Routine perioperative radiographs for early risk stratification of femoral head osteonecrosis after femoral neck fracture fixation: a retrospective cohort study

Abstract Background Femoral head osteonecrosis (ONFH) after internal fixation of femoral neck fracture is difficult to predict before definitive radiographic changes appear. Preoperative and immediate postoperative radiographs may contain prognostic information not fully captured by routine patient characteristics and visual radiographic grades. We compared a four-view radiographic model with a six-variable conventional baseline and explored the value of combining clinical and imaging information. Methods This single-centre retrospective cohort included 877 adults treated with closed reduction and cannulated compression screw fixation (CCS), of whom 222 developed ONFH. The development set comprised 727 patients, and 150 formed a fixed internal test set. The conventional baseline was a logistic regression model using age, sex, body mass index, time from admission to surgery, fracture displacement and reduction quality. The four-view model used preoperative and immediate postoperative anteroposterior and lateral radiographs. The primary measure was the area under the receiver operating characteristic curve (AUROC); the area under the precision–recall curve (AUPRC), Brier score and calibration were also assessed. Uncertainty and paired differences were estimated by patient-level bootstrap resampling; the preoperative–postoperative biplanar comparison used DeLong's test. Results In the test set, the four-view model had an AUROC of 0.864 versus 0.704 for the conventional baseline, a paired difference of 0.160 (95% confidence interval [CI] 0.068–0.256). The corresponding AUPRCs were 0.759 and 0.432, a paired difference of 0.327 (95% CI 0.153–0.461). Brier scores were 0.148 and 0.171, respectively. The uncalibrated four-view probabilities showed calibration error. An exploratory model combining the six variables with the four-view output had an AUROC of 0.863 and showed no additional discrimination. Conclusions A model using four routine radiographic views obtained before and immediately after fixation discriminated subsequent ONFH better than the six-variable conventional baseline, whereas fusion did not improve discrimination. External validation and calibration assessment are needed before using predicted probabilities to guide individual follow-up decisions. Unlike approaches based on later postoperative images, this study assesses prognostic information available before and immediately after fixation, before radiographic signs of ONFH emerge.