JMIR Publications

ABSTRACT

Background:

Cardiac arrest (CA) is the superior cause of death in patients in the intensive care unit (ICU). Many CA prediction models with high sensitivity have been developed to prevent CA in advance, but there was a difficulty in practical use due to the lack of generalization verification. Furthermore, patients with different subtypes of ICU have heterogeneity, but those characteristics were not identified.

Objective:

We propose clinically interpretable ensemble approach for the timely accurate prediction of CA within 24-hour regardless of heterogeneity, including different patient populations and subtypes in the ICU. In addition, subject-independent evaluation was performed to emphasize the generalization performance of the model and we analyzed interpretable results that can adopted by clinicians in real-time.

Methods:

Patients were retrospectively analyzed using data from the Medical Information Mart for Intensive Care-IV (MIMIC-IV) and the eICU Collaborative Research Database (eICU-CRD). To address problem of underperformance, we constructed our framework using vital sign-based, multi-resolution statistical, and Gini index-based feature sets with a 12-hour window to learn the unique characteristics of CA itself. We extracted three types of features from each database to compare the performance of CA prediction between patient groups at high risk for CA using MIMIC-IV and patients without CA using eICU-CRD. After extracting three types of features, we developed TabNet using feature screening with cost-sensitive learning. 10-fold leave-one subject out cross-validation and cross-dataset method used to check real-time CA prediction performance. We evaluated MIMIC-IV and eICU-CRD for different cohort populations and different subtypes of ICU within MIMIC-IV and eICU-CRD, respectively. Finally, external validation using the eICU-CRD and MIMIC-IV was performed to check the generalization ability. The decision mask of the proposed method was used to capture interpretability of the proposed model.

Results:

The proposed method was superior to performance achieved by conventional methods for different cohort populations both MIMIC-IV and eICU-CRD. In addition, the proposed method obtained higher than baseline models for different subtypes within ICU both MIMIC-IV and eICU-CRD. Interpretable prediction results can facilitate clinician’s understanding of CA prediction as a statistical test between non-CA and CA groups. Next, the eICU-CRD and MIMIC-IV was tested using a MIMIC-IV-trained and eICU-CRD-trained model to check generalization ability, and superior performance was achieved compared with baselines.

Conclusions:

Our novel framework to learn unique features provides stable predictive power in the different environments within the ICU. Most of the interpretable global information shows statistical differences between CA and non-CA groups, demonstrating that they are useful indicators for clinical decisions. Therefore, the proposed CA prediction system is a clinically validated algorithm that allows clinicians to intervene early in patients through CA prediction information and can be applied to clinical trials in digital health.