P-036 Artificial neural network modeling of clinical outcomes in subarachnoid hemorrhage. (23rd July 2022)

• Record Type: Journal Article
• Authors: Steed, T; Alawieh, A; Akbik, F; Sadan, O; Samuels, O; Grossberg, J
• Is Part Of: Journal of neurointerventional surgery. Volume 14(2022)Supplement 1
• Page Start: A71
• Page End: A72
• Publication Date: 2022-07-23
• Subjects: Nervous system -- Surgery -- Periodicals; Cerebrovascular disease -- Surgery -- Periodicals; 617.48
• DOI: 10.1136/neurintsurg-2022-SNIS.108
• Languages: English
• ISSNs: 1759-8478

Abstract

Abstract : Background: Aneurysmal subarachnoid hemorrhage is a devastating neurological condition which requires complex neurocritical care and neurosurgical decision-making and results in remarkably variable outcomes. Robust methods of identifying patients who may need operative or critical care intervention, such as those who develop hydrocephalus requiring long term shunting, cerebral vasospasm and delayed cerebral ischemia, or patients with worsened neurological outcome, are lacking. Methods: To this end, we applied artificial neural networks and machine learning techniques to a curated aneurysmal subarachnoid hemorrhage database collected at a large academic medical center between 2015–2020, ( n = 855). Results: Using this data, we have generated artificial neural network models with hyperparameter tuning which predict the need for long-term cerebrospinal fluid (CSF) diversion via placement of a shunt (receiver operating characteristic area under the curve (ROC AUC) = 0.9916), delayed cerebral ischemia (ROC AUC = 0.9903), and vasospasm (ROC AUC = 0.9834) from information available during the first few days of a patient's admission. This data included demographic and medical history information including gender, age, race, Hunt and Hess Scale, hypertension, diabetes, smoking, coronary artery disease, dyslipidemia, and the category of aneurysmal treatment. Additionally, external ventricular drainage variables such as external ventricular drain (EVD) output, EVD level, EVD age, and CSF parameters, such as CSF protein level, were used. The models were trained on a subpopulation of 80% of subjects and then individually validated in the withheld 20% of subjects with high reproducibility and accuracy. Finally, using the same data, another model was generated which accurately predicted categorical disposition at time of discharge (home, subacute rehabilitation, acute rehabilitation, long-term acute care, expiration in hospital) (ROC AUC = 0.9934). Conclusion: We hope the use of artificial intelligence and machine learning techniques will continue to demonstrate power in predicting complex medical outcomes and ultimately help neurosurgeons and neuro-critical care personnel to provide appropriate and timely treatment. Disclosures: T. Steed: None. A. Alawieh: None. F. Akbik: None. O. Sadan: None. O. Samuels: None. J. Grossberg: 1; C; Grant: GRA, EMCF, Neurosurgery Catalyst. 2; C; Cognition.