Parikshit Sah
Sudden cardiac arrest (SCA) accounts for 20% of global deaths annually with an estimated 436,000 annual deaths in the United States alone (Meaney et al., 2013). Intelligent models for predicting out-of-hospital cardiac arrests (OHCA) using pre-event Electrocardiograms (ECGs) were developed using the Subtyping Cardiac Arrest dataset by Nightingale Open Science (Huang et al., 2021). The goal of the study was achieved by fulfilling two primary research objectives. Objective one explored traditional machine learning techniques, utilizing Support Vector Machine (SVM) and XGBoost models across five feature sets and six resampling techniques based on tabular features derived from 12-lead ECGs. Objective two focused on deep learning techniques designed to learn predictive morphologies directly from raw 12-lead ECG signals. The deep learning phase (objective two) evaluated three distinct architectures: a finetuned pre-trained backbone model (HuBERT-ECG), a custom lead-aware transformer encoder-based architecture (Lead-Aware Transformer), and a 1D Convolutional Neural Net model (1D-CNN). The XGBoost model trained with random over-sampling was selected as the optimal configuration from objective one. The chosen model (XGBoost) utilized 624 numerical features, known as the Reference group, and achieved an accuracy of 74%, a sensitivity of 52%, and a specificity of 78%. The 1D-CNN containing four convolutional blocks provided the optimal trade-off, with an accuracy of 69.9%, a sensitivity of 53.6%, and a specificity of 73.4% from objective two. Outcomes from both objectives were compared to external works. Comparison from a previous study showed that the models developed in this study yielded higher balanced accuracy (63.5% vs 59.5%) and sensitivity (54% vs 22.4%), albeit across divergent base rates (17.7% vs 1.63%), while demonstrating inferior specificity (73% vs 96.6%).