In the current study, the mechanical responses of cartilage were simulated within various knee joints under generic gait loading with the HTIPE material model, where material parameters were optimized against to the mechanical responses simulated in experimentally validated FRPVE material model. The previously used TIPE material model [19] was only used to understand the potential limitations of predicting mechanical responses or cartilage degeneration with material models with different stiffness (the softness of TIPE model was not known in advance). Although the optimized HTIPE material model produced nearly identical mechanical responses to the FRPVE material model using a simple joint geometry, it was concluded that only tissue tensile stresses were reproduced with sufficient accuracy compared to the FRPVE knee joint model responses. Furthermore, it was concluded that the variation in cartilage thickness in the knee joint model alters substantially the simulated differences in the mechanical responses between the material models. However, this had only small contribution to predicted cartilage degeneration between the FRPVE and HTIPE models, since the simulated excessive joint loads are considered to be the main mechanisms behind the initiation of OA development and progression [21]–[24]. Therefore, when classifying subjects into the KL01 and KL34 groups, the classification accuracy (moderate) was similar in both material models.
The material parameters of the TIPE model were derived from a previous study [19]. Shortly, it was optimized to a fibril-reinforced poroelastic (FRPE) model without the viscous response of the collagen fibril network. Thus, it is not surprising that the simulated mechanical responses were constantly higher or lower as obtained from the FRPVE material model. In contrast, the HTIPE model was optimized for the mechanical response obtained from the FRPVE material model, and in the simplified joint geometry, primary aim of optimization (tensile stress and cartilage deformation) was reached. However, in knee joint geometry, surprisingly only tensile stresses were moderately reproduced compared to the mechanical responses of the FRPVE material model. This can be explained mainly by the joint shape and cartilage thickness variation at the tibiofemoral contact during gait that generates varying load/strain rates on the cartilage surface between the models. As the FRPVE material is highly sensitive to load/strain-rate, due to viscoelasticity of the model, this causes mechanical responses that are not linearly dependent on load magnitude (this also occurs experimental measurements [25], [26]). For this reason, simulated differences between FRPVE and HTIPE, and FRPVE and TIPE models were not constant, even though the mechanical responses in FRPVE and HTIPE models were the same in the simplified joint geometry using the linear strain rate.
The extent to which the thickness of the existing atlas model was scaled (when the patient specific models were generated) significantly affected the differences in the simulated mechanical responses between FRPVE vs HTIPE, and FRPVE vs TIPE material models. In cases where cartilage thickness of the existing atlas was only slightly scaled (<20%), the average mechanical responses for tissue tensile stresses were in good agreement between FRPVE and HTIPE material models (difference < 1MPa). Other simulated parameters were either overestimated or underestimated. However, importantly the thickness scaling has a (almost) linear response between the simulated mechanical responses of the different models (FRPVE vs HTIPE; FRPVE vs TIPE). This indicates that it is possible to estimate the mechanical response of the FRPVE model from these simple models, knowing how a change in geometry affects the differences in the simulated responses, as simulated in this work. This is important as complex FRPVE material models are computationally heavy compared to simpler material models. For instance, in the current study, simulation of a FRPVE model took ~2 hours, whereas simulation of an HTIPE/TIPE model took ~20min.
When simulating cartilage degeneration, the FRPVE and HTIPE models performed similarly, while the TIPE model underestimated the degeneration. The poorer performance in TIPE model is explained by the age-dependent threshold for initiation of cartilage degeneration that works better in materials that produce higher tensile stresses. Although the AUC value was only moderate to classify subjects between KL01 (who will remain healthy) and KL34 (who will have OA) groups, it must be considered that average BMI’s of the different groups were between overweight and obesity (27-32g/cm2). In other words, in terms of baseline BMI, most of the subjects in each group had already a high risk for the onset and development of knee OA. When this information is reflected to the obtained ROC values, the prediction accuracy seems very promising. Interestingly, how the patients felt current knee pain had no effect on simulated degenerations. This indicates that knee pain is not associated with morphological changes in knee shape or cartilage thickness. This does not exclude the possibility that the pain is caused by a knee injury that has not yet been diagnosed.
Various machine learning (ML) models have been developed for classification of high and low risk subjects [4]–[6]. In these ML models, AUC values ranged from 0.6 to 0.8. However, it should be noted that some of those models require up to 112 predictor variables to make the classification. Collecting of such number of predictors is not clinically feasible timewise. Furthermore, it should be acknowledged that ML models are not capable of simulating quantitatively effects of different interventions such as weight loss or gait retraining that is possible with FEA based simulations, as utilized in the current study. However, it is possible to combine ML models with FEA-based simulations into a single tool [27]. This might be the next route for classification algorithms generation for evaluating subject specific risks for the onset and progression of knee OA.
Despite the encouraging results on the personalized risk of onset and progression of knee OA, there are some limitations to this study. First, only medial compartment of the knee was utilized in the simulation. Although OA usually initiates on the medial compartment, some patients exhibit the initiation on the lateral compartment [28], [29]. Consideration of this aspect might offer an improvement in the classification results (Figure 7). We consider this as critical limitation that should addressed in coming papers. Second, loading conditions were assumed to be identical for all subjects. It is well known that loading conditions vary among different subjects [30], [31]. We utilized a generic loading that was scaled based on the body weights of subjects [8], [32]. This can be considered as a feasible simplification since wrong estimation for personalized loading condition might produce even higher inaccuracies compared with the generic loading conditions. Last, material properties were not subject specific and quantitative measures that indicate cartilage health such as T2 values from MRI were not utilized [9]. In future studies, current joint integrity should be considered when making predictive models. In the current study, an existing degeneration algorithm [8] with high subject number was utilized.
The presented results suggest that simpler material models are capable to predict subject specific progression of knee OA if material parameters are selected properly. In the future, the presented workflow with simpler material model should be tested against to larger cohort data with a wider range of subject characteristics. The contribution of the lateral compartment should also be taken into account when making predictions for the onset and progression of knee OA. Furthermore, FEA based simulations merged with ML models could provide an accurate and fast clinical tool for prediction of osteoarthritis and simulation of different conservative preventative interventions, such as weight loss and gait retraining.