Identifying and characterizing suicide decedent subtypes using deep embedded clustering

Subtypes of suicide decedents have not been studied at a population level using linked clinical and public health surveillance records. Identifying suicide subtypes can help facilitate the development and deployment of population-level prevention strategies. This retrospective study uses the Maryland Suicide Data Warehouse (MSDW). The analyses included 848 individuals who died by suicide as well as 4,161 individuals who died by accident in the state of Maryland between January 1st, 2016, and December 31st, 2019. These individuals had electronic health records from Johns Hopkins Medical Institutes and statewide hospital discharge data. We employed deep embedded clustering and evaluated its performance against traditional clustering approaches. We evaluated different numbers of clusters (k = 2 to 10) and assessed clustering performance using stability metrics, achieving a cross-validated prediction strength of 0.94. We then performed cluster characterization and assessed cluster stability up to 1 year before suicide death. We identified four distinct suicide profiles. Profile 1 (23.2% of suicide cases) included older individuals with high comorbid conditions. Profile 2 (19.2%) was characterized by psychiatric illness, the highest healthcare utilization, and significant social needs. Profile 3 (25.4%) consisted of younger individuals with psychiatric illness, no recorded social needs, and the highest percentage of Medicaid patients. Profile 4 (32.2%) included less clinically engaged individuals with the fewest healthcare visits. Our findings show the effective use of clustering methods to identify meaningful and stable suicide decedent profiles, revealing significant demographic and clinical differences. The identified subtypes can inform population-level suicide prevention strategies.