Automatic Handwriting Recognition with a Minimal EMG Electrodes Setup: A Preliminary Investigation

Electromyographic (EMG)-based human-machine interfaces showed great potential in hand gesture recognition and recently they have been developed for realizing automatic recognition of handwritten characters or digits. Although smart sensors were commercialized for the forearm, the optimal electrode configuration for high accuracy with minimal channels remains debated. A total of six healthy subjects were asked to write a set of thirty common words of the english vocabulary while EMG signals from forearm and wrist were acquired to perform handwriting recognition using five state of the art feature sets and three pattern recognition models, i.e., K-Nearest Neighbors (KNN), Linear Discriminant Analysis (LDA), Support Vector Machine (SVM). Among the three classifiers the KNN outperformed both LDA and SVM showing a mean accuracy of 71.4% and 96.5 % respectively without and with majority voting post-processor when using combined forearm and wrist EMG information. On the other hand, using only wrist electrodes showed significantly performance drop in all classifiers with accuracy lower than 30.0%. Hence, combining forearm and wrist EMG data is crucial for accurate handwriting recognition with sparse electrode configurations, and limiting the sensing area to the wrist alone may not be sufficient for complex myoelectric decoding tasks. Further research is needed to explore alternative feature sets to improve performance with limited electrode setups.