Identifying trajectories of T2 relaxation time over age in the medial tibiofemoral cartilage in participants with pre-radiographic knee osteoarthritis

Abstract

Purpose: The multi-factorial etiology and pathogenesis of knee osteoarthritis (OA) poses a challenge to developing disease-modifying drugs. Identification of the common distinct patterns of disease progression and associating them with the risk of knee OA will deepen our understanding of OA, ultimately enabling the development of targeted treatments. Accordingly, the progression of knee OA is an actively investigated area in the field. Prior studies assessed structural progression based on radiographic joint space narrowing, but cartilage evaluation on magnetic resonance imaging (MRI) features may demonstrate patterns more responsive to structural changes. T2 relaxation times have been suggested as biomarkers of early cartilage degeneration due to their ability to quantify cartilage biochemical composition prior to morphological changes. Although aging is the greatest risk factor for OA, no previous studies have described the age-related change in cartilage compositional variation as yet. This study attempts to address these shortcomings, aiming to (1) identify distinct trajectories of MRI T2 relaxation times in tibiofemoral cartilage with age in knees at high risk of OA and (2) to associate baseline clinical characteristics of participants with the identified trajectory subgroup membership. Methods: Study samplesThis study used data from the Osteoarthritis Initiative (OAI). Quantitative features were derived from radiographs and MRIs and assessed at baseline and the year 1, 2, 3, 4, 6, and 8 follow-up visits. This study was conducted on the entire OAI incidence subcohort (N=3284), participants who had neither frequent symptoms nor radiographic tibiofemoral knee OA at baseline, but their characteristics placed them at increasing risk of getting symptomatic OA. Briefly, the characteristics included knee symptoms, overweight, knee injury history, knee surgery history, family history of total knee replacement (TKR), and Heberden’s nodes at the hands. Participants with more than two measurements of T2 relaxation times during the study were included. Primary outcome: T2 relaxation timesMRIs of the right knee were acquired at 3.0T (Siemens Trio) and included a multi-slice, multi-echo (MSME) spin-echo (SE) sequence for T2 relaxation time mapping. A fully automated model was developed for cartilage segmentation and subsequent T2 relaxation times measurement. A 3D V-net was trained to segment five distinct regions, including patella, medial and lateral femur, and medial and lateral tibia on 3921 manually segmented MR volumes. The optimized model was then used to segment cartilage voxels for the entire OAI cohort’s right knee images (25,110 total MRI exams). The T2 relaxation times were computed by fitting an exponential curve to the multi-echo signals. Since degeneration primarily manifests medially in OA, the median T2 values were computed from the medial femur (MF) and medial tibia (MT) compartment and subsequently used to model the trajectory profiles. Statistical analysisWe first determined trajectory of cartilage T2 relaxation time in MF and MT individually. After individual MF and MT trajectory groups are identified, all pair-wise combinations of MF and MT trajectories groups were examined concurrently. The characterization of individual trajectory over age was performed in a two-stage process. First, acknowledging inherent noise and missingness issue in the dataset, each participant’s T2 values were temporally smoothed using functional principal component expansion (FCPA) for irregularly and sparsely sampled data. We considered different combinations of hyperparameters, the order of basis function (M = 5, 10, 15) and the number of basis functions (r = 2, 3, 4, 5), then selected the best model using cross-validation. The outputs of FPCA were smooth basis functions shared by the samples and a set of FPCA scores per participant. Next, a finite Gaussian mixture model (GMM) was fitted to the FPCA scores to identify clusters following a similar temporal pattern of T2 relaxation time over age. The optimal number of clusters (G = 1-9) and within-cluster covariance parameterization were selected according to Bayesian information criterion (BIC). The identified trajectories of T2 relaxation time in MF and MT against age were depicted graphically. Participants were assigned to the trajectories of the highest posterior probability. The mean posterior probabilities in identified trajectories were computed. We assessed whether baseline characteristics differed by joint trajectory membership in MF and MT, using chi-square test for categorical variables and Kruskal-Wallis for continuous variables with skewed distribution. Results: A total of 2973 participants were included in the final analysis. The number of measurements varied between 2 and 7. The average follow-up was 6.1 years, and the span of the age of the study was 44 years. FPCA selected the model with M = 5, r = 2 and M = 10, r = 2 for MF and MT, respectively. The error variance and eigenvalues are shown in Table 1. Figure 1 shows the fitted mean and estimated eigenfunctions under the selected models. The GMM model identified two clusters (G = 2) with equal diagonal shape but varying volumes as optimal for both MT and MF. The mean posterior probabilities ranged from 0.71-0.74, supporting reliable goodness of clustering analysis. The identified trajectories of T2 relaxation times assessed in MT and MF are shown in Figure 2. The overall temporal patterns identified in MT and MF seemed similar: a continuously increasing trajectory and a rapidly increasing but attenuated trajectory near peak value, which we labeled as “late peak” and “early peak,” respectively. The within-cluster variance was more extensive in the “early peak” trajectory for both MF and MT. The range of variation over age in MF was more comprehensive than those in MT. The “late peak” was the most common trajectory in both MF (83.8%) and MT (76.8%). After trajectory phenotypes were determined separately for MF and MT, we identified four groups of pair-wise joint trajectories. As compared to the trajectory labeled “MF late peak and MT late peak,” comprising 67% of the analytic sample, participants with the “MF late peak and MT early peak” trajectory were more likely to have undergone surgery for repair or debridement of meniscus or cartilage (p = 0.01) (Table 2). Those in “MF late peak and MT early peak” were much more likely to be women (p < 0.001) and had pain, aching, or stiffness over the past 12 months (p = 0.01). Participants in the “MF early peak and MT early peak” group were more likely to have higher WOMAC stiffness scores (p = 0.02). Conclusions: We characterized distinct trajectories of median T2 relaxation times measured in the medial tibiofemoral cartilage compartments. To our knowledge, this study is the first to use T2 relaxation times to capture pre-radiographic changes in cartilage and the first to describe the temporal patterns over age. By utilizing subject similarity, FPCA decomposed noisy and sparse T2 relaxation times series spanning over 44 years of participant age in a data-driven manner. GMM, a model-based clustering approach with minimal a priori assumptions on underlying data structure, discovered the temporal patterns in a large cohort. We found that most (67%) participants demonstrated gradually increasing T2 relaxation, or continuously progressing cartilage degeneration, over age both in the MF and MT, while some more rapidly increase T2 relaxation with varying degree. Our finding is consistent with previous findings - while structural OA progression is highly variable, a small percentage of knees that have begun progressing are likely to experience further worsening. Examining baseline characteristics revealed that the identified four trajectories were significantly different in sex, meniscus or cartilage surgery history, baseline pain, and stiffness severity, suggesting possible categories of MR imaging-based phenotypes of cartilage degeneration.In conclusion, our data-driven method demonstrated that the patterns of T2 relaxation time trajectory were heterogeneous. On average, the majority gradually increased with age in modest magnitude, both in MF and MT. The model identified smaller subgroups that experienced accelerating T2 relaxation time increase with age.View Large Image Figure ViewerDownload Hi-res image Download (PPT)View Large Image Figure ViewerDownload Hi-res image Download (PPT)View Large Image Figure ViewerDownload Hi-res image Download (PPT)