Sex-Specific Machine Learning Improves Prediction of Incident and Prevalent COPD
Kalysta Makimoto1, Benjamin M Smith2, Joseph M Reinhardt3
1Toronto Metropolitan University, Toronto, ON, Canada.
Background:
CT imaging with machine learning can predict incident and prevalent COPD; however, it is unknown if sex-specific models improve performance.
Research Question:
Do sex-specific machine learning models using CT imaging-derived disease features improve prediction of incident and prevalent COPD and identify sex-specific predictors?
Study Design And Methods:
Canadian Cohort Obstructive Lung Disease (CanCOLD) study participants underwent baseline CT imaging and spirometry at baseline and follow-up. Models predicted incident and prevalent COPD using demographics and CT imaging features (lung density, texture, and shape and airway shape) for the combined-sex, male-only, and female-only data sets and externally tested in the Subpopulations and Intermediate Outcome Measures in COPD (SPIROMICS) study. Performance was evaluated using area under the receiver operating characteristic curve (AUC).
Results:
The study included 1,283 participants from the CanCOLD study and 1,840 participants from the SPIROMICS study. For incident COPD, the female-only model outperformed the male-only and combined-sex models in internal tests (AUC, 0.86 vs 0.76 and 0.78; P < .05) and external tests (AUC, 0.83 vs 0.71 and 0.77; P < .05). For prevalent COPD, the female-only model again outperformed the male-only and combined-sex models in internal tests (AUC, 0.84 vs 0.78 and 0.78; P < .01) and external tests (AUC, 0.84 vs 0.70 and 0.73; P < .05). Female-only models selected parenchymal texture and lung shape features not selected in combined-sex models, whereas male-only models selected airway-based features.
Interpretation:
Our results show that sex-specific models outperform combined-sex models for identifying those with and at risk of COPD, particularly in female patients, by capturing different disease-relevant features. These findings highlight that combined-sex models can obscure sex-specific biology, and adopting sex-specific prediction strategies may improve early detection and risk stratification and may allow for treatment targeting.