265 Use of support vector machine to predict discharge destination on an acute stroke ward

Abstract

BackgroundPredicting the discharge destination of stroke patients at admission is important for stream- lining of services, management of patient and carer expectations, and the planning and facilitating of discharge to appropriate destinations.MethodsWe reviewed data from the Sentinel Stroke National Audit Programme (SSNAP) from a single acute stroke ward in South Wales. We analysed age, modified Rankin Scale (mRS) and National Institute of Health Stroke Scale (NIHSS) data using a multi-class Support Vector Machine (SVM) to train and test a model to predict discharge destination.ResultsData was extracted from 1131 patients with acute and subacute stroke admitted to an acute stroke unit between January 1st, 2015, and December 31st, 2016. Per-class sensitivity of the model for final destination or outcome was 0.89 for home, 0.6 community hospital, 0.75 inpatient rehabilitation, 0.45 death. Per-class positive predictive value (PPV) was 0.84 for home, 0.62 community hospital, 0.72 inpatient rehabilitation, 0.55 death.ConclusionFrom data that is available from simple assessments done in routine clinical pratice it is possible to identify patients with acute stroke for whom discharge home or inpatient rehabilitation is highly likely. Collaborations between clinicians and data scientists may help to maximise the value of clinical assessment tools that are already in widespread use.Tom.Hughes2@wales.nhs.uk