CT-based radiomics for modeling surgical decision-making in young and middle-aged patients with ARCO stage III osteonecrosis of the femoral head: an age-stratified retrospective study

BACKGROUND: Osteonecrosis of the femoral head (ONFH) may progress to collapse and disability. While total hip arthroplasty (THA) is standard in older patients, surgical decision-making in younger adults remains heterogeneous. CT provides high spatial resolution for evaluating late-stage ONFH, and radiomics may offer quantitative support. OBJECTIVE: To investigate whether the predictive value of CT-based radiomics for surgical choice under routine practice is age-dependent and whether clinical variables add value across age subgroups. METHODS: We retrospectively analyzed 160 patients (201 hips) with ARCO stage III ONFH who underwent hip-preserving surgery or THA. Radiomics features were extracted from preoperative CT images and used to build radiomics-only and fusion models with four classifiers. Repeated k-fold cross-validation and bootstrap validation were applied. Three-level age stratification (< 65 years, < 50 years, and 20–35 vs. 36–50 years) and SHAP analysis were used. RESULTS: Fusion models outperformed radiomics-only models in patients younger than 65 years. In younger subgroups, radiomics-only models achieved comparable or superior performance, while the contribution of clinical variables decreased. SHAP analysis showed diminishing importance of age and disease duration, with texture-based radiomics features dominating in younger patients. CONCLUSION: CT-based radiomics, combined with interpretable machine learning, may provide quantitative decision support for modeling and supporting routine surgical decision-making between hip-preserving procedures and THA, particularly in younger patients.