The role of preoperative laboratory test indicators in predicting thrombosis risk in elderly hip fracture patients: a random forest approach

Deep vein thrombosis (DVT) poses a common and critical risk for mortality in elderly hip fracture (HF) patients. Venous angiography and ultrasound examinations serve as crucial diagnostic tools but pose challenges in cases with prevalent complications. The extensive training period for technical personnel, coupled with the rapid advancements in machine learning, prompts our research to harness the potential of the random forest algorithm. Our aim is to construct a predictive model that evaluates the risk of thrombosis formation in elderly HF patients upon admission. We conducted a retrospective evaluation of 448 elderly HF patients who received surgical treatment between May 2021 and November 2023. The division of our study cohort ensued, randomly allocating participants into training and test datasets with a meticulous 70:30 ratio. The employment of the Random Forest algorithm in this process led to the creation of a refined and optimized predictive model. Eleven important variables, namely ALB, A/G, PLT, Fib, D-dimer, CREA, Pi, GFR, FDP, fracture onset to surgery time, and cardiopulmonary diseases, were screened based on Random Forest features. In the training set, AUC, the 95% Confidence Interval (CI), Sensitivity, Specificity, Precision, Accuracy, and Balanced Accuracy stand at 1.000, (0.9883, 1), 1.000, 1.000, 1.000, 1.000, and 1.000, respectively. For the test set, AUC, the 95% CI, Sensitivity, Specificity, Precision, Accuracy, and Balanced Accuracy are 0.899, (0.7875, 0.9124), 0.6061, 0.9406, 0.7692, 0.8582, and 0.7733, respectively. A random forest prediction model was developed to anticipate the occurrence of preoperative lower extremity DVT in elderly HF patients. This model demonstrated superior accuracy compared to the logistic regression model. Key preoperative laboratory test indicators proved valuable as variables in the prediction process.