CT-based AI system for quantitative and integrated management of acute respiratory distress syndrome in critical care.

Summary

AutoARDS is a large-scale, multi-center CT AI foundation model that produces a reproducible CT-derived biomarker correlating morphology with severity, achieves high diagnostic AUCs for acute respiratory failure and ARDS, and estimates P/F ratio with strong correlation (PCC=0.83). Trained on >50,000 CTs and validated across 6 centers (6,153 patients), it supports non-invasive, standardized decision support in critical care.

Key Findings

• Developed AutoARDS, a multi-task CT foundation model trained on >50,000 CT volumes with external validation across 6 centers (6,153 individuals).
• Established a reproducible CT-derived biomarker that links morphological lung injury to disease severity and progression.
• Achieved high diagnostic performance for acute respiratory failure (AUC=0.97) and ARDS (AUC=0.87), and accurately estimated PaO2/FiO2 ratio (PCC=0.83).

Clinical Implications

May allow earlier, standardized ARDS recognition and monitoring, reduce need for arterial blood gas sampling to estimate oxygenation, and support prognostication and intervention timing in ICU. Requires prospective clinical workflow integration and regulatory assessment before routine use.

Limitations

• Abstract truncation prevents evaluation of some methodological details (preprocessing, calibration, prospective testing).
• Generalizability across different scanner vendors, acquisition protocols, and patient populations outside validation centers requires further prospective testing.

Future Directions

Prospective interventional studies embedding AutoARDS in ICU workflows, head-to-head comparisons with blood-gas–based P/F measurements, assessment across diverse scanners/protocols, and regulatory/health-economics analyses.