Machine learning-based prediction model for in hospital mortality in sepsis using multimodal structured data fusion:an emergency department cohort study with interpretability analysis
Sun Chengcheng
Wang Ping
Cui Dongliang
Yang Siqi
Xiang Tao
Abstract：Objective This study aims to develop and validate a machine learning model integrating multimodal structured data for predicting in-hospital mortality in sepsis patients within an emergency department cohort.Methods The single-center retrospective cohort study included 177 patients with sepsis in the emergency department form The Third People's Hospital of Chengdu between September 2021 and August 2023.Multimodal clinical data were systematically collected,with missing values imputed using median for continuous variables and mode for categorical variables,with missing indicator variables incorporated.SOFA and qSOFA scores,along with LightGBM,XGBoost,and random forest models,were constructed.Model performance was assessed using AUROC,AUPRC,and Brier score,with SHAP analysis for interpretability.Results LightGBM demonstrated superior discrimination ability(AUROC0.893,95%CI 0.774-0.990),significantly outperforming SOFA(0.535)and qSOFA(0.586)scores(P<0.05).Random forest exhibited optimal performance in positive class prediction(AUPRC 0.687)and calibration(Brier score 0.098).SHAP analysis revealed key predictors:lactate level(>2.0 mmol/L,SHAP value+0.32),PaO/FiO2 ratio(<200 mmHg,SHAP value+0.28),vasopressor use(SHAP contribution+0.25),and age(risk increase of 0.015 per year).The synergistic effect of lactate and PaO/FiO2 ratio yielded a combined SHAP value of+0.60,significantly exceeding the sum of individual effects.Conclusions Machine learning models integrating multimodal data significantly improve the accuracy of in-hospital mortality prediction in sepsis compared to traditional scoring systems.The model's interpretability provides a practical tool for risk stratification,supporting precision decision-making in emergency care and demonstrating substantial translational potential.
Keywords：SepsisIn-hospital mortalityMachine learningMultimodal dataPrecision emergency care
Publication Date：2026-03-10
Online Publishing Date：2026-03-27(First online date of this platform, not the publication date of the document)
Pages：8( 199-206 )