Time-Series modeling for predicting mortality risk in intensive care unit patients with pulmonary inflammation.

Objective: This study aims to evaluate dynamic changes in pneumonia-related mortality risk among ICU patients using time-series models and develop interpretable predictive models for early mortality assessment. Methods: Data were obtained from the MIMIC-IV and eICU Collaborative Research Database. The MIMIC-IV cohort was split into training and testing sets (7:3 ratio), with external validation conducted using the eICU cohort. Missing data were handled using forward filling and multiple imputation. LSTM was used to extract temporal features, which were then entered into five machine-learning classifiers: logistic regression, random forest (RF), Light Gradient Boosting Machine, extreme gradient boosting, and multilayer perceptron. Hyperparameter tuning was performed using cross-validation and grid search. Model performance was assessed using accuracy, precision, recall, F1 score, balanced accuracy, Matthews correlation coefficient (MCC), AUC, Brier score, and Youden's index. The best-performing model was interpreted using Shapley Additive Explanations and Local Interpretable Model-agnostic Explanations. Results: A total of 4,205 patients were included, including 2,751 patients from MIMIC-IV and 1,454 patients from the eICU cohort. The 7-day mortality rates were 7.92% and 11.69%, respectively. The hybrid LSTM-RF model performed exceptionally, achieving an accuracy of 0.961, precision of 0.938, recall of 0.823, F1 score of 0.874, AUC of 0.940, Balanced ACC of 0.914, and MCC of 0.862. The LSTM-RF model yielded a Brier score of 0.020 and Youden's index of 0.818. Validation results showed an AUC of 0.826, a Brier score of 0.070, and a Youden's index of 0.590. Conclusion: Integrating random forest with LSTM-derived features improves early mortality risk prediction in ICU patients with pneumonia.