This study is a pragmatic, cluster-assigned implementation trial designed to evaluate the impact of the digital educational intervention PainTrain-AI on clinical decision-making in Primary Care. The intervention trains health professionals through standardized virtual patient simulations and brief interactive micromodules, using a retrieval-augmented generation (RAG) safety architecture that restricts all outputs to validated clinical content. PainTrain-AI does not generate new clinical information and is used exclusively for educational purposes.
The trial corresponds to Phase 3 of a broader research program described in the project protocol. In this phase, entire Primary Care centers are allocated to either the intervention arm (PainTrain-AI training) or the control arm (usual training/standard practice). A natural comparative setting in specialized care is included for analytical purposes, without access to master datasets and without receiving any financial support. The primary outcome is clinical adequacy, assessed through a validated 0-100 rubric applied to standardized clinical cases at three time points: baseline (T0), immediate post-training (T1), and follow-up (T2). Double-blind scoring and a third-reviewer arbitration process are used to ensure reliability (target kappa ≥0.70).
Secondary outcomes include changes in validated psychometric measures (NPQ-12, PABS-PT, HC-PAIRS, FABQ), adherence to PainTrain-AI micromodules, system usability (SUS), latency/friction metrics derived from platform logs, and indicators of low-value clinical practices. The trial incorporates a sex/gender randomization of clinical vignettes to identify and analyze potential gender bias in clinical decisions. All data collected refer to professionals, not patients. No identifiable clinical data are used; all datasets are pseudonymized according to the approved Data Management Plan (FAIR).
Assignment by clusters prevents contamination between arms, and all centers follow standardized operating procedures (PNT) for recruitment, training, and evaluation. The platform maintains audit logs, human-in-the-loop oversight, and ENS/RGPD-compliant security measures. Analyses will use mixed-effects models and Difference-in-Differences approaches to account for clustering, repeated measures, and baseline differences. Results will inform the feasibility, acceptability, and preliminary effectiveness of a digital simulation-based educational intervention intended to reduce the competency gap, decrease low-value practices, and mitigate gender bias in the management of chronic musculoskeletal pain in Primary Care.
