Advancing atrial fibrillation care: deep learning on ECG for comorbidity detection

Introduction Atrial fibrillation (AF) significantly increases morbidity, mortality, and healthcare costs. Patients with newly diagnosed AF often incur substantially higher healthcare expenditures compared to matched controls and commonly present with comorbidities such as diabetes mellitus (DM), obstructive sleep apnea (OSA), and chronic kidney disease (CKD). These comorbidities exacerbate prognosis and complicate AF management. Although guidelines recommend screening for these conditions, practical challenges—such as long waiting times for specialist evaluations and resource constraints—frequently delay comprehensive assessments. This study investigates the use of 12-lead electrocardiograms (ECGs) combined with deep learning techniques to screen for DM, OSA, and CKD in patients with AF. Methods From our registry, we analyzed 432 patients with paroxysmal or persistent AF. For each patient, a 10-second, 12-lead ECG recorded during sinus rhythm was collected alongside demographic information (age, sex, and body mass index [BMI]). All patients were also tested for DM, CKD, and OSA. These data were processed using a multimodal artificial intelligence (AI) approach. Specifically, we employed a deep learning model combining Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) architectures to analyze the ECGs, while demographic data were integrated to generate final predictions for DM, CKD, and OSA. The model’s robustness and reliability were evaluated using five-fold stratified cross-validation. Results A total of 432 patients were included in this study, of whom 32 had CKD, 45 had DM, and 45 had OSA. The proposed multimodal deep learning model showed promising performance for each condition. For CKD, the area under the receiver operating characteristic curve (AUROC) was 0.77, with an F1-score of 0.70, accuracy of 0.71, precision of 0.73, sensitivity of 0.71, specificity of 0.71, and an average precision of 0.79. For DM, the AUROC was 0.70, F1-score 0.68, accuracy 0.70, precision 0.72, sensitivity 0.70, specificity 0.70, and an average precision of 0.73. For OSA, the AUROC was 0.71, F1-score 0.69, accuracy 0.70, precision 0.71, sensitivity 0.70, specificity 0.70, and an average precision of 0.75. Conclusion These results suggest that a multimodal deep learning approach, integrating ECG data with demographic information, holds promise for accurately detecting key comorbidities—CKD, DM, and OSA—in patients with AF. This strategy could help streamline screening and facilitate earlier intervention for high-risk individuals.