ATS 2025: Machine Learning Prediction of Disease Severity in Patients with Interstitial Lung Disease Based on Lung Ultrasound and CT Imaging
Introduction:
A multidisciplinary approach is needed to diagnose interstitial lung disease (ILD), involving high-resolution computed tomography (HRCT), pulmonary function tests (PFTs), and other clinical data. Evidence has validated point-of-care lung ultrasound (LUS) as a useful tool for diagnosis and monitoring.
Aim:
To evaluate the ability of an artificial intelligence (AI) classifier to predict ILD severity in patients using human-annotated HRCT and LUS features.
Methods:
35 ILD patients participated in the study. The investigators collected LUS videos from eight thoracic locations; additionally, temporally concordant HRCT and PFTs were collected. Investigators who had undergone extensive didactic and practical training on the identification of LUS biomarkers associated with ILD supported with LUS videos annotations, while HRCTs were annotated by 2 board certified pulmonologists to record specific ILD features at the same thoracic locations represented by LUS videos.
Disease severity was categorized normal, mildly reduced, moderately reduced, and severely reduced based on forced vital capacity (FVC) and diffusing capacity for carbon monoxide (DLCO), Random Forest classifiers were trained using annotated HRCT or LUS features from 29 patients, while 6 patients were used for testing. Modles using the LUS annotated features alone, the HRCT annotated features alone, and both features were analysed. Results:
Results:
The accuracy of the model for predicting the correct category of FVC severity was
- 33% using LUS annotations alone,
- 33% using HRCT annotations alone, and
- 50% using combined LUS and HRCT
The accuracy for predicting the correct category of DLCO was
- 50% using LUS annotations alone
- 66% using HRCT annotations alone and
- 66% using combined LUS and HRCT annotations
Conclusion:
The present study showed that PFT ILD severity can be improved by combining LUS and HRCT annotations.
Am J Respir Crit Care Med 2025; 211: A1708
American Thoracic Society 2025 International Conference, May 18-21, San Francisco
