Avance en la Predicción y Comprensión de las Respuestas al Placebo en el Dolor Crónico de Espalda Mediante Modelos de...

Placebo analgesia is a widely studied clinical phenomenon, yet placebo responses vary widely across individuals. Prior research has identified biopsychosocial factors that determine the likelihood of an individual to respond to placebo, yet generalizability and ecological validity in those studies have been limited due to the inability to account for dynamic personal and treatment effects.

We assessed fine-tuned large language models (LLMs) for the prediction of placebo responses in chronic low-back pain using contextual features extracted from patient interviews, as they speak about their lifestyle, pain, and treatment history. Interviews were conducted as part of two RCTs designed to study the placebo effect. These interviews were collected after treatment in the first trial (discovery cohort) and prior to treatment in the second trial (validation cohort).

Semantic features extracted with LLMs can predict which individuals respond to a placebo, with an accuracy of 74% in unseen data, and validating with 70% accuracy in an independent cohort. Furthermore, in contrast to previous work, LLMs eliminated the need for pre-selecting search terms, enabling a fully data-driven approach, and provided interpretable insights into psychosocial factors underlying placebo responses.

These findings expand on prior research by integrating state-of-art NLP techniques to address limitations in interpretability and context sensitivity of the traditional methods in related work. This method highlights the role of language models to link language and psychological states, paving the way for a deeper quantitative exploration of biopsychosocial phenomena, and to understand how they relate to treatment outcomes.

This study paves the way for a deeper yet quantitative exploration of biopsychosocial phenomena through language, and to understand how they relate to treatment outcomes, namely placebo. In this case it highlights nuanced linguistic patterns linked to responder status, which tap into semantic dimensions such as "anxiety," "resignation," and "hope".