Mortality Prediction in Diabetic Sepsis

Evidence

This retrospective cohort study analyzed electronic medical records in the Medical Information Mart for Intensive Care IV database for adult intensive care unit patients with concurrent diagnoses of sepsis and diabetes mellitus. The database contains records from patients treated at Beth Israel Deaconess Medical Center in Boston, Massachusetts. The reported analytic cohort comprised 7,716 patients, of whom 1,357 (17.6%) experienced in-hospital mortality. The investigators framed the work as prognostic model development and comparison for mortality during the same hospitalization, not as an evaluation of a treatment or clinical intervention. pmid:42907983

The comparison included logistic regression, K-nearest neighbors, random forest, gradient boosting decision tree, XGBoost, and LightGBM, alongside the Sequential Organ Failure Assessment score. Investigators assessed discrimination, calibration, and precision-recall performance. Hyperparameters were selected with grid search and five-fold cross-validation, and the records were randomly split into training and validation cohorts for internal validation. This sequence describes computational development and testing within one retrospective database, with routine-record data as the underlying source. pmid:42907983

Gradient Boosting Decision Tree had the reported highest overall predictive performance among the evaluated machine-learning models and exceeded the conventional Sequential Organ Failure Assessment score in the study comparison. Its area under the receiver operating characteristic curve was 0.792 (95% confidence interval, 0.769–0.814). The reported result measures discrimination—how the model separated patients who did and did not experience in-hospital mortality in this dataset—and does not itself measure changes in care, patient outcomes, or clinical benefit. pmid:42907983

In the validation cohort, the Gradient Boosting Decision Tree model had reported accuracy of 0.700, sensitivity of 0.767, specificity of 0.686, precision of 0.342, and F1-score of 0.473. These threshold-dependent measures characterize the model under the study's validation setup; they should be read alongside the discrimination, calibration, and precision-recall assessments rather than as a single universal performance statement. The study did not test the model in an independently assembled external cohort. pmid:42907983

For model interpretation, the investigators applied SHapley Additive exPlanations analysis to the best-performing Gradient Boosting Decision Tree model to examine the relative importance and directional impact of predictors. The Charlson Comorbidity Index was identified as the top predictor, with higher scores associated with increased mortality risk in the model, and older age was associated with a higher likelihood of mortality. These are model-linked associations in the analyzed cohort; they do not establish that comorbidity burden or age causes in-hospital mortality. pmid:42907983

Analysis — Internal prognostic performance

The observed pattern is consistent with a retrospective, selected intensive care unit cohort in which a nonlinear Gradient Boosting Decision Tree model captured relationships in routinely collected variables better than the other evaluated approaches and the Sequential Organ Failure Assessment score on internal testing. Its reported discrimination is relevant to comparative model performance in these records, while the threshold-dependent measures show that different aspects of classification performance did not move identically. The Charlson Comorbidity Index and age results add interpretability about features contributing to predictions, but feature importance is not a causal ranking of modifiable clinical drivers. pmid:42907983

Because training, tuning, and validation were all conducted within the same single-center database, the result is an internal assessment rather than evidence that performance will persist across institutions, data-collection practices, case mix, or management patterns. The retrospective design cannot establish causation. In particular, the reported predictor–mortality relationships reflect the study population, available measurements, and selection process for intensive care unit records with both sepsis and diabetes mellitus. The study therefore reports an association between model inputs and mortality classification, not a demonstrated effect of using the model on patients. pmid:42907983

Limitations

Selection and information bias are plausible because the analysis relied on routinely collected intensive care unit data in a retrospective observational design. The authors state that pre-intensive-care-unit diabetes control, infection source, microbiological characteristics, and detailed treatment strategies were not fully incorporated into the models, although those factors may influence predictor values and mortality risk. Differences in clinical management across intensive care units and over time within the database may also affect the measured relationships and model calibration. pmid:42907983

The single-center Medical Information Mart for Intensive Care IV source constrains direct generalization because case mix, practice patterns, and data-collection processes can differ across institutions and healthcare systems. Random splitting for internal validation can assess performance within the source data but does not resolve uncertainty about external performance or calibration. The authors also identify interpretability as a challenge for complex models such as Gradient Boosting Decision Tree despite use of SHapley Additive exPlanations values. pmid:42907983

The reported cohort construction includes an initial sepsis identification process and exclusions for shorter intensive care unit stays, indicating that the analyzed population was defined through database eligibility criteria rather than sampled to represent all people with sepsis and diabetes mellitus. Accordingly, the reported metrics are bounded by the outcome definition, variables, model-building choices, and selected records described here. No patient-specific conclusion, prediction, or care recommendation follows from this retrospective internal model comparison. pmid:42907983