Machine learning model predicts in-hospital complications after traumatic brain injury

Patients admitted to hospital with acute traumatic brain injury often face a second wave of danger: complications that develop during the hospital stay itself, from prolonged recovery to poorer neurological outcomes. A new retrospective study published in Brain Informatics reports that a simple model built from four admission-level variables can predict which patients are at high risk of these in-hospital events, with a logistic regression version of the model achieving an area under the curve of 0.952 (95% CI, 0.905–0.992).

The study, led by Mei-Hua Wang and colleagues, retrospectively enrolled 186 adult patients with acute traumatic brain injury who were admitted to a Neurocritical Care Unit. The researchers set out to develop and validate a clinically practical model for predicting in-hospital complications in this population, using data available at the time of admission.

Of the 186 patients enrolled, 136 — or 73.12% — experienced in-hospital complications. That high event rate underscores how commonly acute traumatic brain injury is complicated by in-hospital events that contribute to prolonged hospitalization and poor neurological outcomes, the authors note in the study’s background.

To build the predictive models, the team compared five modeling approaches: logistic regression, Lasso regression, random forest, support vector machine, and decision tree. These methods were used both to screen candidate predictors and to construct prognostic models. Model performance was then assessed using discrimination, calibration, and decision-curve analysis, a combination intended to capture not only how well the models separated high-risk from low-risk patients but also how accurately their predicted probabilities matched observed outcomes and whether the models offered meaningful clinical value.

Four admission-level variables emerged as independent predictors of in-hospital complications: HbA1c, the Glasgow Coma Scale (GCS) score, stress-induced hyperglycemia (SIH), and severe extracranial injuries. Each of these can be measured or assessed shortly after a patient arrives, which is what makes the model practical for early risk stratification rather than a tool that depends on information gathered over subsequent days.

The final model incorporating these four variables showed strong discrimination, reliable calibration, and meaningful clinical utility, according to the study’s results. Among the five approaches tested, logistic regression performed best, achieving an AUC of 0.952 with a 95% confidence interval of 0.905 to 0.992. An AUC in that range indicates a high degree of separation between patients who went on to develop complications and those who did not in this retrospective dataset.

The authors conclude that a simple and practical model incorporating HbA1c, GCS score, SIH, and severe extracranial injuries effectively predicts the risk of in-hospital complications in acute TBI patients, and that the model may assist clinicians in early risk stratification and clinical decision-making.

Several caveats accompany these findings. The study was retrospective and single-center, enrolling patients from one Neurocritical Care Unit, and the authors themselves state that further multicenter prospective studies are needed to validate and refine the model. A retrospective design can identify associations and demonstrate predictive performance within the studied cohort, but it cannot establish that the model will perform equally well in other hospitals, health systems, or patient populations. The reported AUC reflects performance in this specific dataset, and external validation remains an open question.

The study was published as open access research in Brain Informatics on 18 September 2026, and the version shared is an accepted manuscript released early to provide faster access to peer-reviewed research; it is subject to further edits before the final Version of Record appears.

Subject of Research: Prediction of in-hospital complications after acute traumatic brain injury using machine learning models

Article Title: Early prediction of in-hospital complications following acute traumatic brain injury: a retrospective study using multiple machine learning models

Article References: Wang, M.-H., Li, R., Fang, J., Deng, X., Zhang, J., Guo, Y., Chen, L., Chen, L., Yu, C., Chen, Y., Wu, G., Hu, J., Yao, H., Yuan, Q., & Yang, L. (2026). Early prediction of in-hospital complications following acute traumatic brain injury: a retrospective study using multiple machine learning models. Brain Informatics. https://doi.org/10.1186/s40708-026-00334-w

Image Credits: AI Generated

DOI: 10.1186/s40708-026-00334-w

Keywords: traumatic brain injury, in-hospital complications, machine learning, logistic regression, Glasgow Coma Scale, HbA1c, stress-induced hyperglycemia, risk prediction, Neurocritical Care Unit, retrospective study, Early, prediction