Performance Metrics of Deep Learning Models in EHR Analysis: A Systematic Review

Abstract

The rapid growth of Electronic Health Records (EHRs), generating 30% of global data with an annual increase of 36%, presents both opportunities and challenges for healthcare analytics. Traditional methods struggle with the complexity of EHR data, prompting the adoption of deep learning (DL) to uncover patterns and improve patient care. This systematic review, adhering to PRISMA guidelines, analyzes 82 peer-reviewed studies from 2017 to 2024 to evaluate the performance of five DL architectures Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, hybrid CNN-LSTM models, and Transformer-based models in EHR applications. CNNs achieved high accuracy (90–95%) in medical imaging, while LSTMs excelled in sequential tasks like ICU readmission prediction (88–93%). Hybrid CNN-LSTM models outperformed others, reaching up to 96% accuracy in multimodal tasks such as sepsis prediction [79]. Transformers, though computationally intensive, showed strong potential for clinical note analysis. Despite these advances, challenges like model interpretability, data privacy, and computational demands persist. This study provides a comprehensive benchmark of DL performance metrics across EHR tasks, offering insights into their practical applications and highlighting future directions, such as explainable AI and federated learning, to enable scalable, privacy-preserving healthcare solutions. By synthesizing these findings, this review guides researchers and clinicians in leveraging DL to enhance data-driven healthcare.