Sepsis-associated acute kidney injury (SA-AKI) is a frequent, severe complication among critically ill patients and carries a poor prognosis. Patients with diabetes are at elevated risk of adverse outcomes due to increased infection susceptibility, immune dysfunction, and renal vulnerability. Existing mortality prediction models specifically for patients who have both diabetes and SA-AKI are limited. The reported study therefore aimed to develop and validate an interpretable machine learning model for early prediction of in-hospital mortality in this high-risk population, with a focus on variables available within the first 24 hours after ICU admission.
The investigators identified a cohort of 6,929 patients with concurrent diabetes and SA-AKI from the Medical Information Mart for Intensive Care IV (MIMIC‑IV) database. The cohort was randomly divided into training and validation sets in a 7:3 ratio. An external dataset, the eICU Collaborative Research Database, was used later for independent testing of the final model. The study timeframe, inclusion/exclusion criteria, and detailed cohort flow were reported in the source.
Ninety-four candidate variables were extracted from the first 24 hours after ICU admission. These variables encompassed demographics, diagnostic information, clinical parameters, laboratory results, and medication records. To reduce dimensionality and identify the most informative features, the authors used recursive feature elimination (RFE). RFE yielded a final set of 32 key predictors that were clinically accessible and routinely collected in ICU settings. Important predictors included urine output rate, platelet count, lactate, weight, blood glucose, SOFA score, arterial pH, blood urea nitrogen, vital signs, coagulation indices, and vasopressor use, among others.
Twelve different machine learning algorithms were developed and compared using the training set. The candidate models included diverse approaches (the source lists the count of algorithms but does not enumerate all models). The optimal-performing algorithm was a categorical boosting approach, referred to as CatBoost in the source. Model development used the selected predictors and standard training procedures as described in the study. The optimal model was then evaluated on the held-out validation set and subsequently on the external eICU dataset.
To increase transparency and clinical interpretability, the authors applied SHapley Additive exPlanations (SHAP) to the final CatBoost model. SHAP values were used to quantify the contribution of each predictor to individual predictions and to the model overall. This interpretability approach allows clinicians to see which features most strongly influenced risk estimates and supports more informed clinical reasoning when model outputs are used for decision support.
On the internal validation set, the CatBoost model demonstrated good discrimination and clinically relevant performance metrics. Reported results included an area under the receiver operating characteristic curve (AUC) of 0.828. Additional metrics were: accuracy 70.9%, sensitivity 78.7%, specificity 69.0%, F1 score 0.509, positive predictive value (PPV) 33.7%, and negative predictive value (NPV) 93.1%. These metrics indicate the model had higher sensitivity and NPV, meaning it was more effective at identifying survivors versus producing false negatives, while PPV was lower, reflecting that positive predictions were less frequently true positives in this cohort.
External testing was performed using data from the eICU Collaborative Research Database to assess generalizability. The CatBoost model retained good discriminatory ability on the external dataset, with an AUC of 0.793. The source reports this external validation result as evidence that the model's predictive performance was maintained across a separate multi-center critical care database.
The authors deployed the final CatBoost model as a web-based tool to facilitate bedside risk estimation and early risk stratification for diabetic patients with SA-AKI. By using 32 routinely available ICU variables and applying interpretable ML methods such as SHAP, the model is positioned to support clinical decision-making without relying on obscure inputs. The study suggests the tool may help identify high-risk patients early after ICU admission and thus inform monitoring intensity, diagnostic evaluation, and therapeutic decisions.
The abstract highlights the model’s development, internal validation, and successful external testing, indicating good predictive performance for in-hospital mortality in this specific patient population. The source emphasizes interpretability (SHAP) and accessibility of predictors. Detailed information on certain methodological aspects, such as the full list of the twelve compared algorithms, exact training procedures, calibration metrics, and prospective clinical impact assessment, are not expanded in the abstract and would require consultation of the full article. In conclusion, a CatBoost-based machine learning model using 32 clinically accessible features showed good discrimination (AUC 0.828 internally; AUC 0.793 externally) and was made available as a web tool to support early risk stratification in diabetic patients with SA-AKI.