243: What Might Have Helped? Estimating Counterfactual Effects Using AI in Chronic Pain

Topic: Clinical Science
Subtopic: Non-specific to any diagnosis, Psychology, Non-pharmacological treatments, Treatment/management

Background and Aims: Treatment response in chronic musculoskeletal pain (CMSP) is highly heterogeneous, yet clinical decision-making remains largely guided by group-level effects rather than patient-specific benefit. Predictive models in pain research have primarily focused on outcome prediction, while estimation of counterfactual treatment responses—essential for personalized allocation—has rarely been embedded within randomized trial settings. We therefore evaluated the feasibility of causal, deep-learning–based models to estimate individual treatment effects (ITEs) and associated uncertainty across multiple psychological interventions.

Methods: Using prospectively collected data from a stratified randomized controlled trial, we retrospectively trained and internally validated causal-inference models capable of estimating both factual and counterfactual outcomes for individual patients. Multiple established and novel architectures were implemented and compared, including a graph–neural-network–based model integrating clinically structured domain knowledge. Importantly, prediction uncertainty was quantified alongside point estimates to enable uncertainty-aware treatment recommendations appropriate for clinical translation.

Results: Conventional analyses showed no superiority of the original stratified allocation strategy. However, counterfactual modeling revealed substantial individual-level heterogeneity in treatment response that was not detectable at the group level. The optimized deep-learning model demonstrated improved internal calibration and yielded higher estimated individual treatment effects relative to the original allocation strategy, reflecting methodological signal rather than evidence of differential clinical efficacy. Predicted outcomes aligned closely with observed effects, and uncertainty estimates provided interpretable confidence bounds for patient-specific predictions.

Conclusions: These findings demonstrate the feasibility of integrating uncertainty-aware counterfactual effect estimation into pain treatment research using prospectively randomized data. Embedding deep-learning–based causal models within clinical trial datasets offers a transparent and reproducible framework for developing clinically deployable decision-support tools. This approach enables data-driven optimization of treatment allocation while preserving methodological rigor and represents a critical step toward AI-assisted personalized pain management.

References: Sirazitdinov A, Löffler M, Vock S, Beiner E, Löffler A, Schick A, Ader L, Hopp M, Ruckes C, Reininghaus U, Flor H, Hesser J, Tesarz J. Personalised treatment allocation in chronic pain using causal machine learning: A stratified randomised controlled trial. OSF Preprints 2025.

Presenting Author: Jonas Tesarz
Submitting Author: Jonas Tesarz

Senior Author: Wolfgang Eich

IASP 2026 Global Year theme of Neuropathic Pain: No