Are Machine Learning methods effective in detecting undiagnosed atrial fibrillation in primary care settings using electronic health records? A systematic review.

Atrial fibrillation (AF) increases the risk of stroke, heart failure and mortality. Current screening guidelines fail to detect AF effectively, and existing models have limited applicability in primary care. Electronic health records (EHRs) provide an opportunity to apply machine learning (ML) for automated AF detection; however, their performance relative to standard care remains unclear. We conducted a systematic review to evaluate the effectiveness, quality, and applicability of EHR-based ML models for detecting AF in primary care. The review is informed by Joanna Briggs Institute and Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) guidelines. We searched seven databases from inception to May 2023. Eligible studies involved adults in primary care where ML models using EHRs were compared to standard care. The primary outcome was the detection of undiagnosed AF; secondary outcomes examined impacts on patients, healthcare providers, and systems. Data were extracted using CHARMS, risk of bias and applicability were evaluated through PROBAST and MI-CLAIM checklists. This review was registered in International Prospective Register of Systematic Reviews (CRD42023390603). From 4,536 references screened, 16 studies were included. Among these, 14 (87%) were retrospective cohort studies, one (6%) was prospective, and one (6%) was a randomized controlled trial. Random forest classifiers were the most common ML model (7 studies, 43%). Only 4 studies (25%) underwent external validation, and 8 (53%) were at high risk of bias. Model discrimination (AUROC) ranged from 0.71 to 0.948, with 8 (50%) outperforming controls. Combining ML with clinical tools (3 studies, 19%) significantly improved discrimination compared to ML models alone. Reviewed models identified gout as a nontraditional predictor of AF and demonstrated that dynamic measures of BMI, blood pressure, and heart failure diagnosis were stronger predictors than static measures. EHR-based ML models show promise for improving AF detection in primary care compared to standard care. Their clinical applicability, however, is limited by insufficient external validation, high risk of bias, and variable performance. Future research should prioritize external validation, evaluation in clinical trials and the integration of predictors routinely available in primary care.