Deep learning to diagnose left ventricular hypertrophy from standard, 12-lead ECG signals: a proof-of-concept study

Abstract

Abstract Funding Acknowledgements Type of funding sources: None. Background Although several sets of voltage and non-voltage criteria have been employed to detect left ventricular hypertrophy (LVH) in 12-lead electrocardiogram (ECG) signals, their accuracy has been proven suboptimal, while their use is limited by the lack of expertise of most non-specialised caregivers. Deep learning (DL) can be employed as an integrated, automated solution, yielding high predictive power for this task. Purpose To train a DL model to diagnose LVH from 12-lead ECG signals and evaluate its diagnostic performance. Methods A total of 21,837 previously annotated ECG samples were retrieved from the PTB-XL database. Of them, 2,137 (9.8%) had a diagnosis of LVH, while the remaining 90.2% presented with other abnormalities, including non-specific ST changes and conduction disturbances. After applying discrete wavelet transformation to denoise signals (remove the baseline drift and fine noise), we applied a dedicated convolutional, time-series-dedicated neural network (InceptionTime), which we trained on 90% of the data. We reserved the remaining 10% of "unseen" samples for testing its performance and quantifying its recall (sensitivity), precision, F1-score (harmonic mean of recall and precision) and area under the curve (AUC). We employed k-fold cross-validation and repeated the same procedure 10 times, to obtain distributions for each metric score. Results The 10 training/testing iterations yielded a median recall of 0.78 (interquartile range, IQR: 0.70-0.91) and a median precision of 0.78 (IQR: 0.75-0.87), indicating the predictive power of the model to both correctly detect LVH when it exists and also not falsely overdiagnose it, missing less than 1 out of 4 cases, in both scenarios. The obtained F1-score was 0.8 (IQR: 0.76-0.83) and the median AUC was 0.86 (IQR: 0.83-0.91). Conclusions Although a proof-of-concept study, this attempt demonstrates the potential of DL approaches to accurately detect heart abnormalities, such as LVH, from standard ECG recordings. Such strategies can not only support non-specialist physicians in picking diagnoses hidden "in plain sight", but they can also offer high diagnostic performance, given that attempts are made to optimize them. Toward this direction, our concept work shows that even a simple DL approach, without rigorous optimization, can display high predictive power, misdiagnosing less than 22% of the cases, paving the way for the improvement of such techniques and their integration in clinical practice.