Abstract/Summary

Psoriatic arthritis (PsA) is a chronic inflammatory disease characterised by unpredictable flare-ups that are difficult to forecast, particularly in patients without an acute phase response. In this paper, we propose and apply an explainable, multimodal machine learning framework that jointly leverages structured temporal electronic patient records (EPRs) – sequential blood tests, disease activity scores, comorbidity burden, medications, and demographics – and unstructured clinical referral letters pre-processed with large language models ((LLMs, (Qwen-2.5 family)) to predict PsA flares. Gradient boosting models, Light Gradient Boosting Machine (LGBM) and eXtreme Gradient Boosting (XGBoost) were used to predict PsA flares, achieving the highest predictive performance 3 months before a clinic visit (accuracy = 92.8 %, AUROC = 0.94). Model performance gradually declined for longer timeframes (6 months: 78.2 %, AUROC = 0.80; 9 months: 76.6 %, AUROC = 0.78; 12 months: 72.2 %, AUROC = 0.75). LLMs applied to unstructured GP referral letters had limited standalone predictive value, but enhanced sensitivity and specificity when combined with the structured models in an ensemble approach. SHapley Additive exPlanations (SHAP) helped explain the prediction and demonstrated comorbidity count, disease scores, and immunosuppressive medications as the top predictors. Our results show that integrating both structured longitudinal data with unstructured clinical narratives using interpretable multimodal artificial intelligence can enable time-sensitive, personalised management of PsA flares and early clinical intervention.