Training an AI model to identify LVSD from single-lead ECG data would allow more reliable, earlier diagnosis.
Left ventricular systolic dysfunction (LVSD) is often diagnosed only after patients develop symptoms. Akshay Khunte and colleagues from Yale University suggest artificial intelligence could play a broader role in screening.
Even though AI is already a promising screening tool, the algorithms have been designed in clinically obtained 12-lead ECGs. Also, wearable devices often get noisy data due to, for instance, poor electrode contact with the skin, movement and muscle contraction during the ECG, and external electrical interference.
Training an AI model to accurately identify LVSD from single-lead ECG data while being resilient to significant noisy artifacts would allow more reliable, earlier diagnosis. LVSD is associated with significantly higher risk of heart failure and premature death.
In their study, published in NPJ Digital Medicine on July 11, the researchers used 385,601 ECGs, representing 116,210 patients with complete 12-lead ECG recordings, to develop two models: a standard one and a noise-adapted model. They trained the noise-adapted model by augmenting ECGs with custom noises in four frequency ranges emulating real-world noise sources, such as artifacts due to motion. Both models were trained to detect left ventricular ejection fraction below 40%.
Both models performed similarly on standard ECGS, but on tests with wearable device noise, the noise-adapted model detected LVSD significantly better, even on ECGs containing twice as much noise as signal. In fact, it handled noises the model had not encountered before, while preserving the model’s robustness in discerning complex hidden labels—an important benefit for wearable devices, which capture ECGs in settings with varying types and magnitudes of noise. Notably, the algorithm was developed and validated in a diverse population and demonstrated consistent performance across age, sex, and race subgroups.
Due to the lack of publicly available wearable device ECG datasets, training models using wearable ECG data directly is challenging, the researchers say. Current 12-lead ECG-based models are limited to investments by health systems to incorporate tools into digital ECG repositories, “and thereby limited to individuals who already seek care in those systems.” The technology may not be available or cost effective for smaller hospitals and clinics with limited access to digital ECGs.
Wearable devices, they point out, allow for community-wide screening, “an important next step in the early detection of common and rare cardiomyopathies.”
Their approach represents a major advancement from a methodological and clinical standpoint, the researchers say. Among other things, it demonstrates that through noise-augmentation single-lead ECG models can retain the prognostic performance of 12-lead ECG models.
It also avoids the unnecessary exclusion of collected data, which increases its generalizability across different device platforms--particularly important for community-wide screening programs, which may not be able to collect multiple ECGs per person and consistently meet high signal quality thresholds for each participant.