Early Warning of In-Hospital Cardiac Arrest from Photoplethysmography Using Deep Residual Networks

Early detection of in-hospital cardiac arrest remains a critical challenge for improving patient outcomes. We propose a deep learning framework that leverages continuous photoplethysmography (PPG) signals to predict cardiac arrest within a 24-hour window. Models were developed using the SCOPE dataset, a recently released collection of paired ECG and PPG waveforms from 4,517 ICU admissions across 3,785 patients at Seoul National University Hospital. We trained a residual 1D convolutional neural network on 5-minute PPG segments sampled at 125 Hz, and evaluated performance using patient-level stratified 5-fold cross-validation. The model achieved strong discrimination with both AUROC and AUPRC, demonstrating that PPG signals contain predictive signatures of impending deterioration. These findings highlight the feasibility of scalable, non-invasive waveform-based risk prediction, and position PPG monitoring as a promising digital biomarker in critical care. Our results also underscore the value of large, publicly available waveform datasets for advancing robust and generalizable prognostic modeling.