Measurement and recognition of traditional Chinese medicine massage therapy maneuvers

ObjectiveTraditional Chinese Medicine (TCM) massage produces remarkable therapeutic outcomes, yet its standardization is hindered by operator-dependent force patterns and ill-defined contact anatomy. An identical maneuver executed with the thumb tip, finger pad or knuckle evokes markedly different biomechanical effects on acupoints, but current multi-axis force sensors cannot locate the contacting anatomical segment. A portable technology that simultaneously acquires six-axis force/torque and the contacting body part in real time is therefore essential for objective teaching, accreditation and robotic replication of TCM massage.MethodsAn optical–tactile sensor was fabricated (Fig.1). The core element is a 96 mm × 96 mm × 10 mm translucent elastic gel slab embossed with a high-contrast black-dot array (Fig.2). Internal red-green-blue (RGB) LEDs illuminate the markers, while a 640 pixel× 480 pixel CMOS camera captures subsurface deformation at 60 Hz. A transparent acrylic base and a 3-D-printed black-resin housing ensure thermal stability from −10 ℃ to 80 ℃ (Tab.1). Thirty-two thousand tactile images were synchronously recorded with ground-truth six-axis force/torque data obtained from an ATI Nano17 load cell. A modified VGG16 network was trained to map the marker displacement field to force and torque vectors, whereas a YOLOv8s classifier was optimized to distinguish five anatomical segments: Thumb tip, fingertip, finger pad, lateral pad and finger joint (Fig.7). For maneuver recognition, temporal and spectral features—mean, extrema, standard deviation, skewness, kurtosis, dominant frequency, spectral entropy, composite magnitude and correlation—were extracted from 50 trials of each of seven fundamental TCM techniques: pressing, rolling, rubbing, kneading, pushing, striking and vibrating (Fig.4 and Fig.5). A Random-Forest ensemble was subsequently employed to identify the maneuver type.Results and DiscussionsThe six-axis force-estimation network converged after 100 epochs, achieving a mean absolute error of 0.27 N for force and 4.3 N·mm for torque, with coefficients of determination R2 ≥ 0.87 for all axes. Over the operational range of 0-12 N, the uncertainties in force and torque remained below 0.3 N and 5 N·mm, respectively (Fig.6). The YOLOv8s anatomical-site classifier attained an overall accuracy of 97% on 2500 test images; the only confusion exceeding 4% was between thumb tip and fingertip (Tab.2). Random-Forest maneuver recognition reached 97.3% accuracy on 350 unseen sequences (Tab.4). Feature-importance analysis revealed that the standard deviation of z-axis torque, the mean composite torque, and the standard deviation of y-axis torque were the three most discriminative variables for maneuver identification. Comparative experiments further demonstrated that, while delivering equivalent force precision, the proposed system additionally provides anatomical localization—an ability unavailable in existing commercial devices (Tab.3).ConclusionsA compact optical—tactile framework that integrates a 96 mm × 96 mm elastomeric sensor, a deep regression–classification network and a Random-Forest maneuver identifier was developed to deliver synchronized six-axis force/torque data, anatomical contact-site labels and maneuver identities during TCM massage. Within the 0-12 N range, force and torque uncertainties remained below 0.27 N and 4.3 N·mm, respectively, while contact-site and maneuver recognition accuracies exceeded 97%. By eliminating the long-standing limitation of conventional multi-axis sensors—the inability to localize the anatomical segment delivering the stimulus—the low-cost, portable system offers an objective and scalable tool for digital teaching, remote assessment and robotic replication of TCM massage, thereby accelerating the standardization and clinical translation of manual therapeutic techniques.