The study was a retrospective cohort and included 340 patients after total ankle replacement, divided into a training (230 patients) and a validation group in order to maintain a similar distribution of results. Preoperative CT parameters of the subchondral bone (e.g. bone mineral density - BMD, trabecular separation - Tb. Sp, talar necrosis volume) were evaluated together with clinical indicators (e.g. talar inclination angle, Charlson comorbidity index - CCI). Univariate analysis showed that subchondral BMD, Tb. Sp, talar inclination angle, CCI, and volume of preoperative talar necrosis are significantly related to the need for revision prosthesis surgery. In multivariable Cox regression, Tb. Sp, talar inclination angle, and volume of preoperative talar necrosis confirmed as independent risk factors for long-term clinical deterioration, while higher subchondral BMD and higher CCI acted as protective factors. Three machine learning models (Random Forest, Support Vector Machine, Gradient Boosting) were created and their performance was evaluated using ROC curves. In the validation group, the AUC models achieved 0.897 (Random Forest), 0.790 (SVM) and 0.815 (GB). The Random Forest model was statistically significantly better compared to both SVM (AUC difference 0.107; p = 0.032) and GB (AUC difference 0.082; p = 0.041), while the difference between SVM and GB was not significant. The authors report that a Random Forest model combining CT-parameters of subchondral bone and clinical data effectively predicts long-term adverse outcomes after total ankle replacement, with key predictors being subchondral BMD, Tb. Sp and preoperative thallus necrosis volume.