(HealthDay News) — A machine learning model based on high-ranking biomarkers can predict mortality in patients with hip fracture, according to a study published online Sept. 20 in the Journal of Orthopedic Research.
George Asrian, MD, PhD, from the University of Pennsylvania in Philadelphia, and colleagues examined whether a machine learning model trained on basic blood and lab test data as well as basic demographic data can predict mortality following a hip fracture. Key variables most associated with one-, five- and 10-year mortality were identified, and their clinical significance was investigated. Data were included for 3,751 hip fracture patients.
The researchers found that the one-year mortality rate was 21% for all patients studied and 29% for those aged 80 years and older. Ten different machine learning classification models were assessed; LightGBM had the strongest prediction performance for one-year mortality, with accuracy of 91%, an area under the receiver operating characteristic curve of 0.79, and sensitivity and specificity of 0.34 and 0.98, respectively, on the test set. For the one-year model, the strongest-weighted features included age, glucose, red blood cell distribution width, mean corpuscular hemoglobin concentration, white blood cells, urea nitrogen, prothrombin time, platelet count, calcium levels and partial thromboplastin time. The top 10 features of the LightGBM five- and 10-year prediction models included most of these features.
“LightGBM is a robust and powerful tool for predicting mortality across short and long timeframes, allowing for simple analysis of the most important input variables,” the authors write.
