Automated Estimation of Computed Tomography-Derived Left Ventricular Mass Using Sex-specific 12-Lead ECG-Based Temporal Convolutional Network

Abstract

BackgroundLeft ventricular hypertrophy (LVH) is characterized by increased left ventricular myocardial mass (LVM) and is associated with adverse cardiovascular outcomes. Traditional LVH diagnosis based on rule-based criteria using limited electrocardiogram (ECG) features lacks sensitivity. Accurate LVM evaluation requires imaging techniques such as magnetic resonance imaging or computed tomography (CT) and provides prognostic information beyond LVH. This study proposed a novel deep learning-based method, the eLVMass-Net, together with sex-specific and various processing procedures of 12-lead ECG, to estimate CT-derived LVM.Methods1,459 ECG-LVM paired data were used in this research to develop a deep-learning model for LVM estimation, which adopted ECG signals, demographic information, QRS interval duration and absolute axis values as the input data. ECG signals were encoded by a temporal convolutional network (TCN) encoder, a deep neural network ideal for processing sequential data. The encoded ECG features were concatenated with non-waveform features for LVM prediction. To evaluate the performance of the predicting model, we utilized a 5-fold cross-validation approach with the evaluation metrics, mean absolute error (MAE) and mean absolute percentage error (MAPE).ResultsThe eLVMass-Net has achieved an MAE of 14.33±0.71 and an MAPE of 12.90%±1.12%, with input of single heartbeat ECG waveform and lead-grouping. The above results surpassed the performance of best state-of-the-art method (MAE 19.51±0.82, P = 0.04; MAPE 17.62%±0.78%; P = 0.07) in 292(±1) test data under 5-fold cross-validation. Adding the information of QRS axis and duration did not significantly improve the model performance (MAE 14.33±0.71, P = 0.82; MAPE 12.90%±1.12%; P = 0.85). Sex-specific models achieved numerically lower MAPE for both males (−2.71%, P=0.48) and females (−2.95%, P=0.71), respectively. The saliency map showed that T wave in precordial leads and QRS complex in limb leads are important features with increasing LVM, with variations between sexes.ConclusionsThis study proposed a novel LVM estimation method, outperforming previous methods by emphasizing relevant heartbeat waveforms, inter-lead information, and non-ECG demographic features. Furthermore, the sex-specific model is a rational approach given the distinct habitus and features in saliency map between sexes.Clinical PerspectivesWhat is new?The eLVMass-Net used ECG encoders with lead grouping, a unique feature that more properly reflects the electrical orientation of left ventricle.The sex-specific deep learning model is able to discriminate inter-gender differences of ECG features as shown by saliency maps.What are the clinical implications?The eLVMass-Net outperforms current state-of-the-art deep learning models for estimating left ventricular mass.A more accurate estimation of left ventricular mass could improve quality of care for comorbidities such as hypertension from easily accessible ECG.