An action recognition method for manual acupuncture techniques using a tactile array finger cot

The current measurement systems for the physical parameters (rotation frequency, and amplitude) of Traditional Chinese Medicine (TCM) manual acupuncture tend to cause disturbance and inconvenience in clinical application and do not accurately capture the tactile signals from the physician's finger during manual acupuncture operations. In addition, the literature rarely discusses classification of the four basic manual acupuncture techniques (reinforcing by twirling and rotating (RFTR), reducing by twirling and rotating (RDTR), reinforcing by lifting and thrusting (RFLT), and reducing by lifting and thrusting (RDLT)). To address this problem, we developed a multi-PVDF film-based tactile array finger cot to collect piezoelectric signals from the acupuncturist's finger-needle contact during manual acupuncture operations. In order to recognize the four typical TCM manual acupuncture techniques, we developed a method to capture piezoelectric signals in related "windows" and subsequently extract features to model acupuncture techniques. Next, we created an ensemble learning-based action classifier for manual acupuncture technique recognition. Finally, the proposed classifier was employed to recognize the four types of manual acupuncture techniques performed by 15 TCM physicians based on the piezoelectric signals collected using the tactile array finger cot. Among all the approaches, our proposed feature-based CatBoost ensemble learning model achieved the highest validation accuracy of 99.63% and the highest test accuracy of 92.45%. Moreover, we provide the efficiency and limitations of using this action recognition method.