Chronic lung allograft dysfunction phenotype and prognosis by machine learning CT analysis.

Area of Science:

• Radiology
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Chronic lung allograft dysfunction (CLAD) is a primary cause of lung transplant graft failure.
• Prognosis in CLAD is dependent on its specific phenotype.
• Accurate phenotyping and prognostication are crucial for patient outcomes.

Purpose of the Study:

• To compare machine learning (ML) computed tomography (CT) lung texture analysis with radiologist scoring for CLAD phenotyping and prognostication.
• To evaluate the utility of ML-derived parameters, including pulmonary vessel volume (PVV), in predicting graft failure.

Main Methods:

• Retrospective analysis of adult first double lung transplant patients diagnosed with CLAD.
• Utilized an ML tool for CT texture analysis quantifying ground-glass opacity, reticulation, hyperlucent lung, and PVV.
• Radiologists scored CTs for various features; statistical analysis included ROC curves and Cox regression.

Main Results:

• ML and radiologist parameters differentiated CLAD phenotypes, with PVV being the strongest indicator for restrictive allograft syndrome/mixed (AUC 0.85).
• ML-derived hyperlucent lung feature effectively phenotyped bronchiolitis obliterans syndrome (AUC 0.76).
• Both ML and radiologist parameters predicted graft failure, with PVV showing the strongest association (HR 1.23).

Conclusions:

• ML-based CT texture analysis can effectively discriminate between CLAD phenotypes.
• ML, particularly PVV, offers a powerful tool for both phenotyping and prognostication of CLAD.
• ML-derived parameters, alongside radiologist scoring, are associated with graft failure risk.