The main findings of this study were: Patients with low BMD and high osteoporosis grades at any site could progress to three competitive OAs, whereas patients with high BMD and low osteosaccharide grades at any site could progress to unit OA. Between tricyte OA knees, patients with BMD, old age, and low high MTF JSN grades could progress to trispore-type JSN dominant OA, whereas patients with metabolic disease and relatively younger ages could progress to tricyte O-class OA.
OA is a heterogeneous, multifactorial disease known to progress due to a variety of factors, and recent studies have categorized OA phenotypes based on the factors affected7,9. Several studies have used ML approaches to predict knee OA progression and identify contributing factors for the OA phenotype13,15. Here, we used an LGBM ML model with better performance than traditional LR to predict OA progression patterns focused on the contributors of ML model interpretation and SHAP methodology.
In the unit cell/tri-compartment OA classification model, patients with low BMD could progress to tri-compartment OA. The association between knee OA and BMD has not yet been established, and there is continued debate about this. Zhang et al. Hart et al. Low BMD in the hip joint appears weakly related to progression of OA16,17. In this study, low BMD was identified as a contributor to the progression of three-compartment OA. In patients with low BMD, the imbalance between bone resorption and formation can impair subchondral bone remodeling. This weakens the microstructure of the subcartilage layer and worsens biomechanical properties.18. Low BMD affects the body bone in the subchondral bone, causing microfluids in all three compartments of the knee joint, increasing the risk of progression to tricholitic OA19,20.
In contrast, patients with higher BMD could progress to uniform OA. Previous studies showed that patients with higher BMD were at a higher risk of incident OA, while patients with lower BMD were associated with progression of OA.21,22,23. This obvious contradiction with previous findings may be explained by differences in the dominant mechanisms of OA development. Local biomechanical stress is more relevant in compartment-specific OA, but systemic changes in bone quality, such as low BMD, may contribute to the pathogenesis of generalized OA. Furthermore, Bergink et al. We demonstrated that there is a positive association between BMD and radiation OA in heavy joints due to local biomechanical factors.twenty four. They focused on the incidence and progression of hypertrophic OA, a phenotype characterized by elevated BMD. In this study, patients with high BMD may exhibit increased subchondral osteostility, which may amplify mechanical stress, particularly in the presence of lower extremity inconsistencies. Such inaccuracies can put excessive stress on the unit's articular cartilage and subchondral bone. Repeated forces can result in dense, hard bone properties with reduced load absorption capabilities, which can lead to increased degradation of the cartilage above.24,25. As alignment factors were not evaluated in this study, future studies using full-length radiographs can help predict OA progression.
The association between OA and BMD may vary depending on whether the patient has a generalized or localized form of OA. OA may be suitable for tailored treatments targeting specific phenotypes13. Unicompartment OA has been suggested to be an OA phenotype associated with high BMD, characterized by mechanical overload and cartilage damage.9. In contrast, it has been suggested that tricompartmental OA is a low BMD-associated OA phenotype characterized by abnormal subchondral bone remodeling that can improve with osteoporosis treatment.
BMI is also an important contributor to the classification model, with the tricompartment mental OA group showing higher BMI values compared to the unicompartment OA group. However, the linear relationship between BMI and each class was not clear. This pattern may reflect a nonlinear dose-response association between BMI and the risk of knee OA, where BMI primarily affects OA risk beyond certain thresholds.26,27. Furthermore, obesity affects knee OA locally and systemically as an intermediate mediator, causing mechanical overload at specific joints, resulting in local effects, and also contributes to low-grade systemic inflammation at multiple joints. Fat mass accumulation and dysregulation promote inflammation and extracellular matrix degradation in the musculoskeletal system, and adipokines play an important role in the development of osteoarthritis. However, the interactions between the inflammatory pathways, mechanical and metabolic processes of cartilage and bone disorders remain unknown28.
In the JSN-preferred OA classification model of the trichocompartment, BMD and old age were low demographic factors. OA in osteoporosis is characterized by subchondral bone strength and inappropriate structural support of articular cartilage, which places greater stress on the medial platform compared to the lateral platform under physiological conditions.29,30. Osteoporosis should be considered a major risk factor for varus deformation in knee OA, both due to stenosis in the joint space and different pathogenic mechanisms and alterations in peri-joint bone structure via unicompartment OA with high BMD. This may approach a bone-driven OA phenotype with structural benefits from treatment with bone-acting agents. Furthermore, age can be related to long-term repetitive forces of the knee joint, which can cause JSN.
The O-dominant OA classification model of the trichocompartment shows the lowest AUC of the OA progression model, and interpreting the model is less convincing. Rather than being applied immediately to clinical practice, it is better interpreted as an initial exploration model for characterization of OA subtypes. However, metabolic diseases emerged as a key contributor with linear relationships. Several studies have shown that metabolic diseases increase the risk of OA31,32,33. Subchondral bone ischemia caused by hypertension-induced atheroma lesions can promote damage to joint tissue. The accumulation of advanced glycation end products or cholesterol activates inflammation mechanisms and increases oxidative stress within the joints. They contribute to chronic low-grade inflammation in all three knee joints, increasing synovial activation and bone phytogenesis in early stage OA. This classification and increasing patient awareness of comorbidities may allow interventions to prevent the progression of OA. Although this study did not include laboratory markers showing systemic inflammation, future studies with large sample sizes and biochemical and inflammatory marker uptake would be valuable in confirming these associations and further characterizing the metabolic OA phenotype.
Many studies of clinical phenotypes and endotypes of OA have examined the prevention and treatment of early stage OA7,13,33. However, despite many studies using biomechanical data to classify phenotypes, structural phenotype studies using radiographic data are limited.12,13. This study sought to develop an ML algorithm using patient-specific information, such as demographics, comorbidities, and radiographic data, which are readily available from clinics. By interpreting the importance of the functionality of each classification model using SHAP methodology, we were able to identify factors contributing to each OA class and predict the corresponding OA phenotype. Evaluating OA progression patterns and identifying patient-specific information during clinical visits may enable targeted interventions and support more effective, personalized treatment and prevention strategies. However, the application of ML in predicting OA progression requires careful consideration of potential impacts on clinical decision-making and communication of prognostic information to patients. The current model was primarily trained in participants who had OA progressed for research design, but could serve as the basis for future clinical tools. Additional training on larger and diverse datasets, including non-progressors, allows models to be adapted and validated for use in clinical prediction. This will help clinicians to prior identifying OA progression patterns and support personalized surveillance and intervention strategies.
This study has several limitations. First, the dataset is obtained from a single institution and the number of datasets for both the three tricompartment JSN dominant OA classification model and the trichogenic O-preferred OA classification model is not sufficient to implement optimal gradient boosting, which can lead to the bias of the outcome. However, recent review articles consider an AUC between 0.7 and 0.8 to have acceptable discrimination, suggesting that the performance indicators of this study are not overabundant.34. Second, patients who progressed to bicompartment OA were excluded from the analysis. This approach allowed us to focus on clear OA progression patterns, but it is possible that the generalization of our findings is limited to a wider patient population. Multicenter studies with large sample sizes may be valuable for future analysis. Third, biomechanical and lifestyle-related factors such as clinical symptoms, pain severity, lower extremity alignment, physical activity, and occupation were not included due to a lack of available data. This limits the comprehensive properties of OA progression. Including these factors in future research could improve clinical relevance and facilitate the identification of contributors to a specific OA phenotype. Fourth, the study population exhibits a clear sexual imbalance, with approximately 80% of patients being female. The dominance of female patients may limit the applicability of our findings to the male population. However, gender showed minimal contributions in classification models, and in subgroup analysis there was no significant difference between male and female patients. Fifth, this study was retrospective in nature, with follow-up varying between patients. Retrospective design limits control over potential confounding factors and requires careful attention when interpreting findings. However, no significant differences were observed in the mean follow-up period between classes.
In conclusion, this study provides insight into the relationship between patient-specific characteristics, radiographic characteristics, and OA phenotypes. Patients with osteoporosis could progress to trichocytic OA with JSN, while patients with higher BMD could progress to unit OA. Metabolic disease-related OAs were associated with tribrain inverted OAs, including large bone plants.
