A practical approach to predicting long-term outcomes in traumatic brain injury: Enhancing clinical decision-making with machine learning

Abstract

BACKGROUND: Traumatic brain injury (TBI) is among the most prevalent causes of emergency department visits globally. TBI leads to high morbidity and mortality rates, which poses a noteworthy burden on the medical system regarding both patients and economics. In this study, we aimed to enhance clinical decision-making and resource allocation by predicting the six-month outcome of patients with TBI based on an extended Glasgow outcome scale using CatBoost, a deep-learning model based on gradient-boosted decision trees. METHODS: Data were obtained from the TBI registry at Shahid Rajaee Trauma Hospital, Shiraz. Data included a total of 3132 patients admitted from 2016 to 2021. The primary target of this study was to predict the six-month outcomes after discharge, which were categorized into three categories: death, disability-dependent outcome, and favorable outcome. Secondary goals were the prediction of favorable versus unfavorable outcomes as well as patient survival. Data augmentation using CTGAN was used to improve model performance by balancing the data. Ten clinically relevant features available at the time of admission, including patient age, gender, Motor part of GCS, Rotterdam index, blood sugar level, pupil state, INR, systolic blood pressure, and need for emergency laparotomy and decompressive craniotomy, were included. The model's performance was evaluated using accuracy, precision, recall, F1 score, and area under the receiver operating characteristic curve (AUC). RESULTS: The CatBoost model achieved an overall accuracy of 76.91 % on the primary dataset, improving cross-categorical metrics when using data augmentation. Precision and recall for predicting favorable outcomes reached 80.66 and 91.16 percent, respectively, with an AUCμ of 0.76. Data augmentation improved the metrics, especially in under-represented categories, improving the precision and recall in the disability-dependent outcome categories from 12.50 to 70.50 and 1.54 to 68.04 percent. AUCμ was improved to 0.88 as well. CONCLUSION: Curation of an applicable interface to readily assess and predict the long-term outcome of patients with TBI can help clinical decision-making. This study utilizes a practical approach to predict the outcome of such patients and provides an interface to be easily applied in the clinical setting.

Ähnliche Arbeiten