296 Can Machine Learning Help to Predict Delays in Acute Stroke Care?

Abstract

Time to revascularization is a key factor affecting the outcome of acute stroke care. We aimed to assess the utility of machine learning with a neural network for predicting the Door to Needle time (DNT) in acute stroke care. This intervention would help alert the treatment team of potential delays associated with the arrival of new patients with an acute cerebrovascular accident (CVA). Historical patient medical records of acute stroke thrombolysis available at our hospital from 2003-2018 (15 years) were used. Data from 2003-2015, consisting of 221 cases, were used in training. Data from 2016-2018 were used for validation. A 3-layered fully connected neural network was implemented using the Tensor Flow framework (An open source machine learning framework). Duration, severity by National institute of health stroke scale (NIHSS), age in years and hour of day at arrival were identified as the predictive features. A door to needle time of above 45 minutes was set as the target for training the neural network. The network after training showed a positive predictive value of 63% and sensitivity 67% on validation, with an overall accuracy of 60%, defined as the percentage of delay > 45 minutes, which was correctly predicted. With the above 4 parameters, though individually having a weak association, it was possible to develop a prediction model for DNT with 60% accuracy and 63% positive predictive value. This type of prediction models can be a potential solution to optimize and help identify errors or delays in various acute care pathways including those for acute stroke, acute myocardial infarction, and sepsis. Application of machine learning in this context of optimizing acute care based on a predictive model has not been reported in the literature so far.