Development and External Validation of a Machine Learning–Based Model for Early Prediction of Multiple Organ Dysfunction Syndrome in Critically Ill Patients with Sepsis

Development and External Validation of a Machine Learning–Based Model for Early Prediction of Multiple Organ Dysfunction Syndrome in Critically Ill Patients with Sepsis | 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 and External Validation of a Machine Learning–Based Model for Early Prediction of Multiple Organ Dysfunction Syndrome in Critically Ill Patients with Sepsis Jinbin Yang, Linying Cai, Xuyang Liu, Kaihuan Zhou, Junyu Lu, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8681490/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 13 You are reading this latest preprint version Abstract Background Multiple organ dysfunction syndrome (MODS) is a key determinant of prognosis in sepsis, yet conventional severity scoring systems based on linear assumptions and static variables fail to capture complex nonlinear physiological disturbances and dynamic inter organ interactions. Although machine learning has shown promise in outcome prediction among critically ill patients, studies focusing on MODS while ensuring interpretability and external validation remain limited. Methods This retrospective cohort study used data from the Medical Information Mart for Intensive Care IV and the eICU Collaborative Research Database. Adult patients meeting Sepsis 3 criteria and admitted to the ICU for the first time were included. Feature selection was performed using least absolute shrinkage and selection operator regression. Multiple machine learning models were developed, including logistic regression, random forest, gradient boosting machine, extreme gradient boosting, Light Gradient Boosting Machine, artificial neural networks, and support vector machines. Model performance was evaluated using the area under the receiver operating characteristic curve, calibration curves, and decision curve analysis. Shapley additive explanations were used for model interpretation, and external validation was conducted in an independent eICU cohort. Results Among 23,018 patients with sepsis, 4,931 (21.4%) developed MODS during ICU hospitalisation. All models showed acceptable discrimination, with LightGBM achieving the highest AUC (0.829), followed by GBM (0.824), random forest (0.823), and XGBoost (0.822). Logistic regression and elastic net showed moderate performance (both AUC 0.802), the neural network showed intermediate discrimination (AUC 0.803), whereas support vector machines (0.759) and k nearest neighbours (0.727) performed less well. LightGBM demonstrated stable discrimination, good calibration, and greater clinical net benefit in both internal testing and external validation. SHAP analysis identified the Sequential Organ Failure Assessment score, respiratory rate, lactate, coagulation indices including international normalised ratio, acid base status, and vasoactive agent use as key predictors with pronounced nonlinear effects. Conclusion Among the evaluated models, the gradient boosting based LightGBM showed the most robust performance for predicting MODS risk in sepsis, supporting early risk stratification and individualised ICU management. Prospective multicentre studies are warranted to confirm its clinical impact. MODS Sepsis Machine learning Risk prediction External validation Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 1 Introduction Sepsis is defined as life-threatening organ dysfunction caused by a dysregulated host response to infection [ 1 , 2 ] , and remains one of the leading causes of mortality and healthcare resource utilisation in intensive care units (ICUs) worldwide [ 3 – 5 ] . Although advances in infection control, organ support, and evidence-based therapeutic strategies have improved outcomes to some extent in recent years, sepsis-associated mortality remains unacceptably high. This is particularly evident when the disease progresses to multiple organ dysfunction syndrome (MODS), a stage at which the risk of death increases sharply [ 6 ] . MODS represents an advanced and severe stage in the clinical course of sepsis, characterised by the progressive accumulation and mutual amplification of dysfunction across multiple organ systems. Its development is not only closely associated with increased short-term mortality but also exerts a substantial adverse impact on long-term survival and functional outcomes [ 7 , 8 ] . Epidemiological studies have demonstrated that MODS is highly prevalent among patients with sepsis admitted to the ICU and is frequently accompanied by prolonged hospitalisation, increased requirements for organ support, and a markedly elevated risk of death [ 5 , 6 ] . Consequently, accurate identification of patients at high risk of developing MODS at an early stage of disease, thereby enabling timely, targeted monitoring and intervention, remains a critical and unresolved clinical challenge in contemporary critical care medicine. Currently, risk assessment in clinical practice relies primarily on severity scoring systems such as the Sequential Organ Failure Assessment (SOFA), Simplified Acute Physiology Score II (SAPS II), and Acute Physiology and Chronic Health Evaluation (APACHE) scores [ 9 – 11 ] . Although these tools provide a certain degree of prognostic value at the population level, they are inherently based on linear assumptions and static combinations of variables, which limits their ability to fully capture the complex non-linear physiological disturbances and dynamic inter-organ interactions observed in patients with sepsis [ 12 ] . Moreover, these scoring systems were originally designed to quantify overall disease severity rather than to generate individualised risk predictions for specific outcomes such as MODS,thereby constraining their utility for precise clinical decision support. With the rapid expansion of electronic health records and large-scale critical care databases, the application of machine learning (ML) approaches to outcome prediction in critical care has attracted increasing attention [ 13 ] . Compared with traditional statistical models, ML algorithms are capable of automatically learning non-linear relationships and complex interaction effects among variables in high-dimensional data settings, thereby demonstrating potential advantages across a range of ICU outcome prediction tasks [ 12 , 14 – 16 ] . Previous studies have shown that ensemble-based and gradient boosting models achieve favourable discriminative performance in predicting sepsis-related mortality, organ failure, and associated complications [ 17 , 18 ] . However, most existing machine-learning studies have primarily focused on sepsis identification or mortality risk prediction [ 12 , 17 , 19 , 19 – 22 ] . In many of these investigations, organ dysfunction has been treated as an intermediate variable or incorporated as a component of composite scores, rather than being modelled directly as MODS, a highly heterogeneous outcome characterised by dynamic temporal evolution. Consequently, model-based studies that explicitly target MODS as the primary endpoint remain relatively scarce. In addition, among published ML studies, relatively few have simultaneously addressed model discrimination, interpretability, and rigorous external validation, which to some extent limits the clinical credibility and broader applicability of their findings. Building on these gaps in the existing literature, the present study leveraged the Medical Information Mart for Intensive Care IV (MIMIC-IV) database to systematically develop and compare multiple machine-learning models for predicting the risk of MODS in patients with sepsis during ICU hospitalisation. Through regularised feature selection, comprehensive multi-model performance evaluation, decision curve analysis, and Shapley additive explanations (SHAP)–based interpretability analyses, followed by external validation in an independent the eICU Collaborative Research Database (eICU) cohort, we aimed to develop a MODS risk prediction model that integrates strong predictive performance, interpretability, and clinical applicability. This data-driven approach seeks to support early risk stratification and individualised management of patients with sepsis. 2 Methods 2.1 Data source The data for this study were obtained from the MIMIC-IV (version 3.0) and eICU [ 23 ] , The MIMIC-IV database is developed and maintained by the Laboratory for Computational Physiology at the Massachusetts Institute of Technology and contains detailed clinical information on 546,028 hospitalised patients admitted to the Beth Israel Deaconess Medical Center in the United States between 2008 and 2022. The eICU database is a multicentre critical care repository that includes de-identified clinical data from adult ICU patients treated at more than 200 hospitals across the United States between 2014 and 2015. Prior to data access, all members of the research team completed the required training provided by the US National Institutes of Health and passed the examination on Protecting Human Research Participants (Record ID: 71642466). All patient data were fully de-identified to protect personal privacy; therefore, this study did not require additional informed consent or approval from an institutional review board. Eligibility Criteria Inclusion criteria were as follows: (1) patients with sepsis diagnosed according to the Sepsis-3 criteria; (2) age ≥ 18 years; and (3) first admission to ICU. Exclusion criteria included: (1) ICU length of stay < 24 h; (2) presence of MODS at the time of ICU admission; (3) severe pre-existing comorbidities, including advanced malignancy, end-stage renal disease, or advanced acquired immunodeficiency syndrome; and (4) prior long-term organ replacement therapy, such as chronic dialysis or prolonged mechanical ventilation. Outcome Definition MODS was defined as the first occurrence of dysfunction in at least two organ systems, each with a Sequential Organ Failure Assessment (SOFA) score ≥ 2, occurring 24 h after ICU admission. The organ systems assessed included the respiratory, cardiovascular, coagulation, hepatic, renal, and central nervous systems. 2.2 Data extraction Data on patients with sepsis who met the predefined inclusion and exclusion criteria were extracted from the MIMIC-IV (version 3.0) and eICU databases. The extracted variables included demographic characteristics, baseline comorbidities, vital sign parameters, laboratory measurements, disease severity scores, such as the SOFA score and the Glasgow Coma Scale (GCS) score, organ support therapies (mechanical ventilation, use of vasoactive agents, and continuous renal replacement therapy [CRRT]), as well as other relevant clinical information. For each variable, the first available value within the first 24 h after ICU admission was used, and for patients with multiple hospitalisations, only data from the first hospital admission were retained for analysis. 2.3 Missing Data Imputation and Outlier Management Strategies To minimise potential bias arising from missing data, variables with more than 20% missing values were excluded from the final cohort. For variables with a missingness rate between 5% and 20%, multiple imputation by chained equations (MICE) was applied [ 24 ] , generating m = 5 imputed datasets. To prevent information leakage, the imputation procedures were conducted independently in the training and test sets. Continuous variables were imputed using predictive mean matching, whereas categorical variables were imputed using logistic regression models. Final parameter estimates were pooled across the imputed datasets according to Rubin’s rules. In addition, during data preprocessing, outliers in continuous variables, excluding selected variables such as age, SOFA score, and GCS score, were handled using the interquartile range (IQR) method. Specifically, the first (Q1) and third (Q3) quartiles were calculated for each variable, and the IQR was defined as Q3 − Q1. Data points exceeding the upper bound (Q3 + 1.5 × IQR) or falling below the lower bound (Q1 − 1.5 × IQR) were considered outliers. After being identified, these outliers were imputed using regression-based methods to preserve the accuracy and robustness of subsequent analyses. 2.4 Statistical analysis 2.4.1 Baseline Characteristics Normality of continuous variables in the baseline characteristics was assessed using the Kolmogorov–Smirnov test. Continuous variables with a normal distribution are presented as mean ± standard deviation (SD), whereas non-normally distributed variables are expressed as median (interquartile range, IQR). Categorical variables are reported as counts and percentages. Comparisons of normally distributed continuous variables were performed using the t test or one-way analysis of variance (ANOVA), as appropriate, while categorical variables were compared using the Pearson χ² test or Fisher’s exact test. 2.4.2 Model Development and Validation Using sepsis patient data from the MIMIC-IV (version 3.0) database, the cohort was randomly divided into a training set and a test set at a ratio of 70:30 to ensure model generalisability and predictive accuracy. Prior to model construction, least absolute shrinkage and selection operator (LASSO) regression was applied to perform feature selection among candidate predictors in order to reduce model redundancy and enhance prediction stability [ 25 ] . By introducing an L1 regularisation penalty into the regression framework, LASSO retains the most informative variables associated with the outcome while shrinking coefficients of less contributory variables towards zero, thereby enabling variable selection and reducing the risk of overfitting. The optimal penalty parameter (λ) was determined through cross-validation, and only variables with non-zero regression coefficients were retained for subsequent multi-model predictive analyses. This feature selection procedure ensured a consistent predictor space across different modelling frameworks, thereby facilitating fair performance comparisons and improving interpretability of the results. To comprehensively evaluate predictive performance, multiple machine-learning algorithms were implemented, including decision trees, k-nearest neighbours (kNN), logistic regression, random forest, artificial neural networks, gradient boosting machine (GBM), extreme gradient boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and support vector machines (SVM). To mitigate overfitting, 10-fold cross-validation was applied, whereby 90% of the data were used for training and the remaining 10% for validation in each iteration, repeated across 10 folds. Model performance was assessed using accuracy, precision, recall, F1 score, and the area under the receiver operating characteristic (ROC) curve (AUC). To ensure model interpretability, we not only constructed models using features selected by LASSO regression but also applied SHAP [ 26 ] to the optimal LightGBM model. This approach enabled quantification of the specific contribution of each feature to the model’s predictions. Comparisons of model performance were conducted using the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. All statistical analyses were performed using IBM SPSS Statistics version 26.0 and R version 4.5. Two-sided P values < 0.05 were considered statistically significant. 2.4.3 External Validation To evaluate the generalisability and robustness of the predictive models across different clinical settings, external validation was performed using an independent cohort from the eICU Collaborative Research Database. This multicentre ICU database comprises critically ill patients from diverse healthcare institutions and was not involved in feature selection, model training, or parameter tuning during model development. Using the same variable definitions, data preprocessing procedures, and outcome criteria as those applied in the internal analyses, the final model was directly applied to the eICU cohort. Model discrimination was assessed using the AUC. Calibration performance was evaluated with calibration curves by comparing predicted risks with observed outcome frequencies, and the results were contrasted with those from internal validation to assess model stability. 3 Results 3.1 Baseline Characteristics A total of 23,018 patients with sepsis were included in this study (Fig. 1 ), among whom 4,931 (21.4%) developed MODS during their ICU stay. Compared with patients who did not develop MODS, those in the MODS group exhibited overall greater disease severity and more pronounced features of multi-organ dysfunction. With respect to demographic characteristics, patients in the MODS group were slightly older and had a lower proportion of males (both P < 0.001). Regarding comorbidities, chronic kidney disease, heart failure, and chronic obstructive pulmonary disease were more prevalent in the MODS group, whereas the distribution of diabetes mellitus was comparable between the two groups. In terms of physiological and laboratory parameters, patients who developed MODS exhibited more pronounced circulatory, respiratory, and metabolic derangements, including higher heart rate and respiratory rate, lower mean arterial pressure and oxygen saturation, as well as more severe renal dysfunction and metabolic acidosis (all P < 0.001). In addition, patients in the MODS group had significantly SOFA scores, Oxford Acute Severity of Illness Score (OASIS), and Simplified Acute Physiology Score II (SAPS II), and were more likely to receive vasoactive agents, mechanical ventilation, and renal replacement therapy (all P < 0.001). 3.2 Feature Selection Using LASSO Prior to model development, LASSO logistic regression was applied to the candidate predictors to reduce model complexity and mitigate the risks of multicollinearity and overfitting. As shown in Fig. 2 , the optimal regularisation parameter (λ) was determined using 10-fold cross-validation. The model achieved the minimum cross-validated error at λ_min, whereas at λ_1se the predictive performance remained stable with a further reduction in model complexity. Balancing robustness and generalisability, variables with non-zero regression coefficients at the selected optimal λ were retained for subsequent model development. The final set of selected features encompassed demographic characteristics, baseline comorbidities, vital signs, laboratory parameters, disease severity scores, and variables related to organ support therapies, collectively reflecting the multidimensional clinical profile associated with the development of MODS in patients with sepsis. 3.3 Model Performance To compare the predictive performance of different machine-learning models for the development of MODS in patients with sepsis, nine models were developed and evaluated in the training cohort (Fig. 3 ). The ROC curves of all models were clearly above the diagonal reference line, indicating that each model demonstrated a certain degree of discriminative ability. In terms of discriminative performance, ensemble-based tree models consistently outperformed the other approaches. LightGBM achieved the highest discrimination (AUC = 0.829), followed by the GBM(AUC = 0.824), random forest (AUC = 0.823), and XGBoost (AUC = 0.822). In contrast, the traditional linear logistic regression model (AUC = 0.802) and its regularised counterpart, elastic net (AUC = 0.802), demonstrated comparatively lower predictive performance. SVM (;AUC = 0.759) and kNN ( AUC = 0.727) exhibited weaker discrimination, whereas the neural network model showed intermediate performance between tree-based and linear models (AUC = 0.803). In the test cohort, calibration curves revealed notable differences in calibration performance across models (Fig. 4 A). Overall, tree-based ensemble models, including LightGBM, XGBoost, and random forest, showed good agreement with the ideal reference line across most ranges of predicted risk, indicating relatively stable probability estimation. In contrast, logistic regression tended to underestimate risk in the intermediate- to high-risk probability ranges. Based on these comparative results, LightGBM was selected as the final model and its calibration performance was further evaluated separately (Fig. 4 B). The calibration curve for LightGBM demonstrated close alignment with the ideal reference line across the overall risk spectrum, with only minor deviations observed in the extreme high-risk range. These findings indicate that LightGBM not only provides strong discriminative ability but also yields accurate individual-level estimates of MODS risk. 3.4 Decision Curve Analysis Decision curve analysis (DCA) was performed to evaluate the clinical net benefit of different models across a range of threshold probabilities (Fig. 5 ). In the test cohort, all predictive models demonstrated greater net benefit across a broad range of threshold probabilities compared with the “treat-all” and “treat-none” strategies, indicating that model-based risk stratification can provide tangible clinical value for decision-making. The LightGBM model consistently achieved the highest and most stable net benefit across the majority of threshold probabilities. In particular, within the low-to-moderate threshold range (approximately 10%–60%), LightGBM substantially outperformed the traditional logistic regression model. As the threshold probability increased, the net benefit of logistic regression gradually approached or even fell below zero, suggesting limited clinical utility at higher risk thresholds. In contrast, gradient boosting– and ensemble-based models continued to demonstrate positive net benefit across these ranges. 3.5 Feature Importance Analysis To assess the stability and consistency of feature importance across different machine-learning models, we compared the relative contributions of key predictors derived from multiple models, including XGBoost, random forest, LightGBM, GBM, elastic net, and logistic regression (Fig. 6 ). To enhance transparency of the internal decision-making processes of the models, we present the feature importance rankings for each machine-learning algorithm in the Supplementary Materials (Supplementary Figures S1 A–S1F). Across models, the Sequential Organ Failure Assessment (SOFA) score, respiratory rate (RR), lactate concentration, coagulation-related indices (international normalised ratio, INR), and indicators of metabolic acid–base status (pH and bicarbonate) consistently ranked among the most important features in the majority of algorithms. This cross-model concordance indicates that these variables provide robust and consistent predictive value for assessing the risk of MODS. In contrast, variables such as age, the GCS score, and mechanical ventilation showed greater variability in importance rankings across different models, reflecting differences in model architecture with respect to capturing feature interactions and non-linear relationships. Overall, this cross-model consistency analysis supports the stability of the selected features across different algorithmic frameworks, thereby strengthening the credibility of the model’s interpretability. 3.6 SHAP Analysis of the LightGBM Model To enhance model interpretability and elucidate the non-linear effects of key predictive variables, we further applied SHAP to the optimal model (LightGBM) (Fig. 7 ). Global SHAP feature importance analysis indicated that the SOFA score was the most influential predictor of MODS risk, exhibiting the widest distribution of SHAP values. This was followed by RR, INR, bicarbonate concentration, and use of vasoactive agents. Lactate, platelet count, creatinine, age, GCS score, mechanical ventilation, and pH also demonstrated consistent, albeit relatively secondary, contributions to risk prediction. SHAP dependence plots further revealed pronounced non-linear relationships between key variables and the risk of MODS. Lactate contributed minimally to risk at lower concentrations; however, when levels increased beyond approximately 8–10 mmol/L, SHAP values rose sharply, indicating a clear threshold effect. Platelet count showed a non-monotonic association with SHAP values, reaching a peak within a moderate range before plateauing. In contrast, the SOFA score exhibited a characteristic non-linear dose–response relationship, with risk increasing markedly as the score rose and approaching saturation at very high levels. 3.7 External Validation To evaluate model generalisability, the final model was externally validated in an independent cohort from the eICU Collaborative Research Database. The results showed that the LightGBM model maintained stable discriminative performance in the external cohort, with ROC curve characteristics comparable to those observed in the internal validation (Fig. 8 A). In terms of calibration, the predicted risks across different probability strata were generally consistent with the observed incidence of MODS, with no evident systematic overestimation or underestimation (Fig. 8 B). Additional results from model comparisons and clinical decision curve analyses are provided in the Supplementary Materials (Supplementary Figures S4A–S4E). Although overall model performance in the external validation cohort was modestly attenuated compared with the internal validation, the relative ranking among models remained consistent, indicating that the model exhibits acceptable stability across different ICU populations. 4 Discussion This study leveraged the large-scale, real-world MIMIC-IV database to systematically develop and validate multiple machine-learning models for predicting the risk of MODS in patients with sepsis during ICU hospitalisation. As one of the most clinically consequential outcomes of sepsis, MODS is highly prevalent in ICU populations and is associated with poor prognosis, remaining a critical and unresolved challenge in contemporary critical care medicine [5,6] . Our findings demonstrate that ensemble-based machine-learning models markedly outperform traditional linear models in terms of discriminative ability, calibration performance, and potential clinical decision-making value. Among these approaches, the LightGBM showed the most robust and consistent performance and was further validated in an independent external cohort. 4.1 Main Findings The primary objective of this study was to develop and validate a prediction model based on large-scale real-world data that can identify the risk of MODS in patients with sepsis at an early stage of ICU admission. In pursuit of this objective, several key findings emerged from our analyses. First, the multi-model prediction framework developed using the MIMIC-IV database effectively discriminated between patients with sepsis who did and did not develop MODS during ICU hospitalisation, indicating that multidimensional clinical information carries quantifiable predictive value for MODS risk identification. Compared with traditional logistic regression, gradient boosting– and ensemble-based models demonstrated superior overall discriminative performance and greater stability. Second, among the various machine-learning models evaluated, LightGBM consistently exhibited the most stable predictive performance across the training set, test set, and the independent external eICU cohort. This finding indicates that LightGBM possesses strong generalisability across different patient populations and clinical settings, supporting tree-based ensemble learning as a preferred modelling strategy for predicting sepsis-associated MODS risk. Third, SHAP-based interpretability analyses demonstrated that the key predictors identified by the model were predominantly related to the severity of organ dysfunction, circulatory and coagulation abnormalities, and metabolic derangements. The importance ranking of these variables closely aligns with established pathophysiological mechanisms underlying MODS. This concordance not only strengthens the credibility of the model’s predictions but also provides clinically interpretable insights into the model’s decision-making process. Overall, this study successfully developed and externally validated a MODS risk prediction model that integrates strong discriminative performance, robustness, and interpretability, thereby providing a data-driven foundation to support early risk stratification and precision management of patients with sepsis. 4.2 Pathophysiological Interpretation of Key Predictors At the level of model interpretation, the key predictors identified in this study demonstrated a high degree of consistency across different modelling frameworks, and their clinical relevance closely aligns with established pathophysiological concepts of MODS. Notably, the SOFA score consistently ranked as the most important feature in both feature importance analyses and SHAP interpretations, underscoring the central role of baseline organ dysfunction severity in the progression of MODS. Importantly, this finding does not simply replicate existing scoring systems; rather, it highlights that within a data-driven modelling context, SOFA remains a pivotal integrative variable that encapsulates multi-organ functional status and serves as a critical hub in MODS risk evolution [27–29] . In addition, the high importance assigned to respiratory rate, lactate concentration, and coagulation-related indices ( INR) highlights the synergistic dysregulation of the circulatory–respiratory–coagulation–metabolic axis in the development of MODS. The SHAP dependence pattern for lactate demonstrates clear non-linear and threshold effects, supporting its role not merely as a static marker of tissue hypoperfusion but also as an indicator of cumulative mitochondrial dysfunction and metabolic stress [30,31] . Abnormalities in coagulation parameters further suggest an amplifying role of sepsis-associated coagulopathy in MODS progression. The reciprocal activation of inflammatory and coagulation pathways is widely recognised as a key mechanism driving microcirculatory dysfunction and the progression of organ injury [32,33] . Notably, the key variables identified by the model do not operate in isolation but instead cluster into risk feature sets at the level of functional systems. This system-level aggregation of predictors reflects the intrinsic nature of MODS as a composite outcome and helps explain the limitations of traditional single-marker approaches or linear models in predicting this complex clinical endpoint. 4.3 Advantages of Machine-Learning Models Over Traditional Approaches From a methodological perspective, our findings indicate that machine-learning models offer clear advantages in predicting MODS, a highly heterogeneous and non-linearly driven clinical outcome. Compared with traditional logistic regression, gradient boosting– and ensemble-based models can autonomously learn complex non-linear relationships and higher-order interactions among variables without imposing a priori linear assumptions, thereby achieving superior predictive performance in high-dimensional, complex clinical datasets [13,34,35] . It is important to emphasise that this study did not merely compare models based on differences in AUC. Instead, model robustness and clinical applicability were comprehensively evaluated through a combination of internal validation, external validation, calibration analyses, and decision curve analysis across different risk strata and patient populations. The consistently strong performance of the LightGBM across these multiple evaluation dimensions supports its selection as a preferred algorithm for MODS risk prediction, rather than attributing its superiority to chance or isolated performance gains. At the same time, by integrating regularised regression with tree-based models, this study achieved a balance between predictive performance and model interpretability, thereby mitigating the trust barriers commonly associated with so-called “black-box” models in clinical practice. This strategy provides a generalisable methodological paradigm for machine-learning modelling of complex ICU outcomes. 4.4 Clinical Decision-Making Value and Potential Applications From a clinical application perspective, the prediction model developed in this study is primarily intended for early risk identification and stratification rather than for replacing clinical decision-making. Decision curve analysis demonstrated that, across a broad range of threshold probabilities, ensemble-based models provide greater clinical net benefit, with particularly clear advantages over traditional logistic regression in the low-to-moderate risk ranges. These findings suggest that the model has potential value in assisting clinicians to identify patients at high risk of developing MODS. Moreover, DCA offers a more intuitive representation of the potential benefits of predictive models within real-world clinical decision-making contexts [36,37] . In real-world ICU settings, this model could potentially be integrated as a component of a clinical decision support system, in conjunction with existing severity scores and bedside assessments, to enhance monitoring intensity for high-risk patients, optimise resource allocation, and support risk stratification in clinical trial design. However, the realisation of its clinical value depends on appropriate integration and use strategies rather than on reliance on model outputs in isolation. Therefore, the broader significance of this study lies in demonstrating that, by integrating multidimensional clinical data with interpretability analyses, machine-learning models can provide incremental information for MODS risk assessment without departing from established clinical reasoning. This approach lays a foundation for future prospective investigations and clinical translation. 4.5 Interpretability and Model Credibility The clinical adoption of machine-learning models depends not only on predictive performance but also critically on the interpretability of their decision-making processes and their acceptability to clinicians. In this study, SHAP-based analyses were applied to the LightGBM model to quantify the relative contribution of each predictor to MODS risk estimation at both the individual and population levels, while also revealing non-linear and threshold relationships between several key variables and the outcome [38,39] . This interpretability-oriented methodological design allows model predictions to be directly contrasted with established pathophysiological understanding, thereby avoiding a “cognitive disconnect” between algorithmic decisions and clinical experience. By illustrating the direction and magnitude of each variable’s effect across different risk levels, the model not only delivers risk stratification results but also provides transparent insights into the sources of risk, which may help enhance clinicians’ trust in the model outputs. Accordingly, interpretability analyses not only strengthen the alignment between model predictions and clinical pathophysiological understanding but also enhance the credibility and acceptability of the model in real-world clinical settings, thereby providing an essential prerequisite for clinical translation. 4.6 Challenges in the Clinical Translation of Machine-Learning Models Although machine-learning models have demonstrated considerable potential in risk prediction for critically ill patients [40] , their translation into routine clinical practice faces multiple challenges. These challenges extend beyond technical performance alone and encompass issues related to generalisability, interpretability, and integration into clinical workflows [41] . First, substantial performance heterogeneity exists across different models, and their generalisability and stability require systematic comparison and rigorous external validation [42] . Prior reviews have emphasised that strong performance in internal validation does not necessarily translate into comparable performance in independent external cohorts [43] . This limitation is particularly pronounced in ICU populations, which are highly heterogeneous; variations in healthcare settings, patient case mix, and data distributions across centres can markedly influence model performance [44] , thereby constraining stable implementation in real-world clinical environments. The lack of comprehensive external validation and cross-regional assessment means that the clinical reliability of many models remains to be fully established. Second, complex models are often regarded as “black boxes,” with internal decision logic that is difficult for clinicians to intuitively understand. This lack of transparency and interpretability substantially limits their credibility and acceptability in clinical decision-making [20,45] . These shortcomings in interpretability and general acceptance have prompted ongoing efforts to develop more interpretable model architectures and robust explanation methods to enhance clinical trust. Moreover, although machine-learning approaches have been applied to early sepsis detection and mortality risk prediction, studies specifically targeting MODS, a highly heterogeneous and dynamically evolving composite clinical outcome, remain relatively limited. At present, there is still a lack of systematic investigations based on large-scale real-world data that simultaneously address predictive performance, interpretability, and clinical decision-making value. Future research should further integrate temporal dynamics and clinically interpretable frameworks to better characterise the complex evolution of organ dysfunction. Overall, although machine-learning models show considerable promise for individual risk prediction, achieving reliable clinical translation requires overcoming challenges related to limited generalisability, insufficient interpretability, and narrow outcome definitions. Future studies should prioritise extensive cross-centre and cross-system validation, the development of more transparent model structures and explanation mechanisms, and methodological innovations that enable deeper integration with clinical decision-making workflows. 4.7 Study Limitations This study has several limitations. First, it employed a retrospective observational design. Although confounding was mitigated as much as possible through regularised feature selection and external validation, the influence of unmeasured or residual confounders cannot be fully excluded. Consequently, the model’s predictions should not be interpreted as evidence of causal relationships. Second, the definition of MODS relied on organ function indicators and scoring systems available within the databases, which may not fully capture the dynamic evolution of organ dysfunction during sepsis or the reversible changes following clinical interventions. This reliance on static or time-window–based measures may lead to underestimation or overestimation of true organ functional status in certain patients. Future studies could address this limitation by developing more precise, dynamically updated risk models. Third, although the model incorporated multidimensional clinical variables, limitations inherent to the database structure precluded inclusion of certain potentially important information, such as more granular temporal data, dynamic biomarker trajectories, and bedside decision-making factors. This constraint may limit the model’s ability to fully characterise complex pathophysiological processes. In addition, although external validation was performed using the independent eICU database, both datasets were derived from ICU populations in the United States. Therefore, the generalisability of the model to healthcare systems in other countries, settings with different levels of medical resources, and varying clinical practice patterns remains uncertain and warrants further confirmation through prospective, multicentre studies. Finally, the model developed in this study is primarily intended for risk prediction and stratification rather than for directly guiding specific therapeutic decisions. Its effectiveness within real-world clinical workflows, as well as its potential impact on improving patient outcomes, still requires evaluation through prospective clinical trials. 5 Conclusion This multicentre study systematically developed and validated multiple machine-learning models to predict the risk of MODS in patients with sepsis. The results demonstrated that the gradient boosting–based LightGBM achieved the most favourable and robust overall performance, with superior discrimination, good calibration, greater clinical net benefit, and consistent performance on external validation. By integrating multidimensional clinical information with interpretability analyses, the model was able to capture non-linear and threshold effects between key predictors and MODS risk, thereby providing data-driven support for early risk identification and stratification in sepsis. Importantly, the risk features identified by the model closely align with established pathophysiological understanding, enhancing its credibility in clinical application. Collectively, these findings suggest that the proposed model represents a potentially valuable tool for early risk stratification and individualised management of patients with sepsis in the ICU. Future studies should evaluate the performance of this model in real-world clinical settings through prospective, multicentre investigations and explore its potential impact on clinical decision-making processes and patient outcomes. Such efforts will be essential to facilitate the standardised and responsible integration of machine-learning models into the precision management of sepsis. Declarations Conflict of Interest : The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. Availability of data and materials: Publicly available data sets were analyzed in this study.These data can be found here: https://mimic.mit.edu/docs/iv/, https://physionet.org/content/eicu-crd/2.0/. Ethics approval and consent to participate ： The research involving human participants was approved by the Massachusetts Institute of Technology (Cambridge, MA) and Beth Israel Deaconess Medical Center (Boston, MA). Informed consent for participation was not required for this study by national legislation and institutional requirements. A researcher who has completed the Collaborative Institutional Training Initiative examination (Certification number 71642466 for author Yang) can access this database.The requirement for individual informed consent was waived due to the retrospective and observational nature of the study. The study complied with the ethical standards of the Declaration of Helsinki. Clinical trial number: not applicable. Author Contributions : JY: Writing – original draft, Data curation, Formal analysis. LC: Writing – original draft, Data curation, Formal analysis. XL: Methodology, Supervision, Writing – review & editing. JL: Conceptualization, Funding acquisition, Supervision, Validation, Writing – review & editing. KZ: Validation, Resources, Writing – review & editing. YY: Methodology, Visualization, Writing – review & editing. All authors were involved in writing the paper and had final approval of the submitted and published versions. Funding : This study was financially supported by the National Natural Science Foundation of China (Grant No.82360372), the Key Research & Development Program of Guangxi (Grant No. GuiKeAB22080088), the Joint Project on Regional High-Incidence Diseases Research of Guangxi Natural Science Foundation (Grant No. 2023GXNSFDA026023), the First-class Discipline Innovation-driven Talent Program of Guangxi Medical University, the Guangxi Zhuang Autonomous Region Health Commission Self-funded Research Program (Grant No. Z-R20251623), and the Guigang Science and Technology Project (GKJ2203022, GKG2300032). Acknowledgments : Not applicable. References Singer M, Deutschman C S, Seymour C W, et al. The third international consensus definitions for sepsis and septic shock (sepsis-3)[J]. JAMA, 2016, 315(8): 801-825. DOI:10.1001/jama.2016.0287. Evans L, Rhodes A, Alhazzani W, et al. Surviving sepsis campaign: international guidelines for management of sepsis and septic shock 2021[J]. Intensive Care Medicine, 2021, 47(11): 1181-1247. DOI:10.1007/s00134-021-06506-y. Fleischmann-Struzek C, Mellhammar L, Rose N, et al. Incidence and mortality of hospital- and ICU-treated sepsis: results from an updated and expanded systematic review and meta-analysis[J]. Intensive Care Medicine, 2020, 46(8): 1552-1562. DOI:10.1007/s00134-020-06151-x. Markwart R, Saito H, Harder T, et al. Epidemiology and burden of sepsis acquired in hospitals and intensive care units: a systematic review and meta-analysis[J]. Intensive Care Medicine, 2020, 46(8): 1536-1551. DOI:10.1007/s00134-020-06106-2. Rudd K E, Johnson S C, Agesa K M, et al. Global, regional, and national sepsis incidence and mortality, 1990–2017: analysis for the global burden of disease study[J]. Lancet, 2020, 395(10219): 200-211. DOI:10.1016/S0140-6736(19)32989-7. Srdić T, Đurašević S, Lakić I, et al. From molecular mechanisms to clinical therapy: understanding sepsis-induced multiple organ dysfunction[J]. International journal of molecular sciences, 2024, 25(14): 7770. DOI:10.3390/ijms25147770. Tang F, Zhao X L, Xu L Y, et al. Endothelial dysfunction: pathophysiology and therapeutic targets for sepsis-induced multiple organ dysfunction syndrome[J]. Biomedicine & Pharmacotherapy, 2024, 178: 117180. DOI:10.1016/j.biopha.2024.117180. Wang L, Chen J, Zhou X, et al. Multiple organ dysfunction and the mortality in sepsis: a multicenter retrospective study[J]. Chinese Medical Journal, 2025, 138(16): 2037-2039. DOI:10.1097/CM9.0000000000003673. Lambden S, Laterre P F, Levy M M, et al. The SOFA score—development, utility and challenges of accurate assessment in clinical trials[J]. Critical Care, 2019, 23(1): 374-383. DOI:10.1186/s13054-019-2663-7. Quintairos A, Pilcher D, Salluh J I F. ICU scoring systems[J]. Intensive Care Medicine, 2023, 49(2): 223-225. DOI:10.1007/s00134-022-06914-8. Pellathy T P, Pinsky M R, Hravnak M. Intensive care unit scoring systems[J]. Critical Care Nurse, 2021, 41(4): 54-64. DOI:10.4037/ccn2021613. Fleuren L M, Klausch T L T, Zwager C L, et al. Machine learning for the prediction of sepsis: a systematic review and meta-analysis of diagnostic test accuracy[J]. Intensive Care Medicine, 2020, 46(3): 383-400. DOI:10.1007/s00134-019-05872-y. Beam A L, Kohane I S. Big data and machine learning in health care[J]. JAMA, 2018, 319(13): 1317. DOI:10.1001/jama.2017.18391. Rajkomar A, Dean J, Kohane I. Machine learning in medicine[J]. New England Journal of Medicine, 2019, 380(14): 1347-1358. DOI:10.1056/NEJMra1814259. Johnson A E W, Ghassemi M M, Nemati S, et al. Machine learning and decision support in critical care[J]. Proceedings of the IEEE, 2016, 104(2): 444-466. DOI:10.1109/JPROC.2015.2501978. Shillan D, Sterne J A C, Champneys A, et al. Use of machine learning to analyse routinely collected intensive care unit data: a systematic review[J]. Critical Care, 2019, 23(1): 284-295. DOI:10.1186/s13054-019-2564-9. Zhou S, Lu Z, Liu Y, et al. Interpretable machine learning model for early prediction of 28-day mortality in ICU patients with sepsis-induced coagulopathy: development and validation[J]. European Journal of Medical Research, 2024, 29(1): 14-28. DOI:10.1186/s40001-023-01593-7. Muralitharan S, Nelson W, Di S, et al. Machine learning–based early warning systems for clinical deterioration: systematic scoping review[J]. Journal of Medical Internet Research, 2021, 23(2): e25187. DOI:10.2196/25187. Zhang Y, Xu W, Yang P, et al. Machine learning for the prediction of sepsis- related death: a systematic review and meta- analysis[J]. BMC Med Inform Decis Mak. 2023 Dec 11;23(1):283. doi: 10.1186/s12911-023-02383-1. Smith L A, Oakden-Rayner L, Bird A, et al. Machine learning and deep learning predictive models for long-term prognosis in patients with chronic obstructive pulmonary disease: a systematic review and meta-analysis[J]. The Lancet DIGITAL HEALTH, 2023, 5(12): e872-e881. DOI:10.1016/S2589-7500(23)00177-2. Moor M, Rieck B, Horn M, et al. Early prediction of sepsis in the ICU using machine learning: a systematic review[J]. Frontiers in Medicine, 2021, 8: 607952. DOI:10.3389/fmed.2021.607952. Yang Z. Predicting sepsis onset in ICU using machine learning models: a systematic review and meta-analysis[J]. BMC Infect Dis. 2023 Sep 27;23(1):635. doi: 10.1186/s12879-023-08614-0. Johnson A E W, Bulgarelli L, Shen L, et al. MIMIC-IV, a freely accessible electronic health record dataset[J]. Scientific Data, 2023, 10(1): 1-10. DOI:10.1038/s41597-022-01899-x. Austin P C, White I R, Lee D S, et al. Missing data in clinical research: a tutorial on multiple imputation[J]. Canadian Journal of Cardiology, 2021, 37(9): 1322-1331. DOI:10.1016/j.cjca.2020.11.010. Fira M, Goras L, Costin H N. Evaluating sparse feature selection methods: a theoretical and empirical perspective[J]. Applied Sciences, 2025, 15(7): 3752. DOI:10.3390/app15073752. Chowdhury S U, Sayeed S, Rashid I, et al. Shapley-additive-explanations-based factor analysis for dengue severity prediction using machine learning[J]. Journal of Imaging, 2022, 8(9): 229-254. DOI:10.3390/jimaging8090229. Ranzani O T, Singer M, Salluh J I F, et al. Development and validation of the sequential organ failure assessment (SOFA)-2 score[J]. JAMA, 2025, 334(23): 2090. DOI:10.1001/jama.2025.20516. Pérez-Torres D, Merino-García P A, Canas-Pérez I, et al. Real-world inter-observer variability of the sequential organ failure assessment (SOFA) score in intensive care medicine: the time has come for an update[J]. Critical Care, 2023, 27(1): 160-163. DOI:10.1186/s13054-023-04449-y. Moreno R, Rhodes A, Piquilloud L, et al. The sequential organ failure assessment (SOFA) score: has the time come for an update?[J]. Critical Care, 2023, 27(1): 15. DOI:10.1186/s13054-022-04290-9. Vincent J L, Bakker J. Blood lactate levels in sepsis: in 8 questions[J]. Current Opinion in Critical Care, 2021, 27(3): 298-302. DOI:10.1097/MCC.0000000000000824. Weinberger J, Klompas M, Rhee C. What is the utility of measuring lactate levels in patients with sepsis and septic shock?[J]. Seminars in Respiratory and Critical Care Medicine, 2021, 42(5): 650-661. DOI:10.1055/s-0041-1733915. Iba T, Helms J, Connors J M, et al. The pathophysiology, diagnosis, and management of sepsis-associated disseminated intravascular coagulation[J]. Journal of Intensive Care, 2023, 11(1): 24. DOI:10.1186/s40560-023-00672-5. Iba T, Helms J, Levy J H. Sepsis-induced coagulopathy (SIC) in the management of sepsis[J]. Annals of Intensive Care, 2024, 14(1): 148-158. DOI:10.1186/s13613-024-01380-5. Sikora A, Zhang T, Murphy D J, et al. Machine learning vs. traditional regression analysis for fluid overload prediction in the ICU[J]. Scientific Reports, 2023, 13(1): 19654. DOI:10.1038/s41598-023-46735-3. Benedetto U, Dimagli A, Sinha S, et al. Machine learning improves mortality risk prediction after cardiac surgery: systematic review and meta-analysis[J]. Journal of Thoracic and Cardiovascular Surgery, 2022, 163(6): 2075-2087.e9. DOI:10.1016/j.jtcvs.2020.07.105. Vickers A J, Holland F. Decision curve analysis to evaluate the clinical benefit of prediction models[J]. Spine Journal, 2021, 21(10): 1643-1648. DOI:10.1016/j.spinee.2021.02.024. Van Calster B, Wynants L, Verbeek J F M, et al. Reporting and interpreting decision curve analysis: a guide for investigators[J]. European Urology, 2018, 74(6): 796-804. DOI:10.1016/j.eururo.2018.08.038. Bifarin O O. Interpretable machine learning with tree-based shapley additive explanations: application to metabolomics datasets for binary classification[J]. PLOS One, 2023, 18(5): e0284315. DOI:10.1371/journal.pone.0284315. Lundberg S M, Erion G, Chen H, et al. From local explanations to global understanding with explainable AI for trees[J]. Nature Machine Intelligence, 2020, 2(1): 56-67. DOI:10.1038/s42256-019-0138-9. Bomrah S, Uddin M, Upadhyay U, et al. A scoping review of machine learning for sepsis prediction- feature engineering strategies and model performance: a step towards explainability[J]. Critical Care, 2024, 28(1): 180-196. DOI:10.1186/s13054-024-04948-6. Yang H S. Machine learning for sepsis prediction: prospects and challenges[J]. Clinical Chemistry, 2024, 70(3): 465-467. DOI:10.1093/clinchem/hvae006. Slieker R C, Van Der Heijden A A W A, Siddiqui M K, et al. Performance of prediction models for nephropathy in people with type 2 diabetes: systematic review and external validation study[J]. BMJ, 2021: n2134. DOI:10.1136/bmj.n2134. Rockenschaub P, Akay E M, Carlisle B G, et al. External validation of AI-based scoring systems in the ICU: a systematic review and meta-analysis[J]. BMC Medical Informatics and Decision Making, 2025, 25(1): 5-15. DOI:10.1186/s12911-024-02830-7. Rockenschaub P, Hilbert A, Kossen T, et al. The impact of multi-institution datasets on the generalizability of machine learning prediction models in the ICU[J]. Critical Care Medicine, 2024, 52(11): 1710-1721. DOI:10.1097/CCM.0000000000006359. Ciobanu-Caraus O, Aicher A, Kernbach J M, et al. A critical moment in machine learning in medicine: on reproducible and interpretable learning[J]. Acta Neurochirurgica, 2024, 166(1): 14-21. DOI:10.1007/s00701-024-05892-8. Table Table 1 is available in the Supplementary Files section. Additional Declarations No competing interests reported. Supplementary Files Table1.docx SupplementaryFigures.docx Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 13 Apr, 2026 Reviews received at journal 20 Mar, 2026 Reviews received at journal 19 Mar, 2026 Reviews received at journal 04 Mar, 2026 Reviewers agreed at journal 04 Mar, 2026 Reviews received at journal 03 Mar, 2026 Reviewers agreed at journal 01 Mar, 2026 Reviewers agreed at journal 28 Feb, 2026 Reviewers agreed at journal 27 Feb, 2026 Reviewers invited by journal 27 Feb, 2026 Editor assigned by journal 02 Feb, 2026 Submission checks completed at journal 02 Feb, 2026 First submitted to journal 23 Jan, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8681490","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":600101382,"identity":"19a81358-7582-4be9-8e3d-e13296bf3694","order_by":0,"name":"Jinbin Yang","email":"","orcid":"","institution":"The Second Affiliated Hospital of Guangxi Medical University","correspondingAuthor":false,"prefix":"","firstName":"Jinbin","middleName":"","lastName":"Yang","suffix":""},{"id":600101383,"identity":"056c9828-d555-4942-ac69-05440373eeec","order_by":1,"name":"Linying Cai","email":"","orcid":"","institution":"Gangbei District People,s Hospital","correspondingAuthor":false,"prefix":"","firstName":"Linying","middleName":"","lastName":"Cai","suffix":""},{"id":600101384,"identity":"ecf226bd-df04-4f05-bf1e-9b7b27afce79","order_by":2,"name":"Xuyang Liu","email":"","orcid":"","institution":"The Second Affiliated Hospital of Guangxi Medical University","correspondingAuthor":false,"prefix":"","firstName":"Xuyang","middleName":"","lastName":"Liu","suffix":""},{"id":600101385,"identity":"0fbd3d33-fbe4-47b6-894e-7b1b5963074a","order_by":3,"name":"Kaihuan Zhou","email":"","orcid":"","institution":"The First Affiliated Hospital of Guangxi Medical University","correspondingAuthor":false,"prefix":"","firstName":"Kaihuan","middleName":"","lastName":"Zhou","suffix":""},{"id":600101386,"identity":"9102b587-09fa-4e96-96b1-435c3d4d6e6e","order_by":4,"name":"Junyu Lu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAzklEQVRIiWNgGAWjYBACxgYQWcAgB+UzE6vFgMGYeC0QYMCQ2EC0FuYZucekeQwOp8+PyE6TYKiwTmxgP3sAv8Nm5KWBtORuPHN2mwTDmfTEBp68BAJacswgWtp7t0kwth1ObJDgMSBKS7phMy9Qyz8StCTIs4NsaSBGS88bY8s5BumGG3jObrZIOJZu3MaTg1+LYXuO4Y03Fdby8jNyN974UGMt289+hoCWBgYWCRDD4ACQSABiNrzqgUAeGDUfwIwGQkpHwSgYBaNgxAIA13xAkXyequkAAAAASUVORK5CYII=","orcid":"","institution":"The Second Affiliated Hospital of Guangxi Medical University","correspondingAuthor":true,"prefix":"","firstName":"Junyu","middleName":"","lastName":"Lu","suffix":""},{"id":600101387,"identity":"938e6155-8bec-4852-9888-7cb2e5ecea9d","order_by":5,"name":"Yegui Yang","email":"","orcid":"","institution":"The Second Affiliated Hospital of Guangxi Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yegui","middleName":"","lastName":"Yang","suffix":""}],"badges":[],"createdAt":"2026-01-23 17:23:26","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8681490/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8681490/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":104175996,"identity":"81bf1af3-a752-48a7-9849-0b9509b7afa1","added_by":"auto","created_at":"2026-03-08 16:33:55","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":358893,"visible":true,"origin":"","legend":"\u003cp\u003eLegend not included with this version\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8681490/v1/90409ce54312072f42ee573d.png"},{"id":104175988,"identity":"b79c2bbb-8b4d-4783-903d-512df5ebd584","added_by":"auto","created_at":"2026-03-08 16:33:55","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":603975,"visible":true,"origin":"","legend":"\u003cp\u003eFeature selection was performed using a logistic regression–based LASSO approach. The optimal penalty parameter (λ) was determined via 10-fold cross-validation, and the shrinkage trajectories of variable regression coefficients across varying levels of penalisation were examined (Figure 2A–B).\u003cstrong\u003eLASSO-based feature selection for MODS.\u003c/strong\u003e\u003cbr\u003e\n(A) Ten-fold cross-validation curve for the LASSO regression model.\u003cbr\u003e\n(B) Shrinkage trajectories of regression coefficients for candidate variables as the penalisation parameter (λ) increases.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-8681490/v1/3d92d9e7cdc8a23dae46e405.png"},{"id":104175992,"identity":"ee347f37-9187-4169-b8ea-bd009e41c7e8","added_by":"auto","created_at":"2026-03-08 16:33:55","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":2833767,"visible":true,"origin":"","legend":"\u003cp\u003eROC curves of machine learning models for predicting MODS in patients with sepsis.\u003c/p\u003e\n\u003cp\u003eROC curves illustrate the discriminative performance of nine machine learning models for MODS prediction in the training cohort. The evaluated models include logistic regression, elastic net, random forest, XGBoost, GBM, LightGBM, neural network, support vector machine with a radial basis function kernel, and k-nearest neighbours. The dashed diagonal line denotes random classification. The area under the curve (AUC) for each model is reported and annotated in the figure.\u003c/p\u003e","description":"","filename":"Figure3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8681490/v1/4cb40920d6d3354b5300a9bf.jpeg"},{"id":104175994,"identity":"46903262-0bf8-42ce-803b-16b11b84f2fd","added_by":"auto","created_at":"2026-03-08 16:33:55","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":973155,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCalibration performance of machine-learning models for predicting MODS.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(A) Calibration curves on the test set.\u003c/strong\u003e\u003cbr\u003e\nPanel A illustrates the calibration curves of multiple machine-learning models for predicting MODS in patients with sepsis on the test set. The x-axis represents the mean predicted probability, and the y-axis represents the observed proportion of MODS within each risk stratum. The dashed diagonal line indicates perfect calibration. Overall, LightGBM, XGBoost, and random forest demonstrate close agreement with the ideal reference line across most risk strata, indicating good probability-level calibration. In contrast, logistic regression shows modest deviations in the moderate- to high-risk ranges.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(B) Calibration curve of the final LightGBM model.\u003c/strong\u003e\u003cbr\u003e\nPanel B presents the calibration curve of the final LightGBM model on the test set. Predicted probabilities closely align with the observed incidence of MODS across the full range of predicted risk, with only minor deviations at the extreme high-risk end. These findings indicate that the LightGBM model provides well-calibrated probability estimates in addition to strong discriminative performance.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-8681490/v1/cfb87fb7cefa86f072eec22a.png"},{"id":104175989,"identity":"f53d1c40-cb31-46eb-9f13-937e09c3b9b0","added_by":"auto","created_at":"2026-03-08 16:33:55","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":165835,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDecision curve analysis for predicting MODS in the test cohort.\u003c/strong\u003e\u003cbr\u003e\nDecision curve analysis illustrates the net clinical benefit of each prediction model across a range of threshold probabilities, compared with the default strategies of treating all patients or treating none. The x-axis represents the threshold probability, and the y-axis represents net benefit. A higher net benefit indicates greater potential clinical utility. Overall, gradient boosting– and ensemble-based models, particularly the LightGBM, demonstrate consistently higher net benefit across a broad range of clinically relevant thresholds, suggesting superior decision-support value compared with traditional logistic regression.\u003c/p\u003e","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-8681490/v1/c3260fcbef259e20be48e97d.png"},{"id":104175995,"identity":"604428b1-4c84-4930-bd6a-3a10762ce4fc","added_by":"auto","created_at":"2026-03-08 16:33:55","extension":"jpeg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":1844197,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCross-model comparison of feature importance for predicting MODS.\u003c/strong\u003e\u003cbr\u003e\nThe heatmap displays the relative importance of key clinical variables across six machine-learning models, including LightGBM, XGBoost, random forest, GBM, elastic net, and logistic regression. Feature importance scores were normalised within each model on a 0–100 scale to facilitate cross-model comparison. Several variables, most notably the SOFA score, RR, lactate concentration, INR, and mechanical ventilation status—consistently ranked highly across different modelling approaches, indicating robust and model-independent contributions to MODS risk prediction. In contrast, variability in the importance of other features reflects differences between linear and non-linear models in capturing prognostic information and feature interactions.\u003c/p\u003e","description":"","filename":"Figure6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8681490/v1/cf3333234fbe7cb2d44bac84.jpeg"},{"id":104175993,"identity":"fc94b183-29ab-444b-ac5a-33089d751c7a","added_by":"auto","created_at":"2026-03-08 16:33:55","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":1166021,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSHAP-based interpretability analysis of the LightGBM model for predicting MODS .\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(A)\u003c/strong\u003e SHAP summary plot illustrating global feature importance in the LightGBM model. Features are ranked according to their mean absolute SHAP values, reflecting their overall contribution to MODS risk prediction.\u003cbr\u003e\n \u003cstrong\u003e(B)\u003c/strong\u003e SHAP dependence plot for serum lactate, demonstrating a pronounced non-linear association with MODS risk. SHAP values increase sharply beyond a critical lactate range, indicating a clear threshold effect.\u003cbr\u003e\n \u003cstrong\u003e(C)\u003c/strong\u003e SHAP dependence plot for platelet count, showing a non-linear and asymmetric relationship with MODS risk, with diminishing marginal effects at higher platelet levels.\u003cbr\u003e\n \u003cstrong\u003e(D)\u003c/strong\u003e SHAP dependence plot for the SOFA score, revealing a strong dose–response relationship between increasing organ dysfunction burden and predicted MODS risk, with evidence of saturation at extreme values.\u003c/p\u003e\n\u003cp\u003eCollectively, these analyses demonstrate that the LightGBM model captures clinically meaningful non-linear relationships and threshold effects between key physiological variables and MODS risk, thereby enhancing interpretability beyond traditional linear modelling approaches.\u003c/p\u003e","description":"","filename":"Figure7.png","url":"https://assets-eu.researchsquare.com/files/rs-8681490/v1/755e36c2f0fb0f48ab452ca9.png"},{"id":104175997,"identity":"239d4812-7b28-4ab1-bb5e-6c64a5fa42ec","added_by":"auto","created_at":"2026-03-08 16:33:55","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":1006029,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eExternal validation of the final model in the eICU cohort.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(A)\u003c/strong\u003e ROC curves comparing the discriminative performance of the final model ( LightGBM) with baseline models, including logistic regression, random forest, and XGBoost, in the external eICU validation cohort. The LightGBM model demonstrates consistently superior discrimination across the full range of thresholds.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(B)\u003c/strong\u003e Calibration curve of the final LightGBM model in the eICU cohort. Predicted probabilities were grouped into deciles, with the mean predicted risk plotted against the observed proportion of MODS. The dashed diagonal line denotes perfect calibration. Overall, predicted risks show good agreement with observed outcomes across most probability ranges, supporting the robustness and generalisability of the model in an independent external population.\u003c/p\u003e","description":"","filename":"Figure8.png","url":"https://assets-eu.researchsquare.com/files/rs-8681490/v1/bc8877fa4f9f280e941fe0d9.png"},{"id":104409390,"identity":"f0314099-00f2-48cf-a0b9-27a38769f56a","added_by":"auto","created_at":"2026-03-11 12:44:57","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":9774733,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8681490/v1/e9041b9c-f0b6-49af-acc5-226a6e586806.pdf"},{"id":104404564,"identity":"72d614de-a39f-4c78-823d-f267f2c755ae","added_by":"auto","created_at":"2026-03-11 12:20:32","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":24953,"visible":true,"origin":"","legend":"","description":"","filename":"Table1.docx","url":"https://assets-eu.researchsquare.com/files/rs-8681490/v1/b9da0bcbf9e989719a64a6c1.docx"},{"id":104175990,"identity":"21e1ef56-dd2b-4b00-b7b3-f054c4f2d91f","added_by":"auto","created_at":"2026-03-08 16:33:55","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":1202885,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigures.docx","url":"https://assets-eu.researchsquare.com/files/rs-8681490/v1/8232ed98c43819c548adef63.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Development and External Validation of a Machine Learning–Based Model for Early Prediction of Multiple Organ Dysfunction Syndrome in Critically Ill Patients with Sepsis","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eSepsis is defined as life-threatening organ dysfunction caused by a dysregulated host response to infection \u003csup\u003e[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]\u003c/sup\u003e, and remains one of the leading causes of mortality and healthcare resource utilisation in intensive care units (ICUs) worldwide \u003csup\u003e[\u003cspan additionalcitationids=\"CR4\" citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]\u003c/sup\u003e. Although advances in infection control, organ support, and evidence-based therapeutic strategies have improved outcomes to some extent in recent years, sepsis-associated mortality remains unacceptably high. This is particularly evident when the disease progresses to multiple organ dysfunction syndrome (MODS), a stage at which the risk of death increases sharply \u003csup\u003e[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/sup\u003e .\u003c/p\u003e \u003cp\u003eMODS represents an advanced and severe stage in the clinical course of sepsis, characterised by the progressive accumulation and mutual amplification of dysfunction across multiple organ systems. Its development is not only closely associated with increased short-term mortality but also exerts a substantial adverse impact on long-term survival and functional outcomes \u003csup\u003e[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]\u003c/sup\u003e. Epidemiological studies have demonstrated that MODS is highly prevalent among patients with sepsis admitted to the ICU and is frequently accompanied by prolonged hospitalisation, increased requirements for organ support, and a markedly elevated risk of death \u003csup\u003e[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/sup\u003e. Consequently, accurate identification of patients at high risk of developing MODS at an early stage of disease, thereby enabling timely, targeted monitoring and intervention, remains a critical and unresolved clinical challenge in contemporary critical care medicine.\u003c/p\u003e \u003cp\u003eCurrently, risk assessment in clinical practice relies primarily on severity scoring systems such as the Sequential Organ Failure Assessment (SOFA), Simplified Acute Physiology Score II (SAPS II), and Acute Physiology and Chronic Health Evaluation (APACHE) scores \u003csup\u003e[\u003cspan additionalcitationids=\"CR10\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]\u003c/sup\u003e. Although these tools provide a certain degree of prognostic value at the population level, they are inherently based on linear assumptions and static combinations of variables, which limits their ability to fully capture the complex non-linear physiological disturbances and dynamic inter-organ interactions observed in patients with sepsis \u003csup\u003e[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]\u003c/sup\u003e. Moreover, these scoring systems were originally designed to quantify overall disease severity rather than to generate individualised risk predictions for specific outcomes such as MODS,thereby constraining their utility for precise clinical decision support.\u003c/p\u003e \u003cp\u003eWith the rapid expansion of electronic health records and large-scale critical care databases, the application of machine learning (ML) approaches to outcome prediction in critical care has attracted increasing attention \u003csup\u003e[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]\u003c/sup\u003e. Compared with traditional statistical models, ML algorithms are capable of automatically learning non-linear relationships and complex interaction effects among variables in high-dimensional data settings, thereby demonstrating potential advantages across a range of ICU outcome prediction tasks \u003csup\u003e[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan additionalcitationids=\"CR15\" citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]\u003c/sup\u003e. Previous studies have shown that ensemble-based and gradient boosting models achieve favourable discriminative performance in predicting sepsis-related mortality, organ failure, and associated complications \u003csup\u003e[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eHowever, most existing machine-learning studies have primarily focused on sepsis identification or mortality risk prediction \u003csup\u003e[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan additionalcitationids=\"CR20 CR21\" citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]\u003c/sup\u003e. In many of these investigations, organ dysfunction has been treated as an intermediate variable or incorporated as a component of composite scores, rather than being modelled directly as MODS, a highly heterogeneous outcome characterised by dynamic temporal evolution. Consequently, model-based studies that explicitly target MODS as the primary endpoint remain relatively scarce. In addition, among published ML studies, relatively few have simultaneously addressed model discrimination, interpretability, and rigorous external validation, which to some extent limits the clinical credibility and broader applicability of their findings.\u003c/p\u003e \u003cp\u003e Building on these gaps in the existing literature, the present study leveraged the Medical Information Mart for Intensive Care IV (MIMIC-IV) database to systematically develop and compare multiple machine-learning models for predicting the risk of MODS in patients with sepsis during ICU hospitalisation. Through regularised feature selection, comprehensive multi-model performance evaluation, decision curve analysis, and Shapley additive explanations (SHAP)\u0026ndash;based interpretability analyses, followed by external validation in an independent the eICU Collaborative Research Database (eICU) cohort, we aimed to develop a MODS risk prediction model that integrates strong predictive performance, interpretability, and clinical applicability. This data-driven approach seeks to support early risk stratification and individualised management of patients with sepsis.\u003c/p\u003e"},{"header":"2 Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Data source\u003c/h2\u003e \u003cp\u003eThe data for this study were obtained from the MIMIC-IV (version 3.0) and eICU \u003csup\u003e[\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]\u003c/sup\u003e, The MIMIC-IV database is developed and maintained by the Laboratory for Computational Physiology at the Massachusetts Institute of Technology and contains detailed clinical information on 546,028 hospitalised patients admitted to the Beth Israel Deaconess Medical Center in the United States between 2008 and 2022. The eICU database is a multicentre critical care repository that includes de-identified clinical data from adult ICU patients treated at more than 200 hospitals across the United States between 2014 and 2015.\u003c/p\u003e \u003cp\u003ePrior to data access, all members of the research team completed the required training provided by the US National Institutes of Health and passed the examination on \u003cem\u003eProtecting Human Research Participants\u003c/em\u003e (Record ID: 71642466). All patient data were fully de-identified to protect personal privacy; therefore, this study did not require additional informed consent or approval from an institutional review board.\u003c/p\u003e \u003cp\u003eEligibility Criteria\u003c/p\u003e \u003cp\u003eInclusion criteria were as follows: (1) patients with sepsis diagnosed according to the Sepsis-3 criteria; (2) age\u0026thinsp;\u0026ge;\u0026thinsp;18 years; and (3) first admission to ICU.\u003c/p\u003e \u003cp\u003eExclusion criteria included: (1) ICU length of stay\u0026thinsp;\u0026lt;\u0026thinsp;24 h; (2) presence of MODS at the time of ICU admission; (3) severe pre-existing comorbidities, including advanced malignancy, end-stage renal disease, or advanced acquired immunodeficiency syndrome; and (4) prior long-term organ replacement therapy, such as chronic dialysis or prolonged mechanical ventilation.\u003c/p\u003e \u003cp\u003eOutcome Definition\u003c/p\u003e \u003cp\u003eMODS was defined as the first occurrence of dysfunction in at least two organ systems, each with a Sequential Organ Failure Assessment (SOFA) score\u0026thinsp;\u0026ge;\u0026thinsp;2, occurring 24 h after ICU admission. The organ systems assessed included the respiratory, cardiovascular, coagulation, hepatic, renal, and central nervous systems.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Data extraction\u003c/h2\u003e \u003cp\u003eData on patients with sepsis who met the predefined inclusion and exclusion criteria were extracted from the MIMIC-IV (version 3.0) and eICU databases. The extracted variables included demographic characteristics, baseline comorbidities, vital sign parameters, laboratory measurements, disease severity scores, such as the SOFA score and the Glasgow Coma Scale (GCS) score, organ support therapies (mechanical ventilation, use of vasoactive agents, and continuous renal replacement therapy [CRRT]), as well as other relevant clinical information. For each variable, the first available value within the first 24 h after ICU admission was used, and for patients with multiple hospitalisations, only data from the first hospital admission were retained for analysis.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Missing Data Imputation and Outlier Management Strategies\u003c/h2\u003e \u003cp\u003eTo minimise potential bias arising from missing data, variables with more than 20% missing values were excluded from the final cohort. For variables with a missingness rate between 5% and 20%, multiple imputation by chained equations (MICE) was applied \u003csup\u003e[\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]\u003c/sup\u003e, generating \u003cem\u003em\u003c/em\u003e\u0026thinsp;=\u0026thinsp;5 imputed datasets. To prevent information leakage, the imputation procedures were conducted independently in the training and test sets. Continuous variables were imputed using predictive mean matching, whereas categorical variables were imputed using logistic regression models. Final parameter estimates were pooled across the imputed datasets according to Rubin\u0026rsquo;s rules.\u003c/p\u003e \u003cp\u003eIn addition, during data preprocessing, outliers in continuous variables, excluding selected variables such as age, SOFA score, and GCS score, were handled using the interquartile range (IQR) method. Specifically, the first (Q1) and third (Q3) quartiles were calculated for each variable, and the IQR was defined as Q3\u0026thinsp;\u0026minus;\u0026thinsp;Q1. Data points exceeding the upper bound (Q3\u0026thinsp;+\u0026thinsp;1.5 \u0026times; IQR) or falling below the lower bound (Q1\u0026thinsp;\u0026minus;\u0026thinsp;1.5 \u0026times; IQR) were considered outliers. After being identified, these outliers were imputed using regression-based methods to preserve the accuracy and robustness of subsequent analyses.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Statistical analysis\u003c/h2\u003e \u003cdiv id=\"Sec7\" class=\"Section3\"\u003e \u003ch2\u003e2.4.1 Baseline Characteristics\u003c/h2\u003e \u003cp\u003eNormality of continuous variables in the baseline characteristics was assessed using the Kolmogorov\u0026ndash;Smirnov test. Continuous variables with a normal distribution are presented as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation (SD), whereas non-normally distributed variables are expressed as median (interquartile range, IQR). Categorical variables are reported as counts and percentages. Comparisons of normally distributed continuous variables were performed using the t test or one-way analysis of variance (ANOVA), as appropriate, while categorical variables were compared using the Pearson χ\u0026sup2; test or Fisher\u0026rsquo;s exact test.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e2.4.2 Model Development and Validation\u003c/h2\u003e \u003cp\u003eUsing sepsis patient data from the MIMIC-IV (version 3.0) database, the cohort was randomly divided into a training set and a test set at a ratio of 70:30 to ensure model generalisability and predictive accuracy. Prior to model construction, least absolute shrinkage and selection operator (LASSO) regression was applied to perform feature selection among candidate predictors in order to reduce model redundancy and enhance prediction stability \u003csup\u003e[\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]\u003c/sup\u003e. By introducing an L1 regularisation penalty into the regression framework, LASSO retains the most informative variables associated with the outcome while shrinking coefficients of less contributory variables towards zero, thereby enabling variable selection and reducing the risk of overfitting.\u003c/p\u003e \u003cp\u003eThe optimal penalty parameter (λ) was determined through cross-validation, and only variables with non-zero regression coefficients were retained for subsequent multi-model predictive analyses. This feature selection procedure ensured a consistent predictor space across different modelling frameworks, thereby facilitating fair performance comparisons and improving interpretability of the results.\u003c/p\u003e \u003cp\u003eTo comprehensively evaluate predictive performance, multiple machine-learning algorithms were implemented, including decision trees, k-nearest neighbours (kNN), logistic regression, random forest, artificial neural networks, gradient boosting machine (GBM), extreme gradient boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and support vector machines (SVM). To mitigate overfitting, 10-fold cross-validation was applied, whereby 90% of the data were used for training and the remaining 10% for validation in each iteration, repeated across 10 folds. Model performance was assessed using accuracy, precision, recall, F1 score, and the area under the receiver operating characteristic (ROC) curve (AUC).\u003c/p\u003e \u003cp\u003eTo ensure model interpretability, we not only constructed models using features selected by LASSO regression but also applied SHAP \u003csup\u003e[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]\u003c/sup\u003e to the optimal LightGBM model. This approach enabled quantification of the specific contribution of each feature to the model\u0026rsquo;s predictions.\u003c/p\u003e \u003cp\u003eComparisons of model performance were conducted using the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. All statistical analyses were performed using IBM SPSS Statistics version 26.0 and R version 4.5. Two-sided P values\u0026thinsp;\u0026lt;\u0026thinsp;0.05 were considered statistically significant.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003e2.4.3 External Validation\u003c/h2\u003e \u003cp\u003eTo evaluate the generalisability and robustness of the predictive models across different clinical settings, external validation was performed using an independent cohort from the eICU Collaborative Research Database. This multicentre ICU database comprises critically ill patients from diverse healthcare institutions and was not involved in feature selection, model training, or parameter tuning during model development.\u003c/p\u003e \u003cp\u003eUsing the same variable definitions, data preprocessing procedures, and outcome criteria as those applied in the internal analyses, the final model was directly applied to the eICU cohort. Model discrimination was assessed using the AUC. Calibration performance was evaluated with calibration curves by comparing predicted risks with observed outcome frequencies, and the results were contrasted with those from internal validation to assess model stability.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"3 Results","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Baseline Characteristics\u003c/h2\u003e \u003cp\u003eA total of 23,018 patients with sepsis were included in this study (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e), among whom 4,931 (21.4%) developed MODS during their ICU stay. Compared with patients who did not develop MODS, those in the MODS group exhibited overall greater disease severity and more pronounced features of multi-organ dysfunction.\u003c/p\u003e \u003cp\u003eWith respect to demographic characteristics, patients in the MODS group were slightly older and had a lower proportion of males (both \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Regarding comorbidities, chronic kidney disease, heart failure, and chronic obstructive pulmonary disease were more prevalent in the MODS group, whereas the distribution of diabetes mellitus was comparable between the two groups. In terms of physiological and laboratory parameters, patients who developed MODS exhibited more pronounced circulatory, respiratory, and metabolic derangements, including higher heart rate and respiratory rate, lower mean arterial pressure and oxygen saturation, as well as more severe renal dysfunction and metabolic acidosis (all \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001).\u003c/p\u003e \u003cp\u003eIn addition, patients in the MODS group had significantly SOFA scores, Oxford Acute Severity of Illness Score (OASIS), and Simplified Acute Physiology Score II (SAPS II), and were more likely to receive vasoactive agents, mechanical ventilation, and renal replacement therapy (all \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Feature Selection Using LASSO\u003c/h2\u003e \u003cp\u003ePrior to model development, LASSO logistic regression was applied to the candidate predictors to reduce model complexity and mitigate the risks of multicollinearity and overfitting.\u003c/p\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e2\u003c/span\u003e, the optimal regularisation parameter (λ) was determined using 10-fold cross-validation. The model achieved the minimum cross-validated error at λ_min, whereas at λ_1se the predictive performance remained stable with a further reduction in model complexity. Balancing robustness and generalisability, variables with non-zero regression coefficients at the selected optimal λ were retained for subsequent model development.\u003c/p\u003e \u003cp\u003eThe final set of selected features encompassed demographic characteristics, baseline comorbidities, vital signs, laboratory parameters, disease severity scores, and variables related to organ support therapies, collectively reflecting the multidimensional clinical profile associated with the development of MODS in patients with sepsis.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Model Performance\u003c/h2\u003e \u003cp\u003eTo compare the predictive performance of different machine-learning models for the development of MODS in patients with sepsis, nine models were developed and evaluated in the training cohort (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e3\u003c/span\u003e). The ROC curves of all models were clearly above the diagonal reference line, indicating that each model demonstrated a certain degree of discriminative ability.\u003c/p\u003e \u003cp\u003eIn terms of discriminative performance, ensemble-based tree models consistently outperformed the other approaches. LightGBM achieved the highest discrimination (AUC\u0026thinsp;=\u0026thinsp;0.829), followed by the GBM(AUC\u0026thinsp;=\u0026thinsp;0.824), random forest (AUC\u0026thinsp;=\u0026thinsp;0.823), and XGBoost (AUC\u0026thinsp;=\u0026thinsp;0.822). In contrast, the traditional linear logistic regression model (AUC\u0026thinsp;=\u0026thinsp;0.802) and its regularised counterpart, elastic net (AUC\u0026thinsp;=\u0026thinsp;0.802), demonstrated comparatively lower predictive performance. SVM (;AUC\u0026thinsp;=\u0026thinsp;0.759) and kNN ( AUC\u0026thinsp;=\u0026thinsp;0.727) exhibited weaker discrimination, whereas the neural network model showed intermediate performance between tree-based and linear models (AUC\u0026thinsp;=\u0026thinsp;0.803).\u003c/p\u003e \u003cp\u003eIn the test cohort, calibration curves revealed notable differences in calibration performance across models (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e4\u003c/span\u003eA). Overall, tree-based ensemble models, including LightGBM, XGBoost, and random forest, showed good agreement with the ideal reference line across most ranges of predicted risk, indicating relatively stable probability estimation. In contrast, logistic regression tended to underestimate risk in the intermediate- to high-risk probability ranges.\u003c/p\u003e \u003cp\u003eBased on these comparative results, LightGBM was selected as the final model and its calibration performance was further evaluated separately (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e4\u003c/span\u003eB). The calibration curve for LightGBM demonstrated close alignment with the ideal reference line across the overall risk spectrum, with only minor deviations observed in the extreme high-risk range. These findings indicate that LightGBM not only provides strong discriminative ability but also yields accurate individual-level estimates of MODS risk.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Decision Curve Analysis\u003c/h2\u003e \u003cp\u003eDecision curve analysis (DCA) was performed to evaluate the clinical net benefit of different models across a range of threshold probabilities (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e5\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eIn the test cohort, all predictive models demonstrated greater net benefit across a broad range of threshold probabilities compared with the \u0026ldquo;treat-all\u0026rdquo; and \u0026ldquo;treat-none\u0026rdquo; strategies, indicating that model-based risk stratification can provide tangible clinical value for decision-making.\u003c/p\u003e \u003cp\u003eThe LightGBM model consistently achieved the highest and most stable net benefit across the majority of threshold probabilities. In particular, within the low-to-moderate threshold range (approximately 10%\u0026ndash;60%), LightGBM substantially outperformed the traditional logistic regression model. As the threshold probability increased, the net benefit of logistic regression gradually approached or even fell below zero, suggesting limited clinical utility at higher risk thresholds. In contrast, gradient boosting\u0026ndash; and ensemble-based models continued to demonstrate positive net benefit across these ranges.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Feature Importance Analysis\u003c/h2\u003e \u003cp\u003eTo assess the stability and consistency of feature importance across different machine-learning models, we compared the relative contributions of key predictors derived from multiple models, including XGBoost, random forest, LightGBM, GBM, elastic net, and logistic regression (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e6\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eTo enhance transparency of the internal decision-making processes of the models, we present the feature importance rankings for each machine-learning algorithm in the Supplementary Materials (Supplementary Figures \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003eA\u0026ndash;S1F).\u003c/p\u003e \u003cp\u003eAcross models, the Sequential Organ Failure Assessment (SOFA) score, respiratory rate (RR), lactate concentration, coagulation-related indices (international normalised ratio, INR), and indicators of metabolic acid\u0026ndash;base status (pH and bicarbonate) consistently ranked among the most important features in the majority of algorithms. This cross-model concordance indicates that these variables provide robust and consistent predictive value for assessing the risk of MODS.\u003c/p\u003e \u003cp\u003eIn contrast, variables such as age, the GCS score, and mechanical ventilation showed greater variability in importance rankings across different models, reflecting differences in model architecture with respect to capturing feature interactions and non-linear relationships.\u003c/p\u003e \u003cp\u003eOverall, this cross-model consistency analysis supports the stability of the selected features across different algorithmic frameworks, thereby strengthening the credibility of the model\u0026rsquo;s interpretability.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e3.6 SHAP Analysis of the LightGBM Model\u003c/h2\u003e \u003cp\u003eTo enhance model interpretability and elucidate the non-linear effects of key predictive variables, we further applied SHAP to the optimal model (LightGBM) (Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e7\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eGlobal SHAP feature importance analysis indicated that the SOFA score was the most influential predictor of MODS risk, exhibiting the widest distribution of SHAP values. This was followed by RR, INR, bicarbonate concentration, and use of vasoactive agents. Lactate, platelet count, creatinine, age, GCS score, mechanical ventilation, and pH also demonstrated consistent, albeit relatively secondary, contributions to risk prediction.\u003c/p\u003e \u003cp\u003eSHAP dependence plots further revealed pronounced non-linear relationships between key variables and the risk of MODS. Lactate contributed minimally to risk at lower concentrations; however, when levels increased beyond approximately 8\u0026ndash;10 mmol/L, SHAP values rose sharply, indicating a clear threshold effect. Platelet count showed a non-monotonic association with SHAP values, reaching a peak within a moderate range before plateauing. In contrast, the SOFA score exhibited a characteristic non-linear dose\u0026ndash;response relationship, with risk increasing markedly as the score rose and approaching saturation at very high levels.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e3.7 External Validation\u003c/h2\u003e \u003cp\u003eTo evaluate model generalisability, the final model was externally validated in an independent cohort from the eICU Collaborative Research Database. The results showed that the LightGBM model maintained stable discriminative performance in the external cohort, with ROC curve characteristics comparable to those observed in the internal validation (Fig.\u0026nbsp;\u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e8\u003c/span\u003eA).\u003c/p\u003e \u003cp\u003eIn terms of calibration, the predicted risks across different probability strata were generally consistent with the observed incidence of MODS, with no evident systematic overestimation or underestimation (Fig.\u0026nbsp;\u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e8\u003c/span\u003eB). Additional results from model comparisons and clinical decision curve analyses are provided in the Supplementary Materials (Supplementary Figures S4A\u0026ndash;S4E).\u003c/p\u003e \u003cp\u003e Although overall model performance in the external validation cohort was modestly attenuated compared with the internal validation, the relative ranking among models remained consistent, indicating that the model exhibits acceptable stability across different ICU populations.\u003c/p\u003e \u003c/div\u003e"},{"header":"4 Discussion","content":"\u003cp\u003eThis study leveraged the large-scale, real-world MIMIC-IV database to systematically develop and validate multiple machine-learning models for predicting the risk of MODS in patients with sepsis during ICU hospitalisation. As one of the most clinically consequential outcomes of sepsis, MODS is highly prevalent in ICU populations and is associated with poor prognosis, remaining a critical and unresolved challenge in contemporary critical care medicine \u003csup\u003e[5,6]\u003c/sup\u003e. Our findings demonstrate that ensemble-based machine-learning models markedly outperform traditional linear models in terms of discriminative ability, calibration performance, and potential clinical decision-making value. Among these approaches, the LightGBM showed the most robust and consistent performance and was further validated in an independent external cohort.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.1 Main Findings\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe primary objective of this study was to develop and validate a prediction model based on large-scale real-world data that can identify the risk of MODS in patients with sepsis at an early stage of ICU admission. In pursuit of this objective, several key findings emerged from our analyses.\u003c/p\u003e\n\u003cp\u003eFirst, the multi-model prediction framework developed using the MIMIC-IV database effectively discriminated between patients with sepsis who did and did not develop MODS during ICU hospitalisation, indicating that multidimensional clinical information carries quantifiable predictive value for MODS risk identification. Compared with traditional logistic regression, gradient boosting\u0026ndash; and ensemble-based models demonstrated superior overall discriminative performance and greater stability.\u003c/p\u003e\n\u003cp\u003eSecond, among the various machine-learning models evaluated, LightGBM consistently exhibited the most stable predictive performance across the training set, test set, and the independent external eICU cohort. This finding indicates that LightGBM possesses strong generalisability across different patient populations and clinical settings, supporting tree-based ensemble learning as a preferred modelling strategy for predicting sepsis-associated MODS risk.\u003c/p\u003e\n\u003cp\u003eThird, SHAP-based interpretability analyses demonstrated that the key predictors identified by the model were predominantly related to the severity of organ dysfunction, circulatory and coagulation abnormalities, and metabolic derangements. The importance ranking of these variables closely aligns with established pathophysiological mechanisms underlying MODS. This concordance not only strengthens the credibility of the model\u0026rsquo;s predictions but also provides clinically interpretable insights into the model\u0026rsquo;s decision-making process.\u003c/p\u003e\n\u003cp\u003eOverall, this study successfully developed and externally validated a MODS risk prediction model that integrates strong discriminative performance, robustness, and interpretability, thereby providing a data-driven foundation to support early risk stratification and precision management of patients with sepsis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.2 Pathophysiological Interpretation of Key Predictors\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAt the level of model interpretation, the key predictors identified in this study demonstrated a high degree of consistency across different modelling frameworks, and their clinical relevance closely aligns with established pathophysiological concepts of MODS. Notably, the SOFA score consistently ranked as the most important feature in both feature importance analyses and SHAP interpretations, underscoring the central role of baseline organ dysfunction severity in the progression of MODS. Importantly, this finding does not simply replicate existing scoring systems; rather, it highlights that within a data-driven modelling context, SOFA remains a pivotal integrative variable that encapsulates multi-organ functional status and serves as a critical hub in MODS risk evolution\u0026nbsp;\u003csup\u003e[27\u0026ndash;29]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eIn addition, the high importance assigned to respiratory rate, lactate concentration, and coagulation-related indices ( INR) highlights the synergistic dysregulation of the circulatory\u0026ndash;respiratory\u0026ndash;coagulation\u0026ndash;metabolic axis in the development of MODS. The SHAP dependence pattern for lactate demonstrates clear non-linear and threshold effects, supporting its role not merely as a static marker of tissue hypoperfusion but also as an indicator of cumulative mitochondrial dysfunction and metabolic stress\u0026nbsp;\u003csup\u003e[30,31]\u003c/sup\u003e. Abnormalities in coagulation parameters further suggest an amplifying role of sepsis-associated coagulopathy in MODS progression. The reciprocal activation of inflammatory and coagulation pathways is widely recognised as a key mechanism driving microcirculatory dysfunction and the progression of organ injury\u0026nbsp;\u003csup\u003e[32,33]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eNotably, the key variables identified by the model do not operate in isolation but instead cluster into risk feature sets at the level of functional systems. This system-level aggregation of predictors reflects the intrinsic nature of MODS as a composite outcome and helps explain the limitations of traditional single-marker approaches or linear models in predicting this complex clinical endpoint.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.3 Advantages of Machine-Learning Models Over Traditional Approaches\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFrom a methodological perspective, our findings indicate that machine-learning models offer clear advantages in predicting MODS, a highly heterogeneous and non-linearly driven clinical outcome. Compared with traditional logistic regression, gradient boosting\u0026ndash; and ensemble-based models can autonomously learn complex non-linear relationships and higher-order interactions among variables without imposing a priori linear assumptions, thereby achieving superior predictive performance in high-dimensional, complex clinical datasets\u0026nbsp;\u003csup\u003e[13,34,35]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eIt is important to emphasise that this study did not merely compare models based on differences in AUC. Instead, model robustness and clinical applicability were comprehensively evaluated through a combination of internal validation, external validation, calibration analyses, and decision curve analysis across different risk strata and patient populations. The consistently strong performance of the LightGBM across these multiple evaluation dimensions supports its selection as a preferred algorithm for MODS risk prediction, rather than attributing its superiority to chance or isolated performance gains.\u003c/p\u003e\n\u003cp\u003eAt the same time, by integrating regularised regression with tree-based models, this study achieved a balance between predictive performance and model interpretability, thereby mitigating the trust barriers commonly associated with so-called \u0026ldquo;black-box\u0026rdquo; models in clinical practice. This strategy provides a generalisable methodological paradigm for machine-learning modelling of complex ICU outcomes.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.4 Clinical Decision-Making Value and Potential Applications\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFrom a clinical application perspective, the prediction model developed in this study is primarily intended for early risk identification and stratification rather than for replacing clinical decision-making. Decision curve analysis demonstrated that, across a broad range of threshold probabilities, ensemble-based models provide greater clinical net benefit, with particularly clear advantages over traditional logistic regression in the low-to-moderate risk ranges. These findings suggest that the model has potential value in assisting clinicians to identify patients at high risk of developing MODS. Moreover, DCA offers a more intuitive representation of the potential benefits of predictive models within real-world clinical decision-making contexts\u0026nbsp;\u003csup\u003e[36,37]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eIn real-world ICU settings, this model could potentially be integrated as a component of a clinical decision support system, in conjunction with existing severity scores and bedside assessments, to enhance monitoring intensity for high-risk patients, optimise resource allocation, and support risk stratification in clinical trial design. However, the realisation of its clinical value depends on appropriate integration and use strategies rather than on reliance on model outputs in isolation.\u003c/p\u003e\n\u003cp\u003eTherefore, the broader significance of this study lies in demonstrating that, by integrating multidimensional clinical data with interpretability analyses, machine-learning models can provide incremental information for MODS risk assessment without departing from established clinical reasoning. This approach lays a foundation for future prospective investigations and clinical translation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.5 Interpretability and Model Credibility\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe clinical adoption of machine-learning models depends not only on predictive performance but also critically on the interpretability of their decision-making processes and their acceptability to clinicians. In this study, SHAP-based analyses were applied to the LightGBM model to quantify the relative contribution of each predictor to MODS risk estimation at both the individual and population levels, while also revealing non-linear and threshold relationships between several key variables and the outcome\u0026nbsp;\u003csup\u003e[38,39]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eThis interpretability-oriented methodological design allows model predictions to be directly contrasted with established pathophysiological understanding, thereby avoiding a \u0026ldquo;cognitive disconnect\u0026rdquo; between algorithmic decisions and clinical experience. By illustrating the direction and magnitude of each variable\u0026rsquo;s effect across different risk levels, the model not only delivers risk stratification results but also provides transparent insights into the sources of risk, which may help enhance clinicians\u0026rsquo; trust in the model outputs.\u003c/p\u003e\n\u003cp\u003eAccordingly, interpretability analyses not only strengthen the alignment between model predictions and clinical pathophysiological understanding but also enhance the credibility and acceptability of the model in real-world clinical settings, thereby providing an essential prerequisite for clinical translation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.6 Challenges in the Clinical Translation of Machine-Learning Models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAlthough machine-learning models have demonstrated considerable potential in risk prediction for critically ill patients\u0026nbsp;\u003csup\u003e[40]\u003c/sup\u003e, their translation into routine clinical practice faces multiple challenges. These challenges extend beyond technical performance alone and encompass issues related to generalisability, interpretability, and integration into clinical workflows\u0026nbsp;\u003csup\u003e[41]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eFirst, substantial performance heterogeneity exists across different models, and their generalisability and stability require systematic comparison and rigorous external validation\u0026nbsp;\u003csup\u003e[42]\u003c/sup\u003e. Prior reviews have emphasised that strong performance in internal validation does not necessarily translate into comparable performance in independent external cohorts \u003csup\u003e[43]\u003c/sup\u003e. This limitation is particularly pronounced in ICU populations, which are highly heterogeneous; variations in healthcare settings, patient case mix, and data distributions across centres can markedly influence model performance \u003csup\u003e[44]\u003c/sup\u003e, thereby constraining stable implementation in real-world clinical environments. The lack of comprehensive external validation and cross-regional assessment means that the clinical reliability of many models remains to be fully established.\u003c/p\u003e\n\u003cp\u003eSecond, complex models are often regarded as \u0026ldquo;black boxes,\u0026rdquo; with internal decision logic that is difficult for clinicians to intuitively understand. This lack of transparency and interpretability substantially limits their credibility and acceptability in clinical decision-making\u0026nbsp;\u003csup\u003e[20,45]\u003c/sup\u003e. These shortcomings in interpretability and general acceptance have prompted ongoing efforts to develop more interpretable model architectures and robust explanation methods to enhance clinical trust.\u003c/p\u003e\n\u003cp\u003eMoreover, although machine-learning approaches have been applied to early sepsis detection and mortality risk prediction, studies specifically targeting MODS, a highly heterogeneous and dynamically evolving composite clinical outcome, remain relatively limited. At present, there is still a lack of systematic investigations based on large-scale real-world data that simultaneously address predictive performance, interpretability, and clinical decision-making value. Future research should further integrate temporal dynamics and clinically interpretable frameworks to better characterise the complex evolution of organ dysfunction.\u003c/p\u003e\n\u003cp\u003eOverall, although machine-learning models show considerable promise for individual risk prediction, achieving reliable clinical translation requires overcoming challenges related to limited generalisability, insufficient interpretability, and narrow outcome definitions. Future studies should prioritise extensive cross-centre and cross-system validation, the development of more transparent model structures and explanation mechanisms, and methodological innovations that enable deeper integration with clinical decision-making workflows.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.7 Study Limitations\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study has several limitations. First, it employed a retrospective observational design. Although confounding was mitigated as much as possible through regularised feature selection and external validation, the influence of unmeasured or residual confounders cannot be fully excluded. Consequently, the model\u0026rsquo;s predictions should not be interpreted as evidence of causal relationships.\u003c/p\u003e\n\u003cp\u003eSecond, the definition of MODS relied on organ function indicators and scoring systems available within the databases, which may not fully capture the dynamic evolution of organ dysfunction during sepsis or the reversible changes following clinical interventions. This reliance on static or time-window\u0026ndash;based measures may lead to underestimation or overestimation of true organ functional status in certain patients. Future studies could address this limitation by developing more precise, dynamically updated risk models.\u003c/p\u003e\n\u003cp\u003eThird, although the model incorporated multidimensional clinical variables, limitations inherent to the database structure precluded inclusion of certain potentially important information, such as more granular temporal data, dynamic biomarker trajectories, and bedside decision-making factors. This constraint may limit the model\u0026rsquo;s ability to fully characterise complex pathophysiological processes.\u003c/p\u003e\n\u003cp\u003eIn addition, although external validation was performed using the independent eICU database, both datasets were derived from ICU populations in the United States. Therefore, the generalisability of the model to healthcare systems in other countries, settings with different levels of medical resources, and varying clinical practice patterns remains uncertain and warrants further confirmation through prospective, multicentre studies.\u003c/p\u003e\n\u003cp\u003eFinally, the model developed in this study is primarily intended for risk prediction and stratification rather than for directly guiding specific therapeutic decisions. Its effectiveness within real-world clinical workflows, as well as its potential impact on improving patient outcomes, still requires evaluation through prospective clinical trials.\u003c/p\u003e"},{"header":"5 Conclusion","content":"\u003cp\u003eThis multicentre study systematically developed and validated multiple machine-learning models to predict the risk of MODS in patients with sepsis. The results demonstrated that the gradient boosting\u0026ndash;based LightGBM achieved the most favourable and robust overall performance, with superior discrimination, good calibration, greater clinical net benefit, and consistent performance on external validation. By integrating multidimensional clinical information with interpretability analyses, the model was able to capture non-linear and threshold effects between key predictors and MODS risk, thereby providing data-driven support for early risk identification and stratification in sepsis. Importantly, the risk features identified by the model closely align with established pathophysiological understanding, enhancing its credibility in clinical application. Collectively, these findings suggest that the proposed model represents a potentially valuable tool for early risk stratification and individualised management of patients with sepsis in the ICU.\u003c/p\u003e \u003cp\u003eFuture studies should evaluate the performance of this model in real-world clinical settings through prospective, multicentre investigations and explore its potential impact on clinical decision-making processes and patient outcomes. Such efforts will be essential to facilitate the standardised and responsible integration of machine-learning models into the precision management of sepsis.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eConflict of Interest\u003c/strong\u003e\u003cstrong\u003e:\u0026nbsp;\u003c/strong\u003eThe authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials:\u0026nbsp;\u003c/strong\u003ePublicly available data sets were analyzed in this study.These data can be found here: https://mimic.mit.edu/docs/iv/, https://physionet.org/content/eicu-crd/2.0/.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003cstrong\u003e：\u003c/strong\u003eThe research involving human participants was approved by the Massachusetts Institute of Technology (Cambridge, MA) and Beth Israel Deaconess Medical Center (Boston, MA). Informed consent for participation was not required for this study by national legislation and institutional requirements. A researcher who has completed the Collaborative Institutional Training Initiative examination (Certification number 71642466 for author Yang) can access this database.The requirement for individual informed consent was waived due to the retrospective and observational nature of the study. The study complied with the ethical standards of the Declaration of Helsinki.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical trial number:\u0026nbsp;\u003c/strong\u003enot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contributions\u003c/strong\u003e\u003cstrong\u003e:\u0026nbsp;\u003c/strong\u003eJY: Writing – original draft, Data curation, Formal analysis. LC: Writing – original draft, Data curation, Formal analysis. XL: Methodology, Supervision, Writing – review \u0026amp; editing. JL: Conceptualization, Funding acquisition, Supervision, Validation, Writing – review \u0026amp; editing. KZ: Validation, Resources, Writing – review \u0026amp; editing. YY: Methodology, Visualization, Writing – review \u0026amp; editing. All authors were involved in writing the paper and had final approval of the submitted and published versions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003eThis study was financially supported by the National Natural Science Foundation of China (Grant No.82360372), the Key Research \u0026amp; Development Program of Guangxi (Grant No. GuiKeAB22080088), the Joint Project on Regional High-Incidence Diseases Research of Guangxi Natural Science Foundation (Grant No. 2023GXNSFDA026023), the First-class Discipline Innovation-driven Talent Program of Guangxi Medical University, the Guangxi Zhuang Autonomous Region Health Commission Self-funded Research Program (Grant No. Z-R20251623), and the Guigang Science and Technology Project (GKJ2203022, GKG2300032).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eSinger M, Deutschman C S, Seymour C W, et al. The third international consensus definitions for sepsis and septic shock (sepsis-3)[J]. JAMA, 2016, 315(8): 801-825. DOI:10.1001/jama.2016.0287.\u003c/li\u003e\n\u003cli\u003eEvans L, Rhodes A, Alhazzani W, et al. Surviving sepsis campaign: international guidelines for management of sepsis and septic shock 2021[J]. Intensive Care Medicine, 2021, 47(11): 1181-1247. DOI:10.1007/s00134-021-06506-y.\u003c/li\u003e\n\u003cli\u003eFleischmann-Struzek C, Mellhammar L, Rose N, et al. Incidence and mortality of hospital- and ICU-treated sepsis: results from an updated and expanded systematic review and meta-analysis[J]. Intensive Care Medicine, 2020, 46(8): 1552-1562. DOI:10.1007/s00134-020-06151-x.\u003c/li\u003e\n\u003cli\u003eMarkwart R, Saito H, Harder T, et al. Epidemiology and burden of sepsis acquired in hospitals and intensive care units: a systematic review and meta-analysis[J]. Intensive Care Medicine, 2020, 46(8): 1536-1551. DOI:10.1007/s00134-020-06106-2.\u003c/li\u003e\n\u003cli\u003eRudd K E, Johnson S C, Agesa K M, et al. Global, regional, and national sepsis incidence and mortality, 1990\u0026ndash;2017: analysis for the global burden of disease study[J]. Lancet, 2020, 395(10219): 200-211. DOI:10.1016/S0140-6736(19)32989-7.\u003c/li\u003e\n\u003cli\u003eSrdić T, Đura\u0026scaron;ević S, Lakić I, et al. From molecular mechanisms to clinical therapy: understanding sepsis-induced multiple organ dysfunction[J]. International journal of molecular sciences, 2024, 25(14): 7770. DOI:10.3390/ijms25147770.\u003c/li\u003e\n\u003cli\u003eTang F, Zhao X L, Xu L Y, et al. Endothelial dysfunction: pathophysiology and therapeutic targets for sepsis-induced multiple organ dysfunction syndrome[J]. Biomedicine \u0026amp; Pharmacotherapy, 2024, 178: 117180. DOI:10.1016/j.biopha.2024.117180.\u003c/li\u003e\n\u003cli\u003eWang L, Chen J, Zhou X, et al. Multiple organ dysfunction and the mortality in sepsis: a multicenter retrospective study[J]. Chinese Medical Journal, 2025, 138(16): 2037-2039. DOI:10.1097/CM9.0000000000003673.\u003c/li\u003e\n\u003cli\u003eLambden S, Laterre P F, Levy M M, et al. The SOFA score\u0026mdash;development, utility and challenges of accurate assessment in clinical trials[J]. Critical Care, 2019, 23(1): 374-383. DOI:10.1186/s13054-019-2663-7.\u003c/li\u003e\n\u003cli\u003eQuintairos A, Pilcher D, Salluh J I F. ICU scoring systems[J]. Intensive Care Medicine, 2023, 49(2): 223-225. DOI:10.1007/s00134-022-06914-8.\u003c/li\u003e\n\u003cli\u003ePellathy T P, Pinsky M R, Hravnak M. Intensive care unit scoring systems[J]. Critical Care Nurse, 2021, 41(4): 54-64. DOI:10.4037/ccn2021613.\u003c/li\u003e\n\u003cli\u003eFleuren L M, Klausch T L T, Zwager C L, et al. Machine learning for the prediction of sepsis: a systematic review and meta-analysis of diagnostic test accuracy[J]. Intensive Care Medicine, 2020, 46(3): 383-400. DOI:10.1007/s00134-019-05872-y.\u003c/li\u003e\n\u003cli\u003eBeam A L, Kohane I S. Big data and machine learning in health care[J]. JAMA, 2018, 319(13): 1317. DOI:10.1001/jama.2017.18391.\u003c/li\u003e\n\u003cli\u003eRajkomar A, Dean J, Kohane I. Machine learning in medicine[J]. New England Journal of Medicine, 2019, 380(14): 1347-1358. DOI:10.1056/NEJMra1814259.\u003c/li\u003e\n\u003cli\u003eJohnson A E W, Ghassemi M M, Nemati S, et al. Machine learning and decision support in critical care[J]. Proceedings of the IEEE, 2016, 104(2): 444-466. DOI:10.1109/JPROC.2015.2501978.\u003c/li\u003e\n\u003cli\u003eShillan D, Sterne J A C, Champneys A, et al. Use of machine learning to analyse routinely collected intensive care unit data: a systematic review[J]. Critical Care, 2019, 23(1): 284-295. DOI:10.1186/s13054-019-2564-9.\u003c/li\u003e\n\u003cli\u003eZhou S, Lu Z, Liu Y, et al. Interpretable machine learning model for early prediction of 28-day mortality in ICU patients with sepsis-induced coagulopathy: development and validation[J]. European Journal of Medical Research, 2024, 29(1): 14-28. DOI:10.1186/s40001-023-01593-7.\u003c/li\u003e\n\u003cli\u003eMuralitharan S, Nelson W, Di S, et al. Machine learning\u0026ndash;based early warning systems for clinical deterioration: systematic scoping review[J]. Journal of Medical Internet Research, 2021, 23(2): e25187. DOI:10.2196/25187.\u003c/li\u003e\n\u003cli\u003eZhang Y, Xu W, Yang P, et al. Machine learning for the prediction of sepsis- related death: a systematic review and meta- analysis[J]. BMC Med Inform Decis Mak. 2023 Dec 11;23(1):283. doi: 10.1186/s12911-023-02383-1.\u003c/li\u003e\n\u003cli\u003eSmith L A, Oakden-Rayner L, Bird A, et al. Machine learning and deep learning predictive models for long-term prognosis in patients with chronic obstructive pulmonary disease: a systematic review and meta-analysis[J]. The Lancet DIGITAL HEALTH, 2023, 5(12): e872-e881. DOI:10.1016/S2589-7500(23)00177-2.\u003c/li\u003e\n\u003cli\u003eMoor M, Rieck B, Horn M, et al. Early prediction of sepsis in the ICU using machine learning: a systematic review[J]. Frontiers in Medicine, 2021, 8: 607952. DOI:10.3389/fmed.2021.607952.\u003c/li\u003e\n\u003cli\u003eYang Z. Predicting sepsis onset in ICU using machine learning models: a systematic review and meta-analysis[J]. BMC Infect Dis. 2023 Sep 27;23(1):635. doi: 10.1186/s12879-023-08614-0.\u003c/li\u003e\n\u003cli\u003eJohnson A E W, Bulgarelli L, Shen L, et al. MIMIC-IV, a freely accessible electronic health record dataset[J]. Scientific Data, 2023, 10(1): 1-10. DOI:10.1038/s41597-022-01899-x.\u003c/li\u003e\n\u003cli\u003eAustin P C, White I R, Lee D S, et al. Missing data in clinical research: a tutorial on multiple imputation[J]. Canadian Journal of Cardiology, 2021, 37(9): 1322-1331. DOI:10.1016/j.cjca.2020.11.010.\u003c/li\u003e\n\u003cli\u003eFira M, Goras L, Costin H N. Evaluating sparse feature selection methods: a theoretical and empirical perspective[J]. Applied Sciences, 2025, 15(7): 3752. DOI:10.3390/app15073752.\u003c/li\u003e\n\u003cli\u003eChowdhury S U, Sayeed S, Rashid I, et al. Shapley-additive-explanations-based factor analysis for dengue severity prediction using machine learning[J]. Journal of Imaging, 2022, 8(9): 229-254. DOI:10.3390/jimaging8090229.\u003c/li\u003e\n\u003cli\u003eRanzani O T, Singer M, Salluh J I F, et al. Development and validation of the sequential organ failure assessment (SOFA)-2 score[J]. JAMA, 2025, 334(23): 2090. DOI:10.1001/jama.2025.20516.\u003c/li\u003e\n\u003cli\u003eP\u0026eacute;rez-Torres D, Merino-Garc\u0026iacute;a P A, Canas-P\u0026eacute;rez I, et al. Real-world inter-observer variability of the sequential organ failure assessment (SOFA) score in intensive care medicine: the time has come for an update[J]. Critical Care, 2023, 27(1): 160-163. DOI:10.1186/s13054-023-04449-y.\u003c/li\u003e\n\u003cli\u003eMoreno R, Rhodes A, Piquilloud L, et al. The sequential organ failure assessment (SOFA) score: has the time come for an update?[J]. Critical Care, 2023, 27(1): 15. DOI:10.1186/s13054-022-04290-9.\u003c/li\u003e\n\u003cli\u003eVincent J L, Bakker J. Blood lactate levels in sepsis: in 8 questions[J]. Current Opinion in Critical Care, 2021, 27(3): 298-302. DOI:10.1097/MCC.0000000000000824.\u003c/li\u003e\n\u003cli\u003eWeinberger J, Klompas M, Rhee C. What is the utility of measuring lactate levels in patients with sepsis and septic shock?[J]. Seminars in Respiratory and Critical Care Medicine, 2021, 42(5): 650-661. DOI:10.1055/s-0041-1733915.\u003c/li\u003e\n\u003cli\u003eIba T, Helms J, Connors J M, et al. The pathophysiology, diagnosis, and management of sepsis-associated disseminated intravascular coagulation[J]. Journal of Intensive Care, 2023, 11(1): 24. DOI:10.1186/s40560-023-00672-5.\u003c/li\u003e\n\u003cli\u003eIba T, Helms J, Levy J H. Sepsis-induced coagulopathy (SIC) in the management of sepsis[J]. Annals of Intensive Care, 2024, 14(1): 148-158. DOI:10.1186/s13613-024-01380-5.\u003c/li\u003e\n\u003cli\u003eSikora A, Zhang T, Murphy D J, et al. Machine learning vs. traditional regression analysis for fluid overload prediction in the ICU[J]. Scientific Reports, 2023, 13(1): 19654. DOI:10.1038/s41598-023-46735-3.\u003c/li\u003e\n\u003cli\u003eBenedetto U, Dimagli A, Sinha S, et al. Machine learning improves mortality risk prediction after cardiac surgery: systematic review and meta-analysis[J]. Journal of Thoracic and Cardiovascular Surgery, 2022, 163(6): 2075-2087.e9. DOI:10.1016/j.jtcvs.2020.07.105.\u003c/li\u003e\n\u003cli\u003eVickers A J, Holland F. Decision curve analysis to evaluate the clinical benefit of prediction models[J]. Spine Journal, 2021, 21(10): 1643-1648. DOI:10.1016/j.spinee.2021.02.024.\u003c/li\u003e\n\u003cli\u003eVan Calster B, Wynants L, Verbeek J F M, et al. Reporting and interpreting decision curve analysis: a guide for investigators[J]. European Urology, 2018, 74(6): 796-804. DOI:10.1016/j.eururo.2018.08.038.\u003c/li\u003e\n\u003cli\u003eBifarin O O. Interpretable machine learning with tree-based shapley additive explanations: application to metabolomics datasets for binary classification[J]. PLOS One, 2023, 18(5): e0284315. DOI:10.1371/journal.pone.0284315.\u003c/li\u003e\n\u003cli\u003eLundberg S M, Erion G, Chen H, et al. From local explanations to global understanding with explainable AI for trees[J]. Nature Machine Intelligence, 2020, 2(1): 56-67. DOI:10.1038/s42256-019-0138-9.\u003c/li\u003e\n\u003cli\u003eBomrah S, Uddin M, Upadhyay U, et al. A scoping review of machine learning for sepsis prediction- feature engineering strategies and model performance: a step towards explainability[J]. Critical Care, 2024, 28(1): 180-196. DOI:10.1186/s13054-024-04948-6.\u003c/li\u003e\n\u003cli\u003eYang H S. Machine learning for sepsis prediction: prospects and challenges[J]. Clinical Chemistry, 2024, 70(3): 465-467. DOI:10.1093/clinchem/hvae006.\u003c/li\u003e\n\u003cli\u003eSlieker R C, Van Der Heijden A A W A, Siddiqui M K, et al. Performance of prediction models for nephropathy in people with type 2 diabetes: systematic review and external validation study[J]. BMJ, 2021: n2134. DOI:10.1136/bmj.n2134.\u003c/li\u003e\n\u003cli\u003eRockenschaub P, Akay E M, Carlisle B G, et al. External validation of AI-based scoring systems in the ICU: a systematic review and meta-analysis[J]. BMC Medical Informatics and Decision Making, 2025, 25(1): 5-15. DOI:10.1186/s12911-024-02830-7.\u003c/li\u003e\n\u003cli\u003eRockenschaub P, Hilbert A, Kossen T, et al. The impact of multi-institution datasets on the generalizability of machine learning prediction models in the ICU[J]. Critical Care Medicine, 2024, 52(11): 1710-1721. DOI:10.1097/CCM.0000000000006359.\u003c/li\u003e\n\u003cli\u003eCiobanu-Caraus O, Aicher A, Kernbach J M, et al. A critical moment in machine learning in medicine: on reproducible and interpretable learning[J]. Acta Neurochirurgica, 2024, 166(1): 14-21. DOI:10.1007/s00701-024-05892-8.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Table","content":"\u003cp\u003eTable 1 is available in the Supplementary Files section.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"european-journal-of-medical-research","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ejmr","sideBox":"Learn more about [European Journal of Medical Research](http://eurjmedres.biomedcentral.com)","snPcode":"40001","submissionUrl":"https://submission.nature.com/new-submission/40001/3","title":"European Journal of Medical Research","twitterHandle":"@BioMedCentral","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"MODS, Sepsis, Machine learning, Risk prediction, External validation","lastPublishedDoi":"10.21203/rs.3.rs-8681490/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8681490/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eMultiple organ dysfunction syndrome (MODS) is a key determinant of prognosis in sepsis, yet conventional severity scoring systems based on linear assumptions and static variables fail to capture complex nonlinear physiological disturbances and dynamic inter organ interactions. Although machine learning has shown promise in outcome prediction among critically ill patients, studies focusing on MODS while ensuring interpretability and external validation remain limited.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eThis retrospective cohort study used data from the Medical Information Mart for Intensive Care IV and the eICU Collaborative Research Database. Adult patients meeting Sepsis 3 criteria and admitted to the ICU for the first time were included. Feature selection was performed using least absolute shrinkage and selection operator regression. Multiple machine learning models were developed, including logistic regression, random forest, gradient boosting machine, extreme gradient boosting, Light Gradient Boosting Machine, artificial neural networks, and support vector machines. Model performance was evaluated using the area under the receiver operating characteristic curve, calibration curves, and decision curve analysis. Shapley additive explanations were used for model interpretation, and external validation was conducted in an independent eICU cohort.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eAmong 23,018 patients with sepsis, 4,931 (21.4%) developed MODS during ICU hospitalisation. All models showed acceptable discrimination, with LightGBM achieving the highest AUC (0.829), followed by GBM (0.824), random forest (0.823), and XGBoost (0.822). Logistic regression and elastic net showed moderate performance (both AUC 0.802), the neural network showed intermediate discrimination (AUC 0.803), whereas support vector machines (0.759) and k nearest neighbours (0.727) performed less well. LightGBM demonstrated stable discrimination, good calibration, and greater clinical net benefit in both internal testing and external validation. SHAP analysis identified the Sequential Organ Failure Assessment score, respiratory rate, lactate, coagulation indices including international normalised ratio, acid base status, and vasoactive agent use as key predictors with pronounced nonlinear effects.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eAmong the evaluated models, the gradient boosting based LightGBM showed the most robust performance for predicting MODS risk in sepsis, supporting early risk stratification and individualised ICU management. Prospective multicentre studies are warranted to confirm its clinical impact.\u003c/p\u003e","manuscriptTitle":"Development and External Validation of a Machine Learning–Based Model for Early Prediction of Multiple Organ Dysfunction Syndrome in Critically Ill Patients with Sepsis","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-08 16:33:50","doi":"10.21203/rs.3.rs-8681490/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-04-13T17:18:30+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-20T20:43:45+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-19T19:00:06+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-05T03:40:29+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"270606405683585545398495635506496630226","date":"2026-03-05T02:48:02+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-03T15:51:13+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"234735212041526183829125768275525033197","date":"2026-03-01T16:33:48+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"207709226041603960717398129441545890553","date":"2026-02-28T13:39:42+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"201074233371245001327804846880583831420","date":"2026-02-27T14:14:21+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-02-27T14:01:41+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-02-02T11:21:37+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-02-02T11:13:24+00:00","index":"","fulltext":""},{"type":"submitted","content":"European Journal of Medical Research","date":"2026-01-23T17:16:21+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"european-journal-of-medical-research","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ejmr","sideBox":"Learn more about [European Journal of Medical Research](http://eurjmedres.biomedcentral.com)","snPcode":"40001","submissionUrl":"https://submission.nature.com/new-submission/40001/3","title":"European Journal of Medical Research","twitterHandle":"@BioMedCentral","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"b7c7ef86-c313-40fd-8e73-5c194af3a335","owner":[],"postedDate":"March 8th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-08T07:53:33+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-08 16:33:50","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8681490","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8681490","identity":"rs-8681490","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}