Document resource
Introduction Brugada syndrome is a rare inherited arrhythmia syndrome predisposing affected individuals to sudden death. Diagnosis requires detection of the Brugada ECG pattern, but non-expert clinicians can mistake this for mimics such as right bundle branch block (RBBB) or ischaemic ST elevation. Artificial intelligence models developed using machine learning (AI-ECG) could automate recognition of the Brugada ECG, to overcome this problem. However, developing AI-ECG models for rare conditions like Brugada syndrome is challenging due to limited ECG data for model training. We developed a Brugada classification AI-ECG model and evaluated the impact of three machine learning (ML) techniques on performance in the context of severely restricted training data.Methods The baseline model was trained using a dataset of 704 12-lead ECGs (352 normal, 176 RBBB and 176 Brugada) from Zhongshan Hospital, with a 25%:75% train-test split to artificially simulate data scarcity. To enhance training, we incorporated 3 additional datasets: (i) a separate ECG dataset of 9648 normal and 9824 RBBB ECGs from Zhongshan Hospital (Zhongshan-pretrain), (ii) an unlabelled ECG dataset from electrophysiology procedures at Hammersmith Hospital, including Brugada and non-Brugada ECGs (Imperial), (iii) an open-access labelled 21799 ECG dataset (PTB-XL). Three ML approaches were assessed: (1) supervised pretraining, (2) self-supervised pretraining (SimCLR and MoCo-V2 frameworks), and (3) oversampling (SMOTE: synthetic minority oversampling technique). All models were also re-trained using a conventional 80%:20% train-test split as part of the secondary analysis. Model performance was evaluated using accuracy, F1-score, and AUC (area under the ROC curve), with statistical comparisons via the DeLong non-parametric test.Results The 25% training data baseline model achieved 89.9% accuracy (F1-score 0.815, AUC 0.937). Supervised pretraining improved performance, with models pretrained on PTB-XL and Zhongshan-pretrain reaching F1-scores of 0.884 and 0.897 respectively, while the SimCLR self-supervised model on PTB-XL achieved a comparable F1-score of 0.882. Application of SMOTE oversampling further enhanced mean accuracy by 1.1% and increased the mean F1-score by 0.014. When models were re-trained using an 80%:20% train-test split, overall performance improved for all models but the added yield on baseline model performance diminished.Conclusions Pretraining and oversampling improve AI-ECG Brugada classification model performance under conditions of limited training data. Self-supervised pretraining performs comparably to supervised pretraining, potentially reducing the need for extensive manual ECG labelling. These approaches are broadly applicable for developing robust AI-ECG models for other rare cardiac conditions, ultimately aiding non-expert clinicians in accurate ECG interpretation.Abstract 4-025 Figure 1SMOTE-generated Brugada ECGsAbstract 4-025 Figure 2ROC curves of the baseline and best performing pretrained models using SMOTE