Systematic review and meta-analysis of artificial intelligence models for diagnosing and subphenotyping ARDS in adults

Background

Artificial intelligence (AI) has emerged as a promising tool to improve the diagnosis and characterization of ARDS, including the identification of subphenotypes.

Objectives

To evaluate the diagnostic performance and methodological quality of AI models for identifying ARDS and its subphenotypes in adults.

Methods

We conducted a systematic review and meta-analysis of 63 studies (n = 135,762) published between 2013 and 2024 in PubMed, Embase, and the Cochrane Library. Extracted outcomes included sensitivity, specificity, AUROC, and validation methods. Risk of bias was assessed with PROBAST, and AI-specific metrics (overfitting, generalization, interpretability, discrimination, calibration) were reported.

Results

Pooled sensitivity was 0.89 (95 % CI 0.84–0.93), specificity 0.88 (95 % CI 0.83–0.92), and AUROC 0.90 (95 % CI 0.86–0.94), with high heterogeneity (I² > 85 %). Twenty-two studies (31 %) were rated high quality, with sensitivity 0.86 (95 % CI 0.82–0.89) and specificity 0.82 (95 % CI 0.78–0.85). Deep learning models (n = 14) achieved sensitivity 0.91, while machine learning models (n = 19) showed 0.87. Imaging-based models (n = 15) outperformed non-imaging approaches. COVID-19 studies (n = 9) reported sensitivity 0.90 with comparable AUROC and specificity. Only seven studies (18 %) investigated subphenotyping, identifying hyperinflammatory and hypoinflammatory profiles with potential therapeutic relevance. Calibration reporting was missing in 47 % and external validation in most (29/63).

Conclusion

AI models for ARDS demonstrate promising diagnostic accuracy but are limited by poor calibration and scarce external validation. Subphenotyping remains exploratory but suggests opportunities for real-time patient stratification. Prospective validation and standardized reporting are essential for clinical adoption.