Using the IVW method as the primary analysis, genetically predicted liability to HF showed evidence of an association with an increased risk of sepsis (OR = 1.295, 95% CI: 1.091–1.538, P = 0.003). Consistent estimates were observed using the weighted median method (OR = 1.305, 95% CI: 1.069–1.593, P = 0.009). Estimates derived from MR-Egger and mode-based methods were directionally consistent with the primary analysis but less precise, with wider confidence intervals (Fig. 2 A). The MR-Egger intercept test did not indicate evidence of directional horizontal pleiotropy (intercept = − 0.005, P = 0.749). Moderate heterogeneity was observed (IVW Q P = 0.042; MR-Egger Q P = 0.032); however, the MR-PRESSO global test was not significant (P = 0.07), suggesting that the observed heterogeneity was unlikely to be driven by strong directional pleiotropy. Scatter plots showed generally positive slopes across MR methods (Fig. 2 B). Leave-one-out analyses indicated that the IVW estimate was not driven by any single SNP (Fig. 2 C), and visual inspection of funnel plots did not suggest marked asymmetry (Fig. 2 D). 2. Reverse MR analysis of sepsis on HF risk In the reverse Mendelian randomisation (MR) analysis, sepsis-associated genetic variants were selected as instrumental variables to evaluate whether sepsis exerts a potential causal influence on HF risk. Five independent single-nucleotide polymorphisms (SNPs) were ultimately retained as instrumental variables. Across multiple MR approaches, no statistically significant evidence was found to support a causal effect of genetically predicted sepsis liability on HF risk. The IVW analysis yielded an odds ratio (OR) of 1.170 (95% confidence interval [CI]: 0.953–1.435, P = 0.134). Consistent null or near-null estimates were obtained using complementary methods, including the weighted median (OR = 1.036, 95% CI: 0.892–1.202, P = 0.645), simple mode (OR = 1.025, 95% CI: 0.869–1.209, P = 0.781), and weighted mode (OR = 1.027, 95% CI: 0.881–1.197, P = 0.751) (Fig. 3 A). Scatter plots did not show a consistent pattern across MR methods (Fig. 3 B). MR-Egger regression similarly did not indicate a statistically significant association (OR = 0.880, 95% CI: 0.667–1.161, P = 0.433). The MR-Egger intercept test did not provide evidence of directional horizontal pleiotropy (intercept = 0.049, P = 0.101). Cochran’s Q statistics indicated heterogeneity in the IVW model (Q P = 0.008), whereas heterogeneity was not evident under the MR-Egger model (Q P = 0.183). The MR-PRESSO global test suggested overall heterogeneity (P = 0.032), while no outlier SNPs were identified by the outlier test. Leave-one-out sensitivity analyses indicated that the overall null association was not driven by any single instrumental variant (Figure S2). Motivated by the potential causal direction suggested by the bidirectional MR analyses, we subsequently evaluated whether routinely collected clinical features could be used to develop a ML–based model to predict sepsis risk in patients with HF. 3.1 Feature selection Feature selection was performed to identify a stable subset of predictors for downstream model development. Boruta feature filtering confirmed 27 predictors with importance exceeding the corresponding shadow features (Fig. 4 A). Subsequently, using LASSO regression with cross-validation, 36 predictors with non-zero coefficients were retained (Fig. 4 B–C). Meanwhile, recursive feature elimination (RFE) retained 43 predictors. The intersection of predictors consistently selected across all three methods was used for subsequent model construction, yielding a final set of 17 predictors (Fig. 4 D). These predictors encompassed antimicrobial use (glycopeptides, β-lactam antibiotics, macrolides, and quinolones), disease severity scores (SOFA and SAPS II), clinical interventions (mechanical ventilation, vasoactive agents, continuous renal replacement therapy, sedatives, and albumin administration), laboratory indicators (white blood cell count, serum calcium, and chloride), and baseline characteristics (age, atrial fibrillation, and furosemide use). 3.2 Model development, internal testing, and external validation Using the 17 selected predictors, multiple ML models were evaluated in the internal testing cohort. ROC analysis demonstrated good discrimination across models, with AUCs ranging from 0.703 to 0.852 (Fig. 5 A). NeuralNetwork achieved the highest AUC (0.852, 95% CI: 0.833–0.870), followed by GBM (AUC = 0.847, 95% CI: 0.828–0.866) and CatBoost (AUC = 0.845, 95% CI: 0.826–0.864). Logistic regression and Xgboost showed comparable performance (AUCs = 0.843 and 0.839, respectively), while SVM (AUC = 0.835, 95% CI: 0.816–0.855) and LightGBM (AUC = 0.836, 95% CI: 0.817–0.856) had slightly lower but similar discrimination; AdaBoost (AUC = 0.780, 95% CI: 0.758–0.802) and KNN (AUC = 0.703, 95% CI: 0.681–0.725) exhibited relatively lower discrimination, with KNN showing the poorest performance in internal validation. Given the notable performance variability of algorithms in internal testing and the potential overfitting risk of models with extreme internal performance, model selection was further based on external validation and clinical utility to ensure generalizability. Model performance was subsequently evaluated in an independent external cohort from MIMIC-III. ROC analysis demonstrated moderate discrimination across models, with AUCs ranging from 0.606 to 0.742 (Fig. 5 B). Among the evaluated algorithms, SVM achieved the highest AUC (0.742, 95% CI: 0.721–0.763), while logistic regression (AUC = 0.732, 95% CI: 0.711–0.754), CatBoost (AUC = 0.715, 95% CI: 0.693–0.736), and Xgboost (AUC = 0.712, 95% CI: 0.690–0.733) showed comparable discrimination. NeuralNetwork (AUC = 0.703, 95% CI: 0.681–0.726) and GBM (AUC = 0.684, 95% CI: 0.661–0.707) had lower performance, and LightGBM (AUC = 0.667, 95% CI: 0.644–0.689), AdaBoost (AUC = 0.636, 95% CI: 0.616–0.656) and KNN (AUC = 0.606, 95% CI: 0.586–0.625) exhibited the poorest discrimination in external validation. Table S2 summarizes the detailed threshold-based performance metrics of all models in external validation, including accuracy, sensitivity, specificity, and other key indicators. The SVM model exhibited balanced operational performance (with Accuracy = 0.699, Sensitivity = 0.666, Specificity = 0.715, Precision = 0.523, F1 = 0.586, and Kappa = 0.31), outperforming all other algorithms across key clinical evaluation indicators.Decision curve analysis indicated that several models provided greater net benefit than the treat-all and treat-none strategies across clinically relevant threshold probabilities, with SVM demonstrating consistently favorable net benefit within intermediate risk thresholds (Fig. 5 C). Based on its overall superior external performance and favorable clinical utility, SVM was identified as the optimal model for subsequent interpretation and explainability analyses. 3.3 Model interpretation using SHAP SHAP analysis was performed for the final SVM model to quantify the contribution of individual predictors to sepsis prediction. The global SHAP beeswarm plot (Fig. 6 A) indicated that established severity scores, including SOFA and SAPS II, were among the most influential predictors. Treatment-related interventions such as vasoactive agent use, mechanical ventilation, and continuous renal replacement therapy also contributed substantially to the model output. In addition, antimicrobial exposures (e.g., glycopeptides, β-lactam antibiotics, and quinolones) and laboratory parameters including white blood cell count, serum calcium, and chloride levels showed notable contributions to the prediction. To further illustrate patient-level interpretability, a representative high-risk case was examined using a SHAP-based local explanation (Fig. 6 B). For this individual, the model output increased markedly from the baseline (E[f(x)] = 0.55) to a much higher value, driven primarily by quinolone exposure, elevated SOFA score, glycopeptide use, and receipt of continuous renal replacement therapy, whereas not receiving mechanical ventilation (MV = 0) contributed negatively. Overall, the combined effects of multiple clinical features jointly resulted in a high predicted sepsis risk. To further examine feature-specific effects, SHAP dependence plots were generated for key continuous predictors in the final SVM model. In the SOFA dependence plot, increasing SOFA values were generally associated with higher SHAP values, indicating greater contributions to predicted sepsis risk (Fig. 6 C). A similar pattern was observed for white blood cell count (WBC), where higher values were associated with increased SHAP values, suggesting that inflammatory burden contributed positively to predicted sepsis risk (Fig. 6 D). Additional SHAP dependence plots for treatment-related variables, antimicrobial exposures, and other laboratory parameters are provided in the Supplementary Materials. Discussion Principal findings In this study, we integrated bidirectional two-sample MR with clinical ML to clarify the directional relationship between HF and sepsis and to translate this insight into individualized risk prediction. The forward MR analysis demonstrated that genetically predicted liability to HF was consistently associated with an increased risk of sepsis across multiple estimators, whereas the reverse MR analysis provided no evidence that genetic liability to sepsis causally increases HF risk. These findings support a direction-specific relationship in which HF represents a predisposing state for sepsis, rather than a consequence of sepsis-related genetic susceptibility. Importantly, the absence of genetic evidence for a reverse causal effect does not contradict the well-recognized occurrence of acute cardiac dysfunction during severe sepsis [ 10 , 17 ] . Sepsis-related myocardial injury and transient ventricular dysfunction are likely driven by short-term inflammatory, metabolic, and hemodynamic insults [ 10 , 17 , 18 ] and are not captured by genetic liability to sepsis or by pathways underlying chronic HF. Our results therefore refine, rather than negate, existing clinical observations by distinguishing long-term causal susceptibility from acute reversible organ dysfunction [ 10 , 17 ] . Potential mechanisms linking HF to sepsis susceptibility Several biological and clinical mechanisms may explain why HF predisposes individuals to sepsis. Chronic HF is characterized by persistent systemic inflammation, neurohormonal activation, endothelial dysfunction, and immune dysregulation, all of which may impair host defense and increase vulnerability to infection [ 19 – 21 ] . Altered innate and adaptive immune responses observed in HF may further facilitate progression from infection to sepsis [ 20 , 22 ] . In addition, HF is frequently accompanied by multi-organ dysfunction, including renal and hepatic impairment, which can reduce physiological reserve and immune competence. Hemodynamic abnormalities inherent to HF, such as tissue hypoperfusion and venous congestion—particularly intestinal congestion—may compromise gut barrier integrity and promote bacterial translocation, amplifying systemic inflammation and increasing the risk of sepsis [ 23 , 24 ] . These mechanisms are broadly consistent with the predictors identified by our model, including disease severity scores, inflammatory markers, and early ICU interventions that reflect clinical deterioration. Implications for risk stratification and early identification Identifying HF as a causal risk-enriching state for sepsis has important implications for prevention and early recognition, particularly among elderly ICU patients. Existing sepsis risk assessment approaches largely focus on acute physiological derangements and infectious triggers, with limited consideration of underlying cardiovascular vulnerability [ 25 , 26 ] . Our findings suggest that incorporating baseline HF-related susceptibility alongside early clinical signals may improve identification of patients at high risk for sepsis, enabling closer monitoring and earlier escalation of care. From genetic evidence to predictive modeling While MR strengthens causal inference by reducing confounding and reverse causation, it does not provide individualized risk estimates suitable for bedside decision-making [ 27 , 28 ] . Accordingly, we developed and externally validated ML models based on routinely collected clinical features to support individualized sepsis risk prediction among patients with HF. The alignment between the MR-supported causal direction and the modeling objective strengthens the coherence of our approach: MR provides etiological justification for the prediction target, whereas the clinical model offers practical utility for risk stratification. Although several algorithms achieved excellent discrimination in internal testing, performance in external validation was more modest, underscoring the importance of independent evaluation and mitigating concerns regarding overfitting [ 29 , 30 ] . Among the evaluated models, the SVM demonstrated the most favorable balance between discrimination and clinical utility in external validation. Importantly, model interpretability using SHAP revealed clinically plausible predictors, including illness severity scores, early ICU interventions, antimicrobial exposures, and laboratory markers, enhancing transparency and supporting clinical insight rather than automated decision-making [ 31 ] . Strengths and limitations This study is supported by several important strengths.We employed a bidirectional MR design using large-scale GWAS summary statistics to clarify causal directionality, followed by the development of externally validated prediction models across two independent critical care databases. The integration of causal inference with interpretable clinical modeling represents a translational framework that links etiological insight with practical risk stratification. Several limitations should be acknowledged. First, the FinnGen sepsis phenotype may not fully correspond to the Sepsis-3 definition used in the clinical cohorts, introducing potential phenotype heterogeneity [ 25 , 32 ] . Second, although extensive sensitivity analyses were performed, key MR assumptions—particularly the exclusion restriction and absence of horizontal pleiotropy—cannot be fully verified [ 28 , 33 ] . Third, some predictors incorpoated in the ML models reflect early ICU interventions and may act as proxies for illness severity or evolving clinical decision-making rather than purely antecedent biological risk. Consequently, model outputs should be interpreted as risk stratification tools rather than causal estimators [ 34 ] . Fourth, the external validation cohort originated from the same medical center as the development dataset, albeit from a different database version, which may limit full geographic generalizability. Finally, our findings are specific to elderly ICU patients with HF and may not extrapolate to younger populations or non-ICU settings. Conclusions In summary, genetic liability to HF was associated with an increased risk of sepsis in bidirectional MR analyses, whereas genetic liability to sepsis was not associated with HF risk. By integrating causal inference with externally validated and interpretable ML models, this study provides a translational framework for sepsis risk stratification among elderly ICU patients with HF. These findings support the concept that HF represents a susceptibility state for sepsis and highlight opportunities for targeted risk assessment in this vulnerable population. Abbreviations AUC Area Under the Curve CI Confidence Interval GWAS Genome-Wide Association Study HF Heart Failure ICD International Classification of Diseases ICU Intensive Care Unit IVW Inverse-Variance Weighted ML Machine Learning MR Mendelian Randomization OR Odds Ratio ROC Receiver Operating Characteristic SAPS II Simplified Acute Physiology Score II SHAP SHapley Additive Explanations SOFA Sequential Organ Failure Assessment SNPs Single-Nucleotide Polymorphisms Declarations Ethics approval and consent to participate This study analysed anonymised, de-identified data obtained from the Medical Information Mart for Intensive Care IV (MIMIC-IV) and Medical Information Mart for Intensive Care III (MIMIC-III) databases. Database access was authorised after the investigators completed the mandatory training and executed the relevant data use agreements. Use of these resources was approved by the institutional review boards of the Massachusetts Institute of Technology and Beth Israel Deaconess Medical Center, and the requirement for informed consent was waived because the study was retrospective and involve Ethics approval and consent to participate only de-identified information. Therefore, no further ethical approval was necessary for the present analysis. Consent for publication Not applicable, as this study does not contain any individual person’s data in any form. Data availability statement The FinnGen GWAS summary statistics used in this study are publicly available from the FinnGen consortium website (https://www.finngen.fi/en/access_results). The MIMIC-IV (version 3.1) and MIMIC-III databases are publicly accessible through PhysioNet (https://physionet.org), subject to completion of the required training (e.g., CITI Program courses) and data use agreements. All data used in this study can be obtained from the original sources under the corresponding access policies. Code availability statement The code used for data preprocessing, statistical analysis, and ML model development in this study is available from the corresponding author upon reasonable request. Conflicts of interest The authors declare that they have no competing interests. Funding Funding: This work was supported by National Natural Science Foundation of China (Grant Nos. U23A20398), Noncommunicable Chronic Diseases-National Science and Technology Major Project(Grant Nos. 2024ZD0537707), Sichuan Science and Technology Program (Grant Nos. 2025YFRG0005), The People's Government of Luzhou Municipality-Southwest Medical University Science and Technology Strategic Cooperation "Science and Technology Climbing Program" (Grant Nos. 2025LZXNYDPD01), Research Start-up Foundation of Southwest Medical University (Grant Nos. 00040155). Authors’ contributions LW and YH contributed to the study conception and design. LW, YH, and YX contributed to data acquisition and curation. LW and YH performed the statistical analysis and machine learning modeling. YX contributed to data interpretation. YY and CZ supervised the study and provided critical revisions. All authors read and approved the final manuscript. Acknowledgements The authors thank the FinnGen consortium for providing access to the genome-wide association study summary statistics and the PhysioNet team for maintaining the MIMIC-IV and MIMIC-III databases. The authors also thank the research team at the Basic Medicine Research Innovation Center for Cardiometabolic Diseases for their support. References Zhang B, Guo S, Fu Z, Wu N, Liu Z. Association between fluid balance and mortality for heart failure and sepsis: a propensity score-matching analysis. BMC Anesthesiol. 2022;22(1):324. Palin V, Brown O, Hamilton F, Lillie P, Kearney M, Cubbon R, et al. Infection in people with heart failure: an overlooked cause of adverse outcomes. Clin Med (Lond). 2025;25(5):100497. Drozd M, Garland E, Walker AMN, Slater TA, Koshy A, Straw S, et al. Infection-related hospitalization in heart failure with reduced ejection fraction: a prospective observational cohort study. Circ Heart Fail. 2020;13(5):e006746. Molinsky RL, Shah A, Yuzefpolskaya M, Yu B, Misialek JR, Bohn B, et al. Infection-related hospitalization and incident heart failure: the Atherosclerosis Risk in Communities Study. J Am Heart Assoc. 2025;14(3):e033877. Corrales-Medina VF, Alvarez KN, Weissfeld LA, Angus DC, Chirinos JA, Chang CC, et al. Association between hospitalization for pneumonia and subsequent risk of cardiovascular disease. JAMA. 2015;313(3):264–274. Murphy SP, Kakkar R, McCarthy CP, Januzzi JL Jr. Inflammation in heart failure: JACC state-of-the-art review. J Am Coll Cardiol. 2020;75(11):1324–1340. McDonagh TA, Metra M, Adamo M, Gardner RS, Baumbach A, Bohm M, et al. 2021 ESC guidelines for the diagnosis and treatment of acute and chronic heart failure. Eur Heart J. 2021;42(36):3599–3726. Tsigkou V, Oikonomou E, Anastasiou A, Lampsas S, Zakynthinos GE, Kalogeras K, et al. Molecular mechanisms and therapeutic implications of endothelial dysfunction in patients with heart failure. Int J Mol Sci. 2023;24(5):4321 Ciccarelli M, Dawson D, Falcao-Pires I, Giacca M, Hamdani N, Heymans S, et al. Reciprocal organ interactions during heart failure: a position paper from the ESC Working Group on Myocardial Function. Cardiovasc Res. 2021;117(12):2416–2433. Sato R, Sanfilippo F, Lanspa M, Duggal A, Dugar S. Sepsis-induced cardiomyopathy: mechanism, prevalence, assessment, prognosis, and management. Chest. 2025;168(6):1383–1394. Beane A, Shankar-Hari M. Long-term ill health in sepsis survivors: an ignored health-care challenge? Lancet. 2024;404(10459):1178–1180. Sattar N, Preiss D. Reverse causality in cardiovascular epidemiological research: more common than imagined? Circulation. 2017;135(24):2369–2372. Liu L, Huang P, Wang C, Liu Y, Gao Y, Yu K. Causal association between heart failure and sepsis: insights from Mendelian randomization and observational studies. Clin Epidemiol. 2024;16:755–767. Zhang Q, Xu L, He W, et al. Survival prediction for heart failure complicated by sepsis: based on machine learning methods. Front Med (Lausanne). 2024;11:1410702. Mou C, Yang J, Wu Q, Qin L, Lu J. Progress in sepsis prediction models: from traditional scoring systems to multimodal intelligence and clinical translation. Front Med (Lausanne). 2026;13:1732164. Zhou H, Li F, Liu X. Early prediction of septic shock in ICU patients using machine learning: development, external validation, and explainability with SHAP. Int J Med Inform. 2026;206:106169. Ehrman RR, Sullivan AN, Favot MJ, Sherwin RL, Reynolds CA, Abidov A, et al. Pathophysiology, echocardiographic evaluation, biomarker findings, and prognostic implications of septic cardiomyopathy: a review of the literature. Crit Care. 2018;22(1):112. Carbone F, Liberale L, Preda A, Schindler TH, Montecucco F. Septic cardiomyopathy: from pathophysiology to the clinical setting. Cells. 2022;11(18):2833. Mann DL. Inflammatory mediators and the failing heart: past, present, and the foreseeable future. Circ Res. 2002;91(11):988–998. Van Linthout S, Tschöpe C. Inflammation – cause or consequence of heart failure or both? Curr Heart Fail Rep. 2017;14(4):251–265. Dick SA, Epelman S. Chronic heart failure and inflammation: what do we really know? Circ Res. 2016;119(1):159–176. Yndestad A, Damas JK, Oie E, Ueland T, Gullestad L, Aukrust P. Systemic inflammation in heart failure—the whys and wherefores. Heart Fail Rev. 2006;11(1):83–92. Sandek A, Bauditz J, Swidsinski A, Buhner S, Weber-Eibel J, von Haehling S, et al. Altered intestinal function in patients with chronic heart failure. J Am Coll Cardiol. 2007;50(16):1561–1569. Krack A, Sharma R, Figulla HR, Anker SD. The importance of the gastrointestinal system in the pathogenesis of heart failure. Eur Heart J. 2005;26(22):2368–2374. Singer M, Deutschman CS, Seymour CW, Shankar-Hari M, Annane D, Bauer M, et al. The third international consensus definitions for sepsis and septic shock (Sepsis-3). JAMA. 2016;315(8):801–810. Schertz AR, Lenoir KM, Bertoni AG, Rosamond WD, Chang PP, Granger CB, et al. Sepsis prediction model for determining sepsis vs SIRS, qSOFA, and SOFA. JAMA Netw Open. 2023;6(8):e2329729. Emdin CA, Khera AV, Kathiresan S. Mendelian randomization. JAMA. 2017;318(19):1925–1926. Davies NM, Holmes MV, Davey Smith G. Reading Mendelian randomisation studies: a guide, glossary, and checklist for clinicians. BMJ. 2018;362:k601. Collins GS, Reitsma JB, Altman DG, Moons KGM. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. BMJ. 2015;350:g7594. Collins GS, Moons KGM. Reporting of artificial intelligence prediction models. Lancet. 2019;393(10181):1577–1579. Zhang D, Yin C, Hunold KM, Banerjee A, Chen Y, Sun J, et al. An interpretable deep-learning model for early prediction of sepsis in the emergency department. Patterns (N Y). 2021;2(2):100196. Kurki MI, Karjalainen J, Palta P, Sipilä TP, Kristiansson K, Donner K, et al. FinnGen provides genetic insights from a well-phenotyped isolated population. Nature. 2023;613(7944):508–518. Bowden J, Davey Smith G, Burgess S. Mendelian randomization with invalid instruments: effect estimation and bias detection through Egger regression. Int J Epidemiol. 2015;44(2):512–525. Wynants L, Van Calster B, Collins GS, Riley RD, Heinze G, Schuit E, et al. Prediction models for diagnosis and prognosis of COVID-19: systematic review and critical appraisal. BMJ. 2020;369:m1328. Additional Declarations No competing interests reported. Supplementary Files TableS1.doc TableS3.doc TableS2.doc Additionalfile.pdf Cite Share Download PDF Status: Under Review Version 1 posted Reviewers invited by journal 15 Apr, 2026 Editor invited by journal 26 Mar, 2026 Editor assigned by journal 23 Mar, 2026 Submission checks completed at journal 23 Mar, 2026 First submitted to journal 21 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-9185668","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":627773232,"identity":"8206bd13-b6a3-484c-8b72-298135fc2934","order_by":0,"name":"Lin Wang","email":"","orcid":"","institution":"Southwest Medical University","correspondingAuthor":false,"prefix":"","firstName":"Lin","middleName":"","lastName":"Wang","suffix":""},{"id":627773233,"identity":"da5e108b-4b5e-4689-a902-15a3211d77b2","order_by":1,"name":"Yiping Hu","email":"","orcid":"","institution":"Southwest Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yiping","middleName":"","lastName":"Hu","suffix":""},{"id":627773234,"identity":"5cc6872b-296d-4c13-9dc3-e659849d13af","order_by":2,"name":"Yu Xiong","email":"","orcid":"","institution":"Southwest Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yu","middleName":"","lastName":"Xiong","suffix":""},{"id":627773235,"identity":"8a0f84b4-f637-4542-a0f2-d9eb0b1db197","order_by":3,"name":"Yang Yu","email":"","orcid":"","institution":"Southwest Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yang","middleName":"","lastName":"Yu","suffix":""},{"id":627773245,"identity":"761df1c3-943b-420a-b947-99d280b61025","order_by":4,"name":"Chunxiang Zhang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAArElEQVRIiWNgGAWjYBADOTb29gNEq2ZsABLGfDxnEkjTkjhPwsGAOPXy7WfMH3yo2ZbeJsGQwPCjYhsRVvSkJTbOOHY7t0268QBjz5nbhLUwSzAfbOZhA2qROZDAzNhGhBY2CcbG5j//bqezSSQYEKeFB2QLUGUC8VokeNISZ/b23TZsAwbyQaL8Agwxgw8/vt2Wl29vP/jgRwURWlDAARLVj4JRMApGwSjABQCV5TuBj7UD5QAAAABJRU5ErkJggg==","orcid":"","institution":"Southwest Medical University","correspondingAuthor":true,"prefix":"","firstName":"Chunxiang","middleName":"","lastName":"Zhang","suffix":""}],"badges":[],"createdAt":"2026-03-21 12:10:03","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9185668/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9185668/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107708441,"identity":"1d1ed29f-08a0-41df-841e-588965c014b8","added_by":"auto","created_at":"2026-04-24 09:27:05","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":25446305,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStudy design and analytical workflow.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLegend:\u003c/strong\u003e Overview of the integrated study design. Bidirectional two-sample Mendelian randomization (MR) was performed to assess causal directionality between heart failure (HF) and septicemia using FinnGen GWAS data. Guided by MR findings, machine-learning models were developed in MIMIC-IV and externally validated in MIMIC-III to predict Sepsis-3–defined sepsis among elderly ICU patients with HF, followed by model interpretation using SHAP.\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-9185668/v1/e5f5dc59a528bf23c4c5612a.png"},{"id":107708454,"identity":"e6d099a9-9c63-4bde-bce1-5a897bc2158a","added_by":"auto","created_at":"2026-04-24 09:27:11","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":14965341,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eForward MR analysis of heart failure liability on septicemia risk.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLegend:\u003c/strong\u003e (A) Forest plot of causal effect estimates across MR methods. (B) Scatter plot of SNP effects on heart failure and septicemia with fitted causal effect lines. (C) Leave-one-out analysis evaluating the influence of individual SNPs on the IVW estimate. (D). Funnel plot of SNP-specific causal estimates for heart failure and sepsis.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-9185668/v1/3ec7b2308205de67b76d0f15.png"},{"id":107709161,"identity":"39aef10f-9739-42f6-bebc-bedf464e162e","added_by":"auto","created_at":"2026-04-24 09:34:52","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":7144995,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eReverse MR analysis of septicemia liability on heart failure risk.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLegend:\u003c/strong\u003e (A) Forest plot of causal effect estimates across MR methods. (B) Scatter plot of SNP effects on septicemia and heart failure.\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-9185668/v1/ef164613768539d3542f322b.png"},{"id":107708486,"identity":"036f88f2-73ee-40e0-a8ad-b2b5e6b88bd3","added_by":"auto","created_at":"2026-04-24 09:27:34","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":14532150,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFeature selection for sepsis prediction in elderly ICU patients with heart failure.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLegend:\u003c/strong\u003e Results of feature selection using Boruta (A), LASSO (B-C), and recursive feature elimination (RFE) (D). The intersection of selected predictors was used for downstream model development.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-9185668/v1/0db40e36469af2c771ca4ac6.png"},{"id":107708460,"identity":"3e3607ac-ae34-4bdf-8aed-3ec0ee96afa9","added_by":"auto","created_at":"2026-04-24 09:27:14","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":15184468,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eModel performance and clinical utility.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e(A) ROC curves of candidate models in internal validation (MIMIC-IV). (B) ROC curves in external validation (MIMIC-III). (C) Decision curve analysis comparing net benefit across clinically relevant risk thresholds.\u003c/p\u003e","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-9185668/v1/7c22d64bc8be80268c694dac.png"},{"id":107708487,"identity":"94aa698d-bca2-4803-bef6-0f38462a8e53","added_by":"auto","created_at":"2026-04-24 09:27:34","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":14526406,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSHAP-based interpretation of the final SVM model.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLegend:\u003c/strong\u003e (A) Global SHAP summary plot showing the most influential predictors. (B) SHAP-based local explanation for a representative patient. (C-D) SHAP dependence plots for key predictors (SOFA score and white blood cell count [WBC]).\u003c/p\u003e","description":"","filename":"Figure6.png","url":"https://assets-eu.researchsquare.com/files/rs-9185668/v1/20813324d18afc8842b915c4.png"},{"id":107650935,"identity":"85a26cfd-c3c8-4aa8-ad12-eb7597bc308f","added_by":"auto","created_at":"2026-04-23 15:03:06","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":237168,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9185668/v1/8fd10883-9767-4dd3-9e3a-85963461f3b5.pdf"},{"id":107708462,"identity":"cb8faae5-8ed5-4560-919b-f6f31f9004de","added_by":"auto","created_at":"2026-04-24 09:27:15","extension":"doc","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":77824,"visible":true,"origin":"","legend":"","description":"","filename":"TableS1.doc","url":"https://assets-eu.researchsquare.com/files/rs-9185668/v1/d008694d317836b9443a2128.doc"},{"id":107708488,"identity":"e6995957-c910-4384-bfb8-6bc75ed3bc89","added_by":"auto","created_at":"2026-04-24 09:27:35","extension":"doc","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":34304,"visible":true,"origin":"","legend":"","description":"","filename":"TableS3.doc","url":"https://assets-eu.researchsquare.com/files/rs-9185668/v1/e39003907600b1a0eb1402a2.doc"},{"id":107708464,"identity":"568b7d9f-87cb-442c-aa5d-bd3d9a216352","added_by":"auto","created_at":"2026-04-24 09:27:16","extension":"doc","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":37376,"visible":true,"origin":"","legend":"","description":"","filename":"TableS2.doc","url":"https://assets-eu.researchsquare.com/files/rs-9185668/v1/63e0906789b5bfb3ba3e7717.doc"},{"id":107708450,"identity":"a7036bc6-28c5-42a9-b876-555b44f6c827","added_by":"auto","created_at":"2026-04-24 09:27:10","extension":"pdf","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":301655,"visible":true,"origin":"","legend":"","description":"","filename":"Additionalfile.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9185668/v1/1660de70bae52122ea8b98b6.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Interpretable Machine Learning for Sepsis Risk Prediction in Elderly ICU Patients With Heart Failure: Integrating Genetic Evidence and Clinical Data","fulltext":[{"header":"Introduction","content":"\u003cp\u003eHeart failure (HF) and sepsis are among the leading causes of morbidity and mortality in critically ill patients, particularly in the elderly population. These conditions frequently coexist in the intensive care unit (ICU), where patients with pre-existing HF often experience infectious complications, and sepsis may precipitate acute cardiac dysfunction.\u003csup\u003e[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]\u003c/sup\u003e Observational studies have consistently reported associations between HF and adverse infectious outcomes\u003csup\u003e[\u003cspan additionalcitationids=\"CR3 CR4\" citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]\u003c/sup\u003e; however, the directionality of this relationship remains uncertain. Chronic HF is characterized by systemic inflammation\u003csup\u003e[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/sup\u003e, neurohormonal activation\u003csup\u003e[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]\u003c/sup\u003e, endothelial dysfunction\u003csup\u003e[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]\u003c/sup\u003e, and multi-organ impairment\u003csup\u003e[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]\u003c/sup\u003e\u0026mdash;factors that may plausibly increase susceptibility to infection and sepsis\u003csup\u003e[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]\u003c/sup\u003e. Conversely, sepsis is well recognized to induce transient myocardial dysfunction and long-term cardiovascular complications\u003csup\u003e[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]\u003c/sup\u003e. Shared risk factors and residual confounding further complicate causal interpretation, making it difficult to determine whether HF predisposes individuals to sepsis or whether observed associations primarily reflect reverse causation and overlapping pathophysiology\u003csup\u003e[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eMendelian randomization (MR) offers a framework for enhancing causal inference by utilizing genetic variants as instrumental variables, thus minimizing confounding and reverse causation bias. Bidirectional two-sample MR is particularly useful in disentangling complex relationships between traits with potential reciprocal associations. Despite growing interest in the interplay between cardiovascular disease and systemic infection, no large-scale bidirectional MR study has comprehensively evaluated the directional relationship between genetic liability to HF and sepsis risk using contemporary genome-wide association data\u003csup\u003e[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]\u003c/sup\u003e. Clarifying this directionality is clinically relevant, as it determines whether HF should be regarded primarily as a downstream consequence of sepsis or as an upstream susceptibility state.\u003c/p\u003e \u003cp\u003eEstablishing causal direction at the population level, however, does not directly translate into individualized clinical decision-making. If HF represents a risk-enriched state for sepsis, patients with HF\u0026mdash;especially elderly ICU patients\u0026mdash;may benefit from targeted risk stratification and early surveillance strategies\u003csup\u003e[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]\u003c/sup\u003e. Existing sepsis prediction models are typically developed in heterogeneous ICU populations and rarely focus specifically on patients with HF\u003csup\u003e[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]\u003c/sup\u003e. Moreover, many models lack independent external validation and interpretable frameworks, limiting their generalizability and clinical adoption. Machine learning (ML) approaches offer the ability to model complex, nonlinear interactions among clinical variables\u003csup\u003e[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]\u003c/sup\u003e, and the incorporation of explainability methods such as SHapley Additive exPlanations (SHAP) enhances transparency and interpretability\u003csup\u003e[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eIn this study, we integrated bidirectional two-sample MR with clinically applicable ML modeling to bridge genetic inference and bedside risk assessment. First, using FinnGen genome-wide association summary statistics, we investigated the directional relationship between genetic liability to HF and sepsis. Guided by the MR findings, we subsequently developed and externally validated ML models to predict Sepsis-3\u0026ndash;defined sepsis among elderly ICU patients with HF, using MIMIC-IV for model development and MIMIC-III for validation. Through this integrated approach, we aimed to clarify the directional association between HF and sepsis and to translate genetic insights into an interpretable framework for individualized sepsis risk stratification.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStudy Design\u003c/h2\u003e \u003cp\u003eThis study utilized an integrated framework that combines genetic and clinical modeling to investigate the relationship between heart failure (HF) and sepsis. To begin, a bidirectional two-sample Mendelian randomization (MR) analysis was performed to assess potential causal connections between genetically predicted susceptibility to HF and the likelihood of developing sepsis. Second, based on the causal direction suggested by the MR findings, ML\u0026ndash;based prediction models were developed and externally validated to estimate the occurrence of sepsis among elderly ICU patients. The overall analytical workflow is summarized in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eData Sources\u003c/h3\u003e\n\u003cp\u003eSummary-level genome-wide association study (GWAS) data were obtained from the 2021 release of the FinnGen consortium, focusing on individuals of European descent. Genetic instruments for all-cause HF were derived from FinnGen (finn-b-I9_HEARTFAIL_ALLCAUSE; 23,397 cases and 194,811 controls), and summary statistics for sepsis were obtained from FinnGen (finn-b-AB1_SEPSIS; 6,164 cases and 197,660 controls). All GWAS datasets were based on the GRCh37 (hg19) reference genome build.\u003c/p\u003e \u003cp\u003eData for model training and internal validation were obtained from version 3.1 of the Medical Information Mart for Intensive Care IV (MIMIC-IV), whereas the MIMIC-III database was used for external validation. MIMIC-IV comprises de-identified health records from patients admitted to the intensive care units of Beth Israel Deaconess Medical Center (Boston, MA, USA) between 2008 and 2022. Access to the databases was granted following completion of the requisite data use agreements and approval by the institutional review board, with a waiver of informed consent (certification number 74299605).\u003c/p\u003e\n\u003ch3\u003eStudy Population and Outcome Measurement\u003c/h3\u003e\n\u003cp\u003eFor the machine-learning (ML) analysis, eligible patients were required to satisfy the following conditions: (1) first ICU admission during the index hospitalisation, (2) an ICU stay of at least 24 h, and (3) age between 65 and 100 years at ICU admission. HF was defined according to International Classification of Diseases (ICD) diagnostic codes documented during the hospital stay.\u003c/p\u003e \u003cp\u003eThe main outcome was the development of sepsis during the intensive care unit stay, defined according to the Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). Sepsis events were identified using the validated MIMIC code repository implementation, based on documented or suspected infection in combination with acute organ dysfunction.\u003c/p\u003e\n\u003ch3\u003ePredictor Variables\u003c/h3\u003e\n\u003cp\u003eBaseline predictive variables were extracted from the MIMIC-IV database via structured query language, with specifications that these values represented the earliest measurements obtained after ICU admission. Collected features included demographic characteristics (e.g., age, sex, and body weight), major comorbidities, medication exposures and clinical interventions, severity scores, and laboratory measurements. The detailed information of all candidate predictor variables\u0026mdash;including specific variable names, operational definitions, data types (continuous/categorical), and measurement timelines\u0026mdash;is summarized in Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eComorbidities encompassed chronic cardiovascular, neurologic, metabolic, respiratory, renal, and hepatic conditions, as well as malignancy. Treatment-related variables included antimicrobial exposures (glycopeptides, macrolides, β-lactam antibiotics, and quinolones), vasoactive agent use, mechanical ventilation, continuous renal replacement therapy, sedatives, albumin administration, and other commonly administered therapies. Disease severity was evaluated using the Sequential Organ Failure Assessment (SOFA) score and the Simplified Acute Physiology Score II (SAPS II). Laboratory variables consisted of routinely collected hematologic, biochemical, coagulation, and arterial blood gas parameters.\u003c/p\u003e\n\u003ch3\u003eMissing Data and Sample Size Considerations\u003c/h3\u003e\n\u003cp\u003ePredictors containing over 20% missing data were excluded from subsequent analyses. For the remaining variables, missing values were addressed through multiple imputation by chained equations.All imputation and preprocessing procedures were performed exclusively within the training dataset to minimize the risk of data leakage, and the same transformations were subsequently applied to the testing and external validation cohorts. Given the retrospective study design and the utilization of large critical care databases, no formal a priori sample size calculation was performed.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analysis\u003c/h2\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003eMR Analysis\u003c/h2\u003e \u003cp\u003eA bidirectional two-sample MR design was employed. In the forward analysis, genetic variants linked to all-cause HF served as instrumental variables, with sepsis designated as the outcome.In the reverse analysis, genetic variants associated with sepsis were used as instrumental variables, with all-cause HF as the outcome. SNPs were selected at P\u0026thinsp;\u0026lt;\u0026thinsp;5 \u0026times; 10⁻⁶, clumped for linkage disequilibrium (r\u0026sup2; \u0026lt; 0.001; 10,000 kb), with all instruments showing F-statistics\u0026thinsp;\u0026gt;\u0026thinsp;10. Summary statistics were standardized by aligning effect alleles across datasets, and palindromic SNPs with ambiguous strand orientation were removed.\u003c/p\u003e \u003cp\u003eCausal effect estimates were primarily obtained using the inverse-variance weighted (IVW) method. Sensitivity analyses were carried out using MR-Egger regression, weighted median, simple mode, and weighted mode approaches. Directional horizontal pleiotropy was evaluated via the MR-Egger intercept test. Leave-one-out analyses and MR-PRESSO were utilized to assess the robustness of causal estimates and identify potential outlier instruments.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e\n\u003ch3\u003eML Analysis\u003c/h3\u003e\n\u003cp\u003eThe MIMIC-IV cohort was randomly split into a training set (70%) and an internal validation set (30%).The independent MIMIC-III cohort was used for external validation. Multiple ML algorithms were developed and compared, including logistic regression, SVM, GBM, neural network, KNN, Xgboost, AdaBoost, LightGBM, and CatBoost. Continuous variables were standardized for algorithms sensitive to feature scaling. Model hyperparameters were adjusted within the training dataset using grid-search procedures combined with cross-validation. The specific parameter grids applied to each machine-learning algorithm are detailed in Supplementary Table S3. Discriminative performance was assessed using the area under the receiver operating characteristic curve (AUC), alongside measures of sensitivity, specificity, and calibration. The identical set of metrics was employed to evaluate performance in the external validation cohort. All statistical analyses were conducted using R software (version 4.4.2).\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e \u003cb\u003e1. Forward MR analysis of HF on sepsis risk\u003c/b\u003e \u003c/p\u003e \u003cp\u003eIn the forward Mendelian randomisation (MR) analysis, HF-related genetic variants were employed as instrumental variables to explore a potential causal relationship between HF and the subsequent risk of sepsis. In total, 24 independent single-nucleotide polymorphisms (SNPs) were selected as instruments.\u003c/p\u003e \u003cp\u003eUsing the IVW method as the primary analysis, genetically predicted liability to HF showed evidence of an association with an increased risk of sepsis (OR\u0026thinsp;=\u0026thinsp;1.295, 95% CI: 1.091\u0026ndash;1.538, P\u0026thinsp;=\u0026thinsp;0.003). Consistent estimates were observed using the weighted median method (OR\u0026thinsp;=\u0026thinsp;1.305, 95% CI: 1.069\u0026ndash;1.593, P\u0026thinsp;=\u0026thinsp;0.009). Estimates derived from MR-Egger and mode-based methods were directionally consistent with the primary analysis but less precise, with wider confidence intervals (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe MR-Egger intercept test did not indicate evidence of directional horizontal pleiotropy (intercept\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;0.005, P\u0026thinsp;=\u0026thinsp;0.749). Moderate heterogeneity was observed (IVW Q P\u0026thinsp;=\u0026thinsp;0.042; MR-Egger Q P\u0026thinsp;=\u0026thinsp;0.032); however, the MR-PRESSO global test was not significant (P\u0026thinsp;=\u0026thinsp;0.07), suggesting that the observed heterogeneity was unlikely to be driven by strong directional pleiotropy. Scatter plots showed generally positive slopes across MR methods (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB). Leave-one-out analyses indicated that the IVW estimate was not driven by any single SNP (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eC), and visual inspection of funnel plots did not suggest marked asymmetry (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eD).\u003c/p\u003e \u003cp\u003e \u003cb\u003e2. Reverse MR analysis of sepsis on HF risk\u003c/b\u003e \u003c/p\u003e \u003cp\u003eIn the reverse Mendelian randomisation (MR) analysis, sepsis-associated genetic variants were selected as instrumental variables to evaluate whether sepsis exerts a potential causal influence on HF risk. Five independent single-nucleotide polymorphisms (SNPs) were ultimately retained as instrumental variables.\u003c/p\u003e \u003cp\u003eAcross multiple MR approaches, no statistically significant evidence was found to support a causal effect of genetically predicted sepsis liability on HF risk. The IVW analysis yielded an odds ratio (OR) of 1.170 (95% confidence interval [CI]: 0.953\u0026ndash;1.435, P\u0026thinsp;=\u0026thinsp;0.134). Consistent null or near-null estimates were obtained using complementary methods, including the weighted median (OR\u0026thinsp;=\u0026thinsp;1.036, 95% CI: 0.892\u0026ndash;1.202, P\u0026thinsp;=\u0026thinsp;0.645), simple mode (OR\u0026thinsp;=\u0026thinsp;1.025, 95% CI: 0.869\u0026ndash;1.209, P\u0026thinsp;=\u0026thinsp;0.781), and weighted mode (OR\u0026thinsp;=\u0026thinsp;1.027, 95% CI: 0.881\u0026ndash;1.197, P\u0026thinsp;=\u0026thinsp;0.751) (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA). Scatter plots did not show a consistent pattern across MR methods (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eMR-Egger regression similarly did not indicate a statistically significant association (OR\u0026thinsp;=\u0026thinsp;0.880, 95% CI: 0.667\u0026ndash;1.161, P\u0026thinsp;=\u0026thinsp;0.433). The MR-Egger intercept test did not provide evidence of directional horizontal pleiotropy (intercept\u0026thinsp;=\u0026thinsp;0.049, P\u0026thinsp;=\u0026thinsp;0.101). Cochran\u0026rsquo;s Q statistics indicated heterogeneity in the IVW model (Q P\u0026thinsp;=\u0026thinsp;0.008), whereas heterogeneity was not evident under the MR-Egger model (Q P\u0026thinsp;=\u0026thinsp;0.183). The MR-PRESSO global test suggested overall heterogeneity (P\u0026thinsp;=\u0026thinsp;0.032), while no outlier SNPs were identified by the outlier test. Leave-one-out sensitivity analyses indicated that the overall null association was not driven by any single instrumental variant (Figure S2).\u003c/p\u003e \u003cp\u003eMotivated by the potential causal direction suggested by the bidirectional MR analyses, we subsequently evaluated whether routinely collected clinical features could be used to develop a ML\u0026ndash;based model to predict sepsis risk in patients with HF.\u003c/p\u003e \u003cp\u003e \u003cb\u003e3.1 Feature selection\u003c/b\u003e \u003c/p\u003e \u003cp\u003eFeature selection was performed to identify a stable subset of predictors for downstream model development. Boruta feature filtering confirmed 27 predictors with importance exceeding the corresponding shadow features (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eA). Subsequently, using LASSO regression with cross-validation, 36 predictors with non-zero coefficients were retained (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eB\u0026ndash;C). Meanwhile, recursive feature elimination (RFE) retained 43 predictors.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe intersection of predictors consistently selected across all three methods was used for subsequent model construction, yielding a final set of 17 predictors (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eD). These predictors encompassed antimicrobial use (glycopeptides, β-lactam antibiotics, macrolides, and quinolones), disease severity scores (SOFA and SAPS II), clinical interventions (mechanical ventilation, vasoactive agents, continuous renal replacement therapy, sedatives, and albumin administration), laboratory indicators (white blood cell count, serum calcium, and chloride), and baseline characteristics (age, atrial fibrillation, and furosemide use).\u003c/p\u003e \u003cp\u003e \u003cb\u003e3.2 Model development, internal testing, and external validation\u003c/b\u003e \u003c/p\u003e \u003cp\u003eUsing the 17 selected predictors, multiple ML models were evaluated in the internal testing cohort. ROC analysis demonstrated good discrimination across models, with AUCs ranging from 0.703 to 0.852 (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eA). NeuralNetwork achieved the highest AUC (0.852, 95% CI: 0.833\u0026ndash;0.870), followed by GBM (AUC\u0026thinsp;=\u0026thinsp;0.847, 95% CI: 0.828\u0026ndash;0.866) and CatBoost (AUC\u0026thinsp;=\u0026thinsp;0.845, 95% CI: 0.826\u0026ndash;0.864). Logistic regression and Xgboost showed comparable performance (AUCs\u0026thinsp;=\u0026thinsp;0.843 and 0.839, respectively), while SVM (AUC\u0026thinsp;=\u0026thinsp;0.835, 95% CI: 0.816\u0026ndash;0.855) and LightGBM (AUC\u0026thinsp;=\u0026thinsp;0.836, 95% CI: 0.817\u0026ndash;0.856) had slightly lower but similar discrimination; AdaBoost (AUC\u0026thinsp;=\u0026thinsp;0.780, 95% CI: 0.758\u0026ndash;0.802) and KNN (AUC\u0026thinsp;=\u0026thinsp;0.703, 95% CI: 0.681\u0026ndash;0.725) exhibited relatively lower discrimination, with KNN showing the poorest performance in internal validation. Given the notable performance variability of algorithms in internal testing and the potential overfitting risk of models with extreme internal performance, model selection was further based on external validation and clinical utility to ensure generalizability.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eModel performance was subsequently evaluated in an independent external cohort from MIMIC-III. ROC analysis demonstrated moderate discrimination across models, with AUCs ranging from 0.606 to 0.742 (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eB). Among the evaluated algorithms, SVM achieved the highest AUC (0.742, 95% CI: 0.721\u0026ndash;0.763), while logistic regression (AUC\u0026thinsp;=\u0026thinsp;0.732, 95% CI: 0.711\u0026ndash;0.754), CatBoost (AUC\u0026thinsp;=\u0026thinsp;0.715, 95% CI: 0.693\u0026ndash;0.736), and Xgboost (AUC\u0026thinsp;=\u0026thinsp;0.712, 95% CI: 0.690\u0026ndash;0.733) showed comparable discrimination. NeuralNetwork (AUC\u0026thinsp;=\u0026thinsp;0.703, 95% CI: 0.681\u0026ndash;0.726) and GBM (AUC\u0026thinsp;=\u0026thinsp;0.684, 95% CI: 0.661\u0026ndash;0.707) had lower performance, and LightGBM (AUC\u0026thinsp;=\u0026thinsp;0.667, 95% CI: 0.644\u0026ndash;0.689), AdaBoost (AUC\u0026thinsp;=\u0026thinsp;0.636, 95% CI: 0.616\u0026ndash;0.656) and KNN (AUC\u0026thinsp;=\u0026thinsp;0.606, 95% CI: 0.586\u0026ndash;0.625) exhibited the poorest discrimination in external validation. Table S2 summarizes the detailed threshold-based performance metrics of all models in external validation, including accuracy, sensitivity, specificity, and other key indicators. The SVM model exhibited balanced operational performance (with Accuracy\u0026thinsp;=\u0026thinsp;0.699, Sensitivity\u0026thinsp;=\u0026thinsp;0.666, Specificity\u0026thinsp;=\u0026thinsp;0.715, Precision\u0026thinsp;=\u0026thinsp;0.523, F1\u0026thinsp;=\u0026thinsp;0.586, and Kappa\u0026thinsp;=\u0026thinsp;0.31), outperforming all other algorithms across key clinical evaluation indicators.Decision curve analysis indicated that several models provided greater net benefit than the treat-all and treat-none strategies across clinically relevant threshold probabilities, with SVM demonstrating consistently favorable net benefit within intermediate risk thresholds (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eC). Based on its overall superior external performance and favorable clinical utility, SVM was identified as the optimal model for subsequent interpretation and explainability analyses.\u003c/p\u003e \u003cp\u003e \u003cb\u003e3.3 Model interpretation using SHAP\u003c/b\u003e \u003c/p\u003e \u003cp\u003eSHAP analysis was performed for the final SVM model to quantify the contribution of individual predictors to sepsis prediction. The global SHAP beeswarm plot (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eA) indicated that established severity scores, including SOFA and SAPS II, were among the most influential predictors. Treatment-related interventions such as vasoactive agent use, mechanical ventilation, and continuous renal replacement therapy also contributed substantially to the model output. In addition, antimicrobial exposures (e.g., glycopeptides, β-lactam antibiotics, and quinolones) and laboratory parameters including white blood cell count, serum calcium, and chloride levels showed notable contributions to the prediction.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo further illustrate patient-level interpretability, a representative high-risk case was examined using a SHAP-based local explanation (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eB). For this individual, the model output increased markedly from the baseline (E[f(x)]\u0026thinsp;=\u0026thinsp;0.55) to a much higher value, driven primarily by quinolone exposure, elevated SOFA score, glycopeptide use, and receipt of continuous renal replacement therapy, whereas not receiving mechanical ventilation (MV\u0026thinsp;=\u0026thinsp;0) contributed negatively. Overall, the combined effects of multiple clinical features jointly resulted in a high predicted sepsis risk.\u003c/p\u003e \u003cp\u003eTo further examine feature-specific effects, SHAP dependence plots were generated for key continuous predictors in the final SVM model. In the SOFA dependence plot, increasing SOFA values were generally associated with higher SHAP values, indicating greater contributions to predicted sepsis risk (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eC). A similar pattern was observed for white blood cell count (WBC), where higher values were associated with increased SHAP values, suggesting that inflammatory burden contributed positively to predicted sepsis risk (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eD). Additional SHAP dependence plots for treatment-related variables, antimicrobial exposures, and other laboratory parameters are provided in the Supplementary Materials.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003ePrincipal findings\u003c/h2\u003e \u003cp\u003eIn this study, we integrated bidirectional two-sample MR with clinical ML to clarify the directional relationship between HF and sepsis and to translate this insight into individualized risk prediction. The forward MR analysis demonstrated that genetically predicted liability to HF was consistently associated with an increased risk of sepsis across multiple estimators, whereas the reverse MR analysis provided no evidence that genetic liability to sepsis causally increases HF risk. These findings support a direction-specific relationship in which HF represents a predisposing state for sepsis, rather than a consequence of sepsis-related genetic susceptibility. Importantly, the absence of genetic evidence for a reverse causal effect does not contradict the well-recognized occurrence of acute cardiac dysfunction during severe sepsis \u003csup\u003e[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/sup\u003e. Sepsis-related myocardial injury and transient ventricular dysfunction are likely driven by short-term inflammatory, metabolic, and hemodynamic insults \u003csup\u003e[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]\u003c/sup\u003e and are not captured by genetic liability to sepsis or by pathways underlying chronic HF. Our results therefore refine, rather than negate, existing clinical observations by distinguishing long-term causal susceptibility from acute reversible organ dysfunction \u003csup\u003e[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003ePotential mechanisms linking HF to sepsis susceptibility\u003c/h2\u003e \u003cp\u003eSeveral biological and clinical mechanisms may explain why HF predisposes individuals to sepsis. Chronic HF is characterized by persistent systemic inflammation, neurohormonal activation, endothelial dysfunction, and immune dysregulation, all of which may impair host defense and increase vulnerability to infection \u003csup\u003e[\u003cspan additionalcitationids=\"CR20\" citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]\u003c/sup\u003e. Altered innate and adaptive immune responses observed in HF may further facilitate progression from infection to sepsis \u003csup\u003e[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]\u003c/sup\u003e. In addition, HF is frequently accompanied by multi-organ dysfunction, including renal and hepatic impairment, which can reduce physiological reserve and immune competence.\u003c/p\u003e \u003cp\u003eHemodynamic abnormalities inherent to HF, such as tissue hypoperfusion and venous congestion\u0026mdash;particularly intestinal congestion\u0026mdash;may compromise gut barrier integrity and promote bacterial translocation, amplifying systemic inflammation and increasing the risk of sepsis \u003csup\u003e[\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]\u003c/sup\u003e. These mechanisms are broadly consistent with the predictors identified by our model, including disease severity scores, inflammatory markers, and early ICU interventions that reflect clinical deterioration.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eImplications for risk stratification and early identification\u003c/h2\u003e \u003cp\u003eIdentifying HF as a causal risk-enriching state for sepsis has important implications for prevention and early recognition, particularly among elderly ICU patients. Existing sepsis risk assessment approaches largely focus on acute physiological derangements and infectious triggers, with limited consideration of underlying cardiovascular vulnerability \u003csup\u003e[\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]\u003c/sup\u003e. Our findings suggest that incorporating baseline HF-related susceptibility alongside early clinical signals may improve identification of patients at high risk for sepsis, enabling closer monitoring and earlier escalation of care.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eFrom genetic evidence to predictive modeling\u003c/h2\u003e \u003cp\u003eWhile MR strengthens causal inference by reducing confounding and reverse causation, it does not provide individualized risk estimates suitable for bedside decision-making \u003csup\u003e[\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]\u003c/sup\u003e. Accordingly, we developed and externally validated ML models based on routinely collected clinical features to support individualized sepsis risk prediction among patients with HF. The alignment between the MR-supported causal direction and the modeling objective strengthens the coherence of our approach: MR provides etiological justification for the prediction target, whereas the clinical model offers practical utility for risk stratification. Although several algorithms achieved excellent discrimination in internal testing, performance in external validation was more modest, underscoring the importance of independent evaluation and mitigating concerns regarding overfitting \u003csup\u003e[\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]\u003c/sup\u003e. Among the evaluated models, the SVM demonstrated the most favorable balance between discrimination and clinical utility in external validation. Importantly, model interpretability using SHAP revealed clinically plausible predictors, including illness severity scores, early ICU interventions, antimicrobial exposures, and laboratory markers, enhancing transparency and supporting clinical insight rather than automated decision-making \u003csup\u003e[\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eStrengths and limitations\u003c/h2\u003e \u003cp\u003e This study is supported by several important strengths.We employed a bidirectional MR design using large-scale GWAS summary statistics to clarify causal directionality, followed by the development of externally validated prediction models across two independent critical care databases. The integration of causal inference with interpretable clinical modeling represents a translational framework that links etiological insight with practical risk stratification. Several limitations should be acknowledged. First, the FinnGen sepsis phenotype may not fully correspond to the Sepsis-3 definition used in the clinical cohorts, introducing potential phenotype heterogeneity \u003csup\u003e[\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]\u003c/sup\u003e. Second, although extensive sensitivity analyses were performed, key MR assumptions\u0026mdash;particularly the exclusion restriction and absence of horizontal pleiotropy\u0026mdash;cannot be fully verified \u003csup\u003e[\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]\u003c/sup\u003e. Third, some predictors incorpoated in the ML models reflect early ICU interventions and may act as proxies for illness severity or evolving clinical decision-making rather than purely antecedent biological risk. Consequently, model outputs should be interpreted as risk stratification tools rather than causal estimators \u003csup\u003e[\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]\u003c/sup\u003e. Fourth, the external validation cohort originated from the same medical center as the development dataset, albeit from a different database version, which may limit full geographic generalizability. Finally, our findings are specific to elderly ICU patients with HF and may not extrapolate to younger populations or non-ICU settings.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusions","content":"\u003cp\u003eIn summary, genetic liability to HF was associated with an increased risk of sepsis in bidirectional MR analyses, whereas genetic liability to sepsis was not associated with HF risk. By integrating causal inference with externally validated and interpretable ML models, this study provides a translational framework for sepsis risk stratification among elderly ICU patients with HF. These findings support the concept that HF represents a susceptibility state for sepsis and highlight opportunities for targeted risk assessment in this vulnerable population.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"502\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 154px;\"\u003e\n \u003cp\u003eAUC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 348px;\"\u003e\n \u003cp\u003eArea Under the Curve\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 154px;\"\u003e\n \u003cp\u003eCI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 348px;\"\u003e\n \u003cp\u003eConfidence Interval\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 154px;\"\u003e\n \u003cp\u003eGWAS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 348px;\"\u003e\n \u003cp\u003eGenome-Wide Association Study\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 154px;\"\u003e\n \u003cp\u003eHF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 348px;\"\u003e\n \u003cp\u003eHeart Failure\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 154px;\"\u003e\n \u003cp\u003eICD\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 348px;\"\u003e\n \u003cp\u003eInternational Classification of Diseases\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 154px;\"\u003e\n \u003cp\u003eICU\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 348px;\"\u003e\n \u003cp\u003eIntensive Care Unit\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 154px;\"\u003e\n \u003cp\u003eIVW\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 348px;\"\u003e\n \u003cp\u003eInverse-Variance Weighted\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 154px;\"\u003e\n \u003cp\u003eML\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 348px;\"\u003e\n \u003cp\u003eMachine Learning\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 154px;\"\u003e\n \u003cp\u003eMR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 348px;\"\u003e\n \u003cp\u003eMendelian Randomization\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 154px;\"\u003e\n \u003cp\u003eOR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 348px;\"\u003e\n \u003cp\u003eOdds Ratio\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 154px;\"\u003e\n \u003cp\u003eROC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 348px;\"\u003e\n \u003cp\u003eReceiver Operating Characteristic\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 154px;\"\u003e\n \u003cp\u003eSAPS II\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 348px;\"\u003e\n \u003cp\u003eSimplified Acute Physiology Score II\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 154px;\"\u003e\n \u003cp\u003eSHAP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 348px;\"\u003e\n \u003cp\u003eSHapley Additive Explanations\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 154px;\"\u003e\n \u003cp\u003eSOFA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 348px;\"\u003e\n \u003cp\u003eSequential Organ Failure Assessment\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd nowrap=\"\" style=\"width: 154px;\"\u003e\n \u003cp\u003eSNPs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd nowrap=\"\" style=\"width: 348px;\"\u003e\n \u003cp\u003eSingle-Nucleotide Polymorphisms\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study analysed anonymised, de-identified data obtained from the Medical Information Mart for Intensive Care IV (MIMIC-IV) and Medical Information Mart for Intensive Care III (MIMIC-III) databases. Database access was authorised after the investigators completed the mandatory training and executed the relevant data use agreements. Use of these resources was approved by the institutional review boards of the Massachusetts Institute of Technology and Beth Israel Deaconess Medical Center, and the requirement for informed consent was waived because the study was retrospective and involve Ethics approval and consent to participate only de-identified information. Therefore, no further ethical approval was necessary for the present analysis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable, as this study does not contain any individual person\u0026rsquo;s data in any form.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe FinnGen GWAS summary statistics used in this study are publicly available from the FinnGen consortium website (https://www.finngen.fi/en/access_results). The MIMIC-IV (version 3.1) and MIMIC-III databases are publicly accessible through PhysioNet (https://physionet.org), subject to completion of the required training (e.g., CITI Program courses) and data use agreements. All data used in this study can be obtained from the original sources under the corresponding access policies.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCode availability statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe code used for data preprocessing, statistical analysis, and ML model development in this study is available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflicts of interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFunding: This work was supported by National Natural Science Foundation of China (Grant Nos. U23A20398), Noncommunicable Chronic Diseases-National Science and Technology Major Project(Grant Nos. 2024ZD0537707), Sichuan Science and Technology Program (Grant Nos. 2025YFRG0005), The People\u0026apos;s Government of Luzhou Municipality-Southwest Medical University Science and Technology Strategic Cooperation \u0026quot;Science and Technology Climbing Program\u0026quot; (Grant Nos. 2025LZXNYDPD01), Research Start-up Foundation of Southwest Medical University (Grant Nos. 00040155).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026rsquo; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eLW and YH contributed to the study conception and design. LW, YH, and YX contributed to data acquisition and curation. LW and YH performed the statistical analysis and machine learning modeling. YX contributed to data interpretation. YY and CZ supervised the study and provided critical revisions. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors thank the FinnGen consortium for providing access to the genome-wide association study summary statistics and the PhysioNet team for maintaining the MIMIC-IV and MIMIC-III databases. The authors also thank the research team at the Basic Medicine Research Innovation Center for Cardiometabolic Diseases for their support.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eZhang B, Guo S, Fu Z, Wu N, Liu Z. Association between fluid balance and mortality for heart failure and sepsis: a propensity score-matching analysis. BMC Anesthesiol. 2022;22(1):324.\u003c/li\u003e\n\u003cli\u003ePalin V, Brown O, Hamilton F, Lillie P, Kearney M, Cubbon R, et al. Infection in people with heart failure: an overlooked cause of adverse outcomes. Clin Med (Lond). 2025;25(5):100497.\u003c/li\u003e\n\u003cli\u003eDrozd M, Garland E, Walker AMN, Slater TA, Koshy A, Straw S, et al. Infection-related hospitalization in heart failure with reduced ejection fraction: a prospective observational cohort study. Circ Heart Fail. 2020;13(5):e006746.\u003c/li\u003e\n\u003cli\u003eMolinsky RL, Shah A, Yuzefpolskaya M, Yu B, Misialek JR, Bohn B, et al. Infection-related hospitalization and incident heart failure: the Atherosclerosis Risk in Communities Study. J Am Heart Assoc. 2025;14(3):e033877.\u003c/li\u003e\n\u003cli\u003eCorrales-Medina VF, Alvarez KN, Weissfeld LA, Angus DC, Chirinos JA, Chang CC, et al. Association between hospitalization for pneumonia and subsequent risk of cardiovascular disease. JAMA. 2015;313(3):264\u0026ndash;274.\u003c/li\u003e\n\u003cli\u003eMurphy SP, Kakkar R, McCarthy CP, Januzzi JL Jr. Inflammation in heart failure: JACC state-of-the-art review. J Am Coll Cardiol. 2020;75(11):1324\u0026ndash;1340.\u003c/li\u003e\n\u003cli\u003eMcDonagh TA, Metra M, Adamo M, Gardner RS, Baumbach A, Bohm M, et al. 2021 ESC guidelines for the diagnosis and treatment of acute and chronic heart failure. Eur Heart J. 2021;42(36):3599\u0026ndash;3726.\u003c/li\u003e\n\u003cli\u003eTsigkou V, Oikonomou E, Anastasiou A, Lampsas S, Zakynthinos GE, Kalogeras K, et al. Molecular mechanisms and therapeutic implications of endothelial dysfunction in patients with heart failure. Int J Mol Sci. 2023;24(5):4321\u003c/li\u003e\n\u003cli\u003eCiccarelli M, Dawson D, Falcao-Pires I, Giacca M, Hamdani N, Heymans S, et al. Reciprocal organ interactions during heart failure: a position paper from the ESC Working Group on Myocardial Function. Cardiovasc Res. 2021;117(12):2416\u0026ndash;2433.\u003c/li\u003e\n\u003cli\u003eSato R, Sanfilippo F, Lanspa M, Duggal A, Dugar S. Sepsis-induced cardiomyopathy: mechanism, prevalence, assessment, prognosis, and management. Chest. 2025;168(6):1383\u0026ndash;1394.\u003c/li\u003e\n\u003cli\u003eBeane A, Shankar-Hari M. Long-term ill health in sepsis survivors: an ignored health-care challenge? Lancet. 2024;404(10459):1178\u0026ndash;1180.\u003c/li\u003e\n\u003cli\u003eSattar N, Preiss D. Reverse causality in cardiovascular epidemiological research: more common than imagined? Circulation. 2017;135(24):2369\u0026ndash;2372.\u003c/li\u003e\n\u003cli\u003eLiu L, Huang P, Wang C, Liu Y, Gao Y, Yu K. Causal association between heart failure and sepsis: insights from Mendelian randomization and observational studies. Clin Epidemiol. 2024;16:755\u0026ndash;767.\u003c/li\u003e\n\u003cli\u003eZhang Q, Xu L, He W, et al. Survival prediction for heart failure complicated by sepsis: based on machine learning methods. Front Med (Lausanne). 2024;11:1410702.\u003c/li\u003e\n\u003cli\u003eMou C, Yang J, Wu Q, Qin L, Lu J. Progress in sepsis prediction models: from traditional scoring systems to multimodal intelligence and clinical translation. Front Med (Lausanne). 2026;13:1732164.\u003c/li\u003e\n\u003cli\u003eZhou H, Li F, Liu X. Early prediction of septic shock in ICU patients using machine learning: development, external validation, and explainability with SHAP. Int J Med Inform. 2026;206:106169.\u003c/li\u003e\n\u003cli\u003eEhrman RR, Sullivan AN, Favot MJ, Sherwin RL, Reynolds CA, Abidov A, et al. Pathophysiology, echocardiographic evaluation, biomarker findings, and prognostic implications of septic cardiomyopathy: a review of the literature. Crit Care. 2018;22(1):112.\u003c/li\u003e\n\u003cli\u003eCarbone F, Liberale L, Preda A, Schindler TH, Montecucco F. Septic cardiomyopathy: from pathophysiology to the clinical setting. Cells. 2022;11(18):2833. \u003c/li\u003e\n\u003cli\u003eMann DL. Inflammatory mediators and the failing heart: past, present, and the foreseeable future. Circ Res. 2002;91(11):988\u0026ndash;998.\u003c/li\u003e\n\u003cli\u003eVan Linthout S, Tsch\u0026ouml;pe C. Inflammation \u0026ndash; cause or consequence of heart failure or both? Curr Heart Fail Rep. 2017;14(4):251\u0026ndash;265.\u003c/li\u003e\n\u003cli\u003eDick SA, Epelman S. Chronic heart failure and inflammation: what do we really know? Circ Res. 2016;119(1):159\u0026ndash;176.\u003c/li\u003e\n\u003cli\u003eYndestad A, Damas JK, Oie E, Ueland T, Gullestad L, Aukrust P. Systemic inflammation in heart failure\u0026mdash;the whys and wherefores. Heart Fail Rev. 2006;11(1):83\u0026ndash;92.\u003c/li\u003e\n\u003cli\u003eSandek A, Bauditz J, Swidsinski A, Buhner S, Weber-Eibel J, von Haehling S, et al. Altered intestinal function in patients with chronic heart failure. J Am Coll Cardiol. 2007;50(16):1561\u0026ndash;1569.\u003c/li\u003e\n\u003cli\u003eKrack A, Sharma R, Figulla HR, Anker SD. The importance of the gastrointestinal system in the pathogenesis of heart failure. Eur Heart J. 2005;26(22):2368\u0026ndash;2374.\u003c/li\u003e\n\u003cli\u003eSinger M, Deutschman CS, Seymour CW, Shankar-Hari M, Annane D, Bauer M, et al. The third international consensus definitions for sepsis and septic shock (Sepsis-3). JAMA. 2016;315(8):801\u0026ndash;810.\u003c/li\u003e\n\u003cli\u003eSchertz AR, Lenoir KM, Bertoni AG, Rosamond WD, Chang PP, Granger CB, et al. Sepsis prediction model for determining sepsis vs SIRS, qSOFA, and SOFA. JAMA Netw Open. 2023;6(8):e2329729.\u003c/li\u003e\n\u003cli\u003eEmdin CA, Khera AV, Kathiresan S. Mendelian randomization. JAMA. 2017;318(19):1925\u0026ndash;1926.\u003c/li\u003e\n\u003cli\u003eDavies NM, Holmes MV, Davey Smith G. Reading Mendelian randomisation studies: a guide, glossary, and checklist for clinicians. BMJ. 2018;362:k601.\u003c/li\u003e\n\u003cli\u003eCollins GS, Reitsma JB, Altman DG, Moons KGM. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. BMJ. 2015;350:g7594.\u003c/li\u003e\n\u003cli\u003eCollins GS, Moons KGM. Reporting of artificial intelligence prediction models. Lancet. 2019;393(10181):1577\u0026ndash;1579.\u003c/li\u003e\n\u003cli\u003eZhang D, Yin C, Hunold KM, Banerjee A, Chen Y, Sun J, et al. An interpretable deep-learning model for early prediction of sepsis in the emergency department. Patterns (N Y). 2021;2(2):100196.\u003c/li\u003e\n\u003cli\u003eKurki MI, Karjalainen J, Palta P, Sipil\u0026auml; TP, Kristiansson K, Donner K, et al. FinnGen provides genetic insights from a well-phenotyped isolated population. Nature. 2023;613(7944):508\u0026ndash;518.\u003c/li\u003e\n\u003cli\u003eBowden J, Davey Smith G, Burgess S. Mendelian randomization with invalid instruments: effect estimation and bias detection through Egger regression. Int J Epidemiol. 2015;44(2):512\u0026ndash;525.\u003c/li\u003e\n\u003cli\u003eWynants L, Van Calster B, Collins GS, Riley RD, Heinze G, Schuit E, et al. Prediction models for diagnosis and prognosis of COVID-19: systematic review and critical appraisal. BMJ. 2020;369:m1328.\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":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Heart failure, Machine learning, Mendelian randomization, Risk prediction, Sepsis","lastPublishedDoi":"10.21203/rs.3.rs-9185668/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9185668/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e: Heart failure (HF) and sepsis frequently coexist in critically ill patients, yet the causal direction between these conditions remains unclear. Clarifying whether genetic susceptibility to HF predisposes individuals to sepsis may improve risk stratification in high-risk populations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods: \u003c/strong\u003eWe performed a bidirectional two-sample Mendelian randomization (MR) analysis using FinnGen genome-wide association study summary statistics to investigate the causal relationship between HF and sepsis. Based on the MR findings, machine learning (ML) models were developed to predict sepsis among elderly intensive care unit (ICU) patients with HF using the MIMIC-IV database and externally validated in the MIMIC-III cohort. Model performance was evaluated using discrimination, calibration, and clinical utility analyses, and SHapley Additive exPlanations (SHAP) were applied to improve model interpretability.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e: Genetically predicted liability to HF was associated with an increased risk of sepsis (inverse-variance weighted odds ratio 1.29, 95% confidence interval 1.09–1.54; P = 0.003), with consistent results across sensitivity analyses and no evidence of directional pleiotropy. Reverse MR analysis showed no causal association between genetic liability to sepsis and HF. In the external validation cohort, ML models demonstrated moderate discrimination (area under the curve 0.606–0.742), with the support vector machine model achieving the best performance (AUC 0.742, 95% CI 0.721–0.763). SHAP analysis identified illness severity scores, early ICU interventions, antimicrobial therapy, and laboratory indicators as key contributors to model predictions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e: Genetic liability to HF is associated with increased susceptibility to sepsis, highlighting a direction-specific relationship between these conditions. Integrating genetic causal inference with interpretable ML models provides a potential framework for sepsis risk prediction among elderly ICU patients with HF.\u003c/p\u003e","manuscriptTitle":"Interpretable Machine Learning for Sepsis Risk Prediction in Elderly ICU Patients With Heart Failure: Integrating Genetic Evidence and Clinical Data","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-23 15:03:00","doi":"10.21203/rs.3.rs-9185668/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewersInvited","content":"","date":"2026-04-15T14:20:51+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-03-26T14:36:45+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-03-23T10:33:29+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-03-23T10:32:28+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Informatics and Decision Making","date":"2026-03-21T11:53:43+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"ae473f2c-1bb3-4fc7-8710-f695373b721b","owner":[],"postedDate":"April 23rd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-04-23T15:03:01+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-23 15:03:00","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9185668","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9185668","identity":"rs-9185668","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}