Prediction Models in Non-Surgical Knee Osteoarthritis Management: A Scoping Review of the Current Evidence and Future Directions

Abstract

OBJECTIVE: While precision is always the goal, non-surgical management of knee osteoarthritis (KOA) is often constrained by heterogeneous patient presentations and practical barriers to personalization. Prediction models could enable precision care by forecasting treatment response, though their development requires robust, standardized datasets. While total knee arthroplasty (TKA) registries enable such modeling, conservative interventions suffer from heterogeneous documentation, limiting prediction model translation. This review maps existing literature on prediction models for non-surgical KOA management to evaluate current evidence and identify research gaps. DESIGN: Scoping review LITERATURE SEARCH: Following PRISMA-ScR guidelines, five electronic databases were systematically searched from inception to 12 December 2025. STUDY SELECTION CRITERIA: Eligible studies included: (1) KOA populations, (2) non-surgical interventions, and (3) models predicting treatment outcomes. Two reviewers independently extracted data; discrepancies were resolved by a third. DATA SYNTHESIS: Data were charted by model type, input/output variables, timepoints, and performance. Narrative synthesis identified similarities, differences, and evidence gaps. RESULTS: Of 3307 records, 38 met inclusion, with one added through citation searching. Interventions included injectables (n = 13), exercise/physiotherapy (n = 11), pharmacological (n = 4), multimodal (n = 6), and others (n = 5). Logistic regression predominated, followed by machine-learning models. Predictors clustered into eight domains: demographic, comorbidities, physical health, clinical, mental health, procedural, radiological, and biomarkers. Discriminative ability was moderate (AUROC 0.66 – 0.88) but variably reported. CONCLUSION: Prediction models for conservative KOA care show potential but limited applicability due to inconsistent reporting and lack of external validation. Future work should test these models in real-world practice.

Ähnliche Arbeiten