Diabetic neuropathy carries a substantial systemic disease burden, but short-term mortality risk among affected intensive care unit (ICU) patients is not well characterized. The authors evaluated whether temporal clinical information captured during the first 24 hours of ICU care improves prediction of subsequent in-hospital death beyond conventional severity scores and static clinical summaries.
This retrospective observational study used the MIMIC-IV v3.1 critical care database. The investigators applied a 24-hour landmark design: only patients who were alive and still hospitalized at 24 hours after ICU admission were included, and the outcome of interest was later in-hospital death. All analyses used de-identified MIMIC-IV data accessed through PhysioNet under standard credentialing and data-use agreements.
Patients were adults (>18 years) with documented diabetic neuropathy identified in MIMIC-IV. After applying the 24-hour landmark, the final analytic cohort comprised 1,347 patients, of whom 83 experienced in-hospital death after the landmark (6.16%). The dataset was partitioned into an 80% development set and a locked 20% test set. All model selection, hyperparameter tuning, calibration, and threshold choices were restricted to the development set; the locked test set was used only for final evaluation.
The study compared predictor sets that included standard severity scores and static clinical summaries with richer temporal representations derived from the first 24 hours of ICU care. The temporal features were intended to capture dynamic trends and repeated measurements within the landmark period. Details of specific variables and preprocessing steps were described in the source but are not restated here beyond noting the distinction between static and temporal predictor representations.
Three supervised algorithms were trained and evaluated: logistic regression, random forest, and XGBoost. Feature selection, hyperparameter optimization, calibration procedures, and threshold selection were performed using only the development data. Cross-validated performance on development data guided model selection for interpretation. The locked 20% test set provided unbiased estimates of final performance.
On the held-out locked test set, the random forest model—selected based on development cross-validated PR-AUC—achieved an area under the receiver operating characteristic curve (AUROC) of 0.851 with a 95% confidence interval of 0.765–0.924. The same model had a precision-recall AUC (PR-AUC) of 0.339 and a Brier score of 0.051. XGBoost and logistic regression achieved AUROCs of 0.847 and 0.806, respectively. These results indicate competitive discrimination across the evaluated methods, with random forest performing best on the development and test metrics reported.
In a strictly nested post hoc analysis, the investigators evaluated the incremental value of adding temporal predictors to models that included only severity scores and static predictors. Across all three algorithms, inclusion of the full temporal representation increased discrimination. For random forest specifically, AUROC improved from 0.815 using severity and static predictors to 0.870 when the full set of temporal features was included. This finding supports the assertion that first-day temporal information adds prognostic value for post-landmark mortality prediction in this patient population.
The study demonstrates that first-day temporal clinical information can improve discrimination for 24-hour landmark prediction of in-hospital mortality among ICU patients with diabetic neuropathy, compared with models using only static summaries and severity scores. However, the authors emphasize that external validation is required before clinical deployment. The report does not claim prospective utility or real-time clinical impact without further validation. Specific limitations inherent to secondary analysis of a single-center aggregated database (MIMIC-IV) and potential selection biases related to the landmark design are acknowledged by the authors in the source.
This research used de-identified patient data from MIMIC-IV v3.1 accessed via PhysioNet after completion of required credentialing and training. Because the work constituted secondary analysis of de-identified records with no direct participant contact, no additional institutional review board approval or individual informed consent was required. The authors declared no competing interests. Patient-level MIMIC-IV records cannot be redistributed by the authors; the analysis code may be obtained from the corresponding author upon reasonable request and subject to PhysioNet data-use regulations.
In this MIMIC-IV cohort of 1,347 ICU patients with diabetic neuropathy and a 24-hour landmark design, adding temporal features from the first 24 hours of ICU care improved model discrimination for subsequent in-hospital mortality across logistic regression, XGBoost, and random-forest models. Random forest had the best development PR-AUC and achieved AUROC 0.851 (95% CI 0.765–0.924) on the locked test set. The authors conclude that first-day temporal information provides additional prognostic value but stress that external validation is necessary before consideration for clinical use.