Background: The exponential growth of electronic health records (EHRs) has created unprecedented opportunities for developing machine learning (ML) models to support clinical decision-making and patient risk stratification. However, translating these models into practical clinical tools remains challenging.
Objective: This review examines current applications of machine learning in clinical decision support systems (CDSS) and risk stratification, evaluates their performance across different healthcare settings, and identifies key barriers to implementation.
Methods: We conducted a systematic review of peer-reviewed literature from 2018-2022, focusing on ML models deployed in real-world clinical environments. We analyzed model architectures, performance metrics, validation approaches, and implementation outcomes across various clinical domains.
Results: We identified 127 studies meeting inclusion criteria, spanning emergency medicine, intensive care, oncology, and primary care settings. Deep learning models demonstrated superior performance for image-based diagnostics (AUC 0.89-0.96), while ensemble methods showed robust results for tabular
EHR data (AUC 0.82-0.91). Key success factors included prospective validation, clinician involvement in development, seamless EHR integration, and interpretable model outputs. Major barriers included data quality issues, algorithmic bias, regulatory uncertainty, and workflow integration challenges.
Conclusions: Machine learning models show substantial promise for enhancing clinical decision support and risk stratification. However, successful implementation requires addressing technical, ethical, and operational challenges through interdisciplinary collaboration, rigorous validation, and careful attention to clinical workflow integration.