Application of MM-Wave Radar and Machine Learning for Post-Stroke Upper Extremity Motor Assessment

This research effort summarizes the results of a pilot study in which 14 healthy participants performed a subset of the Upper Extremity Fugl-Meyer Assessment protocol (FMA-UE) used for post-stroke recovery evaluation. Radar measurements were acquired using a commercially available Texas Instruments millimeter-wave radar sensor and Time-Frequency analysis employed to produce spectrogram images. The data set comprised of spectrogram images of the tasks. Feature vectors were obtained using 2D Principal Component Analysis (2DPCA) and Support Vector Machines (SVM) and k Nearest Neighbors (kNN) algorithm utilized for task classification. The task classification results were in excess of 99% for Baseline, Intermediate and Severe tasks using k-fold cross-validation and dividing data sets into 70% training and 30% testing. For score prediction, the Hankel transform was applied to spectrogram images to form feature vectors while using SVM and kNN for score prediction. Average total baseline score using SVM was 15.4 ± 1.74 points while an expected score was 18 points, average total intermediate score was 8.8 ± 1.6 points while an expected score was 9 points and average total severe score was 4.6 ± 2.15 points while an expected score was 0 points. The average total baseline score using kNN was 13.2 ± 1.47 points while an expected score was 18 points, the average total intermediate score was 9.4 ± 2.65 points while an expected score was 18 points, and the average total severe score was 6.6 ± 1.85 points while an expected score was 18 points.