Machine Learning Model Predicts Delayed Bleeding Risk in Patients With Torso Trauma

A machine-learning model presented in the Journal of Personalized Medicine predicted the risk of delayed bleeding in patients with torso trauma in the intensive care unit (ICU) with an accuracy of 0.816, according to a research team from Chang Gung Memorial Hospital in Taiwan.

“In the current study, we developed the algorithm consisting of 56 elements from baseline hemodynamic parameters before entering the ICU, and regular physiologic and laboratory parameters during the ICU stay,” the authors wrote. “No additional work or examination needed to be performed to attain the prediction results.”

The project accessed electronic health record data for 2218 patients with torso trauma admitted to the surgical ICU. Among them, 46.7% experienced hemorrhage during their stay.

Using patient demographics, vital signs, and lab data, a logistic model tree predicted a drop in hemoglobin 6 hours before it happened with 73.8% precision, 71.3% sensitivity, and 75.9% specificity, with an area under the curve (AUC) of 0.816 for the validation set. Meanwhile, a random forest algorithm achieved 73.6% precision, 73.6% sensitivity, and 73.7% specificity, with an AUC of 0.809 for the validation set, according to the study.

The exploratory research used 10-fold cross-validation to validate results. Researchers plan to conduct external validation for the best algorithm in the future.

“Our study demonstrates that the machine learning-based algorithm can be a promising tool for evaluating trauma patients in predicting delayed bleeding…” researchers wrote. “In this study, a machine learning-based algorithm helps us to identify risky patients with potential delayed hemorrhage events in the next 6 hours. Once patients enter the ICU and the data is automatically attributed to the model, the system will warn us if this patient has a potential bleeding risk in the next 6 hours.”

Reference:
Lee SW, Kung HC, Huang JF, et al. The Clinical application of machine learning-based models for early prediction of hemorrhage in trauma intensive care units. J Pers Med. Published online November 14, 2022. doi:10.3390/jpm12111901