Decoding of fine hand motor control for kinematic acupuncture needling manipulation

Summary

Manual acupuncture depends on precise thumb-index finger coordination, yet the biomechanical principles governing this dexterity remain unresolved. Here, we establish a non-contact, video-based platform integrating marker-labeled motion capture with nonlinear dynamic and multi-modal coupling analyses to decode fine finger coordination across standardized training, animal models, and clinical observation. Tracking eight anatomical joints reveals an asymmetric motor architecture: the distal index finger acts as a temporal pacemaker with high rhythmic stability, whereas the distal thumb functions as an amplitude modulator. Multi-modal coupling resolves two biomechanical modules—an index finger chain with hierarchical phase propagation and a thumb chain with strong intra-module synchronization—linked by millisecond-scale temporal offsets. These patterns are encoded as operator-specific “Acupuncture Kinematic Fingerprints,” stratifying practitioners into index-dominant and thumb-dominant strategies. This framework transforms subjective manipulation styles into quantitative kinematic signatures for objective skill assessment, standardized training, and mechanistic studies.

Graphical abstract

Highlights

• •Non-contact video platform quantifies fine finger coordination in acupuncture manipulation
• •The index finger acts as a temporal pacemaker; the thumb acts as an amplitude modulator in manipulation
• •Index finger-dominant and thumb-dominant strategies capture individual motor habits
• •AKF profiles serve as acupuncture kinematic fingerprints for operator identification

Biomedical engineering; Biomechanics; Biocomputational method; Systems biology

Introduction

The human hand represents one of the most sophisticated biomechanical systems ever evolved in nature, capable of executing fine movements with extraordinary precision and adaptability.¹ Such precise control underlies high-level manual expertise, from microsurgery manipulation to musical performance, where subtle movements must be strictly coordinated.^2,3,4 Within this system, the thumb-index finger module further forms the foundation of dynamic grip, enabling stable and precise operation in controlling most hand activities with force, trajectory, and timing.^5,6

Quantitative characterization of such fine motion, particularly within the thumb-index finger module, is essential for elucidating the relative core positions within the module that governs the manual performance.^7,8 However, capturing the fine movement remains challenge, largely due to the lack of scalable, highly resolved tools capable of deciphering the kinematic processes of hand motion.

Among the many hand movements that demand fine motor control, manual acupuncture may appear simple, yet it carries significant clinical stakes. Acupuncture is one of the wildly adapted external therapeutic approach worldwide, which involves the fine manipulation of needles inserted at specific acupoints to elicit therapeutic effects.⁹ Acupuncture techniques represent a paradigmatic example of such high-level motor skill: Therapeutic efficacy depends critically on the practitioner’s ability to manipulate needles with precise frequency, amplitude and rhythm at specific acupoints.¹⁰ Among classical techniques, twirling-rotating (nian zhuan fa) and lifting-thrusting (ti cha fa), are not isolated manipulations but coordinated motor synergies of finger joints, involving distinct synchronization, coupling and directional patterns.¹¹ Notably, lifting-thrusting manipulation primarily involves vertical movements of the fingers and wrist, whereas twirling-rotating manipulation critically depends on coordinated motion between the thumb and index finger.¹² Accumulating clinical evidence has indicates that acupuncture efficiency highly depends on manipulation parameters, including twirling and lifting-thrusting amplitude and frequency.

The relative motions between the thumb and index finger control the amplitude, velocity, direction, rhythm, and force of needle movement,^13,14 factors considered not merely stylistic differences but key determinants of clinical outcome.^12,15,16 For example, lifting-thrusting frequency varying within the range of 1–3 Hz has been associated with distinct therapeutic effects in both animal models and clinical studies.^17,18 Despite this parameter-dependent efficacy, quantification of manipulation parameters remains insufficient, resulting the short of evaluation standards and inconsistent clinical efficacy.¹⁹

Recent advances in sensing technologies and computer vision have created new opportunities to resolve fine motor behavior. High-precision force sensors and miniaturized piezoelectric devices enable measurement of forces generated by needle-tissue interactions and have helped define core operational parameters of different settings of twirling-rotating manipulation, including twirling frequencies of approximately 40–160 cycles/min (or equally, 0.67–2.67 Hz) and rotational amplitude ranging from <90° (reinforcing) to <180°–360° (reducing).^20,21,22,23 PVDF-based finger-stall sensors enable real-time extraction of these parameters,^15,24,25 propelling acupuncture manipulation from empirical format toward quantifiable, reproducible scientific analysis. However, these approaches require instrumenting the needle or the practitioner’s hand, potentially altering natural motion dynamics. On the other hand, breakthrough in computer vision enable the reconstruction of joint trajectories and extraction of multidimensional kinematic features, including angular displacement, velocity, acceleration, and movement complexity.^11,26,27 Yet, most vision-based studies suffer from limited temporal stability and reduced accuracy in capturing fine-scale finger movements, particularly for the millimeter-scale joint displacements and rapid angular transitions characteristic of skilled twirling sequences.

Here, we develop a low-cost, video-based kinematic platform with nonlinear dynamic analysis to quantitatively decode the spatiotemporal architecture of acupuncture needle manipulation. By capturing coordinated thumb-index finger motion and applying nonlinear dynamic analyses, we characterize the spatiotemporal signatures of five twirling-rotating techniques (90° at 2 Hz, 180° at 0.5 Hz, 180° at 1 Hz, 180° at 2 Hz, and 360° at 2 Hz), identify key joints governing frequency and amplitude control, and reveal modular coordination patterns underlying skilled performance. Integrated kinematic profiles further differentiate techniques and stratify operators by motor consistency, establishing an objective basis for skill evaluation and standardization.

Results

Hand kinematic characterization of twisting-rotating manipulation

To quantitatively characterize the biomechanical features of manipulation, we focused on twirling-rotating manipulation (nian zhuan fa) as a representative technique example, while complementary analyses of lifting-thrusting (ti cha fa) are provided in the Supplementary Materials Figure S2B.

We established an image-based kinematic analysis framework to capture fine-scale hand joint dynamics during needle twirling. Eight anatomically landmarks (ID1−ID8) on the thumb and index finger with high-contrast retroreflective markers and recorded hand motion at 4K/60 fps (Figures 1A–1C). Using a representative twisting-rotating manipulation with a rotation amplitude of 90° and a twirling frequency of 2 Hz, all markers exhibited highly periodic trajectories consistent with cyclic needle rotation (Figures 1D and 1E). In brief, frame-to-frame displacement vectors were computed in both X and Y dimensions, allowing the reconstruction of instantaneous translational dynamics (Figure 1F).

Figure 1.

Image-based workflow for extracting hand kinematics during twisting-rotating needle manipulation

(A) Schematic of the video acquisition and image-processing pipeline for quantitative analysis of hand motion.

(B) Spatial layout of eight high-contrast circular markers positioned on anatomically relevant joints of the thumb and index finger.

(C) Representative time-lapse illustrates automated detection of marker and centroid extraction across sequential frames.

(D) Reconstructed centroid trajectories of all markers (ID1−ID8) over a 20-s recording period.

(E) Enlarged trajectory of marker ID1 during cyclic needle rotation; the color gradient (blue to red) denotes temporal progression from 0 to 20 s.

(F) Definition of frame-to-frame displacement, calculated as the Euclidean distance between centroid positions in adjacent frames, and the resulting displacement series derived from consecutive images.

(G) Sample entropy (SampEn) and multiscale entropy (MSE) analyses determining the optimal temporal lag for kinematic modeling.

(H) Representative kinematic outputs for ID1, including temporal displacement, mean square displacement (MSD), angular velocity, and acceleration; a 12-s segment is shown to enhance visualization of cyclic dynamics.

To determine an optimal observation window (time lag) for subsequent kinematic analyses, sample entropy (SampEn) and multiscale entropy (MSE) were applied to the displacement time series to evaluate signal complexity across increasing time lags (Figure 1G). Based on entropy stabilization, a time lag of 2 frames was selected. Under this condition, multiple kinematic parameters including mean square displacement (MSD), translational velocity, angular displacement, angular velocity, and acceleration were extracted for all markers, revealing highly regular motion patterns and distinct dynamic characteristics of the twisting-rotating manipulation (Figure 1H). Collectively, these parameters revealed a highly ordered dynamic regime marked by periodic acceleration-deceleration cycles synchronized with rotational phases.

Nonlinear dynamic signatures identify the index finger as the primary driver

To further resolve joint-specific functional roles, we performed nonlinear dynamic analyses across all markers under a standardized manipulation paradigm (90°, 2 Hz). MSD was used to describe translational behavior, angular velocity to characterize rotational behavior, and acceleration to reflect motion smoothness and dynamic features (Figure S4). Heatmaps showed that motion amplitude was concentrated in the distal index finger (ID1−ID3) and thumb (ID5−ID6), with ID1 exhibiting the largest displacement (Figure 2A), suggesting a dominant role in torque generation. The entropy analysis of each sequence (Figure 2B) further revealed distinct functional specialization: The index finger joints (ID1−ID3) exhibited lower sample entropy and approximate entropy in angular velocity, suggesting that their rotational patterns have higher temporal regularity; while the thumb joints (ID5−ID6) showed lower entropy in acceleration, suggesting simpler and more predictable force patterns.

Figure 2.

Nonlinear dynamic signatures of hand joint motion during twisting-rotating manipulation

(A) Heatmap of motion amplitude derived from MSD values for each marker (ID1−ID8); higher color saturation denotes greater displacement magnitude.

(B) Entropy metrics of MSD, angular velocity, and acceleration sequences, demonstrating enhanced temporal regularity in index finger rotation and predictable force modulation in the thumb.

(C) Recurrence plots of MSD trajectories illustrate conserved phase-space structures among index finger joints.

(D) Comparison of recurrence quantification analysis (RQA) metrics, including determinism (DET), laminarity (LAM), mean diagonal length (Lmean), and trapping time (TT), highlighting stronger deterministic dynamics in index finger angular velocity.

Recurrence plots of MSD trajectories further illustrated strongly conserved dynamic structures among ID1−ID3 (Figure 2C), consistent with shared attractor-like behavior in phase space. Recurrence quantification analysis (RQA) of angular velocity provided convergent evidence: Index finger joints (ID1 and ID3) exhibited elevated determinism (DET) and higher laminarity (LAM) (especially ID1 (Figure 2D). These features collectively indicate prolonged predictable states, the index finger, especially ID1 as the principal stabilizer of rotational rhythm, while the thumb appears specialized for adaptive force shaping.

Modular and directional coordination defines thumb-index finger coupling architecture

To decode inter-joint coordination during twirling-rotating manipulation, we quantified functional connectivity using five complementary metrics (Figure 3). Phase-locking value (PLV) and phase-phase coupling (PPC) quantify phase synchronization between signals, with higher values indicating stronger and more stable coupling. Coherence measures frequency-specific linear correlation, where higher values indicate stronger coupling within a given frequency band. Granger causality assesses directed predictive influence between signals, indicating directional information flow. Phase lag (τ) quantifies temporal lead-lag relationships, where positive values indicate that one joint leads another in time.

Figure 3.

Multi-modal coupling reveals modular and directionally asymmetric thumb-index finger coordination

(A) Phase-locking value (PLV) matrix shows phase synchronization between joints. Warmer colors indicate stronger phase locking, reflecting more stable temporal coupling.

(B) Phase-phase coupling (PPC) matrix reveals selective connectivity patterns.

(C) Network representation derived from thresholded PPC values (PPC >0.10), resolving two distinct kinematic chains: the index finger chain (ID1→ID2→ID3→ID4) and the thumb chain (ID5→ID6).

(D) Frequency-domain coherence matrix shows frequency-specific coupling between joints. Warmer colors indicate stronger coherence within the same frequency band.

(E) Phase lag (τ) matrix illustrates temporal lead-lag relationships between joints. Positive values (red) indicate that the source joint leads the target joint, whereas negative values (blue) indicate temporal lag.

(F) Granger causality analysis shows directed functional connectivity between joints. Warmer colors indicate stronger predictive influence.

Phase synchronization analysis revealed robust coupling within two distinct biomechanical modules. The PLV matrix shows the strongest intra-module coupling within the thumb (ID5−ID6, PLV = 0.572) and along the index finger chain (ID1−ID4), with comparatively weaker cross-module interactions (Figure 3A). The PPC matrix exhibits a sparser but consistent pattern, again highlighting dominant coupling within the thumb module (ID5−ID6, PPC = 0.336) while most inter-joint connections remain below 0.20 (Figure 3B). Network reconstruction based on PPC further resolved two distinct kinematic chains: an index finger chain (ID1→ID2→ID3→ID4) and a thumb chain (ID5→ID6), supporting a modular biomechanical architecture (Figure 3C). Consistent with these findings, the coherence matrix displays a block-diagonal structure with strong intra-module coupling and weak inter-module interactions (Figure 3D), further reinforcing the presence of two semi-independent but functionally coordinated modules.

Temporal asymmetry analysis revealed directionality within this modular system. Phase lag measurements show near-zero delays within modules (e.g., ID1−ID2, ID2−ID3-ID4, and ID5−ID6), confirming tightly coupled intra-module coordination (Figure 3E). In contrast, a consistent inter-module delay was observed, with the distal index finger (ID1) leading the thumb (ID5) by approximately 10 ms. Furthermore, the thumb precedes proximal index joints (ID3−ID4) by ∼5 ms, indicating a sequential propagation of motion from distal index initiation to thumb engagement and subsequent proximal adjustment. This observation establishes a temporally ordered coordination pattern underlying rhythmic needle rotation.

Granger causality analysis further demonstrated asymmetric predictive interactions among joints (Figure 3F). Within the thumb module, a strong unidirectional influence from ID6 to ID5 was observed, whereas the reverse direction is negligible, indicating proximal-to-distal information flow. Similarly, directional coupling from the distal index finger to the thumb (ID1→ID5) is obvious, while the reverse influence is minimal. Additional interactions involving proximal joints (e.g., ID7→ID2, ID8→ID5) suggest integrative coordination at the module interface.

Collectively, these results demonstrate that twirling-rotating manipulation is governed by a modular and directionally asymmetric coordination system, in which the index finger acts as the primary temporal driver, while the thumb serves as a downstream force modulator to stabilize rhythmic needle motion.

Robustness and precision of the kinematic framework enable extraction of stable manipulation fingerprints

To validate that the video-based framework captures reproducible, operator-specific motor signatures rather than noise-contaminated trajectories, we conducted a multi-level robustness validation. We systematically evaluated its performance across varying acquisition conditions, benchmarking strategies, and operators, hereafter referred operator-specific kinetic signatures as manipulation fingerprints.

First, recordings from three distinct viewing angles demonstrated that angular velocity-derived nonlinear dynamic features, including entropy and DET, were highly consistent across perspectives (Figures 4A and 4B). This invariance indicates that the tracking pipeline is resilient to geometric perturbations and partial occlusion. Among all joints, distal of the index finger (ID1) consistently exhibited the lowest entropy and highest determinism (Figure 4B), confirming its role as the most temporally stable and rhythmically constrained component of the manipulation.

Figure 4.

Functional differentiation and synergistic organization of index finger and thumb joints

(A) Multi-view motion capture configuration enables accurate reconstruction of three-dimensional joint trajectories.

(B) Entropy and DET analyses of angular velocity identifying the proximal index finger joint (ID1) as the most temporally stable and predictable element.

(C) Pearson’s correlation analysis based on integrated kinematic features, revealing a tightly coupled index finger module (ID1−ID3) and a thumb module (ID5−ID6) with an anti-phase relationship.

(D) Spectral characterization of joint motion using MSD-derived trajectories and Fast Fourier Transform, indicating a dominant manipulation frequency near 2 Hz and stable amplitude control at ID1.

(E) Schematic of the cross-validation framework compares three measurement modalities: MediaPipe hand tracking, our marker based video tracking framework, and a flexible resistive pressure sensor, against a common metronome-cued nominal frequency (“See also Figure S3”).

(F) Dominant frequency estimates across the three methods, demonstrating close agreement between our platform and the pressure sensor, whereas OpenCV shows substantially lower accuracy and waveform stability.

(G) MSD fingerprint plots of ID5 (distal thumb) from 10 clinical acupuncture experts performing standardized 90° at 2 Hz twirling, comparing OpenCV-based markerless tracking (left) and video-based marker tracking technology (right).

Inter-joint correlation analysis further verified a modular, functionally segregated architecture comprising three clusters: an index finger module (ID1−ID4), a thumb module (ID5−ID6), and a proximal stabilization module (ID7−ID8) (Figure 4C). Speed correlations demonstrated strong coupling between index finger (ID1−ID4) and thumb distal joints (ID5−ID6), whereas ID7 and ID8 showed weaker association with active rotational motion, consistent with a stabilizing role (Figure 4C, left panel). Notably, motion-angle correlation revealed a clear anti-phase relationship between the index finger and thumb clusters (Figure 4C, middle panel). This anti-phase coordination indicates that the two digits act as antagonistic yet cooperative levers underlying controlled needle rotation: The index finger generates primary rotational torque while the thumb provides counter-stabilization and return-stroke force.

Frequency-domain analysis using fast Fourier transform (FFT) demonstrated a dominant spectral peak near 2 Hz across all joints under standardized manipulation (90°, 2 Hz), consistent with the imposed rhythm (Figure 4D, left panel). Among all markers, ID1, ID2, ID3, and ID5 exhibited the lowest coefficients of variation in dominant frequency 2 Hz, indicating stable temporal regulation. Importantly, ID1 showed the largest spectral amplitude at the dominant frequency, reinforcing its role as the principal temporal driver of the movement (Figure 4D, right panel).

To further evaluate measurement accuracy and temporal fidelity, we benchmarked the marker-based pipeline against two alternative measurement modalities under a metronome-set reference (90°, 2 Hz): markerless deep-learning hand tracking (MediaPipe Hands v0.8) and a flexible piezoelectric pressure sensor (Figures 4E, 4F, and S3). All three approaches accurately recover the dominant frequency. However, markerless MediaPipe tracking exhibits frame dropouts and high-frequency fluctuations, resulting in unstable trajectories and increased signal variance (Figure 4F, left panel). The pressure sensor captures the dominant frequency with high fidelity but lacks kinematic granularity to resolve angular changes and multi-joint coordination (Figure 4F, right panel).

In contrast, the proposed marker-based tracking method produces smooth, periodic trajectories with clearly resolved patterns. Analysis of MSD-derived manipulation fingerprints across 10 clinical experts further highlights this distinction (Figure 4G). Markerless tracking by MediaPipe Hands generates noisy waveforms with poorly defined periodicity under rapid fine-scale finger motion, whereas our proposed method yields more highly stable, repeatable patterns with distinct peaks corresponding to individual manipulation cycles. These stable waveform patterns enable accurate extraction of amplitude envelopes and temporal features, confirming that the proposed framework uniquely captures high-resolution motor dynamics.

Collectively, these results demonstrate that the proposed approach achieves both high measurement accuracy and robustness, while enabling the extraction of stable, operator-specific manipulation fingerprints that encode fine motor control strategies (Figure 4G). Additional robustness validation under perturbation conditions, including bar hand, lifting-thrusting manipulation, is provided in the Supplementary Materials Figure S2A.

Technique-specific and operator-dependent motor fingerprints revealed by ID1 kinematics

Since nonlinear dynamic analysis identified the distal index finger (ID1) as the most temporally stable and deterministic component of the thumb-index finger system (Figure 4), we next examined whether ID1 dynamics alone were sufficient to differentiate manipulation techniques across operators.

To address this, five canonical twirling-rotating techniques spanning clinically relevant ranges of frequency (0.5–2 Hz) and rotational amplitude (90°–360°) were analyzed across eight operators: 90° at 2 Hz, 180° at 0.5 Hz, 180° at 1 Hz, 180° at 2 Hz, and 360° at 2 Hz. These paradigms encompass typical reinforcing (low frequency, small amplitude), reducing (high frequency, large amplitude), and intermediate stimulation patterns.^21,28

Nonlinear dynamic analysis revealed distinct technique-dependent kinematic signatures. Sample entropy and RQA revealed clear technique-dependent nonlinear signatures (Figure 5A). Although individual operator variability was present, the five manipulation techniques exhibited statistically distinct dynamic profiles. In general, weak stimulation techniques were characterized by higher entropy and lower determinism. In contrast, strong stimulation techniques displayed lower entropy and higher determinism.

Figure 5.

Technique- and operator-dependent kinematic features of the proximal index finger joint (ID1)

(A) Entropy and RQA analyses of angular velocity demonstrate that movement regularity is jointly influenced by operator-specific motor habits and manipulation technique. Each scatter represents an individual operator.

(B) Distribution of dominant frequency and movement amplitude in feature space, with distinct clustering corresponding to classical needle manipulation styles.

(C) Operator-level comparison of parameter dispersion across techniques, where narrower distributions indicate greater motor consistency and precision.

To further characterize rhythmic organization, dominant frequency and movement amplitude were extracted via spectral decomposition. Joint mapping of frequency and amplitude parameters segregated the five manipulation techniques into discrete clusters (Figure 5B), each defined by a characteristic rhythm-force combination. This separation demonstrates that the video-based framework captures quantitative biomechanical signatures corresponding to classical manipulation categories.

Despite observable inter-individual differences in entropy and RQA metrics, technique-dependent parameter modulation remained internally consistent within each operator (Figure 5C). The analysis showed that although operators differed in baseline variability and coordination style, the relative scaling of amplitude and rhythmic structure across weak-to-strong manipulations followed reproducible patterns.

Notably, while dominant frequency was relatively conserved across operators for a given technique, amplitude strategies varied substantially. Some operators (e.g., E–H) exhibited narrow amplitude ranges across stimulation intensities, suggesting limited modulation between weak and strong manipulations. Others demonstrated broader amplitude separation (e.g., approximately 90°, 180°, and 360°), reflecting clearer force differentiation. This heterogeneity indicates that although rhythmic timing is tightly regulated, force-scaling strategies represent a major axis of individual variation.

Collectively, these findings establish the distal index finger as a compact yet informative biomechanical readout capable of discriminating manipulation techniques while simultaneously encoding operator-specific coordination strategies. Manipulation kinematics therefore embody a structured motor “fingerprint” shaped jointly by technique-specific constraints and individual neuromuscular habits.

Integrated operator clustering reveals two distinct coordination strategies

To determine whether individual manipulation styles could be classified at the level of global motor control, we constructed integrated kinematic profiles of all five manipulation techniques for eight acupuncture operators spanning two expertise levels (experts and beginners). For each operator, a 96-dimensional feature vector across five skills was generated by aggregating kinematic and nonlinear dynamic metrics across eight-digit markers and five manipulation techniques. These features included the mean and standard deviation of MSD, velocity, and acceleration; mean direction angle and angular velocity; sample entropy; multiscale entropy; and recurrence rate and determinism derived from RQA.

Following Z score normalization, hierarchical clustering revealed three distinct operator groups (Figure 6A). Experienced practitioners (experts) formed a cohesive cluster (Clusters in Figure 6A), whereas beginners were distributed across separate branches, indicating greater heterogeneity in motor organization. Permutational multivariate analysis of variance confirmed that inter-cluster differences were statistically significant (F = 2.0907, p = 0.035), indicating that the systematic divergence in motor control strategies.

Figure 6.

Integrated clustering of operators and joint-level synergy pattern

(A) Dendrogram and heatmap generated from hierarchical clustering of 96-dimensional feature vectors summarizing kinematic metrics across joints and techniques; annotations indicate cluster membership and PERMANOVA results.

(B) Pearson’s correlation matrix and corresponding distance-based clustering of marker features, revealing strong synergy between thumb joints (ID5−ID6) and biomechanical separation between distal functional elements.

(C) Representative comparison of motion frequency and amplitude for index finger (ID1−ID2) and thumb (ID5−ID6) joints between two operators with divergent global kinematic profiles.

Joint-level similarity analysis further verified the conserved modular architecture across operators (Figure 6B), comprising an index finger module (ID1−ID3), a thumb module (ID5−ID6), and a proximal stabilization module (ID4, ID7, ID8). Notably, moderate coupling between the distal index finger (ID1) and distal thumb (ID5) was consistently observed, suggesting that thumb-index finger coordination forms a fundamental structural constraint of twisting-rotating manipulation.

Despite this shared architecture, clear differences emerged in how motion was distributed across the two modules. Representative comparison of two operators with divergent global profiles, operator A and operator H, revealed distinct coordination strategies (Figure 6C). Operator A primarily generated rotational force through coordinated index finger motion (ID1−ID2), with the thumb functioning as a stabilizing axis. In contrast, Operator H relied predominantly on thumb-driven movement (ID5−ID6), characterized by larger amplitude excursions and more regular reciprocating oscillations.

These differences indicate that although global rhythmic structure is conserved, individual manipulation styles are defined by the relative dominance of index finger versus thumb contributions, establishing the existence of coexisting index finger-dominant and thumb-dominant control modes.

Kinematic fingerprints resolve index- and thumb-dominant manipulation strategies across operator

To evaluate whether our proposed framework could resolve fine inter-individual differences under standardized conditions, we analyzed 20 acupuncture practitioners performing two canonical manipulation paradigms (90° at 2 Hz and 180° at 1 Hz). Despite identical metronome pacing, operators segregated into four distinct clusters in the frequency-amplitude space (Figure 7A). The extracted dominant frequencies remained tightly locked to the prescribed rhythm with obvious differences in amplitude.

Figure 7.

Individual motor fingerprints detected by the video-based platform reveal cluster-specific twirling strategies

(A) Frequency-amplitude plot of 20 acupuncture practitioners (10 clinical experts) performing two standardized techniques (90° at 2 Hz, circles; 180° at 1 Hz, triangles). Despite identical metronome pacing, operators segregated into four distinct clusters (colored ellipses) based on ID5 (distal thumb) MSD amplitude.

(B) 3D fingerprint plots of ID5 (distal thumb) MSD grouped by the four clusters from (A) for 90°/2 Hz (B) and 180°/1 Hz (C). Thumb-dominant clusters (1–2) show larger amplitude excursions; index finger-dominant clusters (3–4) exhibit constrained amplitude profiles.

To visualize these cluster-specific amplitude patterns, we reconstructed kinematic fingerprint maps based on MSD trajectories of the distal thumb joint (ID5). MSD fingerprints revealed distinct cluster-specific amplitude signatures that were highly stable within individuals across repeated trials (Figure 7B). Thumb-dominant operators (Clusters 1–2) exhibited larger displacement and high-inertia trajectories, whereas index finger-dominant operators (Clusters 3–4) showed more constrained and precision-grip displacement profiles. Importantly, these signatures remained highly consistent within individuals across repeated trials, demonstrating both intra-operator stability and inter-operator separability.

These results extend the findings from Figure 6 by showing that, while temporal structure is conserved across operators, force allocation strategies vary systematically, giving rise to distinct operator-specific kinematic signatures. Collectively, our framework achieves sufficient sensitivity to resolve individual motor fingerprints and sufficient stability, transforming subjective “twirling styles” into quantifiable kinematic signatures for objective skill assessment and quality control.

A universal platform for quantifying acupuncture manipulation

To translate the video-based analytical framework into a practical and scalable analytical tool, we implemented an integrated web-based platform that enables automated extraction and analysis of hand-motion dynamics from raw video input (web-production-bda79e.up.railway.app). The system streamlines the entire workflow through three sequential modules: (1) video data processing (one-click upload and automated marker tracking), (2) automated data analysis (selectable kinematic and nonlinear dynamic modules), and (3) export of analysis outputs, including compressed image sequences, spreadsheet-ready centroid coordinates, and directly visualizable quantitative reports (Figure 8A).

Figure 8.

A scalable platform for cross-domain quantification of acupuncture manipulation kinematics

(A) Integrated web-based workflow comprising three modules: video processing (marker tracking), automated kinematic and nonlinear dynamic analysis, and export of standardized quantitative outputs.

(B) Clinical application representative diagram. Non-contact capture and analysis of hand kinematics during in-clinic acupuncture procedures. Representative ID1 and ID5 MSD trajectories demonstrate preserved rhythmic structure under real-world conditions.

(C) Preclinical application representative diagram. Quantification of standardized twirling-rotating manipulation (90° at 2 Hz) in animal models, showing consistent periodicity and conserved inter-joint coordination patterns.

(D) Unified quantitative output space enables direct comparison across clinical and experimental datasets.

We next evaluated the performance and applicability of this platform across two complementary domains: controlled preclinical experiments and real-world clinical settings. In clinical scenarios, the platform enables non-contact acquisition and analysis of practitioner hand motion during routine acupuncture procedures without interfering with treatment delivery (Figure 8B). Representative MSD derived from ID1 and ID5 preserve a clear rhythmic pattern and temporal stability, demonstrating that fine motor dynamics can be reliably captured under unconstrained clinical conditions. In addition, operator evaluation of our marker-based framework indicated high operational feasibility, with scores of 4/5 for naturalness, finger dexterity, and manipulation accuracy (see Figure S1 for detailed assessment of marker interference).

To assess cross-species robustness, we further applied the platform to standardized twirling-rotating manipulation (90° at 2 Hz) in both rat and mouse models (Figure 8C). The extracted MSD of ID1 exhibits stable 2 Hz rhythmic periodicity, while inter-joint phase-locking patterns (PLV) are comparable to those observed in human operators. These results indicate that the underlying biomechanical coordination principles are preserved across species and experimental conditions.

Importantly, both clinical and preclinical applications converged on a unified set of quantitative descriptors (Figure 8D), including displacement/MSD, angular velocity and acceleration, entropy-based complexity metrics (SampEn, MSE), recurrence-based measures (RQA), frequency-amplitude decomposition, and multi-modal coupling indices (PLV, PPC). This standardized parameter space enables direct cross-domain comparison of manipulation kinematics, providing a quantitative bridge between clinical skill assessment and mechanistic studies in experimental models.

Discussion

In this study, we established a non-contact, video-based kinematic framework to quantitatively decode fine thumb-index finger coordination during twirling-rotating manipulation (nian zhuan fa), a core technique in manual acupuncture. By integrating marker-based motion capture with nonlinear dynamic and multi-modal coupling analyses, we decoded complex hand movements into high-resolution, reproducible kinematic representations.

A central contribution of this work is the demonstration that the proposed framework accurately resolves the intrinsic coordination logic of finger motion, rather than merely capturing superficial movement trajectories. Across multiple analytical layers, including entropy, recurrence quantification, and phase-based coupling metrics, we consistently observed a structured and non-random coordination pattern. The index finger module (ID1−ID3) exhibited weaker but hierarchically propagated coupling, reflected by sequential phase alignment and moderate PLV/PPC values. The thumb module (ID5−ID6) displayed strong intra-module synchronization, characterized by high PLV and PPC values, indicating tightly coordinated joint behavior. Phase lag analysis revealed systematic temporal offsets (∼5–10 ms) between joints, with the distal index finger (ID1) leading the thumb module in the index-dominant technique. These findings are not arbitrary numerical patterns, but align closely with known neuromuscular control principles, where motor initiation, force transmission, and stabilization are temporally and functionally distributed across different joints. This capability of our framework to resolve such millisecond-scale coordination structures provides strong evidence for its measurement precision and physiological validity.

Beyond methodological validation, our results reveal a directionally asymmetric motor architecture underlying twirling-rotating manipulation. The framework robustly captures operator-specific yet reproducible motor signatures, here termed kinematic fingerprints. Practitioners exhibit two equally valid, stable coordination strategies: an index-dominant pattern, in which rotational torque is generated primarily through the distal index finger (ID1−ID2) with the thumb acting as a low-amplitude stabilizing axis; and a thumb-dominant pattern, in which force is driven by thumb motion (ID5−ID6) with the index finger providing temporal anchoring. Both strategies were observed in experienced practitioners, indicating that expertise reflects stable, low-entropy control within a preferred coordination regime rather than adherence to a single movement pattern.

Interestingly, these quantitatively defined coordination patterns resonate with classical descriptions of acupuncture manipulation. The Great Compendium of Acupuncture and Moxibustion (Zhen Jiu Da Cheng), one of the classical medical literature, emphasizes differences in rhythm and strength between reinforcing (bu) and reducing (xie) manipulations,²⁹ describes twirling-rotating as a coordinated thumb-index action rather than isolated joint movement, implicitly recognizing that subtle variations in force and rhythm arise from the interplay between these two digits. The Hungry Horse Shaking a Bell (E Ma Yao Ling), a canonical reinforcing (bu) method for deficiency patterns, is traditionally described as a slow, gentle motion in which the thumb advances forcefully but retreats weakly. Conversely, the Phoenix Spreading Wings (Feng Huang Zhan Chi), a canonical reducing (xie) method for excess patterns, is described as a rapid, forceful, unidirectional motion, in which the thumb retreats while the index finger drives the needle forward. More broadly, the classical techniques define supplementation as thumb-forward, heavy-force, slow-frequency rotation, and drainage as index-forward, light-force, fast-frequency rotation. From a translational perspective, our framework might transform these classical technique concepts into measurable kinematic signatures, extending traditional empirical knowledge into an evidence-based motor-learning paradigm.

Compared with existing approaches, including flexible force sensors and deep-learning-based markerless tracking (e.g., MediaPipe Hands), our method provides distinct advantages. Cross-modal validation demonstrates that our platform achieves frequency estimation accuracy comparable to physical sensors, while uniquely resolving joint-level motion features and inter-finger coordination. Conventional computer vision approaches often exhibit limited temporal stability and reduced accuracy in capturing fine-scale movements during continuous hand motion. In contrast, our marker-based strategy enables robust, precise tracking of joint trajectories, particularly under conditions of subtle, high-frequency finger coordination. Our approach preserves natural manipulation behavior and avoids the mechanical interference introduced by instrumented sensors, providing a scalable and ecologically valid solution for high-resolution biomechanical analysis.

Furthermore, this framework provides a quantitative platform applicable across three experimental conditions: standardized training modules, animal experiments, and naturalistic clinical observation. In animal experiments, it enables precise recording of manipulation features, facilitating controlled investigation of parameter-dependent biological responses. In clinical settings, the platform allows non-contact capture of expert performance, offering objective metrics for skill assessment, training standardization, and quality control. By transforming qualitative manipulation descriptions into measurable kinematic signatures, this approach supports the development of reproducible and evidence-based acupuncture practice.

In conclusion, this study establishes a quantitative framework for decoding fine motor coordination in acupuncture manipulation. By integrating computer vision-based motion capture technology with nonlinear dynamic analysis, we provide a computational foundation for decoding fine motor expertise. This approach not only advances objective standardization of acupuncture techniques but also establishes a scalable paradigm for studying coordinated human hand movements across clinical and performance domains.

Limitations of the study

Several limitations should be noted. First, the present work focuses primarily on twirling-rotating manipulation. Although the platform is inherently capable of analyzing lifting-thrusting movements (Figure S2B), accurate characterization of such three-dimensional motion may require multi-view acquisition, such as stereo depth imaging, and warrants further investigation. Second, while the lightweight markers used in this study did not measurably interfere with hand motion or subjective manipulation experience (Figure S1), their use may still introduce minor constraints in clinical settings. Future improvements will explore less obtrusive tracking strategies, such as miniaturized markers or markerless motion capture, to enhance usability and clinical applicability. Third, although preliminary analyses indicate that different manipulation techniques exhibit separable kinematic patterns, automated classification remains to be fully developed. Future work will incorporate machine learning approaches to enable end-to-end recognition of manipulation types and more comprehensive modeling of complex motion features. Beyond acupuncture, the identified kinematic features, such as coordination structure and frequency stability, may also serve as sensitive markers for broader applications in motor function assessment, rehabilitation, and skill evaluation.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact: Tian Lan (lantian@sdutcm.edu.cn).

Materials availability

This study did not generate any new physical materials.

Data and code availability

The datasets generated and analyzed during the current study are available in the article, Supplementary Materials. Data S1: Raw video recordings and kinematic analysis outputs for experimental conditions. The analysis scripts used to generate the figures and results are encoded in a platform (web-production-bda79e.up.railway.app). The original analysis code is also deposited in GitHub: https://github.com/wenbo5867-cyber/minisite. All data and analysis code are available upon reasonable request from the corresponding author.

Acknowledgments

We thank Xiqing Xue, Fei Wang, Wanyi Xu, and Wenying Cui for their assistance with acupuncture needling manipulation data collection, and thank doctor Shuzhong Gao for clinical Assistance. This work was supported by the Shandong Provincial Natural Science Foundation Innovation and Development Joint Funding (grant no. ZR2022LZY023), and the Taishan Scholar Foundation of Shandong Province (grant no. tstp20221125).

Author contributions

Conceptualization and study design: H.S., T.L., W.Z., and Y.T.; clinical assitance: Y.Z.; methodology: T.L., H.S., W.Z., and X.W.; data aqutiation: H.S., W.Z., K.W., Y.L., and C.X.; software: T.L., H.S., and W.Z.; data analysis and visualization: H.S. and W.Z.; writing – original draft: H.S. and T.L.; writing – review and editing: T.L. and Y.T.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Experimental models: Organisms/strains
C57BL/6J MiceA	Shandong Pengyue	PY-C-101
Sprague-Dawley rats	Shandong Pengyue	PY-C-201
Software and algorithms
MediaPipe v0.10.33	Google	https://ai.google.dev/edge/mediapipe/
MATLAB R2021b	MathWorks	https://www.mathworks.com/products/matlab.html
Python v3.9	Python Software Foundation	https://www.python.org;
Origin 2024b	OriginLab	https://www.originlab.com/
Custom analysis code	Railway-deployed web application; GitHub	web-production-bda79e.up.railway.app; https://github.com/wenbo5867-cyber/minisite
Portable Metronome	WeiXin, China	v2.3.4
Other
Acupuncture needles (0.18 mm × 25 mm)	Huatuo, China	10053713251589
Contrast black markers	3M adhesive tape	55230
Fexible pressure sensor	Tekscan, USA	FlexiForce A201
iphone	Apple,USA	Iphone14 pro max
Silicone practice pad	ZHONGYANTAIHE	10024009727899

Experimental model and study participant details

Experimental design and overview

This study implemented a three-conditions experimental framework to systematically decode the kinematic and dynamic signatures of manual acupuncture manipulation across controlled, semi-controlled, and naturalistic clinical environments.

• •Condition 1 (Standardized Training Module): utilized a silicone-based acupuncture practice platform to establish reproducible, parameter-defined kinematic baselines.
• •Condition 2 (Animal Model): examined needle manipulation on anesthetized rats to assess kinematic–force coupling within a living tissue environment.
• •Condition 3 (Clinical Naturalistic Observation): captured manipulation during routine clinical treatments performed by expert practitioners, preserving ecological validity.

All procedures involving human participants and animal procedures were approved by the Ethics Committee of the Affiliated Hospital of Shandong University of Traditional Chinese Medicine (Approval No. (2023) 154-KY) and the Institutional Animal Care and Use Committee of Shandong University of Traditional Chinese Medicine (Approval No. JZLLSC2018-0028).

Animals

Twelve-week-old male Sprague-Dawley (SD) rats (280–350 g) and six-week-old male C57BL/6J mice (18–22 g) were purchased from Shandong Pengyue Experimental Animal Technology Co., Ltd. (rat license No. PY-C-201; mouse license No. PY-C-101). Animals were housed in the Experimental Animal Center of Shandong University of Traditional Chinese Medicine under specific pathogen-free (SPF) conditions with controlled environment (26 ± 2°C; 50–60% relative humidity; 12 h light-dark cycle) and unrestricted access to standard chow and water. Rats and mice were anesthetized with inhaled isoflurane to prevent movement during needling. Acupuncture was administered at Zusanli (ST36) according to the WHO Standard Acupuncture Point Locations in the Western Pacific Region using sterile disposable stainless-steel acupuncture needles (0.18 mm × 25 mm; Huatuo, China).

Acupuncture operators

Twenty-eight right-handed operators (10 males and 18 females) representing two different expertise levels were recruited to capture variability in manual performance.

• •Experts: postgraduate trainees with national practitioner certification (n = 18, licensed acupuncturists with ≥2 years of clinical experience).
• •Beginners: undergraduates in acupuncture majors (n = 10, <6 months of training).

All operators performed manipulations under identical experimental constraints to minimize environmental and procedural confounding.

Sex as a biological variable

Given the focused objective of quantifying kinematic parameters of needle manipulation and the standardized right-handed operational protocol applied across all participants, sex-based differences were not analyzed as a primary variable. The sample size and study design were insufficient to evaluate potential sex effects on manipulation kinematics; this represents a limitation acknowledged in the present study. All animal experiments used male rats and mice to minimize hormonal variability; the influence of sex on kinematic–force coupling in living tissue was not assessed.

Method details

Needling manipulation paradigm

Condition 1: Standardized training module

Operators performed bidirectional twirling-rotating manipulation (nian zhuan fa)^21,28 and lifting-thrusting (ti cha fa)^30,31 on a silicone practice pad (ZHONGYANTAIHE, Model 10024009727899) simulating subcutaneous tissue resistance. Stimulation intensity of twirling-rotating manipulation is defined by rotational amplitude and frequency: 90° at 2 Hz, 180° at 0.5 Hz, 180° at 1 Hz, 180° at 2 Hz, and 360° at 2 Hz. Each trial lasts 30 s and is synchronized to an auditory metronome (Portable Metronome v2.3.4). Conditions were randomized across operators with 60 s rest intervlas between trails.

Condition 2: Animal experiment

Operators also applied five standardized twirling-rotating manipulation for 30 s: 90° at 2 Hz, 180° at 0.5 Hz, 180° at 1 Hz, 180° at 2 Hz, and 360° at 2 Hz with metronome guidance.

Condition 3: Clinical Naturalistic Observation

Expert practitioners performed acupuncture under routine clinical conditions without predefined manipulation parameters. Techniques were guided by individual diagnostic assessment and tactile feedback, including acquisition of deqi, using self-selected amplitude and frequency without metronome guidance to approximate natural manipulation under controlled physiological conditions.

Video acquisition

Hand kinematics were captured at 4K resolution (3840 × 2160 pixels) and 60 frames s⁻¹ using a stabilized optical setup to eliminate camera drift and parallax distortion. The imaging axis was oriented perpendicular to the thumb–index finger manipulation plane.

Eight high-contrast black markers (3M, model 55230) were positioned on predefined anatomical landmarks of the operator’s right hand to capture fine motor coordination of the thumb–index finger pinch interface. ID1: distal of the index finger, ID2: distal interphalangeal joint of the index finger, ID3: proximal interphalangeal joint of the index finger, ID4: metacarpophalangeal joint of the index finger, ID5: distal of the thumb, ID6: interphalangeal joint of the thumb, ID7: metacarpophalangeal joint of the thumb and ID8: midpoint between the first and second metacarpals. A stationary reference marker enabled spatial normalization and correction for camera drift.

Frequency standardization

Manipulation frequency was standardized using a digital metronome (Portable Metronome v2.3.4) that delivered continuous auditory cues, ensuring cross-operator consistency and reducing intrinsic motor variability.

Flexible force measurement

To cross-validate kinematic frequency and amplitude metrics, needle-tissue interaction forces were simultaneously recorded using flexible piezoelectric thin-film force sensors (FlexiForce A201, Tekscan, USA; sensing area 9.53 mm diameter; thickness 0.203 mm; force range 0–111 N). Calibration followed the manufacturer’s protocol: sensors were mechanically conditioned by applying 110% of full-scale load (122 N) for 3 s, repeated 5 times. Static calibration was performed using certified standard weights (0.5–10 N range, 0.01 N precision) placed incrementally on the sensing area. Voltage outputs were acquired via a microcontroller-based system (ATmega328P, feedback resistor 20 kΩ, supply voltage 5 V, sampling rate 100 Hz) and converted to force units in MATLAB (R2021b) using a third-order polynomial fit (R² > 0.998 across all sensors; mean linearity error <2.5%) (Figure S3A).

In application, the sensor was affixed to the radial side of the operator’s index finger pad using medical-grade double-sided adhesive tape, positioned to capture the primary contact force during thumb-index finger pinch (Figure S3B). Rotational signals were sampled at 5 Hz and identified via peak detection in the force time series, allowing extraction of dominant frequency, amplitude periodicity, and duty cycle for direct comparison with video-derived kinematic parameters (Figures S3C and S3D).

Marker-based motion tracking

A fixed reference marker established spatial calibration parameters prior to tracking. In the initial frame, marker radii were estimated via Hough circle detection, after which centroids were manually verified to establish persistent marker identities. Local regions of interest (ROIs) were then defined for each marker. Subsequent frames were processed using an adaptive Gaussian-enhanced segmentation algorithm constrained within each ROI. Marker centroids were iteratively localized based on their prior positions, enabling subpixel trajectory reconstruction. For every frame, marker identity, timestamp, and global X-Y coordinates were recorded, generating time-resolved motion trajectories that were visually inspected to confirm tracking fidelity.

MediaPipe markerless tracking approaches

To evaluate alternative markerless tracking strategies, a video-based hand-tracking pipeline was implemented using OpenCV and MediaPipe Hands (v0.8.10.1). MediaPipe’s two-stage architecture (palm detector +21-landmark regression network) extracted normalized 2D coordinates for 21 predefined hand keypoints. Eight anatomical landmarks corresponding to ID1−ID8 were selected from the 21-keypoint set. Normalized coordinates were converted to pixel space (3840 × 2160) and temporally sequenced at 60 Hz. Missing detections (occlusion-induced dropout rate: 3.2 ± 1.1% across all trials) were forward-filled. Kinematic parameters derived from the markerless pipeline were compared against the marker-based gold standard via Bland-Altman analysis.

OpenCV was utilized for video frame acquisition, color space conversion, brightness/contrast adjustment, and real-time visualization of tracked keypoints with ID labels; processed frames were exported as annotated video files for validation. Structured trajectory data were saved in spreadsheet format for subsequent kinematic analysis, including angular displacement, angular velocity, and inter-joint coupling computation.

Needling manipulation experience assessment

To assess whether biomechanical context (with gloves v.s. bare-hand; with markers v.s. markerless) modulates the practitioner’s subjective handling sensation, operators completed a post-session experience rating scale after each manipulation block across all three conditions (Figures S1 and S2). Scale items addressed operational naturalness, finger dexterity, and tactile feedback (full questionnaire see Table S1).

Multidimensional kinematic feature extraction

Kinematic feature quantification

To characterize fine motor behavior during acupuncture manipulation, two-dimensional trajectories of each marker (ID1−ID8) were used to derive core kinematic features, including displacement, velocity, acceleration, and angular motion. Displacement was computed using a lagged frame difference to capture local movement amplitude, while velocity and acceleration were derived as first- and second-order temporal derivatives, respectively. Angular orientation and angular velocity were calculated to describe directional dynamics and rotational transitions during manipulation.

Local displacement (mean squared displacement, MSD)

Spatial movement amplitude was quantified using lagged displacement

$${\text{MSD}}_t^i=\sqrt{{\left(x_t^i-x_{t-\mathrm{\tau }}^i\right)}^2+{\left(y_t^i-y_{t-\mathrm{\tau }}^i\right)}^2}$$

where (${\mathrm{x}}_{\mathrm{t}}^{\mathrm{i}}$, ${\mathrm{y}}_{\mathrm{t}}^{\mathrm{i}}$) denotes the centroid of marker i at frame t, and τ is the temporal lag (in frames). Sliding-window computation was applied to generate temporally aligned displacement sequences. Initial frames (t<τ) were padded as missing values. MSD served as the foundational descriptor for downstream velocity and acceleration metrics.

Instantaneous velocity

Velocity vectors were computed as: (Δt = lag×frame interval):

$${\mathbf{v}}_t^i=\left(\frac{x_t^i-x_{t-\mathrm{\tau }}^i}{\Delta t},\frac{y_t^i-y_{t-\mathrm{\tau }}^i}{\Delta t}\right)$$

Where Δt = τ·Δt_frame. The velocity magnitude (speed) was defined as

$${Speed}_t^i=\Vert {\mathit{v}}_t^i\Vert$$

Instantaneous velocity captures movement rate and rhythmic modulation during manipulation.

Angular motion

Trajectory orientation was derived using the four-quadrant arctangent:

$${\theta }_t^i=\mathrm{atan}2\left(y_t^i-y_{t-\mathrm{\tau }}^i,x_t^i-x_{t-\mathrm{\tau }}^i\right)$$

providing a continuous description ranging from [-π,π] of directional dynamics within the 2D plane.

Angular velocity

Angular velocity was defined as over the lag window:

$${\omega }_t^i=\frac{\Delta {\theta }_t^i}{\Delta t}$$

where cyclic correction prevented angle wrapping artifacts. Elevated values indicate rapid directional transitions associated with torque generation.

Acceleration (jerk index)

Acceleration was approximated as the temporal derivative of velocity:

$${\mathrm{a}}_t^i\approx \frac{{\mathbf{v}}_t^i-{\mathbf{v}}_{t-\mathrm{\tau }}^i}{\Delta t}$$

The magnitude was used as a jerk index to characterize motion smoothness. Lower jerk values denote fluent neuromotor control, whereas higher values reflect abrupt force application.

Entropy-based complexity analysis

Sample entropy

Signal irregularity was quantified using sample entropy (embedding dimension = 2; tolerance = 0.2 SD). Lower entropy indicates structured, repeatable motion; higher values reflect stochastic dynamics.

$$SampEn\left(m,r,N\right)=-\mathit{ln}\frac{A}{B}$$

where m is embedding dimension (set to 2), ris tolerance (0.2 × standard deviation), and A, B represent matched template counts.

Multiscale entropy (MSE)

Coarse-grained time series were generated across scales (τ = 1–15), followed by entropy estimation (m = 2; r = 0.15 SD).

$$y_j^{\left(\tau \right)}=\frac{1}{\tau }\sum _{i=\left(j-1\right)\tau +1}^{j\tau }{x}_i$$

Scale-dependent entropy profiles enabled assessment of long-range temporal organization in motor behavior.

Nonlinear dynamic analysis

To quantify the complexity and stability of movement patterns, nonlinear dynamic analyses were performed. Signal irregularity was assessed using sample entropy, while multiscale entropy was used to evaluate temporal organization across multiple time scales. Phase-space reconstruction was conducted to capture the underlying dynamical structure, followed by recurrence quantification analysis (RQA), including recurrence rate, determinism, entropy, laminarity, and trapping time. These metrics characterize the predictability, stability, and intermittency of motor behavior.

Phase-space reconstruction and recurrence analysis

Phase-space reconstruction

State vectors were reconstructed as:

$$\mathbf{X}\left(t\right)=\left[x\left(t\right),x\left(t+\tau \right),...,x\left(t+\left(m-1\right)\tau \right)\right]$$

where embedding dimension was determined via the Cao method, while time delay τ was optimized using average mutual information combined with the C-C approach. The reconstructed trajectories were used to compute Euclidean distance matrices and generate binary recurrence plots.

Recurrence plot (RP)

The recurrence matrix was defined as:

$$R_{i,j}=\mathrm{\Theta }\left(ϵ-\parallel {\mathrm{X}}_i-{\mathrm{X}}_j\parallel \right)$$

where Θ is the Heaviside function and ϵ is a threshold distance.

Recurrence quantification analysis (RQA)

Nonlinear dynamics were quantified using standard recurrence metrics, including recurrence rate, determinism, longest diagonal line, mean diagonal length, Shannon entropy, laminarity, and trapping time. Collectively, these indices characterize predictability, stability, and intermittency within manipulation dynamics.

Metric Name	Formula	Meaning
Recursion Rate	$\text{REC}=\frac{1}{M^2}\sum R_{ij}$	Overall predictability of the system. Measure the proportion of states that are close to each other across the entire phase space.
Determinism	$\mathrm{DET}=\frac{{\sum }_{lk\ge lmin}{l}_k}{\sum R_{ij}}$	Deterministic components in dynamical processes. Quantify the proportion of recursive points produced by deterministic dynamics
Longest Diagonal Line Length	Lmax = max(lk)	The longest stable evolution period of the system. It reflects the longest period of time a system can continuously and parallelly evolve along a certain trajectory.
Mean Diagonal Line length	$L_{mean}=\frac{1}{K}\sum l_k$	Average stability of system dynamics. It measures how long on average the system state can be maintained without deviating from its original trajectory after being subjected to minor disturbances.
Shannon Entropy	ShanEn = −∑P(lk)log2p(lk)	The complexity of recursive patterns. Measure the uncertainty of the diagonal length distribution. High entropy indicates that the system has multiple evolutionary behaviors over different time lengths, and the dynamic structure is complex.
Laminarity	$\text{LAM}=\frac{{\sum }_{v_k}\ge v_{{\mathit{min}}^{v_k}}}{\sum R_{ij}}$	A tendency for the system to stagnate. Quantifying the probability that the system state will stagnate or change slowly in a certain area.
Trapping Time	$TT=\frac{1}{K_V}\sum v_K$	The average time it takes for the system to get out of stagnation. Also known as capture time, it measures the average time it takes for a system to leave once it enters a state of stagnation.

Inter-joint coordination analysis

Functional coordination among finger joints was assessed using complementary measures of synchronization and coupling. Phase-locking value (PLV)^32,33 and phase-phase coupling (PPC)³⁴ were used to quantify phase synchronization, while frequency-domain coherence evaluated linear coupling within the dominant manipulation frequency band. Directional interactions were inferred using Granger causality, and temporal lead-lag relationships were quantified using phase lag analysis. All analyses were performed on angular velocity time series to capture coordination dynamics underlying rhythmic manipulation.

Phase-locking value (PLV)

The phase-locking value quantifies the consistency of phase differences between two signals over time, providing a measure of phase synchronization independent of amplitude fluctuations. For each pair of markers i and j, the instantaneous phase φ(t) was extracted from the analytic signal via the Hilbert transform. The PLV was computed as:

$${\text{PLV}}_{ij}=\left|\frac{1}{N}{\sum }_{t=1}^N{e}^{i\left[{\varphi }_i\left(t\right)-{\varphi }_j\left(t\right)\right]}\right|$$

where N is the number of time samples. PLV ranges from 0 (no phase synchronization) to 1 (perfect phase locking). A PLV matrix (8 × 8) was constructed for each manipulation trial, and the group-averaged matrix was visualized as a heatmap.

Phase-phase coupling (PPC)

Phase-phase coupling measures the statistical dependence between the phase variables of two signals, reflecting direct phase–phase interactions. PPC was computed as the mean resultant length of the phase difference distribution:

$${\text{PPC}}_{ij}=\left|⟨e^{i\left[{\varphi }_i\left(t\right)-{\varphi }_i\left(t\right)\right]}⟩\right|$$

Where ⟨·⟩ denotes temporal averaging. PPC values close to 1 indicate strong phase coupling, while values near 0 suggest independent phase dynamics. The PPC matrix was generated and thresholded to construct a binary adjacency matrix for network visualization, where edges with PPC >0.1 were retained to illustrate the hand skeleton coupling architecture.

Frequency-domain coherence

Coherence analysis was performed to assess frequency-specific linear coupling between joint pairs. The magnitude-squared coherence was computed as:

$$C_{ij}\left(f\right)=\frac{{\left|S_{ij}\left(f\right)\right|}^2}{S_{ii}\left(f\right)·S_{jj}\left(f\right)}$$

where S_ij(f) is the cross-spectral density between signals i and j at frequency f, and S_ii(f) and S_jj(f) are the respective auto-spectral densities. The coherence value ranges from 0 to 1, with higher values indicating stronger linear correlation at specific frequencies. The mean coherence across the dominant manipulation frequency band (1–3 Hz) was extracted to construct the coherence matrix.

Granger causality

Granger causality was employed to infer directed information flow between joint pairs. For each pair of markers, we tested whether the past values of signal X significantly improved the prediction of signal Y beyond the autoregressive model of Y alone. The bivariate vector autoregressive (VAR) model was fitted:

$$Y_t={\alpha }_0+{\sum }_{k=1}^p{\alpha }_k{Y}_{t-k}+{\sum }_{k=1}^q{\beta }_k{X}_{t-k}+{ϵ}_t$$

The null hypothesis H₀: β₁ = β₂ = … = β_q = 0 was tested using an F-statistic comparing the restricted model (without X) and the unrestricted model (with X). The resulting p values were assembled into an 8 × 8 Granger causality matrix, where lower p values indicate stronger causal influence from the predictor joint (row) to the effector joint (column).

Phase lag (τ)

The phase lag τ quantifies the temporal lead–lag relationship between joint motions, measured in milliseconds. For each joint pair, the time delay maximizing the cross-correlation of the angular velocity signals was identified:

$${\tau }_{ij}={\mathit{arg}\mathit{max}}_{\tau }\left|R_{ij}\left(\tau \right)\right|$$

where R_ij(τ) is the cross-correlation function between signals i and j at lag τ. Positive τ values indicate that joint i leads joint j (phase advance), while negative values indicate that joint i lags joint j (phase delay). The phase lag matrix was computed and visualized using a diverging colormap (blue–white-red) centered at zero to highlight directional temporal asymmetries in the thumb–index finger coordination pattern.

Hierarchical clustering and inter-joint analysis

For each operator, twelve kinematic and nonlinear dynamic parameters were extracted from each of the eight markers (ID1−ID8). These included: mean and standard deviation of MSD, velocity, and acceleration; mean direction angle and angular velocity; sample entropy; multiscale entropy; and recurrence rate and determinism from RQA. By averaging these 12 parameters across the five manipulation techniques, a 96-dimensional integrated feature vector was constructed for each operator (8 markers × 12 parameters). After Z score normalization, pairwise Euclidean distances were computed, and hierarchical clustering was performed using Ward’s minimum-variance method. The statistical significance of the resulting clusters was verified using PERMANOVA with 1,000 permutations.

Feature vectors were aggregated per joint and evaluated using Pearson’s correlation to generate an 8 × 8 similarity matrix. Hierarchical clustering identified functional joint groupings, revealing biomechanical synergy within the thumb-index finger kinetic chain.

Fourier decomposition of motion trajectory

Fast Fourier transform (FFT) was applied to motion trajectories to extract dominant frequency components, corrected amplitudes, and phase-resolved sinusoidal reconstructions for rhythmic component visualization.

Fast Fourier transform (FFT) was applied to obtain the amplitude spectrum and dominant frequency components using peak detection:

$$X\left(f\right)=\sum _{n=0}^{N-1}x\left(n\right)e^{-i2\pi fn/N}$$

Amplitude correction:

$$A=\frac{2\mid X\left(f\right)\mid }{N}$$

Dominant components were identified via peak detection.

The five highest-amplitude components were extracted. Based on their frequency, corrected amplitude, and phase, pure sinusoidal waveforms were reconstructed for each component. All reconstructed waveforms were visualized in a three-dimensional coordinate system (time, frequency, amplitude), illustrating the superposition and evolution of rhythmic components across different techniques and operators.

Quantification and statistical analysis

All video-based marker tracking, trajectory extraction, nonlinear dynamic analysis, hierarchical clustering, and figure generation were performed using custom-developed scripts in MATLAB R2021b (The MathWorks, Inc., Natick, MA, USA, https://www.mathworks.com/). The computation of kinematic parameters was implemented in Python 3.9 (Python Software Foundation, https://www.python.org/) using standard scientific libraries (NumPy, https://numpy.org/; SciPy, https://scipy.org/; pandas, https://pandas.pydata.org/). All statistical analyses are specifically described in figure legends and in the STAR Methods above. Sample sizes are reported in the figure legends where n = operators (Condition 1 and Condition 3) or n = animals (Condition 2). All data are represented as mean ± SEM unless otherwise stated. Origin 2024b (OriginLab, Northampton, Massachusetts, USA, https://www.originlab.com/) was used for all statistical tests and graphical representations. Custom Python scripts (version 3.9; Python Software Foundation, https://www.python.org/) were used for nonlinear dynamics and phase-coupling analyses (see Data and code availability). Detailed implementation and parameter settings are provided in the Supplemental Information.

Published: July 3, 2026

Supplemental information

Data S1. Source data and supplementary datasets

This file contains the raw numerical data, statistical analyses, and processed datasets supporting the figures and supplementary analyses in this manuscript.