Joint clustering of hospitals based on their adminission behavior for different diseases using network of networks data model

Healthcare analytics is a rapidly growing industry where health organizations integrate data-driven insights into their clinical and operational decisions. One such insight involves clustering hospitals based on similarities in their monthly admission behavior, which can help ensure that the supply of healthcare services meets the population’s needs. Since diagnosis plays a crucial role in understanding variations in hospital admissions, clustering hospitals based on their admission behavior for specific diseases provides a more precise approach than clustering them based on their overall admission patterns across all diagnoses. This study aims to jointly cluster hospitals based on their admission behavior for different diseases. To achieve this, the two-layer Network of Networks (NoN) data model and the Network of Networks Clustering (NoNClus) method are utilized. The NoN data model is constructed using more than 7 million discharge records extracted from the California State Inpatient Database (2009–2011). It represents hospital admission similarities for different diseases as multi-domain disease-specific hospital networks at the bottom layer. Each of these networks is then represented as a node within a disease supernetwork at the top layer. The NoNClus method enables multiple underlying clustering structures across different networks by modeling the clustering structure in the disease supernetwork to guide the clustering of disease-specific hospital networks. The study thoroughly investigates the effect of various disease supernetworks in guiding joint clustering. The first disease network used to guide hospital clustering was derived from the Human Disease Symptoms Network (HDSN), based on the premise that disease symptoms play a crucial role in clinical diagnosis and, consequently, in hospital admission decisions {Albarakati2017}. The HDSN was constructed using an extensive medical bibliographic literature database. The second disease network combines literature-based and medical record-based networks, which were fused into a single disease network to explore their impact on guiding the joint clustering of disease-specific hospital networks {Albarakati2019}. Additionally, a medical record-based disease network was extracted by summarizing the various disease-specific hospital networks based on their underlying clustering structures using a graph-matching concept to further guide joint clustering {Albarakati2023}. This approach improved group homogeneity in hospital clustering results. Finally, a unified "meta-disease" network is proposed to integrate information from multiple disease-specific hospital networks, enabling a comprehensive and efficient joint NoN clustering method and supporting improved decision-making {albarakati2025}. This approach simultaneously summarizes and calculates similarities among multiple disease domains while clustering disease-specific hospital networks in a single step, enhancing the clustering process. The algorithm’s performance was evaluated using both synthetic and real-world datasets to assess its effectiveness in capturing meaningful patterns across disease networks. This framework aims to enhance the summarization and joint clustering of these multiple networks more effectively.