PREDICTORS OF LONGITUDINAL MRI-BASED CARTILAGE THICKNESS CHANGE IN THE OBSERVATIONAL MULTICENTER APPROACH COHORT

Abstract

The APPROACH project selected knee OA patients by machine learning-based rankings trained to estimate a high likelihood of joint space width loss and/or increased or sustained knee pain over 2 years from demographic data, pain scores, and radiographic features. In addition, established radiographic (KLG, JSN) and innovative potential predictors of progression were acquired, such as six-month change in laminar cartilage composition (based on cartilage T2 relaxometry), the machine-learning-predicted structural progression scores (structure-score) used for enrollment, and MOAKS cartilage damage / full thickness cartilage damage extent scores. To study the association of established and innovative predictors for cartilage thickness loss. 297 people with knee OA were included in the APPROACH cohort (age: 66.5±7.1 years, BMI: 28.1±5.3 kg/m², 77.5% women, KLG 0/1/2/3/4: n=54/79/67/87/11). Weight-bearing medial (MFTC) and lateral (LFTC) compartment cartilage thickness was measured from 3D SPGR MRI at month 0, 6, 12, and 24 using manual, quality-controlled segmentations. Knees were classified as having progression in the MFTC and/or LFTC when the cartilage thickness change exceeded the respective smallest detectable change (SDC) thresholds (MFTC: -0.132mm, LFTC: -0.120mm). Six-month change in superficial and deep layer T2 times was measured in the MFTC and LFTC of 212 knees from 4 of the 5 centers. MOAKS-atlas cartilage damage scores were assessed for the MFTC and the LFTC at baseline by an experienced radiologist. The association between predictors (KLG, JSN, T2 change, MOAKS cartilage damage / full thickness damage extent, and machine-learning-predicted structure-score) and MFTC/LFTC progression exceeding the SDC threshold was analyzed using binary logistic regression with adjustment for site, age, sex, & BMI. For quantitative predictors, results were presented as odds ratios (OR) per SD. Knee-specific KLG and compartment-specific OARSI-atlas JSN scores at baseline were associated with both MFTC and LFTC progression for all observation periods (OR for 24 months: KLG MFTC/LFTC: 1.67, 95%CI: [1.27, 2.19]/ 2.06 [1.50, 2.83]; JSN MFTC/LFTC: 1.60 [1.16, 2.22] / 4.43 [2.43, 8.05]). Six-month increase in MFTC deep layer T2 times was associated with less MFTC progression over 24 months (OR: 0.57 [0.37, 0.89]). No other associations were observed for change in deep or superficial layer T2 times. A higher structure-score was associated with less MFTC progression over the initial 6 months period (OR: 0.72 [0.53, 0.97]) but not with MFTC progression over 12 and 24 months or LFTC progression. Presence of any MOAKS cartilage damage at baseline was associated with progression in the MFTC and LFTC over 12 and 24 months (24 months OR: MFTC/LFTC: 2.49 [1.17, 5.29] / 3.31 [1.56, 7.03]). Presence of MOAKS full thickness damage was associated with both, MFTC and LFTC progression over all periods (24 months OR: MFTC/LFTC: 5.18 [2.70, 9.95] / 6.54 [3.10, 13.81]). Only the established radiographic predictors and semi-quantitatively assessed MOAKS cartilage damage and MOAKS full-thickness cartilage damage were consistently associated with subsequent cartilage thickness loss whereas change in laminar cartilage T2 times or the structure-score were not. Training the machine learning model with a wider set of features and MRI-based progression labels than that available for screening of Approach participants is, however, likely to result in improved prediction of progression. EU/EFPIA Innovative Medicines Initiative Joint Undertaking (grant n° 115770). AW, SM, FE, WW: Chondrometrics GmbH; FWR: Boston Imaging Core Lab IMI-APPROACH (NCT03883568) participants and investigators CORRESPONDENCE ADDRESS: wolfgang.wirth@pmu.ac.at