Patient-adaptive medical AI: Similarity-based fine-tuning for cross-patient generalization

Medical AI models often face performance degradation when applied to new patients due to inter-patient variability in physiological characteristics, disease manifestations, and clinical histories. This challenge, commonly referred to as the cross-patient problem, limits the generalizability and clinical applicability of machine learning systems.

We introduce a similarity-driven framework for patient-adaptive learning that improves model performance on previously unseen patients. The proposed approach first trains a base model using conventional supervised learning and subsequently estimates the similarity between a target patient and the training population using intermediate model representations. The similarity information is then incorporated into a fine-tuning procedure through patient-dependent weighting, enabling the model to adapt its decision boundaries toward the characteristics of each individual patient.

We demonstrate the effectiveness of this strategy in two medical AI applications, including seizure onset zone classification in epilepsy and medical image classification tasks. Experimental results show consistent improvements over standard cross-patient learning approaches, highlighting the potential of similarity-based adaptation as a practical solution for personalized and generalizable medical AI systems.