Impact of sampling frequency and signal quantization on myoelectric-based hand gesture recognition

The rapid advancement of wearable technologies has facilitated the acquisition of myoelectric signals, which are increasingly used as input for machine learning (ML) architectures to recognize human motion. However, the technical specifications of sensors and the experimental setup can significantly affect signal quality, potentially reducing the reliability of motor command recognition. This study investigates how signal quantization (ADC resolution) and sampling frequency influence the performance of myoelectric hand gesture recognition. Surface EMG was recorded with an armband during 20 gestures performed by 10 healthy subjects. Three acquisition settings were tested: 8-bit/500 Hz, 8-bit/1000 Hz, and 12-bit/500 Hz. A time-domain feature set was extracted and used to train three classifiers: linear discriminant analysis (LDA), linear support vector machine (SVM), and quadratic SVM (SVMQ). Results show that higher sampling frequency consistently improved classification accuracy, both with the full armband configuration and with a reduced sensor setup (4 channels). The linear SVM trained with the complete feature set achieved the best performance, with accuracy up to 90% using all sensors and around 80% with the minimal configuration. Even when trained with a single feature, such as mean absolute value or waveform length, the full configuration yielded accuracy above 80% across conditions. In contrast, ADC resolution had only a marginal impact on performance. Overall, the findings indicate that appropriate feature selection and sensor configuration can mitigate the effects of lower sampling rates, offering practical trade-offs between recognition accuracy and computational efficiency in wearable EMG-based systems.