Automatic early detection of pathological signs following primary total hip arthroplasty using radiographs, clinical scores, and comorbidities.

The growing prevalence of total hip arthroplasty (THA) revisions, along with their generally poorer outcomes compared to primary procedures, emphasizes the urgent need for early detection of primary THA failure. This study proposes a model that integrates radiographic, clinical, and comorbidity data to automatically detect pathological signs within one year after primary THA. The dataset included two independent patient cohorts: the first comprised 400 patients, leading to 801 radiographs with pathological signs and 785 without; the second included 155 patients, resulting in 417 radiographs with pathological signs and 508 without. After preprocessing, the dataset was split into training, validation, and test sets. Three models were developed, one for each data type. A deep learning framework was applied to the radiographic data, while multiple machine learning classifiers were trained on the clinical and comorbidity data. Predictive probabilities were obtained for each subset and data type, and the final combined model was generated by averaging the predicted probabilities from all individual models. The final combined model achieved an F1 score of 0.72 (95% CI: 0.65-0.79), a balanced accuracy of 0.69 (95% CI: 0.63, 0.77), and an area under the curve (AUC) of 0.72 (95% CI: 0.65, 0.79) on the internal test set. On the external validation set, it achieved an F1 score of 0.66 (95% CI: 0.62, 0.70), a balanced accuracy of 0.62 (95% CI: 0.59, 0.66), and an AUC of 0.67 (95% CI: 0.62, 0.72). The results demonstrate the potential of the developed approach to automatically detect early pathological signs of THA, enabling virtual follow-up and potentially reducing the burden on clinicians.