Prediction of Functional Recovery in Patients with Supratentorial Ischemic Stroke by Various Methodology of Machine Learning

Abstract

Objectives: The purpose of the present study was to predict the functional recovery of supratentorial ischemic stroke after post-stroke 3 months with the clinical data of patients obtained within 2 weeks from onset of stroke by using various methodology of machine learning (ML) including artificial neural network (ANN).

Methods: We extracted a list of patients who had been discharged from the Department of Rehabilitation Medicine, a university hospital from Jan. 2000 to Dec. 2017. Afterward, we collected the clinical data of patients meeting inclusion/exclusion criteria. We selected features for the construction of a prediction model among clinical features which has been known to affect post-stroke recovery and expected to affect it. The selected clinical features are age, sex, initial National Institutes of Health Stroke Scale, presence of internal capsule posterior limb involved, strength of shoulder abduction, wrist extension, hip extension, and knee extension, Mini-Mental State Examination, presence of hemorrhagic transformation, aphasia, visuospatial neglect, and depression. We dichotomized post-stroke 3 months functional status assessed with modified Barthel Index, which was used as an outcome label for the prediction model. We optimized the hyperparameters of ANN model and the other method of machine learning by using the grid search with 2-fold cross validation. We repeated the training and validation session 10 times with the different configuration of training and test dataset generated by randomized sampling. The average of performance of 10 individual models was assigned to represent overall performance of the respective method of machine learning.

Results: We screened 5210 patients and eventually enrolled 101 patients with supratentorial ischemic stroke, whose functional recovery was assessed with modified Barthel Index after 3 months post-stroke. The mean age of the enrolled subjects was 62.40 ± 12.67 [19 - 79] years. The patients in group with better functional status after post-stroke 3 months tend to have younger age (59.77 ± 14.18 versus 67.34 ± 7.02) and lower initial NIHSS (7.98 ± 5.26 versus 14.42 ± 5.48), and less likely to have stroke lesion in posterior limb of internal capsule (15.15% versus 57.14%), and have better cognitive function (total MMSE

24.54 ± 5.97 versus 18.86 ± 7.02). The architectures of ANN with optimized hyperparameters was turned out to have 4 hidden layers with from 64 to 4 nodes. The proposed ANN model used rectified linear unit as activation function, Glorot uniform initializer as the way to set the initial weight, 0.3 dropout rate, Adagrad as optimizer, 0.02 learning rate, 5 batch size, 600 epochs and binary crossentropy as loss function. Sigmoid function as the classifier was placed at the last layer for prediction. The accuracy of model constructed by the method of ANN turned out to be 85.38 ± 6.15 (%), which was superior to those by the other method of machine learning.

Conclusion: In the present study, we demonstrated that the prediction of function recovery after supratentorial ischemic stroke can be performed with a high degree of accuracy by the various methodology of machine learning, with the highest in ANN.