JMIR Publications

ABSTRACT

Background:

Life-threatening ventricular arrhythmias (LTVAs) are main causes of sudden cardiac arrest and are highly associated with an increased risk of mortality. The prediction model which enables early identification of the high-risk individuals is still lacking.

Objective:

We aimed to build machine learning (ML)-based models to predict in-hospital mortality in patients with LTVA.

Methods:

A total of 3140 patients with LTVA were randomly divided into training (80%, n = 2512) and internal validation sets (20%, n = 628). Moreover, data of 2851 patients from another database were collected as the external validation set. The primary output was the probability of in-hospital mortality. The discriminatory ability was evaluated by the area under the receiver operating characteristic curve (AUC). The prediction performances of five ML algorithms were compared with two conventional scoring systems including the simplified acute physiology score (SAPS-II) and the logistic organ dysfunction system (LODS).

Results:

The prediction performance of the five ML algorithms significantly outperformed the traditional models in predicting in-hospital mortality. CatBoost showed the highest AUC of 90.5% (95% confidence interval [CI]: 87.5% – 93.5%), which was followed by the LightGBM with an AUC of 90.1% (95% CI: 86.8% – 93.4%). Whereas, the predictive values of SAPS-II and LODS were unsatisfactory, with AUCs of 78.0% (95% CI: 71.7% – 84.3%) and 74.9% (95% CI: 67.2% – 82.6%), respectively. The superiority of ML-based models was also shown in the external validation set.

Conclusions:

ML-based models could improve the predictive values of in-hospital mortality prediction for patients with LTVA compared with the traditional scoring systems.