JMIR Publications

ABSTRACT

Background:

Acute respiratory distress syndrome (ARDS) is associated with significant mortality, yet reliable early prediction methods remain unavailable. Machine learning (ML) approaches have garnered interest for predicting ARDS mortality, but evidence supporting their efficacy is limited.

Objective:

This systematic review assessed the value of ML in early ARDS mortality prediction, aiming to inform the development of simplified scoring tools.

Methods:

A comprehensive database search, including PubMed, Web of Science, Cochrane Library, and Embase, was conducted up to April 27, 2024. The included studies were evaluated using the Prediction Model Risk of Bias Assessment Tool (PROBAST), and subgroup analyses were performed on the datasets.

Results:

The review included 21 studies encompassing 31,291 ARDS patients. Meta-analysis revealed a pooled c-index of 0.84 (95% CI: 0.81, 0.86) for training sets and 0.83 (95% CI: 0.81, 0.86) for in-hospital mortality prediction. For validation sets, the pooled c-index was 0.81 (95% CI: 0.78, 0.84), with 0.80 (95% CI: 0.77, 0.84) for in-hospital mortality. The pooled area under the ROC curve (ROC-AUC) for common scoring tools was 0.70 (95% CI: 0.67, 0.72). Among these, the Sequential Organ Failure Assessment (SOFA) and Simplified Acute Physiology Score II (SAPS-II) demonstrated ROC-AUCs of 0.64 (95% CI: 0.62, 0.67) and 0.70 (95% CI: 0.66, 0.74), respectively.

Conclusions:

ML models showed superior predictive performance compared to individual scoring tools. Future research should focus on developing simplified and clinically applicable ML-based tools to enable early ARDS mortality risk identification and precision-prevention strategies.