Continuous Cardiac Arrest Prediction in ICU using PPG Foundation Model.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Critical Care Medicine

Background:

• Non-invasive patient monitoring is crucial for predicting adverse health events.
• In-hospital cardiac arrest (IHCA) prediction remains a significant challenge in critical care.
• Existing methods often rely on multi-modal data or invasive techniques.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for IHCA prediction using only single-channel photoplethysmography (PPG) signals.
• To investigate the efficacy of pre-trained PPG foundation models for extracting robust health representations.
• To assess the performance of the proposed model within a 24-hour prediction window.

Main Methods:

• Development of a two-stage Feature Extractor-Aggregator Network (FEAN) model.
• Leveraging large-scale pre-trained PPG foundation models (e.g., PPG-GPT).
• Utilizing sequential classification models with one-hour ('1H') and 24-hour ('FH') history inputs.

Main Results:

• The FEAN model achieved an average Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.79 over a 24-hour prediction window.
• Peak performance reached an AUROC of 0.82 one hour before the cardiac arrest event.
• This is the first study to report IHCA prediction in ICU patients using only unimodal PPG deep waveform representations.

Conclusions:

• Single-channel PPG signals, when processed by advanced deep learning models like FEAN, can effectively predict IHCA.
• The proposed FEAN model demonstrates promising potential for non-invasive, real-time patient monitoring in intensive care units.
• Further research can explore architectural tuning and latent space visualization for enhanced clinical interpretability.