Machine learning model predicts new-onset deep vein thrombosis of the lower extremities after pelvic floor fracture surgery and targeted diagnosis

BackgroundPostoperative new-onset deep vein thrombosis (PNO-DVT) of the lower extremities represents a prevalent and serious clinical complication following pelvic fractures, which substantially impedes patient rehabilitation and diminishes quality of life. Conventional risk assessment methodologies exhibit inherent limitations, rendering them inadequate for precise prediction and individualized management of DVT. In recent years, machine learning techniques have demonstrated significant advantages in data analysis, emerging as promising tools for predicting postoperative DVT risk. This study sought to investigate the predictive efficacy of machine learning models for the development of new-onset lower extremity deep vein thrombosis following pelvic fracture surgery. MethodsData from 745 patients who underwent pelvic fracture surgery at our hospital between January 2016 and December 2019 were collected. The analysis encompassed demographic information, general patient data, preoperative laboratory test results, surgical details, and scoring systems. Initially, the data were analyzed using univariate logistic regression, least absolute shrinkage and selection operator (LASSO) regression, and multivariate logistic regression to identify 12 independent risk factors, including age, HDL-C, and ApoB. Subsequently, the dataset was partitioned into a training set and a test set at a 7:3 ratio. Six models were employed for analysis, including logistic regression, support vector machine (SVM), random forest, XGBoost, LightGBM, and AdaBoost. ConclusionComparative analysis of the six machine learning models revealed XGBoost exhibited the highest performance (AUC: 0.8633), followed by LightGBM (0.8349), random forest (0.8055), logistic regression (0.7503), SVM (0.7505), and AdaBoost (0.8179). Model sensitivity ranged from 0.3684 to 0.8421, and accuracy ranged from 0.6502 to 0.9238.