Background: Acute respiratory distress syndrome (ARDS) is a life-threatening condition associated with high mortality rates. Despite advancements in critical care, reliable early prediction methods for ARDS-related mortality remain elusive. Accurate risk assessment is crucial for timely intervention and improved patient outcomes. Machine learning (ML) techniques have emerged as promising tools for mortality prediction in patients with ARDS, leveraging complex clinical datasets to identify key prognostic factors. However, the efficacy of ML-based models remains uncertain. This systematic review aims to assess the value of ML models in the early prediction of ARDS mortality risk and to provide evidence supporting the development of simplified, clinically applicable ML-based scoring tools for prognosis.
Objective: This study systematically reviewed available literature on ML-based ARDS mortality prediction models, primarily aiming to evaluate the predictive performance of these models and compare their efficacy with conventional scoring systems. It also sought to identify limitations and provide insights for improving future ML-based prediction tools.
Methods: A comprehensive literature search was conducted across PubMed, Web of Science, the Cochrane Library, and Embase, covering publications from inception to April 27, 2024. Studies developing or validating ML-based ARDS mortality predicting models were considered for inclusion. The methodological quality and risk of bias were assessed using the Prediction Model Risk of Bias Assessment Tool (PROBAST). Subgroup analyses were performed to explore heterogeneity in model performance based on dataset characteristics and validation approaches.
Results: In total, 21 studies involving a total of 31,291 patients with ARDS were included. The meta-analysis demonstrated that ML models achieved relatively high predictive performance. In the training datasets, the pooled concordance index (C-index) was 0.84 (95% CI 0.81-0.86), while for in-hospital mortality prediction, the pooled C-index was 0.83 (95% CI 0.81-0.86). In the external validation datasets, the pooled C-index was 0.81 (95% CI 0.78-0.84), and the corresponding value for in-hospital mortality prediction was 0.80 (95% CI 0.77-0.84). ML models outperformed traditional scoring tools, which demonstrated lower predictive performance. The pooled area under the receiver operating characteristic curve (ROC-AUC) for standard scoring systems was 0.7 (95% CI 0.67-0.72). Specifically, 2 widely used clinical scoring systems, 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-based models exhibited superior predictive accuracy over conventional scoring tools, suggesting their potential use in early ARDS mortality risk assessment. However, further research is needed to refine these models, improve their interpretability, and enhance their clinical applicability. Future efforts should focus on developing simplified, efficient, and user-friendly ML-based prediction tools that integrate seamlessly into clinical workflows. Such advancements may facilitate the early identification of high-risk patients, enabling timely interventions and personalized, risk-based prevention strategies to improve ARDS outcomes.
Keywords: acute respiratory distress syndrome; machine learning; meta-analysis; mortality; systematic review.
©Ruimin Tan, Chen Ge, Zhe Li, Yating Yan, He Guo, Wenjing Song, Qiong Zhu, Quansheng Du. Originally published in the Journal of Medical Internet Research (https://www.jmir.org), 20.05.2025.