Enhancing Stroke Severity Evaluation: A Machine Learning Approach to Mortality Prediction Versus Traditional Scales

The National Institutes of Health Stroke Scale (NIHSS) is the standard tool for assessing the severity of acute ischemic strokes; however, it has significant limitations. There is a growing need to develop new scales and indicators that address the limitations of the NIHSS scale and provide a more comprehensive, holistic assessment of the patients. In this study we explore a novel severity scale based on a machine learning model that classifies the patient outcome (alive or deceased) using data from 5983 data stroke cases from the Vall d'Hebron Hospital from January 2018 to December 2023. We propose using the probability of belonging to the "deceased" class as a metric to measure stroke severity and patient outcome prediction. To evaluate its effectiveness, we compare this metric to the NIHSS scale measured at 24h after patient arrival at the emergency service. Our model achieved an AUC-ROC of 87%, significantly outperforming the NIHSS-based prediction, which yields and AUC-ROC 53%. These results highlight the improved capability of our approach in assessing patient severity and predicting mortality risk. The proposed method provides a more holistic assessment by incorporating a diverse set of variables, and therefore offering a more comprehensive representation of patient status.Clinical Relevance-This establishes an additional measure of stroke severity that can be instrumental in identifying which patients may require more intensive monitoring or specialized care. By quantifying the risk of mortality, the model can help assess the patient's future outcome and prioritize those at higher risk, enabling more targeted interventions and resource allocation, hence improving patient outcomes.