High-Throughput Detection of Risk Factors to Sudden Cardiac Arrest in Youth Athletes: A Smartwatch-Based Screening Platform

TL;DR

The study tackles the limited effectiveness and high cost of traditional pre-participation ECG screening for sudden cardiac arrest risk in youth athletes. It presents a smartwatch-based comprehensive screening system (CSS) that records sequential 4-lead ECG with an Apple Watch, upscales to a full 12-lead ECG using a decomposition-based regression, and classifies beat-level features via a Transformer Auto-Encoder System (TAES) with an OvO SVM. The approach achieves high diagnostic performance (example results: sensitivity ~95.3%, specificity ~99.1%) and demonstrates strong agreement with standard ECG in human trials, while offering substantial cost advantages over traditional methods. Preliminary validation in a 20-subject diagnostic cohort showed no misidentifications, supporting the potential of CSS for high-throughput, low-cost cardiac risk screening in athletic populations. Future work will expand disorder coverage and conduct larger, more diverse trials to confirm clinical utility and generalizability.

Abstract

Sudden Cardiac Arrest (SCA) is the leading cause of death among athletes of all age levels worldwide. Current prescreening methods for cardiac risk factors are largely ineffective, and implementing the International Olympic Committee recommendation for 12-lead ECG screening remains prohibitively expensive. To address these challenges, a preliminary comprehensive screening system (CSS) was developed to efficiently and economically screen large populations for risk factors to SCA. A protocol was established to measure a 4-lead ECG using an Apple Watch. Additionally, two key advances were introduced and validated: 1) A decomposition regression model to upscale 4-lead data to 12 leads, reducing ECG cost and usage complexity. 2) A deep learning model, the Transformer Auto-Encoder System (TAES), was designed to extract spatial and temporal features from the data for beat-based classification. TAES demonstrated an average sensitivity of 95.3% and specificity of 99.1% respectively in the testing dataset, outperforming human physicians in the same dataset (Se: 94%, Sp: 93%). Human subject trials (n = 30) validated the smartwatch protocol, with Bland-Altman analysis showing no statistical difference between the smartwatch vs. ECG protocol. Further validation of the complete CSS on a 20-subject cohort (10 affected, 10 controls) did not result in any misidentifications. This paper presents a mass screening system with the potential to achieve superior accuracy in high-throughput cardiac pre-participation evaluation compared to the clinical gold standard.

Figures (9)

• Figure 1: Common Causes of Sudden Cardiac Arrest within the Athlete Population mejialopez_2019_focus
• Figure 2: Brief Overview of Research Methodology. A sequential 4-lead ECG is taken. It is then preprocessed as described below, upscaled to 12 leads, and then spatial and temporal features are extracted using the TAES model. OvO SVM is used to classify the final result as the N (normal) label or the H, M, L, D, A (hypertrophic cardiomyopathy, myocarditis, long QT syndrome, dilated cardiomyopathy, arrhythmogenic left ventricular hypertrophy) labels.
• Figure 3: Description of S7 Placement in the Sequential ECG Protocol. Leads V2, V5, II, and AvR were chosen to be measured. The left depicts the typical position in which the device is held during measurement.
• Figure 4: Overview of the Training Process. All data is fed into the preprocessing algorithm in MATLAB. Data is then retrieved and used for autoencoder model training in Python. The finalized Transformer-Autoencoder is used to retrieve low-dimensional latent state representations which are then used by a One vs. One SVM for classification. The final results are analyzed by comparing sensitivity, specificity, and F1 score.
• Figure 5: Overview of Transformer Auto-Encoder System. In the model, the transformer serves as the outer layer of the model. Internally, 5 convolutional autoencoders extract spatial information. The model is trained by reconstructing ECGs to learn the features mapped by the model. In this way, the model learns which features are important by reconstructing the ECG and determining which type of feature has the greatest necessity in reconstruction. The latent state representation, following a pass through the encoder block, is taken and fed into the OvO SVM model before a final classification is determined by the model.