Meta-Learning for Fast Adaptation in Intent Inferral on a Robotic Hand Orthosis for Stroke

Abstract

We propose MetaEMG, a meta-learning approach for fast adaptation in intent inferral on a robotic hand orthosis for stroke. One key challenge in machine learning for assistive and rehabilitative robotics with disabled-bodied subjects is the difficulty of collecting labeled training data. Muscle tone and spasticity often vary significantly among stroke subjects, and hand function can even change across different use sessions of the device for the same subject. We investigate the use of meta-learning to mitigate the burden of data collection needed to adapt high-capacity neural networks to a new session or subject. Our experiments on real clinical data collected from five stroke subjects show that MetaEMG can improve the intent inferral accuracy with a small session- or subject-specific dataset and very few fine-tuning epochs. To the best of our knowledge, we are the first to formulate intent inferral on stroke subjects as a meta-learning problem and demonstrate fast adaptation to a new session or subject for controlling a robotic hand orthosis with EMG signals.

Fig. 1.. MetaEMG for fast adaptation on three subjects with different EMG patterns.

Our method trains EMG-based intent inferral models that can quickly and efficiently adapt to new subjects in the context of a wearable robotic orthosis. Shown are EMG signal recordings of three different stroke survivors, and each colored series in the plots represents the reading of one of the eight EMG electrodes. Training models on new stroke subjects is difficult due to the large subject-to-subject variation in EMG signaling.

Fig. 2.. Task visualization in EMG intent inferral.

We define an EMG task as a single uninterrupted recording with the 8-channel EMG armband. During each recording, users close and open their hands three times. The first open-relax-close motion (outlined in blue) is the support set. The third and second open-relax-close motions (outlined in red) consist of the query set. The ground-truth intent (verbal cues) is shaded in blue, green, and pink for relax, open, and close, respectively.

Fig. 3.. MetaEMG for EMG intent inferral.

Our method uses previously collected data from our orthosis to meta-learn models that adapt to new sessions or subjects more quickly and with less data.

Fig. 4.. Classification accuracy with fewer session-specific data.

In the session adaptation experiments, we further reduce the amount of session-specific fine-tuning data to 0.75, 0.5, and 0.25 of the original amount.

Fig. 5.. Classification accuracy with different number of pretraining subjects.

Models are pretrained on 1, 2, 3, and 4 subjects before being fine-tuned on a subject not seen in pretraining. For each number of pretraining subjects, we run experiments with all possible partitions of the meta-training and meta-testing subjects and report the average classification accuracy.