Prediction Models in Aneurysmal Subarachnoid Hemorrhage: Forecasting Clinical Outcome With Artificial Intelligence.

Background: Predicting outcome after aneurysmal subarachnoid hemorrhage (aSAH) is known to be challenging and complex. Machine learning approaches, of which feedforward artificial neural networks (ffANNs) are the most widely used, could contribute to the patient-specific outcome prediction.

Objective: To investigate the prediction capacity of an ffANN for the patient-specific clinical outcome and the occurrence of delayed cerebral ischemia (DCI) and compare those results with the predictions of 2 internationally used scoring systems.

Methods: A prospective database was used to predict (1) death during hospitalization (ie, mortality) (n = 451), (2) unfavorable modified Rankin Scale (mRS) at 6 mo (n = 413), and (3) the occurrence of DCI (n = 362). Additionally, the predictive capacities of the ffANN were compared to those of Subarachnoid Haemorrhage International Trialists (SAHIT) and VASOGRADE to predict clinical outcome and occurrence of DCI.

Results: The area under the curve (AUC) of the ffANN showed to be 88%, 85%, and 72% for predicting mortality, an unfavorable mRS, and the occurrence of DCI, respectively. Sensitivity/specificity rates of the ffANN for mortality, unfavorable mRS, and the occurrence of DCI were 82%/80%, 94%/80%, and 74%/68%. The ffANN and SAHIT calculator showed similar AUCs for predicting personalized outcome. The presented ffANN and VASOGRADE were found to perform equally with regard to personalized prediction of occurrence of DCI.

Conclusion: The presented ffANN showed equal performance when compared with VASOGRADE and SAHIT scoring systems while using less individual cases. The web interface launched simultaneously with the publication of this manuscript allows for usage of the ffANN-based prediction tool for individual data (https://nutshell-tool.com/).