Predicting knee osteoarthritis progression using neural network with longitudinal MRI radiomics, and biochemical biomarkers: A modeling study

Background: Knee osteoarthritis (KOA) worsens both structurally and symptomatically, yet no model predicts KOA progression using Magnetic Resonance Image (MRI) radiomics and biomarkers. This study aimed to develop and test the longitudinal Load-Bearing Tissue Radiomic plus Biochemical biomarker and Clinical variable Model (LBTRBC-M) to predict KOA progression. Methods and findings: Data from the Foundation of the National Institutes of Health Osteoarthritis Biomarkers Consortium were used. We selected 594 participants with Kellgren-Lawrence grades 1–3 and complete biomarker data. The mean age was 61.6 ± 8.9 years, 58.8% were female, and the racial distribution was 79.3% White or White, 18.0% Black or African American, and 2.7% Asian or other non-White. A total of 1,753 knee MRIs were included across the study period, comprising 594 at baseline, 575 at 1-year follow-up, and 584 at 2-year follow-up. Outcomes included (1) both Joint Space Narrowing (JSN) and pain progression (n = 567), (2) only JSN progression (n = 303), (3) only pain progression (n = 295), and (4) non-progression (JSN or pain) (n = 588), corresponding to an approximate ratio of 2:1:1:2. JSN progression was defined as a minimum joint space width (JSW) loss of ≥0.7 mm, and pain progression as a sustained (≥2 time points) increase of ≥9 points on the Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) pain subscale (0–100 scale). Using the eXtreme Gradient BOOSTing (XGBOOST) algorithm, the model was developed in the total development cohort (n = 877) and tested in the total test cohort (n = 876). In the total test cohort, the Area Under the receiver operating characteristic Curve (AUC) of LBTRBC-M for predicting JSN and pain progression, JSN progression, pain progression, and non-progression were 0.880 (95% confidence interval (CI) [0.853, 0.903]), 0.913 (95% CI [0.881, 0.937]), 0.886 (95% CI [0.856, 0.910]), and 0.909 (95% CI [0.888, 0.926]), respectively. The overall accuracy of LBTRBC-M was 70.1%. With LBTRBC-M assistance, the prognostic accuracy of resident physicians (n = 7) improved from 44.7%–49.0% to 64.4%–66.5%. The main limitations include the use of a non-routine MRI sequence, the lack of external validation in independent cohorts, and limited incorporation of all knee joint structures in radiomic feature extraction. Conclusions: In this study, we observed that longitudinal MRI radiomic features of load-bearing knee joint tissues provide potentially informative markers for predicting knee osteoarthritis progression. These findings may help guide future efforts toward early risk stratification and personalized management of KOA. Why was this study done?: What did the researchers do and find?: What do these findings mean?: Ting Wang and Hao Liu develop and test a predictive model integrating longitudinal MRI radiomic features, biochemical biomarkers, and clinical variables to help forecast the progression of knee osteoarthritis.