One Size Does Not Fit All: Improving Breast Cancer Patient Retrieval Using Variable-Tailored Similarity Measures

Clinical decision support systems (CDSSs) face challenges in managing complex patient cases that fall outside clinical practice guidelines. Using similarity-based approaches to retrieve comparable patients and apply the therapeutic care plans they received has the potential to support decision-making for complex cases. However, the choice of similarity metrics is difficult. We developed a hybrid similarity-based method (HM) which combines multiple similarity measures tailored to the specific characteristics of clinical variables aggregated with expert-defined weights. An interactive interface allowing clinicians to adjust weights in real time has been proposed. Assessed on breast cancer management, with expert-annotated datasets, HM achieved a recall of 0.6 for the top 5 most similar patients and retrieved 27 out of 45 clinically relevant cases. These results emphasize the importance of specific similarity measures in providing an effective, explainable, and customizable solution for patient retrieval of complex breast cancer cases.