Internal and External Validation of a Deep Learning-Based Early Warning System of Cardiac Arrest with Variable-Length and Irregularly Measured Time Series Data

The early detection of cardiac arrest (CA) in emergency departments (EDs) is crucial for patient safety. However, existing deep-learning research often neglects irregular time intervals between measurements and the challenge of performance degradation in short sequences. The limited accessibility of medical data further complicates the external validation of models. To address these issues, we developed a deep learning-based early warning system accommodating variable-length and irregularly measured time series data. Our system includes three models: A Time Mask Temporal Convolutional Network (TM-TCN) incorporates a missing value mask to address the problem of missing values in multivariate time series, and univariate time series with time intervals are used to ensure that the model can detect the rapid deterioration of patients. Finally, we use a designed fusion method to enable the system to make better predictions for short sequence samples. Our system achieved an area under the receiver operating characteristic curve (AUROC) of 0.9831 and an area under the precision-recall curve (AUPRC) of 0.2150 in the experiment of 8 h before CA on the National Taiwan University Hospital dataset. In the external validation, the proposed system achieved an AUROC of 0.9734 and an AUPRC of 0.1336 8 h before CA on the Far Eastern Memorial Hospital dataset and obtained an AUROC of 0.8428 and an AUPRC of 0.0533 0 to 8 h before CA on the MIMIC-IV-ED dataset. These results demonstrate the system’s reliability and adaptability across datasets, highlighting its potential to advance healthcare informatics research by addressing critical challenges in time series data modeling.