# Stockholm Score of Lesion Detection on Computed Tomography following Mild Traumatic Brain Injury (SELECT-TBI) Study: Pilot Analysis and Statistical Analysis Plan

## TL;DR

This study explores using data-driven models to predict brain injury risks in patients with mild traumatic brain injury, aiming to improve clinical decision-making.

## Contribution

The study introduces a novel data-driven approach for risk stratification of traumatic intracranial lesions in mild traumatic brain injury patients.

## Key findings

- The Lasso regression model achieved an AUC of 0.807 for any intracranial lesion and 0.903 for clinically significant lesions.
- Key clinical variables included Glasgow Coma Scale, signs of basilar skull fracture, trauma mechanism, and vomiting.

## Abstract

Mild traumatic brain injury (mTBI) is a common cause of emergency department visits. Only a small percentage of mTBI patients develop an intracranial lesion (ICL) and even fewer will require neurosurgical intervention due to their injury. The Stockholm Score of Lesion Detection on Computed Tomography following Mild Traumatic Brain Injury (SELECT-TBI) study aims to provide a data-driven approach to estimate individualized risk for traumatic ICL and clinically significant lesions in mTBI patients.

To provide a statistical analysis plan and pilot data analysis before completion of data collection, as pre-planned in the published study protocol.

Retrospective study of patients ≥ 15 years old who underwent a computed tomography (CT) scan for their mTBI in Stockholm, Sweden, between 2015–2020. Up to 73 variables were collected for each patient. Data analysis of the first 5 000 patients in the cohort was conducted to develop preliminary prediction models using Lasso regression, general linear model and random forest and to perform an optimal population analysis to determine whether the final sample size would be sufficient.

Six data selection strategies were tested, and area under the curve (AUC) receiver operator characteristic (ROC) curves were generated with a 4:1 training/validation data segmentation. The best-performing model was the Lasso regression model which achieved an AUC of 0.807 for any ICL and 0.903 for clinically significant ICL (accuracy of 70% and 97.7%, and Brier scores of 0.3 and 0.023 respectively). Clinical variables identified as key features across all models were Glasgow Coma Scale, signs of basilar skull fracture, trauma mechanism, and vomiting, each with an importance score greater than 0.1 (explaining more than 10% model variance). Finally, the highest end prediction of the necessary population size was found to be 29 667 patients.

Our preliminary results demonstrate the potential for a data-driven approach to generate personalized risk stratification tools. With a final cohort size expected to exceed 40 000 patients, we anticipate being able to create more granular models optimized for integration into clinical decision-making.

ClinicalTrials.gov NCT04995068.

The online version contains supplementary material available at 10.1007/s00701-025-06598-1.