Machine learning and transformer models for prediction of postoperative pneumonia risk in patients with lower limb fractures

Abstract

Postoperative pneumonia, a prevalent complication arising from lower limb fracture surgery, can significantly prolong hospitalization periods and elevate mortality rates. Consequently, early prevention and identification of this condition are crucial in improving patient prognosis. In this study, clinical indicators pertaining to postoperative pneumonia in patients with lower limb fractures at Nantong University Hospital, spanning the years 2016 to 2023, were subjected to a analysis. The patients who encountered postoperative pneumonia subsequent to their lower limb fracture surgeries during hospitalization were categorized as the case group, whereas those who did not develop such a complication served as the control group. To forecast the likelihood of postoperative pneumonia occurrence, both machine learning and deep learning algorithms were employed. The study identified Age, Gender, Fracture type, Venous thromboembolism (VTE), Hypertension, Chronic obstructive pulmonary disease (COPD), Cancer, Atrial fibrillation, Cerebrovascular disease, Hypoalbuminemia, Free fatty acid, Albumin, Albumin to globulin ratio, Calcium, Fibrinogen, D-dimer, Alcohol, Surgical grade and C-reactive protein as significant predictors of postoperative pneumonia. XGBoost and Transformer models have better performance (AUC 0.866 VS 0.946, F1 0.807 VS 0.889), and both models have better substantial prediction ability for the occurrence of postoperative pneumonia. In conclusion, XGBoost and Transformer models serve as potential tools for the prevention and treatment of postoperative pneumonia in patients with lower-extremity fractures. By adopting appropriate health management practices, the risk of developing postoperative pneumonia in this patient population may be reduced.

Ähnliche Arbeiten