Abstract
<jats:p>Objective. To develop an interpretable multimodal machine-learning model for risk stratification of the rapid pain progression phenotype in knee osteoarthritis and to evaluate its performance in the independent PROCOAC cohort. Methods. An elastic-net logistic regression model was trained using Osteoarthritis Initiative (OAI) data. Rapid pain progression was defined over overlapping 24-month windows using normalized WOMAC pain. Harmonized clinical, genetic and proteomic candidates were evaluated, with feature selection by permutation importance. The frozen algorithm was tested in an OAI hold-out set and externally evaluated in PROCOAC. Logistic recalibration corrected prevalence shifts. Clinical utility was assessed by decision curve analysis. Results. OAI comprised 2,934 individuals and 14,488 instances. Feature pruning reduced 159 candidates to a 19-variable clinical-genetic signature driven by Kellgren-Lawrence grade, localized knee pain, BMI and two genetic variants (rs73631790, rs9912678); no proteomic variable was retained. External testing in PROCOAC (582 individuals, 1609 instances) showed ROC-AUC 0.744 (95% CI 0.714 to 0.772) and PR-AUC 0.519. Following recalibration, the sensitive screening threshold yielded NPV 0.875 (95% CI 0.849 to 0.898) and sensitivity 0.804 (95% CI 0.760 to 0.844), whereas the high-specificity threshold achieved PPV 0.610 (95% CI 0.523 to 0.692) and specificity 0.941 (95% CI 0.924 to 0.954). Decision curve analysis showed positive net benefit at both thresholds, supporting a three-tier risk stratification framework. Conclusions. This externally evaluated model identified patients at risk of rapid pain progression using an MRI-free clinical-genetic signature. Recalibrated thresholds may support risk-adapted monitoring, advanced imaging prioritization and trial enrichment.</jats:p>