JMIR Publications

ABSTRACT

Background:

Cardiac arrest (CA) is the leading cause of death in critically ill patients, and clinical research has shown that the early identification of CA reduces mortality. Algorithms capable of predicting CA through multivariate time-series data with a high level of sensitivity have been developed. However, these algorithms suffer from a high level of false alarms, and their results are not clinically interpretable.

Objective:

Patients were retrospectively analyzed using data from Medical Information Mart for Intensive Care (MIMIC)-IV. Based on the multivariate vital signs of a 24-h time window for adults diagnosed with heart failure (HF), we extracted multi-resolution statistical features and cosine similarity-based features for the construction and development of gradient-boosting decision trees (DTs). Therefore, we propose cost-sensitive learning as a solution to this problem. Then, 10-fold cross-validation was performed to check the consistency of the model performance, and the Shapley additive explanations (SHAP) algorithm was used to capture the overall interpretability of the proposed model.

Methods:

From the Medical Information Mart for Intensive Care-IV, a total of patients were retrospectively analyzed. Based on multivariate vital signs of 24 h time window of adults with a diagnosis related to heart failure, we extracted multi-resolution statistical features and cosine similarity-based features for the construction and development of gradient-boosting decision trees. We proposed cost-sensitive learning as a solution to the imbalance problem. The 10-fold cross-validation was carried out to check the consistency of the model performance. To capture the overall interpretability of the proposed model, Shapley additive explanations value was used.

Results:

The proposed method yielded an overall area under the receiver operating characteristic curve (AUROC) of 0.86 and an area under the precision-recall curve (AUPRC) of 0.56. Regarding the early CA prediction performance of the proposed model, it achieved an AUROC above 0.8 in predicting CA events up to 6 h in advance. This result indicated that the prediction performance of the model was superior to those obtained in previous studies. Additionally, we demonstrated the effect of feature importance on the clinical interpretability of the proposed method and inferred differences in importance between the non-CA and CA groups.

Conclusions:

The proposed framework can provide clinicians with more accurate CA prediction results for patients diagnosed with HF. Additionally, through its clinically interpretable prediction results, it can facilitate the understanding of clinicians. Furthermore, the similarity in vital sign changes can provide insights into understanding the temporal pattern changes in CA prediction for patients with HF-related diagnoses. Clinical Trial: N/A