Development of an AI Model for Early Detection of Needing Delayed Craniotomy or Craniectomy in the Setting of Head Trauma

Introduction: Approximately one million individuals are assessed for traumatic brain injury (TBI) in the US, with a significant percentage necessitating neurosurgical evaluation. The decision for emergent decompression is immediate in severe cases. However, predicting the need for delayed neurosurgical intervention in patients with less pronounced deficits remains challenging, often requiring ICU admission, frequent neurologic assessments, and multiple imaging studies. This study aims to assist in the decision-making process for severe TBI patients by introducing an AI-driven model to predict need for delayed craniotomy/craniectomy based on presenting head CT features.

Methods: A single center retrospective study of patients presenting to a Level 1 academic trauma center requiring neurosurgical consultation was performed. Imaging, demographic information, GCS score, injury severity score, and operative procedures were recorded. Patients who were admitted to the hospital without emergent craniotomy/craniectomy, but required a craniotomy/craniectomy secondary to their TBI during their hospital course were classified as requiring a delayed crani. Data were compiled and pre-processed to train a customized convolutional neural network (CNN) model, designed to predict the likelihood of delayed neurosurgical intervention. The model's architecture was tuned to account for the heterogeneity in our dataset.

Results: Over a ten-year period, need for a delayed craniotomy/craniectomy approximated 3%. Our CNN model demonstrated the ability to predict presenting head CT features suggestive of needing a delayed craniotomy despite the low frequency. Further model development is undergoing to further optimize predictive performance.

Conclusion : Despite TBIs being a heterogenous population, anomalous events requiring significant intervention can be assessed based on presenting features. This study demonstrates AI/ML techniques that can be used to assist providers in clinical decision making, incorporating multiple data sources, in risk stratifying this patient population.