Abstract
BACKGROUND: A large number of studies have focused on building different models to predict postoperative mortality in elderly patients with hip fractures, including sarcopenia risk factor models or deep learning models. However, relying on deep learning models alone may not fully capture the key factors that affect patient outcomes, so it may be a more accurate model to construct predictive models combining clinical baseline features. METHODS: A deep learning model (Densenet161) and a deep learning-clinical baseline feature fusion model (LightGBM) were constructed using 221 patients from Institution 1 as the internal training set and 113 from Institution 2 as the external validation set, respectively. We selected the skeletal muscle tissue image of the 12th thoracic vertebral cross section in the chest CT (computerized tomography) scan as the input data of the Densenet161 model. The model's predictive performance was evaluated using AUC (area under the curve), sensitivity, specificity, and F1 scores. RESULTS: The Densenet161 model has an average performance in predicting 1-year postoperative mortality in elderly patients with hip fractures, with an AUC of 0.723 and an F1 score of 0.421 on the external validation set. Compared with Densenet161 model, the predictive performance of LightGBM fusion model has been greatly improved, with AUC of 0.815 and F1 score of 0.819 on the external validation set. CONCLUSION: Combining the image features extracted by the deep learning model with the patient's clinical baseline characteristics, the LightGBM fusion model can better predict the 1-year mortality of elderly hip fracture patients than relying on a single deep learning model.