Aim This study applied Machine Learning model to develop a predictive model for factors associated with mortality in older patients with hip fracture. This will enhance the potential to identify high-risk patients using multiple factors simultaneously that account for the complexity in older patients.
Methods Nation-wide data of 65,709 hip fracture patients were used to develop a model applying Lasso regression method to predict mortality. The data of potential factors including 17 comorbidities were used to generating scores used to create an optimal threshold for classifying patients into high and low risk groups. Subsequence validation of the model by evaluating mortality, length of stay, and hospital costs across the risk groups. Model performance was evaluated by Accuracy, Precision, Recall, F1-score, and Area Under the Curve (AUC) of Receiver Operating Characteristic curves (ROC).
Results Our model classified hip fracture patients into high-risk(n=4,748) and low-risk(n=8,394) groups. The model identified key predictors for high group as male, higher age and non-surgical management. Comorbidities identified in high-risk group included acute myocardial infarction, chronic heart failure, dementia, chronic obstructive pulmonary disease and chronic kidney disease. The high-risk group had significantly higher 30-days mortality (12.68% vs. 2.37%) (p value < 0.05, t-test). The model had reasonable performance with an AUC of 0.71 and recall of 0.75.
Conclusion The model using machine learning effectively classifies hip fracture patients in high-risk group with multiple key factors resulting in higher mortality. Identify the group with high risk to better management of modifiable risk factors might be beneficial.