Single-lead electrocardiogram Artificial Intelligence model with risk factors detects atrial fibrillation during sinus rhythm

Figures

• 

  Figure 1. Residual neural network architecture. Dotted boxes show model variations: squeeze-excite layers in residual units, encoder-layers, additional structured data input. Input either single-lead or 12-lead ECG, output as logit values, convertible to probabilities using the logistic function. AF, atrial fibrillation; BN, batch normalization layer; Conv, convolution layer; COPD, chronic obstructive pulmonary disease; Enc Unit, encoding unit; GlAvg, global averaging layer; H/O, history of; OSAS, obstructive sleep apnoea syndrome; ReLU, rectified linear unit; Res Unit, residual unit; Self-Attent, self-attention; Sq-Ex, squeeze-excite layer; TIA, transient ischaemic attack.

• 

  Figure 2. ECG dataset with train-validation-test split for replication and matched datasets. ECGs are split in an 80–10–10% split by the patient. The replication dataset uses a time window starting 31 days before the first AF–ECG for positive cases, and selects ECGs for the validation and test sets as in Attia et al.14 The matched dataset filters for ages above 40 years, uses a time window of 91 days before up to 365 days after the first AF–ECG for positive cases, and uses a 4-to-1 age-matching approach for the negative cases. ECG, electrocardiogram; pt, patients; +, positive cases; −, negative cases.

• 

  Figure 3. (A) Relative contributions of risk factors for the RF classifier. Four cardiovascular diseases are the most important. SHAP-values calculated for the matched dataset. (B) ROC curves for the single-lead ECG model with an optimal set of risk factors compared with full 12-lead ECG model and RF model using all risk factors. (C) Age- and sex-stratified AUC for single-lead ECG model with optimal risk factors, showing stable results over a large range of ages and both sexes. AUC, area under the curve; CABG, coronary artery bypass graft; CKD, chronic kidney disease; COPD, chronic obstructive pulmonary disease; MI, myocardial infarction; OSAS, obstructive sleep apnoea syndrome; PAD, peripheral artery disease; RF, random forest; TIA, transient ischaemic attack.

• 

  Figure 4. (A) Relative contributions of risk factors for the RF classifier in the replication analysis, using SHAP-values. Age is the dominant risk factor in the classification. (B) Age- and sex-stratified AUC for 12-lead ECG model with and without (all) risk factors. The age-stratified results show large differences depending on the inclusion of risk factors. Results are much more variable over age groups, compared with the models trained on the matched dataset. CABG, coronary artery bypass graft; CKD, chronic kidney disease; COPD, chronic obstructive pulmonary disease; MI, myocardial infarction; OSAS, obstructive sleap apnea syndrome; PAD, peripheral artery disease; TIA, transient ischaemic attack.