Trustworthy Artificial Intelligence for Improvement of Stroke Outcomes

Stroke is a leading cause of death and disability worldwide. The clinical validation of explainable and interpretable Artificial Intelligence (AI) solutions to assist a timely, personalised management of the acute phase of stroke, would have a major impact since it can greatly reduce the disability levels of patients. Also, the prediction of long-term outcomes is a crucial factor as it may determine critical decisions such as the discharge destination for the patient. Moreover, compliance with guideline-based secondary stroke prevention has been demonstrated to reduce stroke recurrence, but currently, only 40% of patients are adherent to preventive treatments 3 months after stroke. Therefore, patients´ outcomes can improve with proper patient communication and engagement packages. AI may have a dramatic impact on stroke patient journey, improving predictions, resulting in a better choice of secondary stroke strategies, as well as using evidence-based information to promote better adherence to treatment and reduction of vascular risk factors. The aim of this multicentre observational prospective study is to develop and validate AI-based tools to predict short and long-term outcomes in ischemic stroke patients. Specifically, this study aims to demonstrate the accuracy of AI models in predicting the functional outcome of ischaemic stroke patients as measured by the National Institutes of health Stroke Scale (NIHSS, 0-42) and the modified Rankin Scale (mRS, 0-6) scores at hospital discharge and at 3, 6 and 12 months after discharge. Prospective ischemic stroke patients from 3 Large European centres will be recruited. The training and testing of local AI models will be performed using hospitalization data, collected during the standard of care procedures for stroke patient pathways, and outpatient monitored data from a remote home-care system (NORA app) during the follow-up after discharge. These local models will then be integrated into a federated learning system, where only a global AI model, derived from combined insights of all local models, is shared across participating hospitals. The individual local models and the original data are not shared, ensuring data privacy and security. The accuracy and performance of prospectively optimized AI models in predicting clinical outcomes over a 12-month follow-up period will be evaluated and compared to the actual outcomes of the patients.