Applying ComBat harmonization to ultrasomics enhances machine learning generalizability and performance in multi-center clinical settings

Ultrasomics features suffer from multi-source batch effects across centers, including scanning parameters, scanners, and operator variability, limiting the generalizability of machine learning models and impeding their clinical translation. In this study, we systematically validated ComBat harmonization for reducing ultrasomics heterogeneity with a tiered validation framework: phantoms, retrospective clinical thyroid studies, and prospective multi-center human studies. ComBat reduced the proportion of heterogeneous features to below 51%. The concordance correlation coefficients were improved by up to 135%. When models were retrained on harmonized data using features selected pre-harmonization, average AUCs were significantly improved; further gains from optimized feature selection post-harmonization increased average AUCs to up to 0.93, with robust effectiveness across both population screening and high-risk referral settings. Our study suggests that ComBat harmonization can effectively reduce technical heterogeneity in ultrasomics features, serving as a promising tool to enhance the generalizability of machine learning models in ultrasomics research.