Optimizing treatment strategies for femoral neck fractures using a machine learning model

BACKGROUND: Recently, advances in machine learning models have allowed for automatic and highly accurate detection of fractures. To date, however, no machine learning models have been developed for the automatic selection of treatment strategies for femoral neck fractures. This study aimed to develop and evaluate the performance of a machine learning model that would recommend either internal fixation or implant replacement for the surgical treatment of femoral neck fractures. METHODS: We constructed a dataset comprising 535 preoperative anteroposterior hip radiographs obtained from 297 patients with femoral neck fractures who underwent surgical treatment (implant replacement, n = 182; internal fixation, n = 115). After splitting the dataset into training, validation, and test sets, binary classification models were developed to recommend the optimal surgical procedure using two convolutional neural network architectures: EfficientNetV2-S and VGG16. Model performance was compared using an independent test set. To investigate the evidence underlying each model's decision-making, regions of interest were visualized using Grad-CAM and Layer-CAM. RESULTS: The agreement rate for surgical procedure selection on the test set was 0.828 ± 0.039 and 0.821 ± 0.045 for EfficientNetV2-S and VGG16, respectively. The F1-scores for EfficientNetV2-S and VGG16 were 0.833 ± 0.047 and 0.839 ± 0.041, whereas the area under the receiver operating characteristic curve (AUC) were 0.876 ± 0.028 and 0.920 ± 0.009, respectively. Grad-CAM and Layer-CAM heatmaps suggested that the VGG16 model more frequently attended to regions outside the fracture site. CONCLUSION: Both VGG16- and EfficientNetV2-S-based models showed good performance in recommending the surgical strategy for femoral neck fractures. VGG16 yielded a significantly higher AUC, whereas EfficientNetV2-S had a more favorable computational cost and explainability.