This study introduces a Bayesian multidimensional hierarchical item response theory (MHIRT) model to improve patient-reported outcome (PRO) assessments in total knee arthroplasty (TKA). Traditional unidimensional scoring fails to capture the multifaceted nature of recovery. Our model uncovers latent traits and inter-item relationships directly from PROMs such as the OKS and the EQ-5D-3L, without relying on predefined subscales. MHIRT flexibly decomposes PROMs into clinically meaningful traits like pain, mobility, self-care, and confidence. These traits captured more domain-specific variation, showed stronger sensitivity to temporal changes, and better reflected demographic factors than traditional total scores. The model was trained on a large NHS dataset and externally validated on PROMs from the moveUP digital platform. In predictive modeling of postoperative outcomes, MHIRT-derived features consistently outperformed unidimensional scores and conventional multidimensional IRT models. These findings suggest that MHIRT offers a potentially interpretable framework for tracking recovery and predicting health outcomes.