Workers’ Compensation Psychosocial Recovery Journeys: A Comparison of Large Language Model Insights (Preprint)

BACKGROUND Artificial intelligence is a potentially useful tool to derive insights from large data sets of both structured and unstructured data. The psychosocial recovery journey of injured workers in compensable environments is complex and Large Language Models (LLMs) are tools that can illuminate relationships otherwise indiscernible. OBJECTIVE The aim of this study was to model factors important in recovery from injury in a compensable environment, compare two service delivery modes for a support service, and use 3 LLMs to derive insights from a data set comprising of psychometric test data and free form notes from telephone coaching interviews of injured workers by generating recovery personas and personas for feigning spectrum behaviour (FSB). METHODS We used Claude 3.5 Sonnet, ChatGPT-4-Turbo and Gemini 1.5 Pro to derive insights from a deidentified data set of 7472 cases who were coached and treated by the Navigator Support Program between 17th June 2020 and 1st May 2024. Each LLM was used to generate natural recovery personas, personas for each month of the recovery course, and personas for different treatment coaching conditions being: “routine”, “escalated”, “intervention” with therapy, and also “feigning” personas. Two versions of the program were compared: “support only” and “support and therapy”. RESULTS Claude 3.5 generated 7 natural personas, ChatGPT-4-Turbo generated 6 natural personas, and Gemini 1.5 Pro generated 7 natural persons. The models used different features and were “sentiment” based, “functional recovery” based, and “adaptation and support” based respectively, with only moderate overlap between ChatGPT-4-Turbo and Gemini 1.5 Pro (0.3994). Forcing the LLMs to generate personas for coaching conditions month on month showed high similarity (0.991 – 0.979) in persona generation. Five sub-personas were identified for FSB showing the diversity of this behaviour. CONCLUSIONS Systematic modeling with LLMs show the models utilize different features and derive significantly different insights. There were slight differences in complexity of cases in service delivery modes, which likely reflected the differences in the insurer portfolios.