CT-based subchondral bone and clinical predictors of long-term total ankle arthroplasty outcomes

This study aimed to develop a machine learning-based predictive model for personalized long-term prognosis assessment in patients undergoing total ankle arthroplasty (TAA) by integrating preoperative computed tomography (CT)-derived subchondral bone structural parameters with clinical indicators.

A retrospective cohort study involving 340 TAA patients was divided into training (n = 238, 70%) and validation (n = 102, 30%) sets through stratified random sampling, ensuring the outcome distribution was preserved. Radiographic features and clinical metrics were systematically collected. Univariate analysis was conducted to identify variables associated with poor prognosis in the training set, followed by feature reduction using the least absolute shrinkage and selection operator (LASSO) regression. To determine independent risk factors, multivariable COX proportional hazards regression (Cox regression) was used. Three machine learning models-Random Forest (RF), Support Vector Machine (SVM), and Gradient Boosting (GB)-were constructed using Python 3.8.5. Model performance was evaluated using receiver operating characteristic (ROC) curve analysis.

Baseline characteristics showed no statistically significant differences between training and validation sets (p > 0.05). Univariate analysis indicated that subchondral bone mineral density (BMD), trabecular separation (Tb. Sp), talar tilt angle, Charlson Comorbidity Index (CCI), and preoperative talar necrosis volume were significantly associated with the need for prosthesis revision surgery. In the multivariable COX regression, Tb. Sp, talar tilt angle, and preoperative talar necrosis volume emerged as independent risk factors for sustained clinical deterioration. Conversely, subchondral BMD and CCI were identified as protective factors. In the validation set, the area under the ROC (AUC) for the RF, SVM, and GB models was 0.897, 0.790, and 0.815, respectively. Pairwise comparisons using the DeLong test revealed a statistically significant difference in AUC between the RF and SVM models (ΔAUC = 0.107, p = 0.032) and between the RF and GB models (ΔAUC = 0.082, p = 0.041). In contrast, the difference between the SVM and GB models was not statistically significant (ΔAUC = 0.025, p = 0.597).

The RF model that incorporates preoperative CT-quantified subchondral bone parameters and clinical indicators effectively predicts long-term adverse outcomes in TAA patients. The top three predictive features identified are subchondral BMD, Tb. Sp, and preoperative talar necrosis volume.