Diagnosis and risk factor mining for valvular heart disease with atrial fibrillation via ensemble machine learning

Abstract Diagnosis and risk assessment of atrial fibrillation (AF) remain challenging. As the gold standard, a 12-lead electrocardiogram (ECG) has the limitations of missing diagnosis and high cost of 24-hour Holter ECG. Existing risk scores for AF are also limited by low accuracy. Although electronic health records (EHR) have rich information of patients and machine learning (ML) has been proven useful in medical applications, it is still an open issue how to utilize ML and EHR to analyze AF in patients with valvular heart disease. We collected the EHR of 2077 patients with valvular heart disease (707 cases of AF), and investigated logistic regression (LR), standard ML, and ensemble ML models combined with Smotenn and Shapley algorithms. In addition, we explored the factors related to the occurrence of AF from two aspects: population-individual level feature analysis and cut-off value estimation. The sensitivity of Stacking and Blending ensemble ML models is 0.851 (0.842-0.860) and 0.876 (0.868-0.883), respectively, significantly higher than that of other models (P value <0.001). According to the Shapley algorithm, the occurrence of AF was strongly correlated with the left atrial diameter (LAD), peak mitral e-wave velocity (Emv), stroke volume (SV), and right atrial diameter (RAD). Finally, three continuous features (LAD=43mm, Emv=1.2m/s, SV=44mL) were identified as potential AF risk cut-off values. The ensemble ML models show high potential to build knowledge-based AF diagnosis systems that can assist doctors in making a precise diagnosis of AF and give early warning to patients.