Smart Triage: A Machine Learning Approach To Hospital Prediction Optimizing Ischemic Stroke Patients Outcome

Background Stroke is a leading cause of death and disability worldwide where “time is brain”. Thrombolysis is the most common treatment and is available in all certified hospitals. Otherwise, thrombectomy is not available in all hospitals, but it is the most effective treatment. Nevertheless, current guidelines determine that stroke patients must be transported to the nearest hospital without considering the availability of thrombectomy. However, previous studies suggested that stroke patients could benefit from being directly transported to the nearest hospital that performs thrombectomy instead, even for a longer driving time. In this study, we built a Machine Learning based approach to predict which hospital optimizes the good outcome for ischemic stroke patients. Methods We used data from a Portuguese hospital from a period 2006-2018 with a total of 556 observations. We compared the performance of different Machine Learning models to predict the hospital that optimizes good outcomes for ischem-ic stroke patients. We used known features at the activation of medical emergen-cies such as age, gender, the driving time to the nearest hospital with throm-bectomy, and the patient delay on activating medical emergencies to predict which hospital is associated with a higher chance of good outcome (the nearest hospital with thrombolysis or the nearest with thrombectomy). Results The results showed that Multilayer Perceptron Neural Network was the most suitable model, with AUC of 0.83 and 0.85 and F1-Score of 0.82 and 0.83 for training and test datasets. Also, after SHapley Additive exPlanations analysis, the driving time to the nearest hospital with thrombectomy was revealed as being the most important feature of decision in our predictive model. Conclusion Our Machine Learning model approach can predict the hospital desti-nation which optimizes good patient outcome. The most suitable model was MLP Neural Network. Furthermore, this model can lead to a most effective triage and treatment.