Application of Deep Learning Algorithm to Monitor Upper Extremity Task Practice

Abstract

Upper extremity hemiplegia is a serious problem affecting the lives of many people post-stroke. Motor recovery requires high repetitions and quality of task-specific practice. Sufficient practice cannot be completed during therapy sessions, requiring patients to perform additional task practices at home on their own. Adherence to and quality of these home task practices are often limited, which is likely a factor reducing rehabilitation effectiveness post-stroke. However, home adherence is typically measured by self-reports that are known to be inconsistent with objective measurement. The objective of this study was to develop algorithms to enable the objective identification of task type and quality. Twenty neurotypical participants wore an IMU sensor on the wrist and performed four representative tasks in prescribed fashions that mimicked correct, compensatory, and incomplete movement qualities typically seen in stroke survivors. LSTM classifiers were trained to identify the task being performed and its movement quality. Our models achieved an accuracy of 90.8% for task identification and 84.9%, 81.1%, 58.4%, and 73.2% for movement quality classification for the four tasks for unseen participants. The results warrant further investigation to determine the classification performance for stroke survivors and if quantity and quality feedback from objective monitoring facilitates effective task practice at home, thereby improving motor recovery.

Keywords: accelerometer; deep learning; inertial measurement unit (IMU); machine learning; rehabilitation; stroke; upper extremity; wearable sensor.

Figure 1

Task identification confusion matrices for results on the validation set (A) and the testing set from the leave-one-person-out cross-validation (B).

Figure 2

Condition classification confusion matrix on four tasks.

Figure 2

Condition classification confusion matrix on four tasks.

Figure 3

Condition classification results combined for each of the three movement qualities. Results of each task are shown from the top to bottom. Results for the validation set are on the left column. Results for the testing set are on the right column.

Figure 4

Quality classification confusion matrix on four tasks which are directly trained using quality labels. Results of each task are shown from the top to bottom. Results for the validation set are on the left column. Results for the testing set are on the right column.