JMIR Publications

Background:

Background : Early detection of cardiac arrest within the hospital is critical to reduce mortality in intensive care units (ICUs), yet current approaches are limited in accuracy and real-time applicability due to fragmented signal integration and inadequate temporal modeling in continuously recorded physiological signals, which are important for identifying early signs of patient deterioration.

Objective:

Objective: This paper proposes a CA-Net (Cardiac Arrest Network) design to develop and evaluate a multimodal deep learning model that uses synchronized physiological waveform data to predict cardiac arrest events 1 hour in advance in real time.

Methods:

Methods: This study develops CA-Net using high-resolution waveform data from the PhysioNet MIMIC-IV Waveform Database, temporally aligned and labeled using MIMIC-IV clinical data accessed via BigQuery. The input consisted of four synchronized physiological signals such as an electrocardiogram (ECG), a photoplethysmogram (PPG), arterial blood pressure (BP) and oxygen saturation (SpO2), formatted into a unified tensor of shape (1500, 4). A hybrid deep learning architecture was employed that incorporates convolutional neural networks (CNN), bidirectional long short-term memory (BiLSTM) units, and a transformer-based attention module to identify local and global temporal patterns. A 3-class output scheme (Normal, At Risk, Cardiac Arrest) was defined using clinically validated thresholds and labeled based on whether cardiac arrest occurred within the subsequent hour. Model interpretability was established through SHAP value analysis, which identified key physiological contributors to individual predictions. CA-Net was evaluated on a balanced dataset of ICU patients with synchronized multimodal waveform data. The model exhibited strong discriminative ability for predicting cardiac arrest events up to 1 hour in advance. The dataset was divided into training, validation, and held-out test cohorts using patient-level separation.

Results:

Results: The model achieved a macro-averaged AUROC of 0.94 and an cardiac arrest AUROC of 0.96 for cardiac arrest on the internal test set. External validation was conducted on an independent ICU cohort of 126 patients from Medicover Hospital, Hyderabad yielding a macro-averaged AUROC of 0.93 and an AUROC of 0.95 for cardiac arrest. SHAP-based analysis identified clinically relevant physiological features that include hypotension, oxygen desaturation, and ECG abnormalities, thereby enhancing model interpretability.

Conclusions:

Conclusion: By integrating continuous multimodal monitoring with advanced temporal attention mechanisms and interpretable outputs, this proposed model exhibits strong clinical utility as a deployable early warning system for timely cardiac arrest prediction in critical care settings.

Clinicaltrial:

Not applicable