VALIDATE APP: A trustworthy-AI powered mobile application for clinical decision support in acute ischemic stroke treatment

VALIDAT A trustworthy AI-based prognostic tool for Ischemic Stroke Introduction and Purpose In acute ischemic stroke, reperfusion treatments offer significant benefits across various stroke subtypes. Yet, outcomes can be influenced by often overlooked patient factors. We developed a tool to predict the functional outcome according to different treatment options, aiming to guide clinicians in personalized decision-making based on individual patient data. Methods A multidisciplinary team conducted thorough stakeholder research and co-creation. An ethical framework guided the development process, ensuring trustworthy AI. A Neural Network model was trained on multi-center retrospective data from over 7000 patients to predict outcomes based on the modified Rankin Scale at 3 months. Results The NN model achieved AUCs of 0.75-0.88 for predicting mRS>2. An iOS app with high usability captures features in the hyperacute setting, integrating the NN model to predict the full mRS distributions. The users can access predictions for various treatment options before decision, together with input feature impact and model confidence. Prospective, observational validation is ongoing in 3 clinical centers while larger scale retrospective validation continues on 10+ datasets from the EU, USA and Japan. Conclusion Our unique solution arms clinicians with individualized information about likely outcomes while our development process provides a case study for trustworthy-AI development, testing and validation in healthcare. Future considerations We  are  looking for new participating centers for retrospective validation ( high interest in Africa, Australia, Central- / South-America ) and clinical professionals for testing usability. In case of interest, learn more and  reach out on the links below. Prediction screen ( left ) shows the predicted distribution of the mRS in four treatment scenarios. D ata impact screen ( right ) depicts importance of the features in a dichotomized setting set by the mRS threshold on the prediction screen. D ichotomized, trichotomized and full-scale confusion matrices on test set Reliability plots, showing good model calibration on test set A. H ilbert (Charité Lab for AI in Medicine, Charité Universitätsmedizin Berlin), I. Werren (Digital Health, IBM iX, Berlin), M . Rubiera (Department of Neurology, Vall d’Hebrón University Hospital), R. L eker (Department of Neurology, Hadassah Medical Center), S. B onekamp (Department of Neuroradiology, Heidelberg University Hospital), J. D . K elleher (School of Computer Science and Statistics Trinity College Dublin), L . M . L opez-Ramos (Simula Metropolitan Center for Digital Engineering Oslo), V . I. M adai (QUESTCentrefor Responsible Research, BerlinInstitute of Health (BIH)), D . F rey (Charité Lab for AI in Medicine, Charité Universitätsmedizin Berlin) V A L I D ATE is an EU H orizonEurope funded pro j ect ( No. 1010572 6 3 ) Scan Q R Code for more InformationContact adam.hilbert @ charite.de © 2024 Hilbert et al.