DOI: 10.1161/circ.148.suppl_1.19045 ISSN: 0009-7322

Abstract 19045: Detection of Left Ventricular Systolic Dysfunction From PDF Outputs of ECGs Obtained on Wearable Devices

Veer Sangha, Akshay Khunte, Lovedeep S Dhingra, Evangelos K Oikonomou, Rohan Khera
  • Physiology (medical)
  • Cardiology and Cardiovascular Medicine

Background: Left Ventricular Systolic Dysfunction (LVSD) affects 5% of adults with available therapies, but most are diagnosed after the onset of symptoms. AI-ECG tools that detect LVSD from clinical ECGs are not suitable for community screening. Wearable devices offer broad access to ECGs but end users rely on PDF outputs of single-lead data, with limited access to underlying signal data.

Methods: We simulated wearable PDF outputs from commercial wearable devices using lead I data from clinical 12-lead ECGs during 2015-2021. We selected the subset with echocardiography within 15 days of the ECG to define LVSD, based on LVEF < 40%. We developed a model to detect LVSD from wearable-adapted PDF outputs using an EfficientNet-B3 convolutional neural network with self-supervised pretraining. The model deployment pipeline was optimized to segment and process PDF outputs across all commercially available devices as inputs to the model (Fig A & B).

Results: 385,601 ECGs with paired echocardiograms were used for model development. In a held-out test set of 11,611 ECGs from unique patients (8.2% with LVEF < 40%) the model demonstrated high discrimination power for detection of LVSD with an AUROC of 0.88, AUPRC of 0.45, sensitivity 0.88, and specificity 0.72. An ECG suggestive of LVSD portended over 18-fold higher odds of LVSD in the held-out set (OR 18.3, 95% CI, 15.0-22.4). It also represents a digital biomarker in those with a false positive vs a true negative prediction at baseline, with a false positive assessment being associated with 3.6-fold (95% CI, 2.97-4.34) increased risk of developing incident LV systolic dysfunction in the future (Fig C).

Conclusions: We developed and validated a deep learning model that can identify LVSD from PDF image outputs of wearable and handheld single-lead devices. This approach broadens access to screening for LVSD to anyone with access to a wearable device with ECG.

More from our Archive