1Research Scholar, Department of Statistics, Malwanchal University, Indore, MP, India. Email: magicofkaran@gmail.com
2Professor, Department of Statistics, Malwanchal University, Indore, MP, India
3Assistant Professor- Statistician, Department of Community Medicine, Government Medical College, Nashik, Maharashtra, India
4Associate Professor, JBMGMC Nandurbar, Maharashtra, India
5Assistant Professor, SBHGMC Dhule, Maharashtra, India
6Officer Biomedical Statistician, Tata Main Hospital, Jamshedpur, Jharkhand, India
Corresponding Author: Karan Jain, Research Scholar, Department of Community Medicine, Malwanchal University, Indore, MP, India. Email: magicofkaran@gmail.com
Received: 16th Dec, 2025; Revised: 8th Feb 2026; Accepted: 12th Feb, 2026; Available Online: 28th Feb, 2026
Background: Sepsis remains a leading cause of in-hospital mortality worldwide, and early identification of high-risk patients is essential for timely intervention. Traditional severity scoring systems such as SOFA and APACHE II have limited predictive accuracy in heterogeneous clinical settings. Machine-learning (ML) approaches offer the potential to improve mortality prediction by modeling complex, nonlinear relationships in routinely collected clinical data.
Methods: This retrospective observational study included 500 adult patients admitted with sepsis to a tertiary care hospital. In-hospital mortality was the primary outcome. Demographic characteristics, clinical variables, laboratory parameters, comorbidities, treatment-related factors, and established severity scores were extracted from medical records. Multiple ML models, including logistic regression, naïve Bayes, k-nearest neighbors, support vector machine, random forest, AdaBoost, and extreme gradient boosting (XGBoost), were developed and evaluated. Model performance was assessed using the area under the receiver operating characteristic curve (AUROC), calibration measures, and decision curve analysis, and was compared with traditional severity scores.
Results: Of the 500 patients, 170 (34%) died while 330 (66%) survived during hospitalization. Non-survivors were older and had higher severity scores and worse physiological and laboratory parameters at admission. Among the ML models, XGBoost demonstrated the best performance, achieving the highest AUROC and superior calibration compared with other ML algorithms and traditional scores. The XGBoost model outperformed SOFA, APACHE II, and NEWS2 in predicting in-hospital mortality. Feature importance analysis identified serum lactate, SOFA score, renal dysfunction, hypotension, vasopressor requirement, thrombocytopenia, and age as key predictors of mortality.
Conclusion: Machine-learning models, particularly gradient-boosting approaches, provide more accurate prediction of in-hospital mortality in sepsis than conventional severity scores. These findings support the potential role of ML-based tools in early risk stratification and personalized decision support for sepsis care.
Keywords: Sepsis; In-hospital mortality; Machine learning; XGBoost; Severity scoring systems; Mortality prediction; Clinical decision support
How to cite this article: Jain K, Bharate PB, Patil A, Wadde S, Kinge A, Bagul D, Prediction of In-Hospital Mortality in Sepsis Using Advanced Predictive Models: Comparison with Traditional Severity Scores. Int J Drug Deliv Technol. 2026; 16(2): 435-441; DOI: 10.25258/ijddt.16.2.48
Source of support: Nil.
Conflict of interest: None