A NEW study suggests that machine learning could support clinicians in diagnosing usual interstitial pneumonia (UIP), a fibrotic lung pattern associated with poor prognosis, by combining radiomics features from high-resolution CT scans with routine clinical data.
UIP is a hallmark of idiopathic pulmonary fibrosis but is also found in other interstitial lung diseases (ILD), and its recognition is crucial for guiding management. However, diagnosis is often inconsistent, even among experienced radiologists and pulmonologists. To test whether computational tools could reduce this variability, researchers analysed 5,321 high-resolution CT datasets from 2,901 patients across three tertiary medical centres. The cohort was predominantly male (63.5%) with a mean age of 61.7 years. Demographics, smoking history, lung function parameters, and comorbidity data were collected alongside whole-lung radiomics features extracted from each scan.
The dataset was divided into training (n=3,639), internal testing (n=785), and external validation (n=789) groups. UIP prevalence was 43.7% across the full dataset, and around 42% in the internal and external validation cohorts. Researchers built a predictive model using advanced machine learning methods and created a nomogram to estimate the likelihood of UIP.
The radiomics-only model achieved an area under the receiver operating characteristic curve (AUC) of 0.790 in the internal testing set and 0.786 in the external validation set. When clinical variables were integrated with imaging features, the AUC rose to 0.802 and 0.794 respectively. These results were comparable to assessments made by pulmonologists with more than 10 years’ experience in ILD. The model also provided prognostic information. Among 522 patients who died during a median follow-up of 3.37 years, machine learning-predicted UIP was strongly associated with all-cause mortality, with a hazard ratio of 2.52 (p<0.001). This suggests the tool could not only improve diagnostic reproducibility but also assist in risk stratification.
These findings highlight the potential of machine learning as a reliable adjunct to current diagnostic pathways for UIP. With further validation, such multimodal approaches could help standardise diagnosis, improve early recognition, and guide treatment planning in interstitial lung disease.
Reference
Wang H et al. Developing and validation of a multimodal-based machine learning model for diagnosis of usual interstitial pneumonia: a prospective multicenter study. Chest. 2025; DOI:10.1016/j.chest.2025.08.034.
