Deep learning-based in-hospital mortality prediction using long-term sequential data in ICU patients: a multi-center validation study.

Introduction: Intensive Care Unit (ICU) patients are at high risk of acute clinical deterioration and in-hospital mortality. Accurate prediction of in-hospital mortality is essential for optimizing clinical decision-making and resource allocation. However, most existing models rely primarily on short-term physiological data from a single hospitalization and lack adequate external validation. This study develops a deep learning-based model that integrates long-term temporal clinical sequences derived from patients' historical ICU records with diagnostic information from the current ICU admission to predict in-hospital all-cause mortality during the current hospitalization. The proposed framework is applicable to both patients with a single ICU admission and those with multiple ICU admissions. Methods: We used current and historical International Classification of Diseases (ICD) codes, temporal features, and basic demographics to construct the prediction models. Single-admission patients were represented as one-event sequences, while repeated-admission patients were represented as longitudinal diagnostic trajectories. Model training and internal validation were conducted based on the MIMIC-IV database, while the eICU and MIMIC-III databases were used for external independent testing. Five deep learning (DL) models, including Informer and Transformer, were constructed for comparative analysis. On this basis, a stacked model integrating these five base models was further developed. The Area Under the Receiver Operating Characteristic Curve (AUROC), calibration curves, and clinical epidemiological analyses involving meta-analyses across different datasets were used to comprehensively evaluate the model's predictive performance, generalizability, and specific clinical application value. Results: A total of 22,176 patients from Medical Information Mart for Intensive Care IV (MIMIC-IV), MIMIC-III, and eICU were included, and five deep learning models as well as one stacked ensemble model were constructed. In the internal validation set, the AUROC of most deep learning models exceeded 0.89, with the exception of the RNN. Among them, the Informer model achieved the best internal performance, with an AUROC of 0.95 (95% CI [0.928-0.967]). In the dual external validation across different databases, model performance exhibited apparent heterogeneity across cohorts. Although no single model consistently outperformed others, the stacked ensemble model showed relatively competitive overall predictive accuracy and stable generalizability. Conclusions: Compared with traditional models based on short-term biochemical indicators, the proposed deep learning models and the stacked ensemble model demonstrated competitive predictive performance for in-hospital mortality. This study demonstrates the potential of deep learning for ICU mortality prediction and proposes a novel framework for integrating longitudinal diagnostic trajectories into predictive modeling.