Comprehensive Summary
The study, being conducted by Marengo et. al., aimed to study how problematic alcohol use could be predicted by large language models (LLMs) analyzing social media posts. The researchers studied a cohort of 208 Italian adults with Facebook accounts, who were split into two groups based on how recently the participants had posted (no recent posts vs. recent posts). LLMs, including GPT-4o and Gemini-1.5 Pro, were used to determine alcohol risk based on the presence of alcohol and alcohol-related content on participants' Facebook profiless. Researchers found that there is potential for LLMs to be used for alcohol risk determination, especially as compared with the Alcohol Use Disorders Identification Test-Consumption (AUDIT-C), which is a self-reported survey that has been shown to have recall and social desirability biases. Particularly, there is a stronger correlation with AUDIT-C when participants have more recent posts on their Facebook account. There was a strong positive correlation between GPT-4o and Gemini-1.5 Pro as well, which suggests that both sources, despite using different methods, are reaching similar conclusions. Both models were able to detect alcohol presence in posts, and this was shown to have a moderate correlation with participants’ self-reported habits related to alcohol use, with people who had more recent posts having a higher correlation. Overall, there is a promising capability of LLMs in predicting alcohol use risk, especially when interpreting recent Facebook posts from participants.
Outcomes and Implications
Alcohol use disorder, and its predecessor, risky drinking, are becoming increasingly important public health risks. However, traditional screening methods, such as surveys, can be unreliable due to biases, so researchers aimed to determine another way to predict potentially problematic alcohol use. Social media provides a non-invasive, real-time source of information and behavioral data that can be used to assess risk from public accounts. While there are limitations in the sample size and population used in this study, future research could be conducted to determine the generalizability of the results, particularly in other cultures and for other ages. This research could be important for the medical community, as it could quickly assess risk for individuals who are at risk for alcohol misuse, based on factors such as socioeconomic status, age, gender, or genetics. By more efficiently detecting problem drinking using LLMs, future implementations may support early detection and intervention efforts to prevent negative health impacts, such as socioeconomic burden, social and interpersonal harm, and morbidity and mortality.