Individual Submission Summary
Share...

Direct link:

Suicidal Intent in Fatal Drug Overdoses: An Application of LLMs to Classify Death through Investigator Narratives

Saturday, November 15, 3:30 to 5:00pm, Property: Hyatt Regency Seattle, Floor: 5th Floor, Room: 511 - Quinault Ballroom

Abstract

This study, Suicidal Intent in Fatal Drug Overdoses: An Application of Large Language Models to Classify Manner of Death through Death Investigator Narratives, is the final work from a string of five studies to analyze suicidal intent within drug overdose deaths. In the United States, death investigators provide free-text narratives to describe the circumstances of individual injury deaths, which include unintentional drug overdoses, suicides, and deaths of undetermined intent. With data from eight states, this project analyzed structured variables and free-text narratives to assess the accuracy of death manner classification in this population of deaths, specifically the extent to which suicides may be underreported in drug overdose deaths. The first study found large differences in the psychological content of death investigator narratives, where suicides had the highest rate of family references and the lowest scores in an Analytic Thinking index. The second study found wide differences in the rates of opioid toxicology information availability in death investigations by death manner, with opioid test information unavailable in 13% of unintentional drug overdose deaths, 32% of poisoning suicides, 59% of undetermined intent deaths, and 65% of non-poisoning suicides. The third study found that the association between death investigator systems (coroner vs. medical examiner) and rates of undetermined death manner classifications disappears after controls for individual death characteristics. The fourth examines differences in classifications and death narratives across defendant demographic characteristics. Finally, this fifth study combines the information on the structured variables of the previous studies with artificial intelligence large language models on free-form text narratives to predict rates of suicide under-classification in drug overdose deaths. We provide estimates of suicide deaths among both undetermined deaths and unintentional drug overdose deaths.

Author