Using Explainable Artificial Intelligence To Predict Future Stroke Risk Using Routinely Collected Historical Investigations

Background: Stroke is a leading cause of mortality and disability in the UK, with 25-30% of stroke cases cryptogenic, highlighting an ever-increasing need for prevention and identification of unknown risk factors. Primary stroke prevention is currently based upon scoring tools such as QRISK3, CHA2DS2-VASc and ABCD2. These tools lack specificity and sensitivity for stroke. Currently, no tool exists to accurately predict the risk of stroke in the general population and those without a diagnosis of AF.Machine learning models are a promising area which holds the potential to predict the future risk of stroke and identify novel risk factors.MethodologyHistorical data will be collected from 11,000 stroke patients using the SSNAP database, along with 120,000 matched control patients. Data collected will include ECGs, MRIs, CT, Echocardiograms, laboratory data and past medical history, using routine clinical databases, national databases, hospital and GP records. Data collected will be anonymised, and identifiers removed. Separate machine learning models will be created for each investigation type, with a future combination of these models for calibration. The models will be calibrated to assess whether the distribution of generated probabilities reflects those who develop a future stroke. Future steps include testing on external datasets for comparable outcomes. Discussion The study aims to use ML techniques to provide personalised percentage stroke risk to patients at 1, and 13 years. This will inform early implementation of primary and secondary prevention methods, such as lifestyle changes or pharmacological intervention, and identify new novel risk factors for stroke.