Development of Machine Learning Algorithms for Predicting Vitamin B12 Levels Using Biochemical Analyte Data

Development of Machine Learning Algorithms for Predicting Vitamin B12 Levels Using Biochemical Analyte 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 Development of Machine Learning Algorithms for Predicting Vitamin B12 Levels Using Biochemical Analyte Data Ferhat Demirci, Oktay YILDIRIM, Pınar AKAN This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8176315/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background: Vitamin B12 deficiency is a common yet frequently underdiagnosed condition due to the limited diagnostic accuracy of serum total B12 and restricted availability of confirmatory biomarkers such as holotranscobalamin and methylmalonic acid. Advances in machine learning (ML) and large-scale laboratory datasets provide new opportunities to leverage routinely collected biochemical and hematological parameters for early detection. This study aimed to develop, optimize, and validate explainable ML models to predict vitamin B12 deficiency using standard laboratory analytes obtained during routine outpatient care. Methods: This retrospective study included 51,630 adult patients from 2015–2025, with an independent temporal validation cohort of 34,744 individuals. Eight supervised ML algorithms—logistic regression, random forest, decision tree, SVM, KNN, XGBoost, CatBoost, and artificial neural networks—were trained within a four-stage experimental framework incorporating default modeling, threshold optimization, hyperparameter tuning, and feature engineering. Performance was assessed using AUC-ROC, AUC-PR, sensitivity, specificity, F1-score, PPV, NPV, accuracy, MCC, and likelihood ratios. Statistical comparisons included DeLong, paired t-tests, McNemar, NRI, and IDI analyses. Model interpretability was evaluated using SHAP, LIME, and Decision Curve Analysis. Results: Across all experiments, CatBoost achieved the most balanced performance, with the F1-maximization threshold-optimized configuration demonstrating the lowest false-negative rate. In the test set, CatBoost yielded sensitivity 0.92, specificity 0.67, F1 0.82, AUC-ROC 0.88, and AUC-PR 0.86. Temporal validation confirmed robust generalizability (sensitivity 0.85, specificity 0.77, AUC-ROC 0.90, AUC-PR 0.91, MCC 0.63). SHAP and LIME consistently identified MCV, HGB, HCT, RBC, RDW, iron, ferritin, CRP, folate, and age as key contributors. DCA demonstrated superior net clinical benefit across a wide threshold range. Conclusion: This study presents the first large-scale, explainable, and clinically validated ML model capable of predicting vitamin B12 deficiency using only routine laboratory parameters. The model exhibits strong discrimination, reliability under temporal shift, and biologically meaningful interpretability, supporting its potential integration into clinical decision-support systems for early detection and optimized laboratory workflows. Figures Figure 1 Figure 2 Figure 3 Figure 4 INTRODUCTION Vitamin B12 (cobalamin) is an essential micronutrient required for DNA synthesis, myelin maintenance, and one-carbon metabolism. Deficiency can result in macrocytic anemia, neurocognitive impairment, and increased cardiovascular risk mediated through hyperhomocysteinemia [ 1 ]. Although serum total B12 concentration is the most frequently used diagnostic test, it often fails to reflect intracellular cobalamin status, leading to “functional B12 deficiency” even when serum levels fall within the low-normal range [ 1 , 2 ]. More specific biomarkers—such as holotranscobalamin (holo-TC), methylmalonic acid (MMA), and homocysteine (Hcy)—provide improved diagnostic accuracy, yet their higher cost, limited availability, and lack of integration into routine laboratory workflows restrict widespread use [ 3 , 4 ]. Consequently, clinically significant deficiency may remain undetected until hematological or neurological manifestations become evident [ 2 ]. In recent years, the growing availability of large-scale laboratory datasets and advances in machine learning (ML) have opened new possibilities for leveraging routinely collected biochemical and hematological parameters to predict micronutrient deficiencies. Tamune et al. demonstrated that ML models using standard blood tests can efficiently predict vitamin B group deficiencies, including B12, in patients with acute psychiatric episodes [ 5 ]. Beyond B12, several studies have applied ML to other micronutrients and lipid biomarkers: Sharifmousavi and Borhani developed an SVM-based model that incorporated selenium, vitamin B12, and vitamin D3 levels for the diagnosis of multiple sclerosis [ 6 ], while Sancar and Tabrizi showed that ML approaches can accurately classify vitamin D status using routine laboratory variables [ 7 ]. Similarly, ML-based models have been proposed for predicting low-density lipoprotein cholesterol (LDL-C) from standard biochemical profiles, achieving good agreement with directly measured and calculated LDL-C values [ 8 ]. Collectively, these findings highlight the feasibility of using routinely available laboratory data to build predictive models for diverse biochemical targets. Another important development has been the increasing emphasis on explainable artificial intelligence (XAI) in clinical prediction models. Hu et al. reported an explainable ML model for predicting spontaneous bacterial peritonitis in cirrhotic patients with ascites, using SHAP-based explanations to identify the most influential laboratory and clinical variables. In a broader perspective, a recent systematic review by Alkhanbouli et al. underscored the role of explainable AI techniques in disease prediction and clinical decision support, emphasizing transparency, trust, and model interpretability as essential requirements for implementation in healthcare settings [ 9 ]. For high-stakes decisions such as the detection of vitamin B12 deficiency—where underdiagnosis may lead to irreversible neurological damage—models that are both accurate and interpretable are particularly valuable. Given the high prevalence of vitamin B12 deficiency, its often subtle or delayed clinical presentation, and the limited accessibility of confirmatory biomarkers, there is a clear need for robust, low-cost, and scalable predictive models based entirely on routine laboratory parameters. Such models could be integrated into hospital information systems as automated decision-support tools, providing real-time alerts for individuals at high risk and reducing unnecessary second-line testing. The present study aims to address these gaps by developing and validating several machine learning algorithms—including artificial neural networks, boosting-based ensemble models, support vector machines, and logistic regression—to predict serum vitamin B12 levels using a large, multiyear dataset of routinely collected biochemical and hematological parameters. Model performance was evaluated using ROC-AUC, PR-AUC, F1-score, accuracy, and the Matthews correlation coefficient. To ensure clinical interpretability, SHAP and LIME techniques were employed to identify the variables contributing most substantially to model predictions. Furthermore, an independent temporal validation cohort was used to assess the generalizability and robustness of the final model under real-world conditions. Overall, this study aims to establish a reliable, scalable, and explainable machine learning framework capable of supporting early identification of individuals at risk for vitamin B12 deficiency within routine clinical practice. MATERIALS AND METHODS Study Population / Subjects This retrospective study included adult patients who underwent serum vitamin B12 testing together with routine hematological and biochemical analyses at Health Sciences University İzmir Tepecik Training and Research Hospital between 1 January 2015 and 1 October 2025. Prior to study initiation, approval was obtained from the hospital’s Non-Interventional Research Ethics Committee (initial ethics approval: 13/07/2023–2023/06–38; additional approval: 30/10/2025–293077648). Laboratory data were extracted from the hospital information system (HIS) and were based on analyzer outputs verified and authorized by a board-certified medical biochemistry specialist. Only the first eligible laboratory encounter for each patient was included in the study, and all records were anonymized during data extraction. Inclusion Criteria Age ≥ 18 years Availability of serum vitamin B12 measurement performed concurrently with all predefined hematological and biochemical tests First eligible outpatient laboratory encounter within the study period Complete numerical laboratory results available for all required parameters Specimen processed within the institution’s standardized pre-analytical workflow and quality limits Exclusion Criteria Age < 18 years Missing, corrupted, or non-numerical laboratory results Multiple encounters for the same patient, in which case only the first eligible record was retained Specimens exhibiting pre-analytical delays exceeding institutional workflow limits Pregnancy Active oncologic disease or hematologic malignancy Forensic/medico-legal cases Emergent conditions associated with spontaneous acute bleeding or requiring immediate emergency evaluation, including but not limited to: – Acute coronary syndrome – Pulmonary embolism – Aortic dissection/rupture – Subacute arachnoid hemorrhage – Transient ischemic attack – Massive gastrointestinal hemorrhage Venous whole-blood samples were collected into K₂-EDTA tubes, and serum samples were obtained in additive-free plain tube, in accordance with routine clinical protocols. In line with the National Health Quality Standards (Version 5.1), all specimens were delivered to the laboratory within 30 minutes, and biochemical analyses were completed within two hours of sample collection. Hematological parameters (HGB, RBC, HCT, MCV, RDW, MCH, MCHC, WBC, NEU, LYM, MONO, BASO, PLT, MPV, PDW) were measured using Beckman Coulter LH-780 (Mannheim, Germany) and Sysmex XN-1000 (Kobe, Japan) analyzers. Biochemical assays (glucose, creatinine, ALT, AST, ALP, GGT, LDH, serum iron, CRP, albumin, total protein, total bilirubin) were performed on a Beckman Coulter AU5800 analyzer (Mannheim, Germany). Chemiluminescence-based tests (vitamin B12, folate, ferritin, fT4, TSH) were conducted on Beckman Coulter DxI800 analyzers (Mannheim, Germany). All reagents, calibrators, and internal quality-control materials were certified and approved by the manufacturer. Standard institutional internal and external quality-control procedures were maintained throughout the analytical process. All laboratory results were reviewed and analytically validated by a medical biochemistry specialist before being transferred to the hospital information system. Study Design To ensure data confidentiality, all patient identifiers were irreversibly removed prior to analysis. A comprehensive dataset containing age, gender, and all hematological and biochemical parameters was generated through the hospital information system (HIS). A total of 87,301 laboratory records from the study period were reviewed and exported to Microsoft Excel 2021 (USA). After applying the predefined exclusion criteria, 51,630 patients were deemed eligible and included in the study, while all other records were excluded. To address class imbalance in the machine-learning analyses and enable the model to effectively learn the minority class, a stratified sampling approach was used to construct the development dataset. This procedure was performed in Python: all observations belonging to the minority class were retained, whereas an equal number of records from the majority class were randomly selected. As a result of this balancing step, 16,354 records were randomly discarded, yielding a final development cohort of 35,276 patients. The development dataset (01 January 2015–30 May 2023) was subsequently split into 80% training and 20% test sets using stratified random partitioning to prevent overfitting and preserve class distribution. Using identical preprocessing and class-balancing principles, an independent temporal validation set consisting of 34,744 patient records (collected between 01 June 2023 and 01 October 2025) was used for external validation. This allowed the model’s temporal generalizability to be evaluated on data from a later clinical period. All data preprocessing, model development, class balancing, training–test partitioning, and reporting procedures were conducted in accordance with the TRIPOD guidelines. A flow diagram summarizing patient selection, exclusion criteria, and the final analytic cohorts is presented in Fig. 1 . Data Preprocessing and Training of Machine-Learning Algorithms Raw laboratory data extracted from the HIS were preprocessed through removal of duplicate patient records, standardization of unit inconsistencies for selected biochemical parameters (e.g., albumin, total protein), and exclusion of entries containing missing parameters. Given that extreme values may represent true biological measurements within clinical practice and can directly influence model performance, outliers were not removed. Age was retained as a continuous variable, whereas all biochemical and hematological analytes were transformed into categorical variables based on clinically defined threshold values. Evidence-based clinical decision limits were used when available; otherwise, manufacturer-provided reference intervals were applied. Each analyte was encoded as 0 (below reference range), 1 (within reference range), or 2 (above reference range). For parameters without a biologically meaningful lower bound, a binary scheme was adopted. Gender was coded as 1 for males and 2 for females. The initial three-category structure of vitamin B12 was converted into a binary variable (0 = low; 1 = normal/high) to reduce class imbalance and improve clinical interpretability. Classification thresholds for all analytes are provided in Supplementary Material 1, Table 1 . Eight supervised machine-learning algorithms—logistic regression, random forest, decision tree, support vector machines, k-nearest neighbors, XGBoost, CatBoost, and artificial neural networks—were developed within a four-stage experimental framework. In experiemnt 1, all algorithms were trained using default hyperparameters. In experiment 2, probability-threshold optimization was performed using the Youden J index and F1-maximization criterion. Experiment 3 involved extensive hyperparameter tuning via GridSearchCV, RandomizedSearchCV, and Optuna-based bayesian optimization. All cross-validation procedures were conducted exclusively on the training set to prevent information leakage. In experiment 4, a large-scale feature-engineering pipeline generated more than 250 derived variables, including variance-based metrics, ratio-based biomarkers, composite indices, and biologically meaningful interaction terms (Supplementary Material 2). All preprocessing and model-building steps were implemented in Python 3.11 (Holland) using Kaggle Notebook and Google Colab environments. Continuous variables were standardized using the StandardScaler function, and the final dataset was randomly split into training (80%) and test (20%) subsets using a stratified approach to preserve the distribution of vitamin B12 classes. After completion of all stages of the experimental framework, the methodological strategy yielding the highest overall performance was selected and applied—without modification—to an independent temporal validation cohort of 34,744 patients from a later time period, in order to assess real-world, time-dependent generalizability. Model interpretability and clinical utility were comprehensively evaluated using SHAP (feature contributions), LIME (local explainability), and Decision Curve Analysis (net clinical benefit). Performance Evaluation Descriptive analyses were performed for both the development and validation cohorts. Continuous variables were summarized as mean ± standard deviation, and between-group differences were assessed using p-values and Cohen’s d. Additional stratified analyses by age and sex were conducted separately for the development and validation datasets. Subsequently, the performance of the eight algorithms included in the four-stage experimental framework was evaluated using sensitivity, specificity, positive and negative predictive values, accuracy, F1 score, Matthews correlation coefficient, area under the ROC curve (AUC-ROC), and area under the precision–recall curve (AUC-PR), the latter being particularly informative for imbalanced outcomes. For each model, confusion-matrix components (TP, FP, TN, FN) were computed, and ROC/PR curves were generated across a range of probability thresholds. After completion of all experimental stages, the methodological strategy demonstrating the highest overall performance on the test set was selected and applied unchanged to the validation cohort. Clinical interpretability was assessed using SHAP beeswarm plots, LIME-based local explanations, and Decision Curve Analysis, providing complementary insights into feature contributions and net clinical utility. Throughout the analyses, the categorical vitamin B12 variable used as an input feature (0 = low, 1 = normal/high) represented the laboratory classification, whereas a predicted value of 1 in the performance assessment indicated a model-predicted “vitamin B12 deficiency.” These two representations were explicitly distinguished across all analytical steps to avoid interpretive ambiguity. Statistical Comparison of Experimental Models Since all machine-learning algorithms were evaluated on the same test set, statistical comparisons accounted for the dependence arising from correlated prediction structures. DeLong’s test was used to compare AUC-ROC values derived from correlated ROC curves. Differences in continuous performance metrics such as F1 score and accuracy were assessed using paired t-tests. To examine false-positive/false-negative imbalance and asymmetry in misclassification patterns across models, the McNemar test was applied. In addition, Net Reclassification Improvement (NRI) and Integrated Discrimination Improvement (IDI) were calculated to quantify the relative gains in reclassification performance between models. These metrics enabled a more comprehensive evaluation of the methodological advantages achieved through model optimization, particularly regarding improvements in accuracy and discrimination. Together, these complementary procedures allowed for a rigorous, multidimensional, and model-agnostic comparison of all experimental configurations. Statistical Analysis Continuous variables were summarized as mean ± standard deviation, whereas categorical variables were presented as counts and percentages. Given the large sample size, parametric methods were preferred. Group comparisons for continuous variables were performed using one-way ANOVA, and Welch’s ANOVA was applied when the assumption of homogeneity of variances was violated. When appropriate, Tukey’s HSD or Games–Howell tests were used for post hoc multiple comparisons. Categorical variables were compared using the Pearson chi-square test. The diagnostic performance of the machine-learning models was assessed using sensitivity, specificity, positive and negative predictive values, accuracy, F1 score, AUC-ROC, AUC-PR, positive and negative likelihood ratios (LR⁺/LR⁻), and the Matthews correlation coefficient. All statistical analyses were conducted in Python 3.11 using the pandas , scipy , scikit-learn , and statsmodels libraries. A two-sided p-value < 0.05 was considered statistically significant. RESULTS Statistically significant differences were observed across numerous hematological and biochemical parameters between the development (n = 35,256) and independent validation (n = 34,744) cohorts (with p < 0.05 for most comparisons). However, Cohen’s d analyses indicated that the magnitude of these differences was predominantly very small from a biological perspective. Age, hemoglobin, hematocrit, erythrocyte count, and MCV values were lower in the validation cohort, demonstrating small to moderate effect sizes (d = 0.18–0.33). In contrast, parameters reflecting inflammatory or hematologic variability—such as RDW, CRP, LDH, and PLT—were higher, corresponding to small effect sizes (d ≈ − 0.16 to − 0.20). Among liver enzymes, only ALP exhibited a moderate biologic effect (d = − 0.546). Although serum vitamin B12 concentrations were higher in the validation period, the biological magnitude of this difference remained very small (d = − 0.143). Means, standard deviations, p-values, and Cohen’s d effect sizes for all variables are presented in Table 1 . Additionally, age- and sex-stratified distributions and comparative statistics for both cohorts are provided in Supplementary Material 1, Tables 2 – 3 . Across the eight machine-learning algorithms evaluated within the four-stage experimental framework, CatBoost demonstrated the most balanced sensitivity–specificity profile and the highest overall classification performance in all experiments. Notably, the threshold-optimized configuration (Experiment 2), despite yielding performance metrics numerically similar to the other strategies, provided a clinically decisive advantage. This approach improved model accuracy not only from a statistical perspective but also in terms of clinical safety, achieving the lowest number of false negatives and thereby minimizing the risk of missing true vitamin B12 deficiency cases. Given the association of vitamin B12 deficiency with neurological sequelae and irreversible damage, this marked reduction in false negatives represents the key outcome defining the model’s clinical applicability rather than merely its methodological strength. For this reason, the primary findings highlighted in the manuscript are derived from the F1-maximization–based threshold-optimized configuration, which is summarized in Table 2 . Performance distributions for all experimental strategies are provided in Supplementary Material 1, Tables 4 –6. In the test set, the threshold-optimized CatBoost model achieved sensitivity 0.92, specificity 0.67, F1 score 0.82, PPV 0.73, and NPV 0.88, yielding the most balanced performance among all models; AUC-ROC was 0.88 and AUC-PR was 0.86. XGBoost and the neural network produced comparable AUC values but demonstrated lower accuracy, specificity, MCC, and PPV. Logistic regression and random forest failed to maintain overall classification balance, whereas SVM, KNN, and decision tree models showed markedly reduced specificity and MCC. Performance metrics for the four CatBoost configurations are summarized in Table 3 . Statistical comparisons revealed no significant differences across configurations for AUC-ROC (DeLong test) or for F1 and accuracy (paired t-tests) (all p > 0.05). In contrast, the McNemar test demonstrated a clear advantage for the threshold-optimized model, particularly through its favorable asymmetry in the distribution of false negatives versus false positives. This model substantially reduced false-negative decisions—the clinically most critical error type—supporting its selection as the final model. For the threshold-optimized CatBoost model, the calculated NRI = 0.359 and IDI = 0.147 further confirmed that the approach produced meaningful improvements in both correct reclassification and discrimination capacity. These findings align with the McNemar test results demonstrating improved false-negative/false-positive asymmetry, collectively indicating that the optimized model provides a substantially enhanced risk-classification framework. Accordingly, the model exhibits strong potential as a reliable and effective clinical decision-support tool for the early identification of vitamin B12 deficiency. The performance of the selected CatBoost model remained robust in the independent temporal validation cohort, with sensitivity 0.85, specificity 0.77, accuracy 0.81, F1 score 0.82, and MCC 0.63. Discrimination improved compared with the internal test set, with AUC-ROC 0.90 and AUC-PR 0.91. The likelihood-ratio profile (PLR = 3.74; NLR = 0.19) also indicated strong clinical utility under real-world conditions. These results are presented in Table 4 and Fig. 2 . SHAP analyses demonstrated that hematologic and inflammatory parameters—such as MCV, HGB, HCT, RBC, RDW, age, iron and ferritin—were the strongest contributors to model predictions. The SHAP beeswarm plot (Fig. 3 ) showed that physiologically abnormal measurements consistently shifted predictions toward deficiency. LIME analyses confirmed the internal consistency of feature contributions for individual-level predictions and reinforced the clinical interpretability of the model (Supplementary Material 1, Fig. 1 ). Decision Curve Analysis (DCA) demonstrated that the model provided a consistently higher net benefit across a wide threshold range compared with “treat-all” or “treat-none” strategies (Fig. 4 ), supporting its potential integration into clinical decision-support systems. DISCUSSION This study represents the first comprehensive machine-learning investigation designed to predict serum vitamin B12 levels directly from routinely collected laboratory parameters. Current literature contains predictive models for vitamin D, folate, lipid metabolism, and other biochemical markers, yet no prior study has focused specifically on estimating B12 status using hematologic and biochemical indices at scale [ 7 , 10 , 11 ]. By integrating routine laboratory features with explainable artificial intelligence (XAI) techniques, the present work offers a novel and clinically actionable framework that bridges biomedical informatics, laboratory medicine, and nutritional neuroscience. Across eight machine-learning algorithms evaluated in four sequential experimental settings, gradient-boosting models consistently demonstrated superior ability to capture the nonlinear and population-level variability inherent in biochemical data. Linear models such as logistic regression and support vector machines showed limited generalizability with AUC values around 0.86, consistent with prior work by Sharifmousavi et al. using SVM-based micronutrient prediction in neurological populations [ 6 ]. Artificial neural networks yielded lower accuracy (0.78) than the 85% reported by Tamune et al. in B6 deficiency [ 5 ], while also suffering from well-recognized limitations in interpretability—an important consideration for clinical deployment. Tree-based ensembles such as random forest similarly demonstrated moderate discrimination but experienced performance instability under external validation, a finding aligned with previously reported limitations of over-partitioning in high-dimensional biomedical data [ 12 ]. XGBoost outperformed these baseline models, reflecting its strong capacity to capture interactions between biochemical variables, consistent with prior lipid profiling work (r = 0.98) by Anudeep et al. [ 8 ]. However, CatBoost provided the most robust and generalizable performance across all experimental conditions—an observation supported by previous reports emphasizing its unbiased handling of categorical variables and resilience in clinical datasets [ 13 , 14 ]. In our study, CatBoost achieved an AUC-ROC of 0.88, AUC-PR of 0.86, F1 of 0.82, and the most balanced sensitivity–specificity profile across both test and validation cohorts. Interpretability analyses further strengthened the biological plausibility of the model. SHAP and LIME consistently highlighted MCV, RDW, HGB, HCT, ferritin, CRP, folate, and age as principal contributors to classification. These findings reinforce well-established hematological consequences of B12 deficiency, including macrocytosis, anisocytosis, and impaired erythropoiesis. Elevated contributions of ferritin and CRP mirror the known inhibitory effects of systemic inflammation on circulating B12 bioavailability via haptocorrin-mediated sequestration, as reported by Jensen et al. [ 15 , 16 ]. The observed interplay between folate and B12 also aligns with their shared role in one-carbon metabolism and DNA synthesis [ 2 , 11 , 17 ]. Age contributed positively, consistent with declining intrinsic factor secretion and reduced gastrointestinal absorption in older adults [ 18 ]. Together, these patterns demonstrate that the model not only predicts deficiency but also captures meaningful pathophysiological signatures associated with B12 metabolism. The sequential experimental framework provided additional insight into how model performance evolved under different analytical strategies. Experiment 1 established baseline performance; Experiment 2 applied F1-maximizing threshold optimization and yielded the most clinically balanced results, increasing sensitivity to 0.92 and reducing false negatives by nearly 45%. Experiment 3 evaluated hyperparameter tuning, which improved discrimination modestly but introduced mild overfitting in the validation cohort—consistent with prior findings on the trade-off between complexity and generalizability in clinical ML models [ 19 ]. Experiment 4 introduced feature engineering; however, CatBoost’s intrinsic ability to manage correlated and categorical variables meant that manual engineering added minimal incremental benefit. Collectively, these findings validated Experiment 2 as the optimal configuration for clinical decision support. When compared with the only available machine-learning study on B12 deficiency, conducted by Tamune et al. in psychiatric patients using a sample of 497 individuals (AUC-ROC 0.62), our model showed substantially higher discrimination (AUC-ROC 0.88) using a dataset nearly 150 times larger, highlighting its superior generalizability and methodological rigor [ 5 ]. In contrast to genetic Mendelian randomization studies such as Lu et al. [ 20 ], which explore upstream determinants of B12 metabolism, our work emphasizes biochemical phenotypes and their downstream hematologic expression—providing complementary but clinically more immediate insights. Similarly, while Duman et al. demonstrated hematologic alterations in pediatric B12 deficiency [ 21 ], their narrower variable set lacked inflammatory and metabolic markers that contributed substantially to our model’s predictive signal. From an implementation perspective, the model has strong potential for integration into laboratory information systems (LIS) or hospital information systems (HIS). Because it relies entirely on existing routine tests (e.g., hemogram, ferritin, CRP, albumin), it imposes no additional cost and can be deployed passively in background systems. LIME-based patient-specific explanations could be displayed to clinicians, allowing early recognition of biochemical patterns suggestive of deficiency before irreversible neurological or hematological complications arise. Early B12 replacement has been shown to reduce hospitalization duration by approximately two days, generating meaningful savings in medium-sized hospitals [ 22 ]. Such a model could therefore contribute to both clinical and economic decision-making. Limitations This study has several limitations. Its retrospective nature prevented inclusion of symptoms, dietary intake, medication history, gastrointestinal disorders, and genetic factors—variables known to influence B12 metabolism. Only adult patients were included, limiting generalizability to pediatric populations. Although data were obtained from the laboratory information system, all results had undergone biochemical specialist approval, ensuring that pre-analytical artifacts such as hemolysis or improper storage were already filtered out. Missing values were removed rather than imputed, eliminating imputation bias but introducing a theoretical selection bias. Classification thresholds were based on the NICE 2024 guideline, ensuring standardization across laboratories [ 23 ]. Nevertheless, total serum B12 is an imperfect marker of functional deficiency; lack of holo-transcobalamin or methylmalonic acid data may lead to misclassification in certain subgroups [ 24 ] Finally, micronutrient patterns may vary across regions and populations, suggesting that model recalibration may be necessary before widespread deployment [ 25 ]. As with all clinical ML models, temporal performance drift remains a potential risk and requires periodic re-evaluation [ 26 ]. Serum total vitamin B12 levels were not measured using the gold-standard isotope-dilution liquid chromatography–tandem mass spectrometry (ID-LC–MS/MS) method. Instead, a chemiluminescence immunoassay was used, which is subject to method-dependent analytical bias, including variability in antibody specificity and differential detection of protein-bound B12. Therefore, assay-related misclassification cannot be fully excluded. Conclusion Despite these limitations, this study introduces the first large-scale, explainable, and clinically validated machine-learning model for predicting vitamin B12 deficiency using routine laboratory parameters. The model demonstrates high discrimination, strong biological interpretability, and stable performance across internal and temporal validation cohorts. When integrated into clinical workflows, it has the potential to support early diagnosis, reduce unnecessary testing, and enhance decision-making in nutritional and hematologic assessments. Future work should focus on multi-center prospective validation, continuous monitoring for model drift, and extension of the framework to other micronutrient deficiencies such as folate, vitamin D, and ferritin. In parallel, evaluating how this explainable AI model behaves when quietly embedded into routine laboratory workflows—augmenting clinical decisions without disrupting established pathways—will be essential for realizing its full translational impact. Declarations Ethics approval and consent to participate This study was conducted in accordance with the Declaration of Helsinki and approved by the Non-Interventional Clinical Research Ethics Committee of the İzmir Tepecik Training and Research Hospital, University of Health Sciences Türkiye (Initial ethics approval: 13/07/2023–2023/06–38; additional approval: 30/10/2025–293077648). The requirement for informed consent was waived by the ethics committee because the study used retrospective, anonymized laboratory data with no identifiable personal information. Consent for publication Not applicable. No individual person’s data in any form (including images, videos, or identifiable information) are included in this manuscript. Competing interests The authors declare that they have no competing interests. Footnotes Not applicable. Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Author Contribution FD (Ferhat Demirci) obtained ethics approval, conceptualized the study, curated and analyzed the data, developed the methodology, wrote and optimized the machine-learning algorithms, performed the statistical analysis, and drafted the original manuscript. OY (Oktay Yıldırım) contributed to data collection, assisted in writing, debugging, and refining the Python code used in data preprocessing and model development, and participated in the review and editing of the manuscript. PA (Pınar Akan) verified the accuracy and integrity of the laboratory data, conducted literature review, and contributed to writing sections of the original manuscript draft. All authors read and approved the final version of the manuscript. Acknowledgement The authors would like to express their sincere gratitude to Prof. Dr. Savaş Yakan, Chief Physician of İzmir Tepecik Training and Research Hospital, and Prof. Dr. Ayfer Çolak, Head of the Department of Medical Biochemistry, for their valuable support and encouragement throughout this study. Data Availability The datasets generated and analyzed during the current study are not publicly available due to institutional data protection regulations but are available from the corresponding author (FD) upon reasonable request. All key summary tables and model coefficients are included in this published article and its supplementary information files. References O’Leary F, Samman S. Vitamin B12 in Health and Disease. Nutrients. 2010;2(3):299–316. Stabler SP. Vitamin B12 Deficiency. N Engl J Med. 2013;368(2):149–60. Hannibal L, Lysne V, Bjørke-Monsen AL, Behringer S, Grünert SC, Spiekerkoetter U et al. Biomarkers and Algorithms for the Diagnosis of Vitamin B12 Deficiency. Front Mol Biosci. 2016;3. Stover P, Discussion. Folate and Vitamin B12 Metabolism: Overview and Interaction with Riboflavin, Vitamin B6, and Polymorphisms. Food Nutr Bull. 2008;29(2suppl1):S17–9. Tamune H, Ukita J, Hamamoto Y, Tanaka H, Narushima K, Yamamoto N. Efficient Prediction of Vitamin B Deficiencies via Machine-Learning Using Routine Blood Test Results in Patients With Intense Psychiatric Episode. Front Psychiatry. 2019;10:1029. Sharifmousavi SS, Borhani MS. Support vectors machine-based model for diagnosis of multiple sclerosis using the plasma levels of selenium, vitamin B12, and vitamin D3. Inf Med Unlocked. 2020;20:100382. Sancar N, Tabrizi SS. Machine learning approach for the detection of vitamin D level: a comparative study. BMC Med Inf Decis Mak. 2023;23(1):219. Kumari PPA, Rajasimman S, Nayak AS, Priyadarsini S. Machine learning predictive models of LDL-C in the population of eastern India and its comparison with directly measured and calculated LDL-C. Annals of Clinical Biochemistry. Int J Lab Med. 2022;59(1):76–86. Alkhanbouli R, Matar Abdulla Almadhaani H, Alhosani F, Simsekler MCE. The role of explainable artificial intelligence in disease prediction: a systematic literature review and future research directions. BMC Med Inf Decis Mak. 2025;25(1):110. Paplomatas P, Kantartzi M, Anagnostopoulos K, Vrahatis A. An Artificial Intelligence and Biochemical Modelling Approach for LDL Cholesterol Prediction. Medinformatics. 2025. Kim JM, Stewart R, Kim SW, Yang SJ, Shin IS, Yoon JS. Predictive value of folate, vitamin B 12 and homocysteine levels in late-life depression. Br J Psychiatry. 2008;192(4):268–74. Mohammed H, Ahamad S. S, Abdul Hameed DrM. Enhancements in Random Forest Algorithms for Improving Healthcare Applications. In: Proceedings of the International Conference on Industrial Engineering and Operations Management. Michigan, USA: IEOM Society International; 2024. Hu Y, Chen R, Gao H, Lin H, Wang J, Wang X, et al. Explainable machine learning model for predicting spontaneous bacterial peritonitis in cirrhotic patients with ascites. Sci Rep. 2021;11(1):21639. Prokhorenkova L, Gusev G, Vorobev A, Dorogush AV, Gulin A. (2018). CatBoost: unbiased boosting with categorical features. Advances in neural information processing systems, 2018 (31). Ismail O, Albdour K, Albdour Z, Jaber K. Differences in Ferritin, Vitamin D, and Vitamin B12 Between Fibromyalgia Patients and Healthy Individuals: A Systematic Review and Meta-Analysis. Musculoskelet Care. 2025;23(1). Jensen HR, Laursen MF, Lildballe DL, Andersen JB, Nexø E, Licht TR. Effect of the vitamin B12-binding protein haptocorrin present in human milk on a panel of commensal and pathogenic bacteria. BMC Res Notes. 2011;4:208. Reynolds E. Vitamin B12, folic acid, and the nervous system. Lancet Neurol. 2006;5(11):949–60. Simonenko SYu, Bogdanova DA, Kuldyushev NA. Emerging Roles of Vitamin B12 in Aging and Inflammation. Int J Mol Sci. 2024;25(9):5044. Wong J, Manderson T, Abrahamowicz M, Buckeridge DL, Tamblyn R. Can Hyperparameter Tuning Improve the Performance of a Super Learner? Epidemiology. 2019;30(4):521–31. Lu T, Paterson AD. Estimating effects of serum vitamin B12 levels on psychiatric disorders and cognitive impairment: a Mendelian randomization study. Commun Med. 2025;5(1):316. Duman D, Uslu G, Malbora B. Retrospective Analysis of Pediatric Patients with Vitamin B12 Deficiency: Vitamin B12 and Children. Chronicles Precision Med Researchers. 2024;5(1):39–45. da Gomes C, von Gunten A, Jopp D, Ribeiro O, Verloo H. Why Centenarians’ Depressive Symptoms Must Become a Priority for Nurses. J Am Med Dir Assoc. 2021;22(5):1118–9. NICE guideline. Vitamin B12 deficiency in over 16s: diagnosis and management [Internet]. 2024 [cited 2025 Oct 4]. Available from: https://www.nice.org.uk/guidance/ng239 Dastidar R, Sikder K. Diagnostic reliability of serum active B12 (holo-transcobalamin) in true evaluation of vitamin B12 deficiency: Relevance in current perspective. BMC Res Notes. 2022;15(1):329. Gallina AL, Otay S, de Frutos-Lucas J, Buso M, Moral Martinez P, Cashman KD et al. Hidden hunger in Europe: a review on determinants, fragmented policy responses, and implementation barriers. Front Nutr. 2025;12. Vela D, Sharp A, Zhang R, Nguyen T, Hoang A, Pianykh OS. Temporal quality degradation in AI models. Sci Rep. 2022;12(1):11654. Tables Table 1 Descriptive Statistics and Effect Size (Cohen’s d) for All Variables Variable Unit Development Dataset Mean ± SD Validation Dataset Mean ± SD p-value Cohen d (Effect) Age years 49.43 ± 16.66 41.42 ± 24.99 < 0.05 0.377 B12 pg/mL 277.27 ± 203.78 308.62 ± 234.34 < 0.05 -0.143 Folate ng/mL 8.3 ± 3.98 8.97 ± 4.51 < 0.05 0.197 Glucose mg/dL 99.71 ± 38.00 99.65 ± 45.98 0.85 0.001 ALT U/L 22.83 ± 26.05 26.12 ± 75.04 < 0.05 -0.059 AST U/L 22.35 ± 21.01 30.39 ± 206.21 < 0.05 -0.055 Ferritin ng/mL 236.00 ± 987.00 99.99 ± 204.05 < 0.05 0.191 ALP U/L 72.48 ± 36.35 112.84 ± 97.94 < 0.05 -0.546 GGT U/L 26.41 ± 40.01 33.10 ± 81.44 < 0.05 -0.104 T. Bilirubin mg/dL 0.69 ± 0.34 0.78 ± 0.93 < 0.05 -0.129 Creatinine mg/dL 0.88 ± 0.36 0.95 ± 0.85 < 0.05 -0.114 C-RP mg/L 6.55 ± 16.10 10.92 ± 29.11 < 0.05 -0.186 Albumin g/L 43.60 ± 3.86 43.21 ± 6.23 < 0.05 0.074 T. Protein g/L 72.83 ± 4.90 72.02 ± 6.12 < 0.05 0.145 Iron µg/dL 176.51 ± 148.63 176.32 ± 153.23 0.87 0.001 LDH U/L 176.47 ± 61.25 208.16 ± 215.55 < 0.05 -0.200 FT4 ng/dL 0.89 ± 0.17 0.92 ± 0.20 < 0.05 -0.137 TSH µIU/mL 2.18 ± 2.46 2.29 ± 2.58 < 0.05 -0.045 HGB g/dL 13.43 ± 1.67 12.84 ± 1.91 < 0.05 0.331 RBC 10⁶/µL 4.69 ± 0.53 4.61 ± 0.64 < 0.05 0.133 HCT % 40.19 ± 4.44 38.75 ± 5.18 < 0.05 0.298 MCV fL 85.89 ± 6.28 84.62 ± 7.78 < 0.05 0.179 MCH pg 28.76 ± 2.60 28.17 ± 3.07 < 0.05 0.208 MCHC g/dL 33.48 ± 1.24 33.32 ± 1.34 < 0.05 0.123 RDW % 14.03 ± 1.80 14.41 ± 2.52 < 0.05 -0.174 WBC 10³/µL 7.26 ± 2.25 8.20 ± 6.79 < 0.05 -0.186 NEU 10³/µL 5.50 ± 1.44 4.07 ± 3.37 < 0.05 0.552 LYM 10³/µL 3.08 ± 0.98 2.46 ± 5.43 < 0.05 0.159 MONO 10³/µL 0.79 ± 0.23 0.64 ± 0.49 < 0.05 0.382 BASO 10³/µL 0.64 ± 0.46 0.05 ± 0.21 < 0.05 1.641 PLT 10³/µL 269.78 ± 75.84 284.84 ± 106.67 < 0.05 -0.163 MPV fL 9.49 ± 1.27 9.73 ± 1.33 < 0.05 -0.183 PDW fL 15.27 ± 2.54 14.22 ± 3.00 < 0.05 0.376 SD: standart deviation Table 2 Performance comparison of eight machine-learning models under the F1-max threshold (experiment 2) Metrics / Model Cat Boost XGB ANN RF LR SVM KNN DT CutOff 0,402 0,378 0,372 0,410 0,396 0,306 0,400 0,500 TP 3233 3265 3268 3184 3098 3216 3297 2468 FP 1160 1198 1325 1197 1452 2006 2473 994 TN 2368 2330 2203 2331 2076 1522 1055 2534 FN 295 263 260 344 430 312 231 1060 Sensitivity 0.92 (90.6-92.53) 0.92 (91.6–93.4) 0.93 (91.7–93.4) 0.90 (89.2–91.2) 0.87 (86.7–88.9) 0.91 (90.2–92.0) 0.93 (92.6–94.2) 0.70 (68.4–71.4) Specificity 0.67 (65.5–68.7) 0.66 (64.5–67.6) 0.62 (60.8–64.0) 0.66 (64.5–67.6) 0.59 (57.2–60.5) 0.43 (41.5–44.8) 0.30 (28.4–31.4) 0.72 (70.3–73.3) PPV 0.73 (72.7–74.5) 0.73 (71.8–74.4) 0.71 (69.8–72.4) 0.73 (71.3–74.0) 0.68 (66.7–69.4) 0.61 (60.3–62.9) 0.57 (55.9–58.4) 0.71 (69.8–72.8) NPV 0.88 (87.8–89.0) 0.89 (88.6–91.0) 0.89 (88.2–90.6) 0.87 (85.8–88.4) 0.82 (81.3–84.3) 0.83 (81.2–84.6) 0.82 (79.8–84.0) 0.70 (69.0–72.0) PLR 2.79 (2.66–2.92) 2.73 (2.60–2.86) 2.47 (2.36–2.58) 2.66 (2.54–2.79) 2.13 (2.05–2.22) 1.60 (1.55–1.65) 1.33 (1.30–1.36) 2.48 (2.35–2.63) NLR 0.12 (3.13–3.52) 0.11 (0.10–0.13) 0.12 (0.10–0.13) 0.15 (0.13–0.16) 0.21 (0.19–0.23) 0.20 (0.18–0.23) 0.22 (0.19–0.25) 0.42 (0.40–0.44) F1 0.82 (0.81–0.83) 0.82 (0.81–0.83) 0.80 (0.80–0.81) 0.81 (0.80–0.81) 0.77 (0.76–0.78) 0.74 (0.72–0.75) 0.71 (0.70–0.72) 0.71 (0.69–0.72) Accuracy 0.79 (78.4–80.3) 0.79 (78.3–80.2) 0.78 (76.5–78.5) 0.78 (77.2–79.1) 0.73 (72.3–74.3) 0.67 (66.0-68.2) 0.62 (60.5–62.8) 0.71 (69.8–71.9) A_ROC 0.88 (0.86–0.90) 0.88 (0.87–0.89) 0.86 (0.85–0.87) 0.86 (0.85–0.87) 0.81 (0.79–0.82) 0.77 (0.76–0.78) 0.72 (0.70–0.73) 0.71 (0.70–0.72) A_PR 0.86 (0.84–0.88) 0.85 (0.84–0.86) 0.84 (0.82–0.85) 0.82 (0.80–0.83) 0.78 (0.76–0.79) 0.75 (0.74–0.77) 0.66 (0.65–0.68) 0.65 (0.63–0.66) MCC 0.61 (0.58–0.62) 0.61 (0.58–0.62) 0.58 (0.55–0.59) 0.58 (0.56–0.6) 0.49 (0.47–0.50) 0.39 (0.37–0.41) 0.30 (0.28–0.32) 0.42 (0.39–0.43) CatBoost: Categorical Gradient Boosting; XGB: Extreme Gradient Boosting (XGBoost); ANN: Artificial Neural Network; RF: Random Forest; LR: Logistic Regression; SVM: Support Vector Machine; KNN: k-Nearest Neighbors; DT: Decision Tree; TP: true positive; FP: false positive; TN: true negative; FN: false negative; PPV: positive predictive value; NPV: negative predictive value; PLR: positive likelihood ratio; NLR: negative likelihood ratio; A_ROC: area under the ROC curve; A_PR: area under the precision–recall curve; MCC: Matthews correlation coefficient. Table 3 Performance Metrics of the CatBoost Model Across All Experimental Configurations Performance Metrics Experiment 1 (default parameter) Experiment 2 (cut-off optimization) Experiment 3 (hyperparameter tuning) Experiment 4 (feature engineering) TP 2993 3233 3028 3014 FP 895 1160 924 908 TN 2633 2368 2604 2620 FN 535 295 500 514 Sensitivity 0.85 (0.84–0.86) 0.92 (0.91–0.93) 0.86 (0.85–0.87) 0.85 (0.84–0.86) Specificity 0.75 (0.73–0.76) 0.67 (0.65–0.69) 0.74 (0.73–0.75) 0.74 (0.73–0.75) PPV 0.77 (0.76–0.78) 0.73 (0.72–0.75) 0.77 (0.76–0.78) 0.77 (0.76–0.78) NPV 0.83 (0.82–0.84) 0.88 (0.87–0.89) 0.84 (0.83–0.85) 0.84 (0.83–0.85) PLR 3.34 (3.15–3.55) 2.79 (2.65–2.94) 3.28 (3.12–3.44) 3.32 (3.24–3.40) NLR 0.20 (0.19–0.22) 0.12 (0.10–0.14) 0.19 (0.18–0.20) 0.20 (0.20–0.20) Accuracy 0.81 (0.80–0.82) 0.79 (0.78–0.80) 0.80 (0.79–0.81) 0.80 (0.79–0.81) F1 0.80 (0.79–0.81) 0.82 (0.81–0.83) 0.81 (0.80–0.82) 0.81 (0.80–0.82) AUC-ROC 0.88 (0.87–0.89) 0.88 (0.87–0.89) 0.88 (0.87–0.89) 0.86 (0.85–0.87) AUC-PR 0.86 (0.85–0.87) 0.86 (0.85–0.87) 0.86 (0.85–0.87) 0.88 (0.87–0.89) MCC 0.60 (0.58–0.62) 0.61 (0.59–0.63) 0.60 (0.58–0.62) 0.60 (0.59–0.61) TP: true positive; FP: false positive; TN: true negative; FN: false negative; PPV: positive predictive value; NPV: negative predictive value; PLR: positive likelihood ratio; NLR: negative likelihood ratio; AUC-ROC: area under the ROC curve; AUC-PR: area under the precision–recall curve; MCC: Matthews correlation coefficient. Table 4 Performance of the Validation Set (Final CatBoost Model) Performance Metric Validation Set TP 15018 FP 3910 TN 1320 FN 2576 Sensitivity 0.85 (0.84–0.86) Specificity 0.77 (0.75–0.79) PPV 0.79 (0.78–0.80) NPV 0.83 (0.82–0.84) PLR 3.74 (3.61–3.87) NLR 0.19 (0.18–0.20) Accuracy 0.81 (0.80–0.82) F1 Score 0.82 (0.81–0.83) AUC-ROC 0.90 (0.89–0.91) AUC-PR 0.91 (0.90–0.92) MCC 0.63 (0.61–0.65) TP: true positive; FP: false positive; TN: true negative; FN: false negative; PPV: positive predictive value; NPV: negative predictive value; PLR: positive likelihood ratio; NLR: negative likelihood ratio; AUC-ROC: area under the ROC curve; AUC-PR: area under the precision–recall curve; MCC: Matthews correlation coefficient. Additional Declarations No competing interests reported. Supplementary Files SupplementaryMaterial1.docx SupplementaryMaterial2.xlsx Cite Share Download PDF Status: Posted Version 1 posted 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-8176315","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":567363961,"identity":"45eea987-8685-49a1-b41c-3b9c9bd3a510","order_by":0,"name":"Ferhat Demirci","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA4ElEQVRIiWNgGAWjYDCCAwwGDCDExsz/8AGQz8NHvBb2HmYDkBY24rSAAM8ZNgkQTVAL3+3DGx/dKLgnzyeRe6zya46dDBsD88NHN/BokTyXVmycY1Bs2CaRl3Zbdlsy0GFsxsY5eLQYnOExk84xSEhgk0gwuy25jRmohYdNmoAW898wLcWS2+qJ0mLGDNbCc8aM8eO2w4S1SJ5hKwY5zLCNvS1ZmnHbcR42ZgJ+4TvDvPFzzp8Eeflm5oMff26rtudnb374GJ8WFMDMAyaJVQ4CjD9IUT0KRsEoGAUjBgAAvdpBL5N0qWoAAAAASUVORK5CYII=","orcid":"","institution":"Health Sciences University","correspondingAuthor":true,"prefix":"","firstName":"Ferhat","middleName":"","lastName":"Demirci","suffix":""},{"id":567363962,"identity":"b401cc4e-42a1-46aa-9ebd-ccf7fdce14c4","order_by":1,"name":"Oktay YILDIRIM","email":"","orcid":"","institution":"Dokuz Eylül University","correspondingAuthor":false,"prefix":"","firstName":"Oktay","middleName":"","lastName":"YILDIRIM","suffix":""},{"id":567363963,"identity":"9bc909e5-fa37-41c2-b23f-2f291a05c2c7","order_by":2,"name":"Pınar AKAN","email":"","orcid":"","institution":"Dokuz Eylül University","correspondingAuthor":false,"prefix":"","firstName":"Pınar","middleName":"","lastName":"AKAN","suffix":""}],"badges":[],"createdAt":"2025-11-21 19:53:11","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8176315/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8176315/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":99371633,"identity":"bb0a3aaa-a10f-4b11-aeab-8eb2aa8ad395","added_by":"auto","created_at":"2026-01-02 05:56:15","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":453512,"visible":true,"origin":"","legend":"","description":"","filename":"Figures.docx","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/db16d016050b571f462f06b9.docx"},{"id":99371632,"identity":"0ca9f09f-0983-44e8-95f3-1519bbddc27a","added_by":"auto","created_at":"2026-01-02 05:56:15","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":119964,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/1791a35fefa68891470975ec.docx"},{"id":99371636,"identity":"de155847-839a-4aff-9ced-f268e0f63fbe","added_by":"auto","created_at":"2026-01-02 05:56:15","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":28094,"visible":true,"origin":"","legend":"","description":"","filename":"Tables.docx","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/d1d3a06643674f53f43579f8.docx"},{"id":99371635,"identity":"d28bef29-3216-4ae3-9399-6e7b8d498e7a","added_by":"auto","created_at":"2026-01-02 05:56:15","extension":"json","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7019,"visible":true,"origin":"","legend":"","description":"","filename":"dd4556a6ebd449c18dc9780eac7dddce.json","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/7955d7fbf8e9c2d379e0207a.json"},{"id":99371639,"identity":"83184440-dfda-4f48-81c2-84f35a0a6311","added_by":"auto","created_at":"2026-01-02 05:56:15","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":123133,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterial1.docx","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/ca02e979e97acaa8886c15dc.docx"},{"id":99371646,"identity":"c6baf8e0-e70b-4302-ad51-51cae17ed483","added_by":"auto","created_at":"2026-01-02 05:56:16","extension":"xlsx","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":17334,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterial2.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/1d0078a0602879c25959e6dc.xlsx"},{"id":99789471,"identity":"21a2709f-a04f-4df2-b358-d088073a5900","added_by":"auto","created_at":"2026-01-08 12:49:46","extension":"xml","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":133388,"visible":true,"origin":"","legend":"","description":"","filename":"dd4556a6ebd449c18dc9780eac7dddce1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/6a5f71605482822ee6ed649b.xml"},{"id":99789151,"identity":"22f55c59-5a5b-4700-8b3b-b98fb7901b03","added_by":"auto","created_at":"2026-01-08 12:48:53","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":91504,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/3ecada36d54775077238bd98.png"},{"id":99789347,"identity":"d3959b61-6b19-4c94-963d-64426873e830","added_by":"auto","created_at":"2026-01-08 12:49:26","extension":"jpeg","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":605274,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/b4759c324bd9abed4b07b6dc.jpeg"},{"id":99371647,"identity":"fb266166-ecd9-4a42-877e-76478f35e4e7","added_by":"auto","created_at":"2026-01-02 05:56:16","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":105495,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/fcaae0defb42892fc433b2b4.png"},{"id":99371649,"identity":"71146f60-85fe-4124-a5bd-88d17b1977d9","added_by":"auto","created_at":"2026-01-02 05:56:16","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":79503,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/e4b591a3ad5bd79674aebf13.png"},{"id":99371638,"identity":"4e7fb8dd-4be8-42d9-b30a-da33d7cb84aa","added_by":"auto","created_at":"2026-01-02 05:56:15","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":38856,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/4909cf1a4bbdd14bf38ba8b9.png"},{"id":99371642,"identity":"7efa64fa-6c00-404d-a082-f9463c1e3505","added_by":"auto","created_at":"2026-01-02 05:56:15","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":122154,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/80a8fd1b82f2d481f196133a.png"},{"id":99789207,"identity":"15548cb5-fc98-4afa-b4a7-9ae7a79555b2","added_by":"auto","created_at":"2026-01-08 12:49:03","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":26355,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/16b61b07c66b5c4b8704cab8.png"},{"id":99371650,"identity":"56f27b24-3458-4294-a05c-e0edde19c00a","added_by":"auto","created_at":"2026-01-02 05:56:16","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":18126,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/bfb1999ce2915b999ca31865.png"},{"id":99371641,"identity":"e135eb1d-625c-469a-8298-c0efddd4a9bb","added_by":"auto","created_at":"2026-01-02 05:56:15","extension":"xml","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":131910,"visible":true,"origin":"","legend":"","description":"","filename":"dd4556a6ebd449c18dc9780eac7dddce1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/97af27228cb2aaffb7e695d8.xml"},{"id":99371652,"identity":"dd2c12b2-6512-4cfc-a7fa-3877fb07874c","added_by":"auto","created_at":"2026-01-02 05:56:16","extension":"html","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":141474,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/7e350b08e81956013bdaf133.html"},{"id":99371630,"identity":"5560f1bc-949e-4c64-a2cc-22ec91c5549f","added_by":"auto","created_at":"2026-01-02 05:56:15","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":74759,"visible":true,"origin":"","legend":"\u003cp\u003eStudy Flow Diagram for Cohort Selection and Data Processing\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/5d08efd70fae2cfc63030255.png"},{"id":99371629,"identity":"41651b1d-1cc0-45c8-8667-8995a106de5e","added_by":"auto","created_at":"2026-01-02 05:56:15","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":99554,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eROC and Precision–Recall curves of the threshold-optimized CatBoost model on the test set.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eROC: Receiver Operating Characteristic; PR: Precision–Recall; AUC: area under the ROC curve; AP: average precision (area under the PR curve). The orange marker indicates the optimal probability threshold selected using the F1-maximization criterion (thr = 0.402). In the ROC plot, this threshold corresponds to the operating point balancing sensitivity and specificity; in the PR curve, the same threshold reflects the precision–recall trade-off at the point of maximal F1 performance.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/df12fcbbefcddf8b19da9f90.png"},{"id":99371644,"identity":"263f7884-09dc-4313-b2bf-a765f049be59","added_by":"auto","created_at":"2026-01-02 05:56:15","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":146739,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eGlobal feature importance visualization using SHAP beeswarm for the final CatBoost classifier.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSHAP: Shapley Additive Explanations. Each point represents a single observation, with color indicating the underlying feature value (red = high, blue = low). The position on the x-axis reflects the SHAP value, which quantifies the direction and magnitude of each feature’s contribution to the model output. Positive SHAP values shift the prediction toward vitamin B12 deficiency, whereas negative values shift the prediction toward the non-deficient class. The vertical spread demonstrates the variability of feature effects across individuals.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/dd89f4424462cdab53fb489e.png"},{"id":99789243,"identity":"b5f6a49a-85d7-421c-85f5-8149e6519dfe","added_by":"auto","created_at":"2026-01-08 12:49:10","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":103034,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDecision curve analysis of the CatBoost and logistic regression models for predicting vitamin B12 deficiency\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe solid blue line represents the CatBoost model, and the solid orange line represents logistic regression. The green dashed line (“Treat All”) assumes all patients would be classified as deficient, whereas the red dotted line (“Treat None”) assumes no intervention. Net benefit is plotted across threshold probabilities for predicting low vitamin B12 levels, demonstrating the clinical utility of each strategy relative to default decision rules.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/7e6ade0144e8e200883a4cfe.png"},{"id":103405521,"identity":"1ffe8247-34e4-4785-8e1b-d62eac365973","added_by":"auto","created_at":"2026-02-25 09:58:26","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1641725,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/573d5c52-fce1-46d6-8fbf-12564c09f95f.pdf"},{"id":99371631,"identity":"9d594b8d-d319-4088-baae-24c5ae0952b6","added_by":"auto","created_at":"2026-01-02 05:56:15","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":123133,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterial1.docx","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/de9f14987e4b27ab2aa907e3.docx"},{"id":99371634,"identity":"c44c53f5-cf83-4c86-942d-a62e6fddb5df","added_by":"auto","created_at":"2026-01-02 05:56:15","extension":"xlsx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":17334,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterial2.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8176315/v1/6901616f5b6d8c73536edfb3.xlsx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Development of Machine Learning Algorithms for Predicting Vitamin B12 Levels Using Biochemical Analyte Data","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eVitamin B12 (cobalamin) is an essential micronutrient required for DNA synthesis, myelin maintenance, and one-carbon metabolism. Deficiency can result in macrocytic anemia, neurocognitive impairment, and increased cardiovascular risk mediated through hyperhomocysteinemia [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Although serum total B12 concentration is the most frequently used diagnostic test, it often fails to reflect intracellular cobalamin status, leading to \u0026ldquo;functional B12 deficiency\u0026rdquo; even when serum levels fall within the low-normal range [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. More specific biomarkers\u0026mdash;such as holotranscobalamin (holo-TC), methylmalonic acid (MMA), and homocysteine (Hcy)\u0026mdash;provide improved diagnostic accuracy, yet their higher cost, limited availability, and lack of integration into routine laboratory workflows restrict widespread use [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Consequently, clinically significant deficiency may remain undetected until hematological or neurological manifestations become evident [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn recent years, the growing availability of large-scale laboratory datasets and advances in machine learning (ML) have opened new possibilities for leveraging routinely collected biochemical and hematological parameters to predict micronutrient deficiencies. Tamune et al. demonstrated that ML models using standard blood tests can efficiently predict vitamin B group deficiencies, including B12, in patients with acute psychiatric episodes [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Beyond B12, several studies have applied ML to other micronutrients and lipid biomarkers: Sharifmousavi and Borhani developed an SVM-based model that incorporated selenium, vitamin B12, and vitamin D3 levels for the diagnosis of multiple sclerosis [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], while Sancar and Tabrizi showed that ML approaches can accurately classify vitamin D status using routine laboratory variables [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Similarly, ML-based models have been proposed for predicting low-density lipoprotein cholesterol (LDL-C) from standard biochemical profiles, achieving good agreement with directly measured and calculated LDL-C values [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Collectively, these findings highlight the feasibility of using routinely available laboratory data to build predictive models for diverse biochemical targets.\u003c/p\u003e \u003cp\u003eAnother important development has been the increasing emphasis on explainable artificial intelligence (XAI) in clinical prediction models. Hu et al. reported an explainable ML model for predicting spontaneous bacterial peritonitis in cirrhotic patients with ascites, using SHAP-based explanations to identify the most influential laboratory and clinical variables. In a broader perspective, a recent systematic review by Alkhanbouli et al. underscored the role of explainable AI techniques in disease prediction and clinical decision support, emphasizing transparency, trust, and model interpretability as essential requirements for implementation in healthcare settings [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. For high-stakes decisions such as the detection of vitamin B12 deficiency\u0026mdash;where underdiagnosis may lead to irreversible neurological damage\u0026mdash;models that are both accurate and interpretable are particularly valuable.\u003c/p\u003e \u003cp\u003eGiven the high prevalence of vitamin B12 deficiency, its often subtle or delayed clinical presentation, and the limited accessibility of confirmatory biomarkers, there is a clear need for robust, low-cost, and scalable predictive models based entirely on routine laboratory parameters. Such models could be integrated into hospital information systems as automated decision-support tools, providing real-time alerts for individuals at high risk and reducing unnecessary second-line testing.\u003c/p\u003e \u003cp\u003eThe present study aims to address these gaps by developing and validating several machine learning algorithms\u0026mdash;including artificial neural networks, boosting-based ensemble models, support vector machines, and logistic regression\u0026mdash;to predict serum vitamin B12 levels using a large, multiyear dataset of routinely collected biochemical and hematological parameters. Model performance was evaluated using ROC-AUC, PR-AUC, F1-score, accuracy, and the Matthews correlation coefficient. To ensure clinical interpretability, SHAP and LIME techniques were employed to identify the variables contributing most substantially to model predictions. Furthermore, an independent temporal validation cohort was used to assess the generalizability and robustness of the final model under real-world conditions. Overall, this study aims to establish a reliable, scalable, and explainable machine learning framework capable of supporting early identification of individuals at risk for vitamin B12 deficiency within routine clinical practice.\u003c/p\u003e"},{"header":"MATERIALS AND METHODS","content":"\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\n \u003ch2\u003eStudy Population / Subjects\u003c/h2\u003e\n \u003cp\u003eThis retrospective study included adult patients who underwent serum vitamin B12 testing together with routine hematological and biochemical analyses at Health Sciences University İzmir Tepecik Training and Research Hospital between 1 January 2015 and 1 October 2025. Prior to study initiation, approval was obtained from the hospital\u0026rsquo;s Non-Interventional Research Ethics Committee (initial ethics approval: 13/07/2023\u0026ndash;2023/06\u0026ndash;38; additional approval: 30/10/2025\u0026ndash;293077648). Laboratory data were extracted from the hospital information system (HIS) and were based on analyzer outputs verified and authorized by a board-certified medical biochemistry specialist. Only the first eligible laboratory encounter for each patient was included in the study, and all records were anonymized during data extraction.\u003c/p\u003e\n \u003cp\u003eInclusion Criteria\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003eAge\u0026thinsp;\u0026ge;\u0026thinsp;18 years\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eAvailability of serum vitamin B12 measurement performed concurrently with all predefined hematological and biochemical tests\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eFirst eligible outpatient laboratory encounter within the study period\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eComplete numerical laboratory results available for all required parameters\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eSpecimen processed within the institution\u0026rsquo;s standardized pre-analytical workflow and quality limits\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n \u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n \u003cp\u003eExclusion Criteria\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003eAge\u0026thinsp;\u0026lt;\u0026thinsp;18 years\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eMissing, corrupted, or non-numerical laboratory results\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eMultiple encounters for the same patient, in which case only the first eligible record was retained\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eSpecimens exhibiting pre-analytical delays exceeding institutional workflow limits\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003ePregnancy\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eActive oncologic disease or hematologic malignancy\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eForensic/medico-legal cases\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eEmergent conditions associated with spontaneous acute bleeding or requiring immediate emergency evaluation, including but not limited to:\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n \u003cp style=\"margin-left: 40px;\"\u003e\u0026ndash; Acute coronary syndrome\u003c/p\u003e\n \u003cp style=\"margin-left: 40px;\"\u003e\u0026ndash; Pulmonary embolism\u003c/p\u003e\n \u003cp style=\"margin-left: 40px;\"\u003e\u0026ndash; Aortic dissection/rupture\u003c/p\u003e\n \u003cp style=\"margin-left: 40px;\"\u003e\u0026ndash; Subacute arachnoid hemorrhage\u003c/p\u003e\n \u003cp style=\"margin-left: 40px;\"\u003e\u0026ndash; Transient ischemic attack\u003c/p\u003e\n \u003cp style=\"margin-left: 40px;\"\u003e\u0026ndash; Massive gastrointestinal hemorrhage\u003c/p\u003e\n \u003cp\u003eVenous whole-blood samples were collected into K₂-EDTA tubes, and serum samples were obtained in additive-free plain tube, in accordance with routine clinical protocols. In line with the National Health Quality Standards (Version 5.1), all specimens were delivered to the laboratory within 30 minutes, and biochemical analyses were completed within two hours of sample collection.\u003c/p\u003e\n \u003cp\u003eHematological parameters (HGB, RBC, HCT, MCV, RDW, MCH, MCHC, WBC, NEU, LYM, MONO, BASO, PLT, MPV, PDW) were measured using Beckman Coulter LH-780 (Mannheim, Germany) and Sysmex XN-1000 (Kobe, Japan) analyzers. Biochemical assays (glucose, creatinine, ALT, AST, ALP, GGT, LDH, serum iron, CRP, albumin, total protein, total bilirubin) were performed on a Beckman Coulter AU5800 analyzer (Mannheim, Germany). Chemiluminescence-based tests (vitamin B12, folate, ferritin, fT4, TSH) were conducted on Beckman Coulter DxI800 analyzers (Mannheim, Germany).\u003c/p\u003e\n \u003cp\u003eAll reagents, calibrators, and internal quality-control materials were certified and approved by the manufacturer. Standard institutional internal and external quality-control procedures were maintained throughout the analytical process. All laboratory results were reviewed and analytically validated by a medical biochemistry specialist before being transferred to the hospital information system.\u003c/p\u003e\n\u003c/div\u003e\n\u003ch3\u003eStudy Design\u003c/h3\u003e\n\u003cp\u003eTo ensure data confidentiality, all patient identifiers were irreversibly removed prior to analysis. A comprehensive dataset containing age, gender, and all hematological and biochemical parameters was generated through the hospital information system (HIS). A total of 87,301 laboratory records from the study period were reviewed and exported to Microsoft Excel 2021 (USA). After applying the predefined exclusion criteria, 51,630 patients were deemed eligible and included in the study, while all other records were excluded.\u003c/p\u003e\n\u003cp\u003eTo address class imbalance in the machine-learning analyses and enable the model to effectively learn the minority class, a stratified sampling approach was used to construct the development dataset. This procedure was performed in Python: all observations belonging to the minority class were retained, whereas an equal number of records from the majority class were randomly selected. As a result of this balancing step, 16,354 records were randomly discarded, yielding a final development cohort of 35,276 patients.\u003c/p\u003e\n\u003cp\u003eThe development dataset (01 January 2015\u0026ndash;30 May 2023) was subsequently split into 80% training and 20% test sets using stratified random partitioning to prevent overfitting and preserve class distribution. Using identical preprocessing and class-balancing principles, an independent temporal validation set consisting of 34,744 patient records (collected between 01 June 2023 and 01 October 2025) was used for external validation. This allowed the model\u0026rsquo;s temporal generalizability to be evaluated on data from a later clinical period.\u003c/p\u003e\n\u003cp\u003eAll data preprocessing, model development, class balancing, training\u0026ndash;test partitioning, and reporting procedures were conducted in accordance with the TRIPOD guidelines. A flow diagram summarizing patient selection, exclusion criteria, and the final analytic cohorts is presented in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\n\u003ch3\u003eData Preprocessing and Training of Machine-Learning Algorithms\u003c/h3\u003e\n\u003cp\u003eRaw laboratory data extracted from the HIS were preprocessed through removal of duplicate patient records, standardization of unit inconsistencies for selected biochemical parameters (e.g., albumin, total protein), and exclusion of entries containing missing parameters. Given that extreme values may represent true biological measurements within clinical practice and can directly influence model performance, outliers were not removed. Age was retained as a continuous variable, whereas all biochemical and hematological analytes were transformed into categorical variables based on clinically defined threshold values. Evidence-based clinical decision limits were used when available; otherwise, manufacturer-provided reference intervals were applied. Each analyte was encoded as 0 (below reference range), 1 (within reference range), or 2 (above reference range). For parameters without a biologically meaningful lower bound, a binary scheme was adopted. Gender was coded as 1 for males and 2 for females.\u003c/p\u003e\n\u003cp\u003eThe initial three-category structure of vitamin B12 was converted into a binary variable (0\u0026thinsp;=\u0026thinsp;low; 1\u0026thinsp;=\u0026thinsp;normal/high) to reduce class imbalance and improve clinical interpretability. Classification thresholds for all analytes are provided in Supplementary Material 1, Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\n\u003cp\u003eEight supervised machine-learning algorithms\u0026mdash;logistic regression, random forest, decision tree, support vector machines, k-nearest neighbors, XGBoost, CatBoost, and artificial neural networks\u0026mdash;were developed within a four-stage experimental framework. In experiemnt 1, all algorithms were trained using default hyperparameters. In experiment 2, probability-threshold optimization was performed using the Youden J index and F1-maximization criterion. Experiment 3 involved extensive hyperparameter tuning via GridSearchCV, RandomizedSearchCV, and Optuna-based bayesian optimization. All cross-validation procedures were conducted exclusively on the training set to prevent information leakage. In experiment 4, a large-scale feature-engineering pipeline generated more than 250 derived variables, including variance-based metrics, ratio-based biomarkers, composite indices, and biologically meaningful interaction terms (Supplementary Material 2).\u003c/p\u003e\n\u003cp\u003eAll preprocessing and model-building steps were implemented in Python 3.11 (Holland) using Kaggle Notebook and Google Colab environments. Continuous variables were standardized using the StandardScaler function, and the final dataset was randomly split into training (80%) and test (20%) subsets using a stratified approach to preserve the distribution of vitamin B12 classes.\u003c/p\u003e\n\u003cp\u003eAfter completion of all stages of the experimental framework, the methodological strategy yielding the highest overall performance was selected and applied\u0026mdash;without modification\u0026mdash;to an independent temporal validation cohort of 34,744 patients from a later time period, in order to assess real-world, time-dependent generalizability. Model interpretability and clinical utility were comprehensively evaluated using SHAP (feature contributions), LIME (local explainability), and Decision Curve Analysis (net clinical benefit).\u003c/p\u003e\n\u003ch3\u003ePerformance Evaluation\u003c/h3\u003e\n\u003cp\u003eDescriptive analyses were performed for both the development and validation cohorts. Continuous variables were summarized as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation, and between-group differences were assessed using p-values and Cohen\u0026rsquo;s d. Additional stratified analyses by age and sex were conducted separately for the development and validation datasets. Subsequently, the performance of the eight algorithms included in the four-stage experimental framework was evaluated using sensitivity, specificity, positive and negative predictive values, accuracy, F1 score, Matthews correlation coefficient, area under the ROC curve (AUC-ROC), and area under the precision\u0026ndash;recall curve (AUC-PR), the latter being particularly informative for imbalanced outcomes. For each model, confusion-matrix components (TP, FP, TN, FN) were computed, and ROC/PR curves were generated across a range of probability thresholds.\u003c/p\u003e\n\u003cp\u003eAfter completion of all experimental stages, the methodological strategy demonstrating the highest overall performance on the test set was selected and applied unchanged to the validation cohort. Clinical interpretability was assessed using SHAP beeswarm plots, LIME-based local explanations, and Decision Curve Analysis, providing complementary insights into feature contributions and net clinical utility.\u003c/p\u003e\n\u003cp\u003eThroughout the analyses, the categorical vitamin B12 variable used as an input feature (0\u0026thinsp;=\u0026thinsp;low, 1\u0026thinsp;=\u0026thinsp;normal/high) represented the laboratory classification, whereas a predicted value of 1 in the performance assessment indicated a model-predicted \u0026ldquo;vitamin B12 deficiency.\u0026rdquo; These two representations were explicitly distinguished across all analytical steps to avoid interpretive ambiguity.\u003c/p\u003e\n\u003ch3\u003eStatistical Comparison of Experimental Models\u003c/h3\u003e\n\u003cp\u003eSince all machine-learning algorithms were evaluated on the same test set, statistical comparisons accounted for the dependence arising from correlated prediction structures. DeLong\u0026rsquo;s test was used to compare AUC-ROC values derived from correlated ROC curves. Differences in continuous performance metrics such as F1 score and accuracy were assessed using paired t-tests.\u003c/p\u003e\n\u003cp\u003eTo examine false-positive/false-negative imbalance and asymmetry in misclassification patterns across models, the McNemar test was applied. In addition, Net Reclassification Improvement (NRI) and Integrated Discrimination Improvement (IDI) were calculated to quantify the relative gains in reclassification performance between models. These metrics enabled a more comprehensive evaluation of the methodological advantages achieved through model optimization, particularly regarding improvements in accuracy and discrimination.\u003c/p\u003e\n\u003cp\u003eTogether, these complementary procedures allowed for a rigorous, multidimensional, and model-agnostic comparison of all experimental configurations.\u003c/p\u003e\n\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\n \u003ch2\u003eStatistical Analysis\u003c/h2\u003e\n \u003cp\u003eContinuous variables were summarized as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation, whereas categorical variables were presented as counts and percentages. Given the large sample size, parametric methods were preferred. Group comparisons for continuous variables were performed using one-way ANOVA, and Welch\u0026rsquo;s ANOVA was applied when the assumption of homogeneity of variances was violated. When appropriate, Tukey\u0026rsquo;s HSD or Games\u0026ndash;Howell tests were used for post hoc multiple comparisons. Categorical variables were compared using the Pearson chi-square test.\u003c/p\u003e\n \u003cp\u003eThe diagnostic performance of the machine-learning models was assessed using sensitivity, specificity, positive and negative predictive values, accuracy, F1 score, AUC-ROC, AUC-PR, positive and negative likelihood ratios (LR⁺/LR⁻), and the Matthews correlation coefficient. All statistical analyses were conducted in Python 3.11 using the \u003cem\u003epandas\u003c/em\u003e, \u003cem\u003escipy\u003c/em\u003e, \u003cem\u003escikit-learn\u003c/em\u003e, and \u003cem\u003estatsmodels\u003c/em\u003e libraries. A two-sided p-value\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"RESULTS","content":"\u003cp\u003eStatistically significant differences were observed across numerous hematological and biochemical parameters between the development (n\u0026thinsp;=\u0026thinsp;35,256) and independent validation (n\u0026thinsp;=\u0026thinsp;34,744) cohorts (with p\u0026thinsp;\u0026lt;\u0026thinsp;0.05 for most comparisons). However, Cohen\u0026rsquo;s d analyses indicated that the magnitude of these differences was predominantly very small from a biological perspective. Age, hemoglobin, hematocrit, erythrocyte count, and MCV values were lower in the validation cohort, demonstrating small to moderate effect sizes (d\u0026thinsp;=\u0026thinsp;0.18\u0026ndash;0.33). In contrast, parameters reflecting inflammatory or hematologic variability\u0026mdash;such as RDW, CRP, LDH, and PLT\u0026mdash;were higher, corresponding to small effect sizes (d\u0026thinsp;\u0026asymp;\u0026thinsp;\u0026minus;\u0026thinsp;0.16 to \u0026minus;\u0026thinsp;0.20). Among liver enzymes, only ALP exhibited a moderate biologic effect (d\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;0.546). Although serum vitamin B12 concentrations were higher in the validation period, the biological magnitude of this difference remained very small (d\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;0.143). Means, standard deviations, p-values, and Cohen\u0026rsquo;s d effect sizes for all variables are presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Additionally, age- and sex-stratified distributions and comparative statistics for both cohorts are provided in Supplementary Material 1, Tables\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e\u0026ndash;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eAcross the eight machine-learning algorithms evaluated within the four-stage experimental framework, CatBoost demonstrated the most balanced sensitivity\u0026ndash;specificity profile and the highest overall classification performance in all experiments. Notably, the threshold-optimized configuration (Experiment 2), despite yielding performance metrics numerically similar to the other strategies, provided a clinically decisive advantage. This approach improved model accuracy not only from a statistical perspective but also in terms of clinical safety, achieving the lowest number of false negatives and thereby minimizing the risk of missing true vitamin B12 deficiency cases. Given the association of vitamin B12 deficiency with neurological sequelae and irreversible damage, this marked reduction in false negatives represents the key outcome defining the model\u0026rsquo;s clinical applicability rather than merely its methodological strength.\u003c/p\u003e \u003cp\u003eFor this reason, the primary findings highlighted in the manuscript are derived from the F1-maximization\u0026ndash;based threshold-optimized configuration, which is summarized in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Performance distributions for all experimental strategies are provided in Supplementary Material 1, Tables\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e\u0026ndash;6.\u003c/p\u003e \u003cp\u003eIn the test set, the threshold-optimized CatBoost model achieved sensitivity 0.92, specificity 0.67, F1 score 0.82, PPV 0.73, and NPV 0.88, yielding the most balanced performance among all models; AUC-ROC was 0.88 and AUC-PR was 0.86. XGBoost and the neural network produced comparable AUC values but demonstrated lower accuracy, specificity, MCC, and PPV. Logistic regression and random forest failed to maintain overall classification balance, whereas SVM, KNN, and decision tree models showed markedly reduced specificity and MCC.\u003c/p\u003e \u003cp\u003ePerformance metrics for the four CatBoost configurations are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Statistical comparisons revealed no significant differences across configurations for AUC-ROC (DeLong test) or for F1 and accuracy (paired t-tests) (all p\u0026thinsp;\u0026gt;\u0026thinsp;0.05). In contrast, the McNemar test demonstrated a clear advantage for the threshold-optimized model, particularly through its favorable asymmetry in the distribution of false negatives versus false positives. This model substantially reduced false-negative decisions\u0026mdash;the clinically most critical error type\u0026mdash;supporting its selection as the final model.\u003c/p\u003e \u003cp\u003eFor the threshold-optimized CatBoost model, the calculated NRI\u0026thinsp;=\u0026thinsp;0.359 and IDI\u0026thinsp;=\u0026thinsp;0.147 further confirmed that the approach produced meaningful improvements in both correct reclassification and discrimination capacity. These findings align with the McNemar test results demonstrating improved false-negative/false-positive asymmetry, collectively indicating that the optimized model provides a substantially enhanced risk-classification framework. Accordingly, the model exhibits strong potential as a reliable and effective clinical decision-support tool for the early identification of vitamin B12 deficiency.\u003c/p\u003e \u003cp\u003eThe performance of the selected CatBoost model remained robust in the independent temporal validation cohort, with sensitivity 0.85, specificity 0.77, accuracy 0.81, F1 score 0.82, and MCC 0.63. Discrimination improved compared with the internal test set, with AUC-ROC 0.90 and AUC-PR 0.91. The likelihood-ratio profile (PLR\u0026thinsp;=\u0026thinsp;3.74; NLR\u0026thinsp;=\u0026thinsp;0.19) also indicated strong clinical utility under real-world conditions. These results are presented in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eSHAP analyses demonstrated that hematologic and inflammatory parameters\u0026mdash;such as MCV, HGB, HCT, RBC, RDW, age, iron and ferritin\u0026mdash;were the strongest contributors to model predictions. The SHAP beeswarm plot (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) showed that physiologically abnormal measurements consistently shifted predictions toward deficiency. LIME analyses confirmed the internal consistency of feature contributions for individual-level predictions and reinforced the clinical interpretability of the model (Supplementary Material 1, Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Decision Curve Analysis (DCA) demonstrated that the model provided a consistently higher net benefit across a wide threshold range compared with \u0026ldquo;treat-all\u0026rdquo; or \u0026ldquo;treat-none\u0026rdquo; strategies (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e), supporting its potential integration into clinical decision-support systems.\u003c/p\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eThis study represents the first comprehensive machine-learning investigation designed to predict serum vitamin B12 levels directly from routinely collected laboratory parameters. Current literature contains predictive models for vitamin D, folate, lipid metabolism, and other biochemical markers, yet no prior study has focused specifically on estimating B12 status using hematologic and biochemical indices at scale [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. By integrating routine laboratory features with explainable artificial intelligence (XAI) techniques, the present work offers a novel and clinically actionable framework that bridges biomedical informatics, laboratory medicine, and nutritional neuroscience.\u003c/p\u003e \u003cp\u003eAcross eight machine-learning algorithms evaluated in four sequential experimental settings, gradient-boosting models consistently demonstrated superior ability to capture the nonlinear and population-level variability inherent in biochemical data. Linear models such as logistic regression and support vector machines showed limited generalizability with AUC values around 0.86, consistent with prior work by Sharifmousavi et al. using SVM-based micronutrient prediction in neurological populations [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Artificial neural networks yielded lower accuracy (0.78) than the 85% reported by Tamune et al. in B6 deficiency [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e], while also suffering from well-recognized limitations in interpretability—an important consideration for clinical deployment. Tree-based ensembles such as random forest similarly demonstrated moderate discrimination but experienced performance instability under external validation, a finding aligned with previously reported limitations of over-partitioning in high-dimensional biomedical data [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eXGBoost outperformed these baseline models, reflecting its strong capacity to capture interactions between biochemical variables, consistent with prior lipid profiling work (r = 0.98) by Anudeep et al. [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. However, CatBoost provided the most robust and generalizable performance across all experimental conditions—an observation supported by previous reports emphasizing its unbiased handling of categorical variables and resilience in clinical datasets [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. In our study, CatBoost achieved an AUC-ROC of 0.88, AUC-PR of 0.86, F1 of 0.82, and the most balanced sensitivity–specificity profile across both test and validation cohorts.\u003c/p\u003e \u003cp\u003eInterpretability analyses further strengthened the biological plausibility of the model. SHAP and LIME consistently highlighted MCV, RDW, HGB, HCT, ferritin, CRP, folate, and age as principal contributors to classification. These findings reinforce well-established hematological consequences of B12 deficiency, including macrocytosis, anisocytosis, and impaired erythropoiesis. Elevated contributions of ferritin and CRP mirror the known inhibitory effects of systemic inflammation on circulating B12 bioavailability via haptocorrin-mediated sequestration, as reported by Jensen et al. [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. The observed interplay between folate and B12 also aligns with their shared role in one-carbon metabolism and DNA synthesis [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Age contributed positively, consistent with declining intrinsic factor secretion and reduced gastrointestinal absorption in older adults [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Together, these patterns demonstrate that the model not only predicts deficiency but also captures meaningful pathophysiological signatures associated with B12 metabolism.\u003c/p\u003e \u003cp\u003eThe sequential experimental framework provided additional insight into how model performance evolved under different analytical strategies. Experiment 1 established baseline performance; Experiment 2 applied F1-maximizing threshold optimization and yielded the most clinically balanced results, increasing sensitivity to 0.92 and reducing false negatives by nearly 45%. Experiment 3 evaluated hyperparameter tuning, which improved discrimination modestly but introduced mild overfitting in the validation cohort—consistent with prior findings on the trade-off between complexity and generalizability in clinical ML models [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Experiment 4 introduced feature engineering; however, CatBoost’s intrinsic ability to manage correlated and categorical variables meant that manual engineering added minimal incremental benefit. Collectively, these findings validated Experiment 2 as the optimal configuration for clinical decision support.\u003c/p\u003e \u003cp\u003eWhen compared with the only available machine-learning study on B12 deficiency, conducted by Tamune et al. in psychiatric patients using a sample of 497 individuals (AUC-ROC 0.62), our model showed substantially higher discrimination (AUC-ROC 0.88) using a dataset nearly 150 times larger, highlighting its superior generalizability and methodological rigor [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. In contrast to genetic Mendelian randomization studies such as Lu et al. [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e], which explore upstream determinants of B12 metabolism, our work emphasizes biochemical phenotypes and their downstream hematologic expression—providing complementary but clinically more immediate insights. Similarly, while Duman et al. demonstrated hematologic alterations in pediatric B12 deficiency [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e], their narrower variable set lacked inflammatory and metabolic markers that contributed substantially to our model’s predictive signal.\u003c/p\u003e \u003cp\u003eFrom an implementation perspective, the model has strong potential for integration into laboratory information systems (LIS) or hospital information systems (HIS). Because it relies entirely on existing routine tests (e.g., hemogram, ferritin, CRP, albumin), it imposes no additional cost and can be deployed passively in background systems. LIME-based patient-specific explanations could be displayed to clinicians, allowing early recognition of biochemical patterns suggestive of deficiency before irreversible neurological or hematological complications arise. Early B12 replacement has been shown to reduce hospitalization duration by approximately two days, generating meaningful savings in medium-sized hospitals [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Such a model could therefore contribute to both clinical and economic decision-making.\u003c/p\u003e "},{"header":"Limitations","content":"\u003cp\u003eThis study has several limitations. Its retrospective nature prevented inclusion of symptoms, dietary intake, medication history, gastrointestinal disorders, and genetic factors—variables known to influence B12 metabolism. Only adult patients were included, limiting generalizability to pediatric populations. Although data were obtained from the laboratory information system, all results had undergone biochemical specialist approval, ensuring that pre-analytical artifacts such as hemolysis or improper storage were already filtered out. Missing values were removed rather than imputed, eliminating imputation bias but introducing a theoretical selection bias. Classification thresholds were based on the NICE 2024 guideline, ensuring standardization across laboratories [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. Nevertheless, total serum B12 is an imperfect marker of functional deficiency; lack of holo-transcobalamin or methylmalonic acid data may lead to misclassification in certain subgroups [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] Finally, micronutrient patterns may vary across regions and populations, suggesting that model recalibration may be necessary before widespread deployment [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. As with all clinical ML models, temporal performance drift remains a potential risk and requires periodic re-evaluation [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. Serum total vitamin B12 levels were not measured using the gold-standard isotope-dilution liquid chromatography–tandem mass spectrometry (ID-LC–MS/MS) method. Instead, a chemiluminescence immunoassay was used, which is subject to method-dependent analytical bias, including variability in antibody specificity and differential detection of protein-bound B12. Therefore, assay-related misclassification cannot be fully excluded.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eDespite these limitations, this study introduces the first large-scale, explainable, and clinically validated machine-learning model for predicting vitamin B12 deficiency using routine laboratory parameters. The model demonstrates high discrimination, strong biological interpretability, and stable performance across internal and temporal validation cohorts. When integrated into clinical workflows, it has the potential to support early diagnosis, reduce unnecessary testing, and enhance decision-making in nutritional and hematologic assessments. Future work should focus on multi-center prospective validation, continuous monitoring for model drift, and extension of the framework to other micronutrient deficiencies such as folate, vitamin D, and ferritin. In parallel, evaluating how this explainable AI model behaves when quietly embedded into routine laboratory workflows\u0026mdash;augmenting clinical decisions without disrupting established pathways\u0026mdash;will be essential for realizing its full translational impact.\u003c/p\u003e "},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eEthics approval and consent to participate\u003c/h2\u003e \u003cp\u003eThis study was conducted in accordance with the Declaration of Helsinki and approved by the Non-Interventional Clinical Research Ethics Committee of the İzmir Tepecik Training and Research Hospital, University of Health Sciences T\u0026uuml;rkiye (Initial ethics approval: 13/07/2023\u0026ndash;2023/06\u0026ndash;38; additional approval: 30/10/2025\u0026ndash;293077648). The requirement for informed consent was waived by the ethics committee because the study used retrospective, anonymized laboratory data with no identifiable personal information.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eConsent for publication\u003c/strong\u003e \u003cp\u003eNot applicable. No individual person\u0026rsquo;s data in any form (including images, videos, or identifiable information) are included in this manuscript.\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eCompeting interests\u003c/h2\u003e \u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eFootnotes\u003c/h2\u003e \u003cp\u003eNot applicable.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThis research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eFD (Ferhat Demirci) obtained ethics approval, conceptualized the study, curated and analyzed the data, developed the methodology, wrote and optimized the machine-learning algorithms, performed the statistical analysis, and drafted the original manuscript. OY (Oktay Yıldırım) contributed to data collection, assisted in writing, debugging, and refining the Python code used in data preprocessing and model development, and participated in the review and editing of the manuscript. PA (Pınar Akan) verified the accuracy and integrity of the laboratory data, conducted literature review, and contributed to writing sections of the original manuscript draft. All authors read and approved the final version of the manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThe authors would like to express their sincere gratitude to Prof. Dr. Savaş Yakan, Chief Physician of İzmir Tepecik Training and Research Hospital, and Prof. Dr. Ayfer \u0026Ccedil;olak, Head of the Department of Medical Biochemistry, for their valuable support and encouragement throughout this study.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets generated and analyzed during the current study are not publicly available due to institutional data protection regulations but are available from the corresponding author (FD) upon reasonable request. All key summary tables and model coefficients are included in this published article and its supplementary information files.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eO\u0026rsquo;Leary F, Samman S. Vitamin B12 in Health and Disease. Nutrients. 2010;2(3):299\u0026ndash;316.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStabler SP. Vitamin B12 Deficiency. N Engl J Med. 2013;368(2):149\u0026ndash;60.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHannibal L, Lysne V, Bj\u0026oslash;rke-Monsen AL, Behringer S, Gr\u0026uuml;nert SC, Spiekerkoetter U et al. Biomarkers and Algorithms for the Diagnosis of Vitamin B12 Deficiency. Front Mol Biosci. 2016;3.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStover P, Discussion. Folate and Vitamin B12 Metabolism: Overview and Interaction with Riboflavin, Vitamin B6, and Polymorphisms. Food Nutr Bull. 2008;29(2suppl1):S17\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTamune H, Ukita J, Hamamoto Y, Tanaka H, Narushima K, Yamamoto N. Efficient Prediction of Vitamin B Deficiencies via Machine-Learning Using Routine Blood Test Results in Patients With Intense Psychiatric Episode. Front Psychiatry. 2019;10:1029.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSharifmousavi SS, Borhani MS. Support vectors machine-based model for diagnosis of multiple sclerosis using the plasma levels of selenium, vitamin B12, and vitamin D3. Inf Med Unlocked. 2020;20:100382.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSancar N, Tabrizi SS. Machine learning approach for the detection of vitamin D level: a comparative study. BMC Med Inf Decis Mak. 2023;23(1):219.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKumari PPA, Rajasimman S, Nayak AS, Priyadarsini S. Machine learning predictive models of LDL-C in the population of eastern India and its comparison with directly measured and calculated LDL-C. Annals of Clinical Biochemistry. Int J Lab Med. 2022;59(1):76\u0026ndash;86.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlkhanbouli R, Matar Abdulla Almadhaani H, Alhosani F, Simsekler MCE. The role of explainable artificial intelligence in disease prediction: a systematic literature review and future research directions. BMC Med Inf Decis Mak. 2025;25(1):110.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePaplomatas P, Kantartzi M, Anagnostopoulos K, Vrahatis A. An Artificial Intelligence and Biochemical Modelling Approach for LDL Cholesterol Prediction. Medinformatics. 2025.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim JM, Stewart R, Kim SW, Yang SJ, Shin IS, Yoon JS. Predictive value of folate, vitamin B 12 and homocysteine levels in late-life depression. Br J Psychiatry. 2008;192(4):268\u0026ndash;74.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMohammed H, Ahamad S. S, Abdul Hameed DrM. Enhancements in Random Forest Algorithms for Improving Healthcare Applications. In: Proceedings of the International Conference on Industrial Engineering and Operations Management. Michigan, USA: IEOM Society International; 2024.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHu Y, Chen R, Gao H, Lin H, Wang J, Wang X, et al. Explainable machine learning model for predicting spontaneous bacterial peritonitis in cirrhotic patients with ascites. Sci Rep. 2021;11(1):21639.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eProkhorenkova L, Gusev G, Vorobev A, Dorogush AV, Gulin A. (2018). CatBoost: unbiased boosting with categorical features. Advances in neural information processing systems, 2018 (31).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIsmail O, Albdour K, Albdour Z, Jaber K. Differences in Ferritin, Vitamin D, and Vitamin B12 Between Fibromyalgia Patients and Healthy Individuals: A Systematic Review and Meta-Analysis. Musculoskelet Care. 2025;23(1).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJensen HR, Laursen MF, Lildballe DL, Andersen JB, Nex\u0026oslash; E, Licht TR. Effect of the vitamin B12-binding protein haptocorrin present in human milk on a panel of commensal and pathogenic bacteria. BMC Res Notes. 2011;4:208.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eReynolds E. Vitamin B12, folic acid, and the nervous system. Lancet Neurol. 2006;5(11):949\u0026ndash;60.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSimonenko SYu, Bogdanova DA, Kuldyushev NA. Emerging Roles of Vitamin B12 in Aging and Inflammation. Int J Mol Sci. 2024;25(9):5044.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWong J, Manderson T, Abrahamowicz M, Buckeridge DL, Tamblyn R. Can Hyperparameter Tuning Improve the Performance of a Super Learner? Epidemiology. 2019;30(4):521\u0026ndash;31.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLu T, Paterson AD. Estimating effects of serum vitamin B12 levels on psychiatric disorders and cognitive impairment: a Mendelian randomization study. Commun Med. 2025;5(1):316.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDuman D, Uslu G, Malbora B. Retrospective Analysis of Pediatric Patients with Vitamin B12 Deficiency: Vitamin B12 and Children. Chronicles Precision Med Researchers. 2024;5(1):39\u0026ndash;45.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eda Gomes C, von Gunten A, Jopp D, Ribeiro O, Verloo H. Why Centenarians\u0026rsquo; Depressive Symptoms Must Become a Priority for Nurses. J Am Med Dir Assoc. 2021;22(5):1118\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNICE guideline. Vitamin B12 deficiency in over 16s: diagnosis and management [Internet]. 2024 [cited 2025 Oct 4]. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.nice.org.uk/guidance/ng239\u003c/span\u003e\u003cspan address=\"https://www.nice.org.uk/guidance/ng239\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDastidar R, Sikder K. Diagnostic reliability of serum active B12 (holo-transcobalamin) in true evaluation of vitamin B12 deficiency: Relevance in current perspective. BMC Res Notes. 2022;15(1):329.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGallina AL, Otay S, de Frutos-Lucas J, Buso M, Moral Martinez P, Cashman KD et al. Hidden hunger in Europe: a review on determinants, fragmented policy responses, and implementation barriers. Front Nutr. 2025;12.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVela D, Sharp A, Zhang R, Nguyen T, Hoang A, Pianykh OS. Temporal quality degradation in AI models. Sci Rep. 2022;12(1):11654.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Tables","content":"\u003cdiv id=\"Fig1\" class=\"Figure\"\u003e\n \u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eDescriptive Statistics and Effect Size (Cohen\u0026rsquo;s d) for All Variables\u003c/div\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eVariable\u003c/div\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eUnit\u003c/div\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eDevelopment Dataset Mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD\u003c/div\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eValidation Dataset Mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD\u003c/div\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003ep-value\u003c/div\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eCohen d (Effect)\u003c/div\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eAge\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eyears\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e49.43\u0026thinsp;\u0026plusmn;\u0026thinsp;16.66\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e41.42\u0026thinsp;\u0026plusmn;\u0026thinsp;24.99\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.377\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eB12\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003epg/mL\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e277.27\u0026thinsp;\u0026plusmn;\u0026thinsp;203.78\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e308.62\u0026thinsp;\u0026plusmn;\u0026thinsp;234.34\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e-0.143\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eFolate\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eng/mL\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e8.3\u0026thinsp;\u0026plusmn;\u0026thinsp;3.98\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e8.97\u0026thinsp;\u0026plusmn;\u0026thinsp;4.51\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.197\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eGlucose\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003emg/dL\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e99.71\u0026thinsp;\u0026plusmn;\u0026thinsp;38.00\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e99.65\u0026thinsp;\u0026plusmn;\u0026thinsp;45.98\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.85\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.001\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eALT\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eU/L\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e22.83\u0026thinsp;\u0026plusmn;\u0026thinsp;26.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e26.12\u0026thinsp;\u0026plusmn;\u0026thinsp;75.04\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e-0.059\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eAST\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eU/L\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e22.35\u0026thinsp;\u0026plusmn;\u0026thinsp;21.01\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e30.39\u0026thinsp;\u0026plusmn;\u0026thinsp;206.21\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e-0.055\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eFerritin\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eng/mL\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e236.00\u0026thinsp;\u0026plusmn;\u0026thinsp;987.00\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e99.99\u0026thinsp;\u0026plusmn;\u0026thinsp;204.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.191\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eALP\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eU/L\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e72.48\u0026thinsp;\u0026plusmn;\u0026thinsp;36.35\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e112.84\u0026thinsp;\u0026plusmn;\u0026thinsp;97.94\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e-0.546\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eGGT\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eU/L\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e26.41\u0026thinsp;\u0026plusmn;\u0026thinsp;40.01\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e33.10\u0026thinsp;\u0026plusmn;\u0026thinsp;81.44\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e-0.104\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eT. Bilirubin\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003emg/dL\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.69\u0026thinsp;\u0026plusmn;\u0026thinsp;0.34\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.78\u0026thinsp;\u0026plusmn;\u0026thinsp;0.93\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e-0.129\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eCreatinine\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003emg/dL\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.88\u0026thinsp;\u0026plusmn;\u0026thinsp;0.36\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.95\u0026thinsp;\u0026plusmn;\u0026thinsp;0.85\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e-0.114\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eC-RP\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003emg/L\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e6.55\u0026thinsp;\u0026plusmn;\u0026thinsp;16.10\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e10.92\u0026thinsp;\u0026plusmn;\u0026thinsp;29.11\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e-0.186\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eAlbumin\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eg/L\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e43.60\u0026thinsp;\u0026plusmn;\u0026thinsp;3.86\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e43.21\u0026thinsp;\u0026plusmn;\u0026thinsp;6.23\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.074\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eT. Protein\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eg/L\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e72.83\u0026thinsp;\u0026plusmn;\u0026thinsp;4.90\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e72.02\u0026thinsp;\u0026plusmn;\u0026thinsp;6.12\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.145\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eIron\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026micro;g/dL\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e176.51\u0026thinsp;\u0026plusmn;\u0026thinsp;148.63\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e176.32\u0026thinsp;\u0026plusmn;\u0026thinsp;153.23\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.87\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.001\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eLDH\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eU/L\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e176.47\u0026thinsp;\u0026plusmn;\u0026thinsp;61.25\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e208.16\u0026thinsp;\u0026plusmn;\u0026thinsp;215.55\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e-0.200\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eFT4\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eng/dL\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.89\u0026thinsp;\u0026plusmn;\u0026thinsp;0.17\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.92\u0026thinsp;\u0026plusmn;\u0026thinsp;0.20\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e-0.137\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eTSH\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026micro;IU/mL\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2.18\u0026thinsp;\u0026plusmn;\u0026thinsp;2.46\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2.29\u0026thinsp;\u0026plusmn;\u0026thinsp;2.58\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e-0.045\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eHGB\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eg/dL\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e13.43\u0026thinsp;\u0026plusmn;\u0026thinsp;1.67\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e12.84\u0026thinsp;\u0026plusmn;\u0026thinsp;1.91\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.331\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eRBC\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e10⁶/\u0026micro;L\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e4.69\u0026thinsp;\u0026plusmn;\u0026thinsp;0.53\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e4.61\u0026thinsp;\u0026plusmn;\u0026thinsp;0.64\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.133\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eHCT\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e%\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e40.19\u0026thinsp;\u0026plusmn;\u0026thinsp;4.44\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e38.75\u0026thinsp;\u0026plusmn;\u0026thinsp;5.18\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.298\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eMCV\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003efL\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e85.89\u0026thinsp;\u0026plusmn;\u0026thinsp;6.28\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e84.62\u0026thinsp;\u0026plusmn;\u0026thinsp;7.78\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.179\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eMCH\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003epg\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e28.76\u0026thinsp;\u0026plusmn;\u0026thinsp;2.60\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e28.17\u0026thinsp;\u0026plusmn;\u0026thinsp;3.07\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.208\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eMCHC\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eg/dL\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e33.48\u0026thinsp;\u0026plusmn;\u0026thinsp;1.24\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e33.32\u0026thinsp;\u0026plusmn;\u0026thinsp;1.34\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.123\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eRDW\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e%\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e14.03\u0026thinsp;\u0026plusmn;\u0026thinsp;1.80\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e14.41\u0026thinsp;\u0026plusmn;\u0026thinsp;2.52\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e-0.174\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eWBC\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e10\u0026sup3;/\u0026micro;L\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e7.26\u0026thinsp;\u0026plusmn;\u0026thinsp;2.25\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e8.20\u0026thinsp;\u0026plusmn;\u0026thinsp;6.79\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e-0.186\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eNEU\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e10\u0026sup3;/\u0026micro;L\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e5.50\u0026thinsp;\u0026plusmn;\u0026thinsp;1.44\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e4.07\u0026thinsp;\u0026plusmn;\u0026thinsp;3.37\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.552\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eLYM\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e10\u0026sup3;/\u0026micro;L\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e3.08\u0026thinsp;\u0026plusmn;\u0026thinsp;0.98\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2.46\u0026thinsp;\u0026plusmn;\u0026thinsp;5.43\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.159\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eMONO\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e10\u0026sup3;/\u0026micro;L\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.79\u0026thinsp;\u0026plusmn;\u0026thinsp;0.23\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.64\u0026thinsp;\u0026plusmn;\u0026thinsp;0.49\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.382\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eBASO\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e10\u0026sup3;/\u0026micro;L\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.64\u0026thinsp;\u0026plusmn;\u0026thinsp;0.46\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.05\u0026thinsp;\u0026plusmn;\u0026thinsp;0.21\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e1.641\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003ePLT\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e10\u0026sup3;/\u0026micro;L\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e269.78\u0026thinsp;\u0026plusmn;\u0026thinsp;75.84\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e284.84\u0026thinsp;\u0026plusmn;\u0026thinsp;106.67\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e-0.163\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eMPV\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003efL\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e9.49\u0026thinsp;\u0026plusmn;\u0026thinsp;1.27\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e9.73\u0026thinsp;\u0026plusmn;\u0026thinsp;1.33\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e-0.183\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003ePDW\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003efL\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e15.27\u0026thinsp;\u0026plusmn;\u0026thinsp;2.54\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e14.22\u0026thinsp;\u0026plusmn;\u0026thinsp;3.00\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e\u0026lt;\u0026thinsp;0.05\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.376\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eSD: standart deviation\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n \u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n \u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cdiv class=\"SimplePara\"\u003ePerformance comparison of eight machine-learning models under the F1-max threshold (experiment 2)\u003c/div\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eMetrics / Model\u003c/div\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eCat\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003eBoost\u003c/div\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eXGB\u003c/div\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eANN\u003c/div\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eRF\u003c/div\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eLR\u003c/div\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eSVM\u003c/div\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eKNN\u003c/div\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eDT\u003c/div\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eCutOff\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0,402\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0,378\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0,372\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0,410\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0,396\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0,306\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0,400\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0,500\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eTP\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e3233\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e3265\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e3268\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e3184\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e3098\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e3216\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e3297\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2468\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eFP\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e1160\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e1198\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e1325\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e1197\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e1452\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2006\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2473\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e994\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eTN\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2368\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2330\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2203\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2331\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2076\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e1522\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e1055\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2534\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eFN\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e295\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e263\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e260\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e344\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e430\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e312\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e231\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e1060\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eSensitivity\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.92\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(90.6-92.53)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.92\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(91.6\u0026ndash;93.4)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.93\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(91.7\u0026ndash;93.4)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.90\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(89.2\u0026ndash;91.2)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.87\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(86.7\u0026ndash;88.9)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.91\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(90.2\u0026ndash;92.0)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.93\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(92.6\u0026ndash;94.2)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.70\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(68.4\u0026ndash;71.4)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eSpecificity\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.67\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(65.5\u0026ndash;68.7)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.66\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(64.5\u0026ndash;67.6)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.62\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(60.8\u0026ndash;64.0)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.66\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(64.5\u0026ndash;67.6)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.59\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(57.2\u0026ndash;60.5)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.43\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(41.5\u0026ndash;44.8)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.30\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(28.4\u0026ndash;31.4)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.72\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(70.3\u0026ndash;73.3)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003ePPV\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.73\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(72.7\u0026ndash;74.5)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.73\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(71.8\u0026ndash;74.4)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.71\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(69.8\u0026ndash;72.4)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.73\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(71.3\u0026ndash;74.0)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.68\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(66.7\u0026ndash;69.4)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.61\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(60.3\u0026ndash;62.9)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.57\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(55.9\u0026ndash;58.4)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.71\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(69.8\u0026ndash;72.8)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eNPV\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.88\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(87.8\u0026ndash;89.0)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.89\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(88.6\u0026ndash;91.0)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.89\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(88.2\u0026ndash;90.6)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.87\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(85.8\u0026ndash;88.4)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.82\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(81.3\u0026ndash;84.3)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.83\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(81.2\u0026ndash;84.6)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.82\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(79.8\u0026ndash;84.0)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.70\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(69.0\u0026ndash;72.0)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003ePLR\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2.79\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(2.66\u0026ndash;2.92)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2.73\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(2.60\u0026ndash;2.86)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2.47\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(2.36\u0026ndash;2.58)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2.66\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(2.54\u0026ndash;2.79)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2.13\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(2.05\u0026ndash;2.22)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e1.60\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(1.55\u0026ndash;1.65)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e1.33\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(1.30\u0026ndash;1.36)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2.48\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(2.35\u0026ndash;2.63)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eNLR\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.12\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(3.13\u0026ndash;3.52)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.11\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.10\u0026ndash;0.13)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.12\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.10\u0026ndash;0.13)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.15\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.13\u0026ndash;0.16)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.21\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.19\u0026ndash;0.23)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.20\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.18\u0026ndash;0.23)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.22\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.19\u0026ndash;0.25)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.42\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.40\u0026ndash;0.44)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eF1\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.82\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.81\u0026ndash;0.83)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.82\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.81\u0026ndash;0.83)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.80\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.80\u0026ndash;0.81)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.81\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.80\u0026ndash;0.81)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.77\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.76\u0026ndash;0.78)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.74\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.72\u0026ndash;0.75)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.71\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.70\u0026ndash;0.72)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.71\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.69\u0026ndash;0.72)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eAccuracy\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.79\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(78.4\u0026ndash;80.3)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.79\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(78.3\u0026ndash;80.2)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.78\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(76.5\u0026ndash;78.5)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.78\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(77.2\u0026ndash;79.1)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.73\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(72.3\u0026ndash;74.3)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.67\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(66.0-68.2)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.62\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(60.5\u0026ndash;62.8)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.71\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(69.8\u0026ndash;71.9)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eA_ROC\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.88\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.86\u0026ndash;0.90)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.88\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.87\u0026ndash;0.89)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.86\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.85\u0026ndash;0.87)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.86\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.85\u0026ndash;0.87)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.81\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.79\u0026ndash;0.82)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.77\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.76\u0026ndash;0.78)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.72\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.70\u0026ndash;0.73)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.71\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.70\u0026ndash;0.72)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eA_PR\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.86\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.84\u0026ndash;0.88)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.85\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.84\u0026ndash;0.86)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.84\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.82\u0026ndash;0.85)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.82\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.80\u0026ndash;0.83)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.78\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.76\u0026ndash;0.79)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.75\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.74\u0026ndash;0.77)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.66\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.65\u0026ndash;0.68)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.65\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.63\u0026ndash;0.66)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eMCC\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.61\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.58\u0026ndash;0.62)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.61\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.58\u0026ndash;0.62)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.58\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.55\u0026ndash;0.59)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.58\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.56\u0026ndash;0.6)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.49\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.47\u0026ndash;0.50)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.39\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.37\u0026ndash;0.41)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.30\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.28\u0026ndash;0.32)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.42\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.39\u0026ndash;0.43)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eCatBoost: Categorical Gradient Boosting; XGB: Extreme Gradient Boosting (XGBoost); ANN: Artificial Neural Network; RF: Random Forest; LR: Logistic Regression; SVM: Support Vector Machine; KNN: k-Nearest Neighbors; DT: Decision Tree; TP: true positive; FP: false positive; TN: true negative; FN: false negative; PPV: positive predictive value; NPV: negative predictive value; PLR: positive likelihood ratio; NLR: negative likelihood ratio; A_ROC: area under the ROC curve; A_PR: area under the precision\u0026ndash;recall curve; MCC: Matthews correlation coefficient.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tab3\" style=\"width: 655px;\" border=\"1\"\u003e\n \u003ccaption\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cdiv class=\"SimplePara\"\u003ePerformance Metrics of the CatBoost Model Across All Experimental Configurations\u003c/div\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth style=\"width: 79.2222px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003ePerformance\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003eMetrics\u003c/div\u003e\n \u003c/th\u003e\n \u003cth style=\"width: 122.778px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eExperiment 1\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(default parameter)\u003c/div\u003e\n \u003c/th\u003e\n \u003cth style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eExperiment 2\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(cut-off optimization)\u003c/div\u003e\n \u003c/th\u003e\n \u003cth style=\"width: 155px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eExperiment 3\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(hyperparameter tuning)\u003c/div\u003e\n \u003c/th\u003e\n \u003cth style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eExperiment 4\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(feature engineering)\u003c/div\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 79.2222px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eTP\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122.778px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2993\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e3233\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 155px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e3028\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e3014\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 79.2222px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eFP\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122.778px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e895\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e1160\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 155px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e924\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e908\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 79.2222px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eTN\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122.778px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2633\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2368\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 155px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2604\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2620\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 79.2222px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eFN\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122.778px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e535\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e295\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 155px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e500\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e514\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 79.2222px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eSensitivity\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122.778px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.85\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.84\u0026ndash;0.86)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.92\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.91\u0026ndash;0.93)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 155px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.86\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.85\u0026ndash;0.87)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.85\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.84\u0026ndash;0.86)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 79.2222px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eSpecificity\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122.778px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.75\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.73\u0026ndash;0.76)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.67\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.65\u0026ndash;0.69)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 155px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.74\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.73\u0026ndash;0.75)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.74\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.73\u0026ndash;0.75)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 79.2222px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003ePPV\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122.778px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.77\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.76\u0026ndash;0.78)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.73\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.72\u0026ndash;0.75)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 155px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.77\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.76\u0026ndash;0.78)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.77\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.76\u0026ndash;0.78)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 79.2222px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eNPV\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122.778px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.83\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.82\u0026ndash;0.84)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.88\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.87\u0026ndash;0.89)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 155px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.84\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.83\u0026ndash;0.85)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.84\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.83\u0026ndash;0.85)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 79.2222px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003ePLR\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122.778px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e3.34\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(3.15\u0026ndash;3.55)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2.79\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(2.65\u0026ndash;2.94)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 155px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e3.28\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(3.12\u0026ndash;3.44)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e3.32\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(3.24\u0026ndash;3.40)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 79.2222px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eNLR\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122.778px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.20\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.19\u0026ndash;0.22)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.12\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.10\u0026ndash;0.14)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 155px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.19\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.18\u0026ndash;0.20)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.20\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.20\u0026ndash;0.20)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 79.2222px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eAccuracy\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122.778px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.81\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.80\u0026ndash;0.82)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.79\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.78\u0026ndash;0.80)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 155px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.80\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.79\u0026ndash;0.81)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.80\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.79\u0026ndash;0.81)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 79.2222px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eF1\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122.778px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.80\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.79\u0026ndash;0.81)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.82\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.81\u0026ndash;0.83)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 155px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.81\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.80\u0026ndash;0.82)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.81\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.80\u0026ndash;0.82)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 79.2222px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eAUC-ROC\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122.778px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.88\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.87\u0026ndash;0.89)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.88\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.87\u0026ndash;0.89)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 155px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.88\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.87\u0026ndash;0.89)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.86\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.85\u0026ndash;0.87)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 79.2222px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eAUC-PR\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122.778px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.86\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.85\u0026ndash;0.87)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.86\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.85\u0026ndash;0.87)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 155px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.86\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.85\u0026ndash;0.87)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.88\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.87\u0026ndash;0.89)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 79.2222px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eMCC\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 122.778px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.60\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.58\u0026ndash;0.62)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.61\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.59\u0026ndash;0.63)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 155px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.60\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.58\u0026ndash;0.62)\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 134px;\" align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.60\u003c/div\u003e\n \u003cdiv class=\"SimplePara\"\u003e(0.59\u0026ndash;0.61)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eTP: true positive; FP: false positive; TN: true negative; FN: false negative; PPV: positive predictive value; NPV: negative predictive value; PLR: positive likelihood ratio; NLR: negative likelihood ratio; AUC-ROC: area under the ROC curve; AUC-PR: area under the precision\u0026ndash;recall curve; MCC: Matthews correlation coefficient.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n \u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n \u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n \u003ctable id=\"Tab4\" border=\"1\"\u003e\n \u003ccaption\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cdiv class=\"SimplePara\"\u003ePerformance of the Validation Set (Final CatBoost Model)\u003c/div\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003ePerformance Metric\u003c/div\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eValidation Set\u003c/div\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eTP\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e15018\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eFP\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e3910\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eTN\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e1320\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eFN\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e2576\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eSensitivity\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.85 (0.84\u0026ndash;0.86)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eSpecificity\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.77 (0.75\u0026ndash;0.79)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003ePPV\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.79 (0.78\u0026ndash;0.80)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eNPV\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.83 (0.82\u0026ndash;0.84)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003ePLR\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e3.74 (3.61\u0026ndash;3.87)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eNLR\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.19 (0.18\u0026ndash;0.20)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eAccuracy\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.81 (0.80\u0026ndash;0.82)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eF1 Score\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.82 (0.81\u0026ndash;0.83)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eAUC-ROC\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.90 (0.89\u0026ndash;0.91)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eAUC-PR\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.91 (0.90\u0026ndash;0.92)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003eMCC\u003c/div\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cdiv class=\"SimplePara\"\u003e0.63 (0.61\u0026ndash;0.65)\u003c/div\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eTP: true positive; FP: false positive; TN: true negative; FN: false negative; PPV: positive predictive value; NPV: negative predictive value; PLR: positive likelihood ratio; NLR: negative likelihood ratio; AUC-ROC: area under the ROC curve; AUC-PR: area under the precision\u0026ndash;recall curve; MCC: Matthews correlation coefficient.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"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":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-8176315/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8176315/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground:\u003c/h2\u003e \u003cp\u003eVitamin B12 deficiency is a common yet frequently underdiagnosed condition due to the limited diagnostic accuracy of serum total B12 and restricted availability of confirmatory biomarkers such as holotranscobalamin and methylmalonic acid. Advances in machine learning (ML) and large-scale laboratory datasets provide new opportunities to leverage routinely collected biochemical and hematological parameters for early detection. This study aimed to develop, optimize, and validate explainable ML models to predict vitamin B12 deficiency using standard laboratory analytes obtained during routine outpatient care.\u003c/p\u003e\u003ch2\u003eMethods:\u003c/h2\u003e \u003cp\u003eThis retrospective study included 51,630 adult patients from 2015\u0026ndash;2025, with an independent temporal validation cohort of 34,744 individuals. Eight supervised ML algorithms\u0026mdash;logistic regression, random forest, decision tree, SVM, KNN, XGBoost, CatBoost, and artificial neural networks\u0026mdash;were trained within a four-stage experimental framework incorporating default modeling, threshold optimization, hyperparameter tuning, and feature engineering. Performance was assessed using AUC-ROC, AUC-PR, sensitivity, specificity, F1-score, PPV, NPV, accuracy, MCC, and likelihood ratios. Statistical comparisons included DeLong, paired t-tests, McNemar, NRI, and IDI analyses. Model interpretability was evaluated using SHAP, LIME, and Decision Curve Analysis.\u003c/p\u003e\u003ch2\u003eResults:\u003c/h2\u003e \u003cp\u003eAcross all experiments, CatBoost achieved the most balanced performance, with the F1-maximization threshold-optimized configuration demonstrating the lowest false-negative rate. In the test set, CatBoost yielded sensitivity 0.92, specificity 0.67, F1 0.82, AUC-ROC 0.88, and AUC-PR 0.86. Temporal validation confirmed robust generalizability (sensitivity 0.85, specificity 0.77, AUC-ROC 0.90, AUC-PR 0.91, MCC 0.63). SHAP and LIME consistently identified MCV, HGB, HCT, RBC, RDW, iron, ferritin, CRP, folate, and age as key contributors. DCA demonstrated superior net clinical benefit across a wide threshold range.\u003c/p\u003e\u003ch2\u003eConclusion:\u003c/h2\u003e \u003cp\u003eThis study presents the first large-scale, explainable, and clinically validated ML model capable of predicting vitamin B12 deficiency using only routine laboratory parameters. The model exhibits strong discrimination, reliability under temporal shift, and biologically meaningful interpretability, supporting its potential integration into clinical decision-support systems for early detection and optimized laboratory workflows.\u003c/p\u003e","manuscriptTitle":"Development of Machine Learning Algorithms for Predicting Vitamin B12 Levels Using Biochemical Analyte Data","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-02 05:56:10","doi":"10.21203/rs.3.rs-8176315/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a7a396ef-97f5-4c75-a0a8-9a06e3553fa0","owner":[],"postedDate":"January 2nd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-02-25T09:57:09+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-02 05:56:10","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8176315","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8176315","identity":"rs-8176315","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}