Personalised Treatment Allocation in Chronic Pain Using Causal Machine Learning: A Stratified Randomised Controlled Trial

Background Chronic musculoskeletal pain (CMSP) remains a major clinical challenge, with many patients achieving suboptimal outcomes despite effective treatments. Although personalised strategies are promising, systematic algorithms for their implementation are scarce. We used data from a stratified randomised controlled trial (RCT) as a prospectively collected dataset to retrospectively develop and evaluate a machine-learning allocation algorithm. Using individual treatment-effect (ITE) estimation and statistical uncertainty quantification, we aimed to optimise treatment matching based on patient-specific profiles.Methods Participants with CMSP—including chronic primary back pain, osteoarthritis, fibromyalgia syndrome, and rheumatoid arthritis—were randomised to a personalised or non-personalised treatment arm. In the personalised arm, participants were classified into three psychological clusters (dysfunctional, interpersonally distressed, and adaptive coping) derived from the Multidimensional Pain Inventory and received one of three corresponding interventions: group-based Pain Extinction and Retraining Therapy, individual Emotional Distress Desensitisation Therapy, or a smartphone-based Ecological Momentary Diary Intervention. Participants in the non-personalised arm received deliberately incongruent assignments. The primary clinical candidate outcome was change in pain severity from baseline to post-intervention. To optimise allocation, we retrospectively trained three causal-inference models—T-Learner, TARNet, and a novel graph-neural-network TARNet (GNN-TARNet)—and incorporated prediction-uncertainty estimation to validate patient-specific responses.Findings Eighty-nine participants (mean age 53 ± 12 years; 63 % female) were included. Standard analyses showed no significant main effects for group (p = 0·34) or treatment condition (p = 0·12). The GNN-TARNet model produced the best internal calibration and suggested a larger estimated individual treatment effect compared to the initial allocation strategy (0·53 ± 1·11 vs. 1·05 ± 0·74; p = 0·02), reflecting a model-based hypothetical improvement rather than demonstrated clinical efficacy. Predicted outcomes closely matched observed effects (95 % within 1 SD; p = 0·98). Pain severity emerged as the most influential predictor.Interpretation This proof-of-concept primarily demonstrates methodological feasibility and internal plausibility of integrating uncertainty-aware individual treatment-effect estimation and predictive uncertainty within a stratified randomized framework. The observed increase in estimated individual treatment effects reflects methodological potential for more accurate allocation development, rather than evidence of clinical efficacy. Clinical efficacy must be prospectively evaluated in future confirmatory trials. Together, these findings establish a transparent, reproducible foundation for upcoming AI-assisted personalisation trials in chronic pain management.