Outcome prediction of endovascular therapy for acute stroke with emphasis on common pitfalls in the application of artificial intelligence in medicine

• Applicant: Dr. Quirin Strotzer
• Subject Area: Nuclear Medicine, Radiotherapy, Radiobiology ; Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing ; Clinical Neurology; Neurosurgery and Neuroradiology ; Medical Informatics and Medical Bioinformatics
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 522669021

Cerebral stroke is one of the leading causes of morbidity and mortality worldwide, with ischemic insult accounting for the largest proportion. Besides intravenous thrombolysis, endovascular therapy by mechanical thrombectomy (interventional removal of the blood clot with wires and catheters inserted, for example, via the groin) is the most important therapeutic option for acute ischemic stroke. Targeted selection of patients suitable for thrombectomy could significantly increase success rates while conserving financial and human resources. However, due to the time-critical nature, it is nearly impossible for clinical staff to consider and correctly classify the amount of all necessary information when making an indication for interventional therapy. A comprehensive solution that provides acceptable predictive values on the likelihood of success of endovascular therapy for acute stroke in a clinical context does not yet exist. Since large amounts of data (clinical parameters, imaging, etc.) must be evaluated for decision-making, applying methods from the field of artificial intelligence (AI) is promising. However, these should be handled with caution due to the potentially far-reaching consequences for patients. Among other things, models can pick up and amplify bias present in the training data or lead to erroneous predictions in underrepresented data. With the help of artificial intelligence, a model will be developed that uses clinical and imaging parameters to estimate the chances of success of mechanical thrombectomy for treating acute stroke and thus supports the indication and therapy planning in the acute setting. In this context, it will also be analyzed which factors have the greatest influence on the outcome. For this purpose, a multicenter database with clinical parameters and image data will be created and expanded prospectively. Since research in the field of AI in medicine often is purely theoretical and gained knowledge and developed models rarely find their way into clinical reality, a subsequent prospective clinical validation will be carried out. Starting with the data collection, via the model development and the evaluation, up to the prospective validation, special attention is paid to common pitfalls (e.g., data leakages), which usually lead to the malfunction of AI models. Thus, a model is obtained that truly adds value.

• DFG Programme: WBP Fellowship
• International Connection: USA
• Host: Professor Rajiv Gupta, Ph.D.