Abstract 166: Supervised Learning Based Detection of Stroke and Stroke Mimic
Background & Purpose: Up to 30% of treated patients for suspected acute ischemic stroke (AIS) could be stroke mimic (SM). Using a balanced clinical dataset, we have developed a supervised learning method from artificial intelligence to classify AIS/transient ischemic attack (TIA) and SM in an emergency room (ER) setting.
Methods: We labeled and randomly partitioned consecutive patients who received intravenous thrombolysis for suspected AIS in three tertiary centers during a three year period. The clinical data includes patients’ demographic information, past medical history, blood pressure and glucose level at presentation, NIH Stroke Scale, and presence of facial weakness. The clinical data were used to create a balanced training and testing set. We developed an Artificial Neural Network (ANN) model on patients’ clinical data, using the training set ( Figure 1). The model was used to classify the diagnosis of AIS/TIA versus SM. The ANN model was subsequently assessed using the testing set. In addition, we compared the ANN model with three regression models (ordinary least squares, logistic, and reduced logistic model).
Results: In total 803 patients (mean age 62±15 years, 52% men, median admission NIHSS-score: 6 points, IQR 3-11) were analyzed. Data from 56 random patients, not included in the training set, were used to test the model. Of 56 predictions, 49 were correct (19 out of 20 correct prediction for SM, and 30 out of 36 correct prediction for AIS/TIA). Additional analysis indicated that the model detects SM with 95% sensitivity (95%, CI: 73-99%) and 83% specificity (95%, CI: 66-93%). We further performed a 10-fold cross validation to better estimate the performance of the model. In addition, the three regression models were comparable to the ANN model at a specific adjusted threshold.
Conclusions: This study highlights a potential for ANN to discriminate AIS/TIA and SM. This model can be further optimized and integrated in patient management software in an ER setting.
Author Disclosures: V. Abedi: None. N. Goyal: None. N. Hosseinichimeh: None. G. Tsivgoulis: None. R. Hontecillas: None. J. Bassaganya-Riera: None. P. Lu: None. J. Chang: None. N. Ghaffarzadegan: None. L. Elijovich: None. J.E. Metter: None. A.W. Alexandrov: None. D.S. Liebeskind: Consultant/Advisory Board; Modest; Medtronic, Stryker. Research Grant; Significant; NIH-NINDS. A. Alexandrov: None. R. Zand: None.
- © 2016 by American Heart Association, Inc.