Advancing the prediction and understanding of placebo responses in chronic back pain using large language models

Placebo analgesia in chronic pain is a widely studied clinical phenomenon, where expectations about the effectiveness of a treatment can result in substantial pain relief when using an inert treatment agent. While placebos offer an opportunity for non-pharmacological treatment in chronic pain, not everyone demonstrates an analgesic response. Prior research has identified biopsychosocial factors that determine the likelihood of an individual to respond to a placebo, yet generalizability and ecological validity in those studies have been limited due to the inability to account for dynamic personal and treatment effects--which are well-known to play a role. Here, we assessed the potential of using fine-tuned large language models (LLMs) to predict placebo responders in chronic low-back pain using contextual features extracted from patient interviews, as they speak about their lifestyle, pain, and treatment history. We re-analyzed data from two clinical trials where individuals performed open-ended interviews and used these to develop a predictive model of placebo response. Our findings demonstrate that semantic features extracted with LLMs accurately predicted placebo responders, achieving a classification accuracy of 74% in unseen data, and validating with 70% accuracy in an independent cohort. Further, LLMs eliminated the need for pre-selecting search terms or to use dictionary approaches, enabling a fully data-driven approach. This LLM method further provided interpretable insights into psychosocial factors underlying placebo responses, highlighting nuanced linguistic patterns linked to responder status, which tap into semantic dimensions such as "anxiety," "resignation," and "hope." These findings expand on prior research by integrating state-of-art NLP techniques to address limitations in interpretability and context sensitivity of standard methods like bag-of-words and dictionary-based approaches. This method highlights the role of language models to link language and psychological states, paving the way for a deeper yet quantitative exploration of biopsychosocial phenomena, and to understand how they relate to treatment outcomes, including placebo.