Deep Learning for Early Prediction of Adverse Events in Intensive Care Units Using Electronic Health Records: A Methodological Review

Abstract

Purpose To provide a methodological synthesis of deep learning (DL) approaches for early prediction of preventable adverse events (AEs) in intensive care units (ICUs) using electronic health record (EHR) data, and to examine how intrinsic EHR data challenges affect model development, predictive performance, and clinical applicability. Methods This review was designed as a methodological synthesis rather than a quantitative meta-analysis or statistical comparison. Seven databases were systematically searched to identify studies on early AE prediction in ICU settings using DL models. Eligible studies were screened based on predefined criteria, and data were extracted and synthesized qualitatively to analyze EHR data characteristics, DL architectures, and associated methodological challenges. Results A total of 31 studies comprising 33 predictive models were included. These studies covered multiple preventable AEs, diverse ICU EHR datasets, and a range of DL architectures. Across studies with different prediction windows, reported AUROC values ranged from 0.75 to 0.95. Four studies used oversampling to address class imbalance. Despite the increasing use of advanced and hybrid architectures, three recurring methodological challenges remained: subtle early-stage risk signals, class imbalance, and heterogeneity. Overall, the effectiveness of DL-based AE prediction depends less on architectural novelty alone and more on whether models are explicitly designed around the data conditions under which preventable AEs emerge. Conclusion Future studies should move beyond architecture-centered optimization and develop methodological strategies, such as multimodal learning, bias-aware modeling, and personalized prediction strategies. More attention should also be given to external and prospective validation, safe clinical use, and real-world effects on patient care and preventable harm.