Cross-site Transportability in Biomedical Machine Learning: Evaluation and Adaptation Strategies

In recent years, machine learning has assumed a relevant role in the biomedical field for clinical, diagnostic, and prognostic applications. However, the effective use of these models in real world clinical settings is often limited by their reduced transferability, defined as the ability to maintain reliable predictive performance when applied to data, populations, or environments different from those used during training. This limitation is particularly critical in multi site biomedical studies, where differences in acquisition protocols, instrumentation, patient populations, and data characteristics may generate domain shift and compromise model performance outside the original training context. The objective of this thesis is to systematically investigate cross-site transportability in biomedical machine learning by comparing different adaptation and harmonization strategies across heterogeneous domains. Beyond comparing predictive performance, the thesis also aims to analyze how dataset and domain characteristics influence model behavior across different settings. The experimental evaluation focused on supervised classification tasks designed to simulate realistic cross-site scenarios by separating source domains from target domain. Five approaches were compared: a baseline, a statistical harmonization method based on ComBat, a domain weighting strategy, a transductive learning approach, and a hybrid strategy combining ComBat harmonization and transductive adaptation. Each approach was evaluated using Logistic Regression, Random Forest, and Histogram Gradient Boosting classifiers under multiple cross-site configurations. To support the interpretation of the results, domain shift characterization and pattern discovery were performed using the dashi framework, while meta learning and SHAP analyses were used to investigate the relationship between domain characteristics and model performance. Performance was assessed using complementary evaluation metrics considering both in-distribution (ID) and out-of-distribution (OOD) behavior. The study was conducted on three biomedical datasets representing different clinical domains and data modalities. UCI Heart Disease included 920 samples from 4 domains and represented a structured clinical setting. ABIDE II included 1114 subjects from 19 sites, with 17 sites eligible as target domains, and represented a multisite neuro clinical setting based on phenotypic variables related to autism spectrum disorder. TCGA-BRCA included 829 samples from 13 sites and represented a high dimensional genomic setting related to breast cancer. Together, these datasets provided heterogeneous scenarios with different levels of dimensionality, domain variability, and cross-site complexity. The results showed that no single adaptation strategy consistently outperformed all others across datasets, models, and evaluation criteria. UCI Heart Disease represented the most favourable transportability scenario, where target aware methods, especially transductive adaptation, provided the clearest benefits. ABIDE II showed a more heterogeneous and target site dependent behaviour, with moderate advantages from adaptation depending on the metric considered. TCGA-BRCA was the most challenging dataset, where the baseline often remained the strongest option for OOD discrimination, although transductive methods improved calibration. Overall, these findings suggest that cross-site transportability is a context dependent property rather than an intrinsic characteristic of a model. Adaptation should therefore be selected according to the type of domain shift, the available target domain information, and the practical objective of the biomedical application.