Machine Learning Outperforms SOFA in Diabetic Sepsis Mortality Prediction

In a retrospective cohort study published in BMC Medical Informatics and Decision Making, researchers demonstrated that machine learning algorithms can predict in-hospital mortality among intensive care unit patients with both sepsis and diabetes mellitus more accurately than the traditional Sequential Organ Failure Assessment score. The study, conducted by Zongxin Shen, Haijiao Zhang, and Yu Wang, utilized data from the Medical Information Mart for Intensive Care database to evaluate six distinct predictive models. The Gradient Boosting Decision Tree model emerged as the most effective tool, achieving an area under the receiver operating characteristic curve of 0.792, a metric that significantly surpassed the performance of the conventional SOFA score, which recorded an area under the curve of 0.617. This finding suggests that advanced analytical techniques may offer a more nuanced approach to risk stratification in this specific and high-risk patient population.

The clinical context for this research is driven by the substantial morbidity and mortality associated with sepsis when complicated by diabetes. Sepsis affects approximately 48.9 million individuals annually, resulting in 11 million deaths globally. Among those diagnosed with sepsis, the prevalence of diabetes mellitus ranges from 17% to 30%, highlighting the frequent co-occurrence of these conditions. The interaction between the two is complex; diabetes can alter immune function and modify the host response to infection, while the dysregulated glucose metabolism characteristic of diabetes may exacerbate the metabolic derangements seen in sepsis. These physiological interactions create unique challenges for diagnosis, treatment, and prognostication, necessitating more sophisticated analytical approaches to understand their combined impact on patient outcomes.

Traditional scoring systems, such as the SOFA score, have long been the standard for risk stratification in sepsis. However, these tools often rely on a limited set of variables and may not fully capture the intricate interactions between sepsis and diabetes. Previous studies have shown conflicting results regarding the impact of pre-existing diabetes on sepsis outcomes, with some reporting increased mortality and others finding no significant difference or even a potential protective effect. To address these gaps, the researchers sought to systematically compare multiple machine learning algorithms against the SOFA score using a broad panel of routinely available clinical variables. The goal was to determine if these models could provide a more accurate and interpretable prediction of in-hospital mortality for this specific subgroup of critically ill patients.

The study cohort comprised 7,716 adult patients admitted to the ICU with concurrent diagnoses of sepsis and diabetes mellitus. Data were extracted from the MIMIC-IV database, which contains medical records for over 65,000 patients treated at the Beth Israel Deaconess Medical Center in Boston, Massachusetts, between 2008 and 2022. The researchers applied strict inclusion and exclusion criteria to ensure a homogeneous study population. Patients were included if they met Sepsis-3 criteria for sepsis and had a pre-existing diagnosis of diabetes confirmed prior to ICU admission. Exclusions were made for patients under 18 years of age, pregnant women, those with AIDS, malignant cancer, or metastatic solid tumors, and those with an ICU length of stay less than 24 hours. Additionally, patients with more than 20% missing data in the variables of interest were excluded to preserve data quality.

For the predictive models, the researchers focused on clinical data collected during the first 24 hours following ICU admission. This time window was chosen because it reflects a clinically meaningful early assessment period during which intensivists make key decisions regarding triage and organ support. Restricting predictors to this interval allows for early risk stratification before substantial post-admission deterioration occurs, thereby reducing the risk of incorporating intermediate outcomes along the causal pathway. The extracted information included demographics, comorbidities, laboratory test results, and disease severity indicators. The dataset was randomly split into a training set comprising 70% of the patients and a validation set comprising the remaining 30%. Missing data were handled using multiple imputation techniques based on the random forest algorithm.

Six machine learning algorithms were evaluated for their predictive performance: Logistic Regression, K-Nearest Neighbors, Random Forest, Gradient Boosting Decision Tree, XGBoost, and LightGBM. These models were compared against the traditional SOFA score to assess their ability to discriminate between patients who survived and those who died during their hospital stay. Model discrimination was quantified using the area under the receiver operating characteristic curve, while calibration was assessed by comparing predicted probabilities with actual outcomes. To enhance the interpretability of the best-performing model, the researchers employed SHapley Additive exPlanations values to elucidate the relative importance of predictor variables. This approach allows for a transparent ranking of the clinical features that most significantly influence the model’s predictions.

The results indicated that the Gradient Boosting Decision Tree model exhibited superior predictive performance, achieving an area under the curve of 0.792 with a 95% confidence interval of 0.769 to 0.814. This performance surpassed that of other machine learning algorithms, including XGBoost and LightGBM, which achieved areas under the curve of 0.785 and 0.784, respectively. All machine learning models outperformed the SOFA score, which had an area under the curve of 0.617. The non-overlapping confidence intervals between the top-performing models and the SOFA score indicate statistically significant improvements in predictive ability. The Gradient Boosting Decision Tree model also showed the highest average precision, further confirming the superior predictive capability of tree-based ensemble models in this context.

SHAP analysis provided valuable insights into the hierarchical importance of various clinical and demographic features contributing to the model’s predictions. The Charlson Comorbidity Index was identified as the top predictor, with higher scores indicating increased mortality risk. Age also emerged as a key factor, with older patients showing a higher likelihood of mortality. Elevated SOFA scores, maximum lactate levels, increased maximum creatinine levels, and prolonged maximum prothrombin time were also important predictors associated with increased risk. The analysis highlighted lower minimum albumin levels and decreased minimum platelet counts as significant factors linked to higher mortality. The relationship between maximum white blood cell count and mortality was complex, indicating that both very high and very low counts could be associated with increased risk.

The superior performance of the machine learning models compared with the SOFA score suggests that these approaches capture more nuanced relationships in the data. This is particularly relevant in diabetic sepsis, where complex physiological interactions may not be fully reflected by traditional scores. The SHAP-based explanations reflect statistical patterns in the model that may at times appear counterintuitive clinically, highlighting the non-linear and multifactorial nature of the associations between clinical parameters and mortality risk. The variables used in the models are routinely collected in the ICU, which facilitates their potential integration into existing electronic health record systems as an automated decision-support tool for early risk stratification.

Despite the promising results, several important limitations must be acknowledged. The study was conducted using a single-center database, and case-mix, practice patterns, and data collection processes may differ substantially in other institutions and healthcare systems. The models were evaluated only by internal validation with a random split into training and validation cohorts, and their performance may change when applied to external populations. The retrospective observational design and reliance on routinely collected data mean that factors such as pre-ICU diabetes control, infection source, and detailed treatment strategies were not fully incorporated. Consequently, the models should be interpreted as capturing statistical associations rather than causal relationships. Further multi-center validation studies and prospective trials are warranted to confirm the generalizability of these results and to explore the clinical impact of implementing such models in diverse healthcare settings.

Subject of Research: Machine learning prediction of in-hospital mortality in diabetic sepsis patients

Article Title: Application of machine learning prediction models for in-hospital mortality in diabetic sepsis patients: a study based on the MIMIC IV database

Article References: Shen, Z., Zhang, H., & Wang, Y. (2026). Application of machine learning prediction models for in-hospital mortality in diabetic sepsis patients: a study based on the MIMIC IV database. BMC Medical Informatics and Decision Making, 26(1), Article 426. https://doi.org/10.1186/s12911-026-03580-4

Image Credits: AI Generated

DOI: 10.1186/s12911-026-03580-4

Keywords: machine learning, sepsis, diabetes, mortality prediction, MIMIC-IV, GBDT, SOFA score, intensive care, prognostication, SHAP analysis, Application, machine