Development and validation of machine learning models to predict mortality in elderly patients with chronic critical illness: a multi-cohort study using MIMIC, eICU, and a Chinese dataset
Year: 2026
Type: article
Study design: Observational Study
Abstract: Background and objective Patients with chronic critical illness (CCI) often require prolonged mechanical ventilation and other ICU interventions. The significant morbidity and mortality linked to CCI make the early identification of risk factors an imperative clinical objective. Methods Data from elderly CCI patients were sourced from the MIMIC-III, MIMIC-IV and eICU databases. An independent external validation cohort of 308 patients was additionally collected from a Chinese hospital. PSM was applied to the eICU cohort to address class imbalance. Feature selection was conducted using LASSO regression and the Boruta algorithm. Twelve machine learning models were developed and validated. To identify the optimal model, we evaluated AUC, accuracy, F1-score, sensitivity, and specificity. Calibration and clinical utility were further assessed using calibration plots and DCA, with SHAP employed for model interpretation. Results A total of 1,628 elderly CCI patients were included across three cohorts (MIMIC derivation: 928; eICU: 392; Chinese local: 308). Thirty variables with significant differences between survivors and non-survivors were identified in the derivation cohort. Eight key predictors were selected: age, AKI stage, hemoglobin, platelets, RDW, potassium, body temperature, and PaCO 2 , all measured on ICU day 14. In the derivation cohort, the top-performing models were GBDT (AUC = 0.872), LGBM (AUC = 0.870), and logistic regression (LR, AUC = 0.865). The LR model was selected as the final model based on its superior clinical net benefit and consistent performance in the eICU PSM cohort (AUC = 0.819) and the full eICU cohort (AUC = 0.719). In the Chinese local validation cohort, the LR model maintained acceptable discrimination (AUC = 0.685) with stable recall, while ensemble models showed further performance degradation. SHAP analysis revealed that lower body temperature, elevated RDW, higher potassium, lower platelet count, and higher PaCO 2 were the strongest predictors of increased mortality risk, with age and AKI stage also contributing. Conclusion The LR model may assist in risk stratification for in-hospital mortality in elderly CCI patients, with consistent discrimination across U.S. and Chinese populations. Its interpretability and generalizability support its potential utility as a clinical tool for identifying high-risk individuals to guide early intervention.
Source: Frontiers in Medicine
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Keywords: chronic critical illness, in-hospital mortality prediction, elderly patients, risk stratification, mortality risk factors, acute kidney injury, body temperature, red cell distribution width, serum potassium, platelet count, PaCO2, logistic regression, SHAP, external validation, mimics, eICU, Chinese patients, generalizability across populations, class imbalance
Open Access status: gold
Estimated APC paid: $3,801
OpenAlex ID: W7236681976
