Artificial intelligence capabilities in identifying atrial fibrillation using baseline sinus rhythm ECG : a systematic review

Abstract

Background Atrial fibrillation (AF) is a prevalent arrhythmia associated with adverse outcomes, often presenting paroxysmally. The lack of an efficient method to promptly detect paroxysmal AF and the absence of a unified screening approach necessitate exploring novel solutions. Artificial intelligence (AI) models show promise in addressing this gap, enabling early intervention. This study assessed the effectiveness of AI in detecting AF using baseline sinus rhythm-ECG (SR-ECG) and factors influencing their performance. Methods A systematic review was conducted across eight databases and registries (International Platform of Registered Systematic Review and Meta-analysis Protocols (INPLASY) registration: INPLASY202530059). References up to May 2024 were double-screened for eligibility. Included studies used AI to detect AF from baseline SR-ECGs in patients without prior AF confirmation. Quality was assessed using the Quality Assessment of Diagnostic Accuracy Studies-2 tool. Performance metrics were summarised using medians with subgroup analyses by AI type and AF confirmation timeframe. Results 14 studies and 33 AI models were analysed. Participant data were available for 13 studies, totalling 1459653 patients, with one study providing only testing dataset data. Median (95% CI) performance metrics were: accuracy 58.0% (55.0 to 62.0), sensitivity 62.0% (57.0 to 70.2), specificity 57.8% (51.0 to 61.1), precision 52.0% (47.0 to 56.0) and area under the receiver operating characteristic curve (AUC) 0.740 (0.630 to 0.830). Deep learning (DL) models outperformed traditional machine learning in sensitivity (72.6% vs 54.5%; q=0.027) and AUC (0.830 vs 0.610; q<0.001). Models using a 31-day confirmation window showed higher accuracy (83.2% vs 56.0%; q=0.010) and AUC (0.851 vs 0.630; q<0.001) than those using a 1-year timeframe. 11 studies (78.6%) cited possible negative cases misclassification as a limitation, and nine (64.3%) were deemed ‘high risk of bias’ in at least one domain. Conclusions AI-enhanced SR-ECG for identifying AF patients holds growing potential. Our findings show that DL and models incorporating a 31-day confirmation window are more effective in this context. Further research is needed to explore clinical benefits and cost-effectiveness.