Community-aware recommender system for personalized patient–physician matchmaking on bipartite graphs

Physicians often create a network of associations and trust among themselves which can impact the patient referral process. However, patients may sometimes disregard these referrals due to the far distances between them and the physicians. In this work, we present an approach to optimize healthcare accessibility in densely populated urban areas by incorporating community structures identified in a medical relationship network into a personalized recommender system. We modified the Louvain community detection algorithm to find partitions in the intra-medical network that are consistent with the locations of medical offices. These communities were then employed as shortcuts within the physician-patient relationship graph, enabling us to leverage the BiRank algorithm to provide recommendations considering both geographic office locations and patients’ preferences. Our approach was developed by considering data from Unimed-BH, a major healthcare provider in Brazil, using nearly 10 million appointments between 2017 and 2023, performed by 1,448,310 patients in the Belo Horizonte metropolitan area. Our proposed recommender system achieves an NDCG@5% of 0.435 and a mean internal distance between recommendations of 0.7 km, highlighting its robustness and emerging as a promising solution for efficient and personalized referrals within the healthcare system. All results for the proposed algorithm are superior in comparison with the baseline and vanilla BiRank methods, leading to efficient matchmaking between patients and physicians while focusing on mitigating distances and improving trust into the doctor-patient relationship.