JMIR Publications

ABSTRACT

Background:

Subtle motor signs are the clues of serious neurological diseases, however, it is hard to detect and objectively assess those symptoms for non-specialists, while those neurological deficits require fast initiation of treatment allowed in the restricted time. In the clinical environment, diagnosis and the decision are based on clinical grading methods, including Medical Research Council (MRC), which have been used to measure motor weakness. Clinical scores determined by sensor measurement and machine learning techniques can assist instant diagnosis of stroke and rapid initiation of treatment through objective evaluation of proximal functions in the various environment without medical staffs. The system should also provide a solution to complement the data set of acute disease in training machine learning model and accurately grade MRC scores of four limbs.

Objective:

In this study, we aimed to develop an autonomous grading system for stroke patients. We investigated the feasibility of the new system to assess motor weakness similar to the clinical examination performed by medical staff.

Methods:

We implemented AutoMRC grading system composed of a measuring unit with wearable sensors on patients and a grading unit with multi-classification of optimized ensemble learning. Inertial sensors were attached to measure subtle weakness caused by unintended drift of upper and lower limbs of stroke patients. To construct an AutoMRC model estimating clinical scores in the clinic, we performed neurological examinations of 15 patients with MRC 7, 8, and 9 grades. 12 kinematic features of motor disorders and 2 demographic features were extracted and training data was generated using synthetic minority oversampling technique (SMOTE) to complement the low number of subjects and imbalanced classes. The proper ensemble learning algorithm and hyper-parameters were searched by Bayes optimization. We customized MRC grading to identical limb; upper-left, upper-right, lower-left, and lower-right MRC grading. The performance of prediction was evaluated in accuracy, sensitivity, specificity, and area under the receiver operating characteristics curve (AUC).

Results:

The feasibility of AutoMRC was demonstrated in predicting MRC grades with the accuracy of 86.67%; 86.7% for upper-left and upper-right MRC grading; 93.3% for lower-left MRC grading; and 80% for lower-right MRC grading. The average sensitivity was 0.858 and specificity was 0.926. The AUC of predicting MRC of 3 grades ranged from 0.919 to 1.

Conclusions:

The proposed system quantifies proximal weakness in real-time and assesses symptoms through automatic grading. The pilot outcomes demonstrated the feasibility of remote monitoring of motor weakness caused by stroke. The system can facilitate rapid diagnosis and initiation of treatment by sharing MRC grades between patients, caregivers, paramedics, and medical staff as an objective observation method.