Introduction: Increased left ventricular myocardial mass (LVM) is associated with adverse cardiovascular outcomes. Various rule-based criteria using limited electrocardiogram (ECG) features lacks sensitivity for evaluating LVM. Recent studies using deep learning methods have made progress in LVM evaluation taking advantage of ECG amplitude data. Hypothesis: We hypothesized that LVM prediction with deep learning models are improved by adding inter-lead information and building sex-specific models. Methods: This study proposed a novel deep learning-based method, the eLVMass-Net, by using ECG-LVM (n=1,459) paired data. ECG signals, QRS duration and axis, demographic features were used as input data. ECG signals were encoded by a temporal convolutional network (TCN) encoder. We adopted a total of four models (non-lead-grouping v. lead-grouping, non-sex-specific v. sex-specific) for LVM estimation. For lead-grouping models, the encoders were constructed based on segregated limb and precordial leads. For sex-specific models, we constructed models on both genders separately. Encoded ECG features and demographic features were concatenated for LVM prediction. To evaluate the performance, we utilized a 5-fold cross-validation approach with the evaluation metrics, mean absolute error (MAE) and mean absolute percentage error (MAPE). Results: For non-sex-specific models, eLVMass-Net has achieved an MAE of 14.33±0.71 and an MAPE of 12.90%±1.12%, outperforming the best state-of-the-art method (MAE 19.51±0.82; MAPE 17.62%±0.78%; P < 0.01). Sex-specific models achieved even lower MAPE for both males and females respectively (male MAPE 12.55%±0.88%; female MAPE 12.52%±0.34%), which also surpassed state-of-the-art methods. Adding the information of QRS axis and duration did not significantly improve the model performance (P = 0.28). The saliency map showed that T wave in precordial leads and QRS complex in limb leads are important features with increasing LVM. Conclusions: This study proposed a novel LVM estimation method, outperforming previous methods by emphasizing relevant heartbeat waveforms, inter-lead information, and non-ECG demographic features. The sex-specific analysis is crucial in improving LVM prediction.
Read full abstract