JMIR Publications

ABSTRACT

Background:

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

Objective:

To enable timely and accurate prediction of CA events, we propose a real-time prediction model based on clinical vital signs. Concurrently, a meta-learning framework is applied to enhance the generalization capability of the model.

Methods:

We reviewed and analyzed 4063 patients from the MIMIC-III database, selecting six clinical features to develop TrGRU, a deep learning-based CA prediction model. The primary evaluation metrics for the model performance include accuracy, sensitivity, AUROC, and AUPRC. Subsequently, its generalization capability was validated on the eICU-CRD.

Results:

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

Conclusions:

The model proposed demonstrates high sensitivity and low false alarm rates. Furthermore, its generalization capability is effectively enhanced through the meta-learning framework, highlighting its value for clinical application.