Deep learning to predict left ventricular hypertrophy from the electrocardiogram.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Left ventricular hypertrophy (LVH) is a significant predictor of cardiovascular disease.
• Previous supervised machine learning models for LVH classification using ECG and clinical data achieved an AUROC of 0.85 but required external validation.
• External validation is crucial for assessing the generalizability of predictive models.

Purpose of the Study:

• To develop a deep learning (DL) model for improved classification of cardiac magnetic resonance (CMR)-derived LVH.
• To externally evaluate the DL model's performance in the Study of Health in Pomerania (SHIP) cohort.
• To assess the feasibility of DL-based ECG screening tools for LVH prediction.

Main Methods:

• A fully convolutional network DL model was developed using 12-lead ECGs and clinical variables from 48,835 UK Biobank participants.
• The model predicted indexed left ventricular mass (iLVM), with logistic regression used for recalibration.
• Performance was evaluated using area under the receiving operating curve (AUROC) in training, validation, and test sets, and externally in the SHIP cohort.

Main Results:

• The DL model achieved an AUROC of 0.97 in the UK Biobank cohort, significantly outperforming previous methods.
• The QRS complex and ventricular rate on ECG were identified as key predictors of LVH.
• The DL model demonstrated modest generalizability to the SHIP cohort (AUROC 0.78), with variations attributed to clinical profiles, ECG acquisition, and CMR labeling.

Conclusions:

• Scalable DL-based screening tools for LVH prediction from ECG are feasible.
• Further model development using larger, more diverse datasets is necessary to enhance generalizability.
• Differences in cohort characteristics and data acquisition impact model performance across different populations.