Lung function prediction using machine learning-assisted breath-holding tests in patients with systemic sclerosis

Interstitial lung disease is a leading cause of mortality in systemic sclerosis (SSc), yet conventional pulmonary function tests are often challenging for some patients. The breath-holding test (BHT) offers a simple bedside alternative to the 6-min walk test (6MWT). Here, we developed a machine learning (ML) model to predict pulmonary function using real-time oxygen saturation (SpO₂) and heart rate data collected during BHT. Two SSc cohorts were analyzed: Cohort 1 ( n = 72) for model training and internal validation, and Cohort 2 ( n = 84) for temporal validation. A random forest classifier was applied to predict the forced vital capacity (%FVC) < 70 and diffusing capacity of the lung for carbon monoxide (%DLCO) < 60. Using BHT data, the model achieved mean AUROC (SD) values of 0.739 (0.043) for %FVC and 0.713 (0.075) for %DLCO. Models using only SpO₂ showed similar performance (AUROC 0.767 (0.096) for %FVC; 0.763 (0.041) for %DLCO), and comparable results were observed using 6MWT data. Temporal validation confirmed the model’s robustness (AUROC 0.686 for %FVC; 0.669 for %DLCO). These findings suggest that SpO₂-based ML models derived from BHT may serve as a practical, noninvasive complementary tool for assessing pulmonary function in SSc patients unable to perform conventional tests. Trial registration number : NCT04484948.