Usual Interstitial Pneumonia Can Be Detected in Transbronchial Biopsies Using Machine Learning

Area of Science:

• Pulmonary Medicine
• Genomics
• Computational Biology

Background:

• Usual interstitial pneumonia (UIP) is the hallmark of idiopathic pulmonary fibrosis.
• High-resolution CT scans can be inconclusive for UIP diagnosis.
• Transbronchial biopsy (TBB) pathology has low sensitivity, often necessitating surgical lung biopsy.

Purpose of the Study:

• To develop a genomic classifier for distinguishing UIP from non-UIP using TBB tissue.
• To train and validate a machine learning algorithm against central pathology as the gold standard.

Main Methods:

• Exome-enriched RNA sequencing on 283 TBBs from 84 subjects.
• Machine learning algorithm trained on 53 subjects and validated on 31 subjects.
• Exploration of combining multiple TBBs per subject to improve diagnostic accuracy.

Main Results:

• The genomic classifier achieved an AUC of 0.86 for single TBB samples (86% specificity, 63% sensitivity).
• Combining 3-5 TBB samples improved performance to an AUC of 0.92.
• Alveolar-specific gene expression did not correlate with classifier accuracy.

Conclusions:

• Genomic analysis combined with machine learning significantly enhances TBB utility for UIP diagnosis.
• Multi-sample TBB testing increases accuracy compared to single-sample testing.
• Further validation in an independent cohort is required before clinical implementation.