Early Prediction of Cardiac Arrest Based on Time-Series Vital Signs Using Deep Learning: Retrospective Study.

backgroundCardiac arrest (CA), characterized by an extremely high mortality rate, remains one of the most pressing global public health challenges. It not only causes a substantial strain on health care systems but also severely impacts individual health outcomes. Clinical evidence demonstrates that early identification of CA significantly reduced the mortality rate. However, the developed CA prediction models exhibit limitations such as low sensitivity and high false alarm rates. Moreover, issues with model generalization remain insufficiently addressed.

objectiveThe aim of this study was to develop a real-time prediction method based on clinical vital signs, using patient vital sign data from the past 2 hours to predict whether CA would occur within the next 1 hour at 5-minute intervals, thereby enabling timely and accurate prediction of CA events. Additionally, the eICU-CRD dataset was used for external validation to assess the model's generalization capability.

methodsWe reviewed and analyzed 4063 patients from the MIMIC-III waveform database, extracting 6 features to develop a deep learning-based CA prediction model named TrGRU. To further enhance performance, statistical features based on a sliding window were also constructed. The TrGRU model was developed using a combination of transformer and gated recurrent unit architectures. The primary evaluation metrics for the model included accuracy, sensitivity, area under the receiver operating characteristic curve (AUROC), and area under the precision-recall curve (AUPRC), with generalization capability validated using the eICU-CRD dataset.

resultsThe proposed model yielded an accuracy of 0.904, sensitivity of 0.859, AUROC of 0.957, and AUPRC of 0.949. The results showed that the predictive performance of TrGRU was superior to that of the models reported in previous studies. External validation using the eICU-CRD achieved a sensitivity of 0.813, an AUROC of 0.920, and an AUPRC of 0.848, indicating excellent generalization capability.

conclusionsThe proposed model demonstrates high sensitivity and a low false-alarm rate, enabling clinical health care providers to predict CA events in a more timely and accurate manner. The adopted meta-learning approach effectively enhances the model's generalization capability, showcasing its promising clinical application.