JAMA Netw Open. 2026 Mar 2;9(3):e264024. doi: 10.1001/jamanetworkopen.2026.4024.
ABSTRACT
IMPORTANCE: Mental health morbidity is a proximal factor in suicide. Using the research domain criteria (RDoC) framework to investigate sex and age differences in psychopathology at the time of death may provide better characterization than focusing on clinical diagnosis alone.
OBJECTIVES: To (1) evaluate whether token-based and large language model scoring of RDoC can be successfully applied to law enforcement and coroner or medical examiner death narratives in the US National Violent Death Reporting System (NVDRS), and (2) investigate sex and age differences in clinically relevant scores to illuminate underidentified dimensions of mental health dysfunction proximal to suicide.
DESIGN, SETTING, AND PARTICIPANTS: This cross-sectional study drew on death records in the restricted-access 2020 to 2021 NVDRS. Participants were limited to suicide decedents aged 12 years and older from all 50 states whose death record included a law enforcement and coroner or medical examiner death narrative of 20 words or more. Analyses were conducted between May 2024 and September 2025.
EXPOSURES: Sex and age of decedents.
MAIN OUTCOMES AND MEASURES: RDoC symptom scores; mental health status measures in death records (mental health diagnosis, current depressed mood, and alcohol or drug misuse).
RESULTS: Using both a token-based system and a large language model approach, law enforcement and coroner or medical examiner narratives of 72 585 suicide decedents were scored (mean [SD] age, 46.3 [19.3] years; 57 770 [80.6%] male). Both methods were previously validated with psychiatric electronic health records. To validate this approach, token density and large language model scores were compared with current NVDRS mental health status measures. Both scoring methods correlated with precoded measures and demonstrated levels of neurobehavioral dysfunction at the time of death similar to psychiatric inpatients on admission. Sex- and age-related differences in clinically relevant dysfunction showed the highest levels among female vs male and younger vs older decedents after adjusting for demographic confounding.
CONCLUSIONS AND RELEVANCE: This cross-sectional study of suicide decedents found that information relevant to RDoC domains is encoded in NVDRS death narratives and can be extracted using large language models. The approach used here observed more pervasive neurobehavioral dysfunction among suicide decedents than that captured by currently employed NVDRS measures of mental health.
PMID:41910970 | DOI:10.1001/jamanetworkopen.2026.4024
