fAI-BRO: a multimodal AI decision-support system to address diagnostic delay in fibromyalgia syndrome

To evaluate the diagnostic accuracy and patient acceptability of fAI-BRO (Fibromyalgia AI-Based Rheumatology Observer), a multimodal artificial intelligence (AI) system integrating video-based visual descriptor analysis and psycholinguistic transcript evaluation, for distinguishing fibromyalgia syndrome (FMS) from other rheumatic and musculoskeletal diseases (RMDs).

In this cross-sectional proof-of-concept study, 100 patients at the University of Bari (50 FMS per ACR (American College of Rheumatology) 2016 criteria, 50 RMD controls including inflammatory arthritis, osteoarthritis and other conditions) underwent standardised interviews in Italian. Audio was processed with OpenAI Whisper in translate mode. Magistral-Small (Mistral AI), deployed locally in zero-shot configuration without FMS-specific fine-tuning or in-context examples, analysed four video frames using OpenCV-derived image-level descriptors (30% weight) and psycholinguistic transcript features (70% weight). Classification thresholds were optimised post hoc via Youden's J index.

Among 100 patients, fAI-BRO achieved 91% accuracy (94% sensitivity (95% CI 83.5% to 98.7%), 88% specificity (95% CI 75.7% to 95.5%), area under the curve 0.96) for FMS detection on a 0-100 composite score. Youden's J identified the empirically optimal cut-off at 47, rounded down to 45 to favour screening sensitivity. All inflammatory arthritis cases (rheumatoid arthritis (RA), psoriatic arthritis (PsA), spondyloarthritis (SpA), systemic lupus erythematosus (SLE); n=36) and mechanical conditions (OA, impingement; n=8) were correctly classified as RMD. All six false positives for FMS occurred in dermatomyositis (n=3) and polymyalgia rheumatica (n=3)-a biologically plausible overlap involving muscle tissue rather than synovial pathology, presented here as a hypothesis-generating observation rather than incidental noise. Perturbation analysis confirmed widespread pain descriptors as the dominant classification drivers. Binary classification was robust to the English translation effect. Patient comfort was high (mean 4.5/5). Mean processing time was 324 s.

fAI-BRO demonstrates promising diagnostic accuracy for FMS screening as a human-in-the-loop decision-support tool, with perfect inflammatory arthritis classification and high patient acceptability. Post hoc optimised thresholds require external validation before clinical deployment.