Application of sparse matrix clustering with convex-adjusted dissimilarity matrix in an ambulatory hospital specialist service

Objective Patients with chronic diseases and complications may frequently visit different specialists. Analytics could help deliver patient-centric seamless care by providing insights on the visit patterns of this group of patients, so that utilization of healthcare resources can be optimized. A new perspective focusing on patients´ specialist utilization records combined with statistical learning methodology can quantify the tightness of links between different specialties and highlight important specialist clusters. Method & Data Cosine angular dissimilarity matrix was used to measure connections among 163 specialties in 3 Singapore general hospitals based on 931,504 specialist attendance visits in 2013. A convex transformation on angular dissimilarity was introduced to solve low similarity problem caused by matrix sparsity and thus improved hierarchical clustering performance. The objective was to improve transformation by maximizing variance of off-diagonal dissimilarity coefficients. Ward´s method was used in clustering with dissimilarity matrix. Interactive visualization of sortable matrix was used to highlight important specialist clusters. Results Through clustering, 20 significant clusters were identified in 3 hospitals. Common clusters such as orthopedics, oncology-surgery, internal medicine, neuroscience, etc. were found among the 3 hospitals. Components of common clusters among hospitals were similar. Conclusion Patient utilization records can bring new and systematic insight of cooperative specialist services alongside traditional clinical research. Convex adjustment improves performance of Ward´s method on low similarity distance matrix significantly. Hierarchical clustering on convex-adjusted dissimilarity matrix is effective in discovering specialist clusters.