Machine Learning of ECG Waveforms to Improve Selection for Testing for Asymptomatic Left Ventricular Dysfunction.

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Subclinical left ventricular dysfunction (LVD) has significant management implications for heart failure (HF) patients.
• Routine echocardiography is not standard for at-risk populations, necessitating alternative screening methods.
• Continuous Wavelet Transforms (CWTs) applied to ECG signals show potential for identifying abnormal myocardial relaxation.

Purpose of the Study:

• To determine if machine learning analysis of "energy waveform" electrocardiograms (ewECGs) derived from CWTs can be integrated with echocardiography for subclinical LVD detection.
• To assess the utility of ewECG in identifying systolic and diastolic left ventricular dysfunction.

Main Methods:

• 398 participants at risk of HF underwent ewECG and echocardiography.
• Left ventricular dysfunction (LVD) was defined by reduced global longitudinal strain (GLS), diastolic abnormalities, or LV hypertrophy.
• A random forest (RF) classifier utilized 643 CWT-derived features and the ARIC heart failure risk score for LVD prediction.

Main Results:

• The RF model, using 18 CWT features and ARIC score, achieved 85% sensitivity and 72% specificity (AUC: 0.83) for LVD prediction in an independent test set.
• Removing the ARIC score slightly altered performance (AUC: 0.78), while models for specific LVD types showed unsuitable sensitivity for screening.
• Conventional screening methods demonstrated inferior discriminative ability compared to the ewECG-based approach.

Conclusions:

• Machine learning applied to ewECG serves as a sensitive screening tool for LVD in individuals at risk of HF.
• Integrating ewECG into screening protocols could reduce echocardiography requirements by approximately 45%, significantly improving efficiency.