Shared Human and Artificial Intelligence to Improve Surgical Decision Making for Brain Tumors

Introduction: A central tenet of neurosurgical oncology is achieving maximal safe resection of the tumor while preserving neurological function in the patient, a concept known as onco-functional balance. Deciding the extent of resection of the tumor poses a challenge because it is not possible to know the patient’s post-surgical neurological function before operating. It is possible that a shared intelligence approach combining a surgeon’s initial estimate on the extent of resection with an artificial intelligence model that returns the probability of introducing a post-operative neurological deficit based on the initial plan could refine surgical decision. We hypothesize that a machine learning model can further predict the post-operative neurological deficit beyond the surgical decision and that providing the surgeon’s initial decision can act as a prior that can boost the model performance.

Methods: Patients undergoing brain tumor surgery (n = 17) were sub-divided into two groups based on the extent of resection: gross total or sub-total resection. In each group, a logistic regression machine learning model was fit to predict the presence of a post-operative focal neurological deficit based on a feature set including age, sex, tumor size, eloquent region infiltration, planned intra-operative mapping, pre-operative focal neurological deficit, and radiological characteristics of a midline crossing tumor, tumor touching the ventricles, multi-gyri involvement, and white matter infiltration.

Results: For the patients who received gross total resection, eloquent region infiltration and planned intra-operative mapping further predicted post-surgical neurological deficit (p < 0.01). For the patients who received sub-total resection, white matter infiltration negatively correlated with post-operative deficit, while age and tumor size positively correlated with post-surgical deficit, which predicted the neurological deficit.

Conclusion : Pre-operative information can predict the likelihood of post-operative neurological deficits, conditioned on the surgical decision, suggesting that an artificial intelligence model can be used to aid the surgical decision.