Effectiveness of generative AI-large language models' recognition of veteran suicide risk: a comparison with human mental health providers using a risk stratification model.

BACKGROUND

With over 6,300 United States military veterans dying by suicide annually, the Veterans Health Administration (VHA) is exploring innovative strategies, including artificial intelligence (AI), for suicide risk assessment. Machine learning has been predominantly utilized, but the application of generative AI-large language models (GAI-LLMs) remains unexplored.

OBJECTIVE

This study evaluates the effectiveness of GAI-LLMs, specifically ChatGPT-3.5, ChatGPT-4o, and Google Gemini, in using the VHA's Risk Stratification Table for identifying suicide risks and making treatment recommendations in response to standardized veteran vignettes.

METHODS

We compared the GAI-LLMs' assessments and recommendations for both acute and chronic suicide risks to evaluations by mental health care providers (MHCPs). Four vignettes, representing varying levels of suicide risk, were used.

RESULTS

GAI-LLMs' assessments showed discrepancies with MHCPs, particularly rating the most acute case as less acute and the least acute case as more acute. For chronic risk, GAI-LLMs' evaluations were generally in line with MHCPs, except for one vignette rated with higher chronic risk by the GAI-LLM. Variation across GAI-LLMs was also observed. Notably, ChatGPT-3.5 showed lower acute risk ratings compared to ChatGPT-4o and Google Gemini, while ChatGPT-4o identified higher chronic risk ratings and recommended hospitalization for all veterans. Treatment planning by GAI-LLMs was predicted by chronic but not acute risk ratings.

CONCLUSION

While GAI-LLMs offers potential suicide risk assessment comparable to MHCPs, significant variation exists across different GAI-LLMs in both risk evaluation and treatment recommendations. Continued MHCP oversight is essential to ensure accuracy and appropriate care.

IMPLICATIONS

These findings highlight the need for further research into optimizing GAI-LLMs for consistent and reliable use in clinical settings, ensuring they complement rather than replace human expertise.