Development of Machine-learning Model to Predict Anticoagulant Use and Type in Geriatric Traumatic Brain Injury Using Coagulation Parameters.

This study aimed to investigate the patterns of anticoagulation therapy and coagulation parameters and to develop a prediction model to predict the type of anticoagulation therapy in geriatric patients with traumatic brain injury. A retrospective analysis was performed using the nationwide neurotrauma database of Japan. Elderly patients (≥65 years) with traumatic brain injury. Patients were divided into 3 groups based on their daily anticoagulant medication (none, direct oral anticoagulant [DOAC], and vitamin K antagonist [VKA]), and coagulation parameters were compared in each group. We then developed a machine-learning model to predict the anticoagulant using coagulation parameters and visualized the pattern using a heat map. A total of 495 patients were enrolled and divided into 3 groups: none (n = 439), DOACs (n = 37), and VKA (n = 19). Comparing none to DOAC and DOAC to VKA for prothrombin time-international normalized ratio (PT-INR), the mean difference and 95% confidence intervals (CIs) were 0.38 (95% CI: 0.59-0.17) and 1.56 (95% CI: 1.21-1.90), and for activated partial thromboplastin time (APTT), the mean difference between none to DOAC and DOAC to VKA was 3.46 (95% CI: 0.98-5.94) and 95% CI was 7.39 (95% CI: 3.29-11.48). A prediction model for the type of anticoagulant used by PT-INR and APTT was developed using machine-learning methods, and a heat map visually revealed their relationship with acceptable predictive ability. This study revealed the characteristic patterns of coagulation parameters in anticoagulants and a pilot model to predict anticoagulant use.