PaperlayerView exact revision on arXiv

Source-linked AI summary

A spinal-inspired spiking neural network for continuous finger-joint kinematics estimation from high-density surface EMG

Cristian David Guerrero-Mendez, Leonardo Abdala Elias

arXiv:2610.33979v1cs.HC

TL;DR

Data-driven decoding of continuous finger kinematics from HD sEMG has limited physiological interpretability, and spinally organized approaches remain less explored. The paper evaluates a spinal-inspired SNN for simultaneous and proportional prediction, reporting improved performance over traditional machine-learning and deep-learning methods. Its internal firing-rate and ISI measures also differed across motor tasks.

  • Problem

    Data-driven methods for finger-kinematics decoding have limited physiological interpretability, while SNN use for continuous finger movement prediction organized around spinal motor-control principles remains less explored.

    Source: §1 · p. 2

  • Method

    The authors evaluate a spinal-inspired SNN that uses spike-encoded HD sEMG, a phenomenological representation of Henneman’s size principle, E/I balance, and input variability to predict 15 hand joint angles.

    Source: §1 · p. 2

  • Results

    The proposed SNN significantly improved simultaneous and proportional finger-kinematics prediction from HD sEMG compared with traditional machine-learning and deep-learning approaches.

    Source: §4 · p. 17

  • Takeaways & Limitations

    Firing rates and ISI statistics served as interpretable latent descriptors and differed significantly across motor tasks.

    Source: §4 · p. 19

    Robustness under more dynamic conditions and in populations with severe motor impairments remains to be validated.

    Source: §4 · p. 20

Abstract

from arXiv · show

Decoding simultaneous and proportional finger kinematics from high-density surface electromyography (HD sEMG) remains challenging. Although data-driven methods have shown promising performance for finger movement decoding, their limited physiological interpretability remains a relevant limitation. This study proposes and evaluates a spiking neural network (SNN) that incorporates spinal-inspired structures and mechanisms to predict multi-joint finger kinematics from HD sEMG signals. The spinal-inspired SNN includes a spike-encoding computational model with input-noise variability, a phenomenological representation of Henneman's size principle, and excitation/inhibition (E/I) balance. HD sEMG and kinematic data were collected from 23 healthy participants performing grasping tasks under different elbow and wrist postures. Performance was assessed using six prediction metrics and compared against six baseline machine learning and deep learning methods. The proposed method showed a large and significant ($p<0.05$, rank-biserial correlation $r_{rb}>0.7$) improvement over all baseline methods across all accuracy metrics. Performance improved significantly with increasing spike encoder firing rates up to approximately 128 Hz, and with more regular input spike trains compared with a Poisson-like encoder. In contrast, the E/I balance did not significantly affect the decoding performance of the algorithm. The firing rate and inter-spike interval statistics of the SNN differed significantly across motor tasks. These results demonstrate that incorporating neurophysiological principles into SNNs improves prediction performance while providing additional physiological interpretability for simultaneous and proportional finger-kinematics decoding. This approach represents a promising step toward physiology-informed neural networks, with potential applications in motor rehabilitation.

1 Introduction

The paper motivates continuous finger-kinematics decoding as relevant to human-machine interfaces and neurorehabilitation, while identifying limited physiological interpretability in data-driven methods. It introduces a spinal-inspired SNN for simultaneous and proportional prediction from HD sEMG.

1 Introduction

  • Accurate decoding of coordinated and individuated finger kinematics is relevant to intuitive human-machine interfaces and neurorehabilitation strategies.The paper frames natural hand movement as involving both independent finger control and synergistic neural commands.Source: §1 · p. 2
  • Data-driven motor decoding faces limited physiological interpretability, performance variability across modeling choices, and reliance on data that can be difficult to obtain clinically.The paper identifies these issues alongside unresolved questions about interpreting and evaluating neurorehabilitation interventions.Source: §1 · p. 2
  • SNN applications in sEMG have largely focused on gesture classification, leaving their ability to predict natural, continuous finger movements less established.Prior continuous-regression work has not organized the network according to spinal motor-control principles, as characterized by the authors.Source: §1 · p. 2
  • Physiologically inspired SNNs are proposed to predict 15 hand joint angles from HD sEMG across three motor tasks and multiple upper-limb postures.The approach combines spike-based sEMG encoding, a phenomenological representation of motor neuron recruitment inspired by Henneman’s size principle, E/I balance, and stochastic input variability.Source: §1 · p. 2

2 Methods

The methods combine recordings of hand kinematics and HD sEMG across participants, tasks, and postures with a spinal-inspired SNN decoder. The model encodes sEMG as spikes, processes them with LIF neurons, and maps neural activity to continuous joint-angle estimates.

2.1 Experimental dataset

  • Twenty-three healthy volunteers aged 20–70 years contributed recordings from the right hand.The sample included 12 men and 11 women; all recordings were performed on the right hand.Source: §2.1 · p. 3

2.1.1 Signal acquisition

  • HD sEMG from forearm flexor and extensor muscles was recorded alongside hand kinematics estimated using five cameras.Two 64-channel electrode grids recorded FDS and EDC activity, while the cameras captured hand motion at 100 Hz.Source: §2.1.1 · p. 3

2.1.2 Procedures

  • Participants performed one-finger pinch, two-finger pinch, and full-hand grasp tasks across five wrist and two elbow postures, yielding 30 conditions.Each condition lasted 30 seconds, and task order was randomized; Figure 1 depicts the setup and protocol.Source: §2.1.2 · p. 3, Figure 1 · p. 4

2.2 Data pre-processing

  • Signals were split into training, validation, and test sets, independently filtered, and scaled using training-set parameters.HD sEMG envelopes were computed from channel-wise RMS; joint angles were reconstructed from camera landmarks and standardized using training data.Source: §2.2 · pp. 4–5

2.3.1 Spike Encoding:

  • The spike encoder converts normalized RMS signals from each sEMG channel into spike trains whose rate follows input amplitude and whose regularity is adjustable.An order parameter controls the transition from irregular Poisson-like firing to more regular Gamma-process firing without changing mean firing rate.Source: §2.3.1 · p. 5

2.3.2 Neuron model:

  • The network uses LIF neurons with membrane leakage and synaptic-current decay, then updates their states in discrete time.The voltage formulation includes a decay mechanism for both membrane potential and input current; the implementation resets membrane potential fully after each spike.Source: §2.3.2 · pp. 5–7

2.3.3 Architecture:

  • The SNN comprises spike-encoded inputs, two hidden LIF layers, and an MLP that produces continuous predictions for 15 finger joints.Membrane potentials are averaged over overlapping 100 ms windows before the MLP maps them to joint angles.Source: Figure 2 · p. 6, §2.3.3 · p. 7

2.4 Physiologically inspired components

The methods implement spinal-inspired neuron heterogeneity and E/I constraints, tune the SNN and comparison methods, and assess ablations, encoder settings, prediction performance, and task-related neural activity.

2.4 Physiologically inspired components

  • The physiological adaptations include graded firing thresholds and time constants inspired by Henneman’s size principle, adjustable encoder firing rate and input variability, and fixed channel-level E/I signs.Lower-threshold neurons receive longer time constants, while encoder order controls spike regularity and noise.Source: §2.4 · p. 8

2.5 Hyperparameter tuning and training

  • Optuna tuning optimized threshold, encoder amplitude and order, and E/I balance using validation-set variance-weighted R2, with 60 trials per study.Tuning used five participants per task under a neutral wrist and extended-elbow posture; training used backpropagation through time with surrogate gradients and early stopping.Source: §2.5 · pp. 8–9

2.6 Data Analysis

  • Six reference methods—RR, KNN, HGB, SVR, MLP, and CNN—were compared with the SNN, alongside three optimized SNN variants for ablation.The ablation variants altered the encoder and/or second-layer neuron heterogeneity; baseline methods used consistent data splits.Source: §2.6 · p. 9, §2.6.1 · p. 9

2.6.1 Ablation analysis

  • The component ablation included all 30 experimental conditions, while exploratory parameter analyses were restricted to full-hand grasp with neutral wrist and extended elbow.Explorations varied encoder firing rate, spike-train variability, and E/I configuration.Source: §2.6.1 · p. 9, §2.6.2 · p. 9
  • The ablation compared the full SNN with a conventional SNN and variants that separately removed the heterogeneous second layer or the proposed spike encoder.The variants shared the LIF architecture and MLP readout, with the conventional SNN using an adaptive temporal encoder and unconstrained weights.Source: §2.6.1 · p. 9

2.6.3 Performance and interpretability assessments

  • Prediction performance was assessed with RMSE, R2, Pearson correlation, MAE, maximum cross-correlation, and estimated lag; internal activity measures included firing rates and ISI statistics.Performance metrics were variance-weighted across joints, and internal activity was compared across tasks at a fixed elbow and wrist posture.Source: §2.6.3 · p. 10

2.6.4 Statistical analysis

  • Non-parametric repeated-measures tests evaluated condition differences, followed by Bonferroni-adjusted Wilcoxon comparisons when significant differences were detected.Analyses used one value per participant (n = 23); effect sizes were reported using Kendall’s W and rank-biserial correlation.Source: §2.6.4 · p. 10

3 Results

Across 30 experimental conditions, the spinal-inspired SNN outperformed the reference methods, while ablations and encoder tests showed that spike encoding was a major contributor to performance. Its firing statistics also differed across motor tasks.

3.1 Data characterization

  • Joint angles and FDS/EDC sEMG envelopes followed the 0.5 Hz command in the recordings.Figure 3 presents example recordings averaged across 23 participants for extended elbow and 90° wrist posture.Source: §3.1 · p. 10, Figure 3 · p. 11

3.2 Ablation of the physiologically inspired components

  • The spike encoder accounted for most of the ablation performance gain; the heterogeneous second layer added improvement only when combined with that encoder.SNNPhysio outperformed the conventional SNN and adaptive-temporal-encoder variant on R2 and ρ, and also outperformed the variant without the heterogeneous second layer.Source: §3.2 · p. 11, Figure 4 · p. 12

3.3 Effects of the spike encoder firing rate

  • Beyond θ=250, no further significant R2 improvement was observed, while ρ showed only small additional gains at the highest amplitudes.Increasing spike-encoder firing rate improved decoding performance; at θ=250, R2 was 0.75 [0.59, 0.83] and ρ was 0.89 [0.82, 0.92] in the tested FG posture.Source: §3.3 · pp. 12–13, Figure 5 · p. 13

3.4 Effects of the variability of input spikes

  • The Poisson-like encoder (CoV 100%) performed worse than more regular inputs, while no significant differences appeared among CoV levels of 26% or below.Median R2 ranged from 0.83 to 0.86 at CoV levels of 26% or below, versus 0.73 [0.55, 0.80] at CoV 100%.Source: §3.4 · pp. 13–14, Figure 6 · p. 14

3.5 Excitation/inhibition balance

  • Imposed E/I balance did not significantly affect R2 or ρ in the tested full-hand grasp condition.After training, excitatory and inhibitory weight proportions matched the imposed configurations; unconstrained weights remained close to 50% each.Source: §3.5 · p. 14, Figure 7 · p. 15

3.6 Performance comparison

  • SNNPhysio used 29,903 trainable parameters, fewer than MLP and CNN but more than RR; statistical comparisons were reported only against SNNPhysio.Supplementary comparisons with HGB, the reference method with the highest R2, found significant improvements in many task–joint combinations, particularly during FG.Source: §3.6 · pp. 15–16
  • Across 30 conditions, SNNPhysio significantly outperformed six reference methods on R2, ρ, X_Cmax, RMSE, and MAE, with large effects (p<0.0001, W>0.82).For cross-correlation lag, it was significantly shorter than MLP, KNN, HGB, and SVR, but not RR or CNN.Source: §3.6 · pp. 14–15, Figure 8 · p. 16

3.7 Interpretability of the SNN internal activity

  • Firing rates and inter-spike-interval statistics differed across tasks in both network layers; 1FP produced the highest firing rates in both layers.Across tasks, ISI distributions were positively skewed and showed high kurtosis, with a concentration of short intervals and a long tail.Source: §3.7 · p. 16

4 Discussion

The spinal-inspired SNN improved simultaneous and proportional finger-kinematics prediction over traditional machine-learning and deep-learning methods across tasks and tested conditions. Ablations favored spike-rate encoding and regular input trains, while E/I balance did not significantly affect performance; internal firing descriptors were task-dependent but not quantitatively physiological.

4 Discussion

  • The SNN significantly improved simultaneous and proportional prediction of 15 finger-joint angles over traditional machine-learning and deep-learning methods across both elbow postures, most wrist positions, and all three tasks.It used substantially fewer trainable parameters than the MLP and CNN.Source: §4 · p. 17
  • Replacing the spike encoder reduced median R^2 by approximately 0.15, while heterogeneous second-layer neurons improved performance only when paired with that encoder.The encoder conveys sEMG activation level through firing rate; the adaptive temporal encoder instead represents changes between consecutive samples.Source: §4 · pp. 18–19
  • No further significant R^2 improvement occurred when encoder firing rates reached approximately 128 Hz, a network input-gain setting rather than a physiologically calibrated rate.These rates exceed typical human motor-unit discharge rates during voluntary contractions.Source: §4 · p. 19
  • Fully noisy input trains with CoV 100% performed significantly worse than more regular inputs, while CoV values of 14–26% preserved performance.The latter range overlaps reported variability for human hand motor units during steady voluntary contractions.Source: §4 · p. 19
  • E/I weight ratios, including more excitatory regimes up to 80/20, did not significantly affect decoding performance.The authors note that such biologically plausible E/I constraints can be incorporated without compromising prediction performance.Source: §4 · p. 19
  • Firing-rate and ISI statistics differed across motor tasks, but task effects cannot be separated from task-specific hyperparameter configurations.The analysis covered one elbow and wrist posture; Figure 9 shows task-wise firing properties, with participant-level boxplots and significant differences marked.Source: Figure 9 · p. 18, §4 · p. 19
  • SNN firing rates exceeded experimentally reported human motor-neuron ranges, limiting direct biological interpretation despite capturing qualitative firing properties.The mismatch was particularly notable in layer 2 during the one-finger pinch task.Source: §4 · pp. 19–20
  • Validation used controlled tasks and postures in healthy participants, so robustness in dynamic conditions and severe motor impairment remains unvalidated.The study also used two HD sEMG grids over forearm flexor and extensor muscles.Source: §4 · pp. 19–20

Funding

The work received consortium, institutional, and individual research support.

Funding

  • The MYOREHAB consortium work was supported by FAPESP grant 22/05104-8.The consortium is part of the 4th EU-LAC Joint Call in STI 2022.Source: Funding · p. 20
  • The Brazilian Institute of Data Science (BI0S), a FAPESP Applied Research Center in Artificial Intelligence, supported the work through grant 20/09838-0.Source: Funding · p. 20
  • Cristian David Guerrero-Mendez held a FAPESP doctoral scholarship, and Leonardo Abdala Elias was a CNPq Research Productivity Fellow.The listed process numbers are 25/03022-2 and 316320/2023-4, respectively.Source: Funding · p. 20

S1 LIF Neuron

The LIF model represents membrane-threshold crossings and reset behavior, while its firing-rate and temporal responses reflect input current and membrane time constant. Its membrane potential integrates input more slowly and smoothly than synaptic current.

S1 LIF Neuron

  • The neuron records threshold crossings at 15 mV and resets the membrane potential to 0 mV, rather than explicitly generating spikes.The illustrated constant input current is approximately 1.45 nA.Source: S1 LIF Neuron · p. 26
  • The f–I curve shows an approximately proportional increase in firing rate with input current, reflecting the absence of an explicit refractory period.Panel C shows the firing-frequency–current relationship.Source: S1 LIF Neuron · p. 26, Figure S1 · p. 26
  • Synaptic current evolves faster than membrane potential, whose slower, smoother dynamics support temporal integration.Smaller membrane time constants produce faster, lower-amplitude charging; larger values produce slower, larger depolarization.Source: S1 LIF Neuron · p. 26, Figure S1 · p. 26

S2 Spike encoding properties

The encoder maps continuous channel inputs into spike trains, with modulation amplitude affecting firing activity and process order affecting spike-train regularity. Its output preserves the input’s temporal structure.

S2 Spike encoding properties

  • Increasing Gamma-process order reduces ISI dispersion, CoV, and standard deviation relative to the more variable Poisson process.Empirical CoV exceeds theoretical values because input modulation varies firing rate over time.Source: S2 Spike encoding properties · p. 26
  • At θ=50, the firing-rate distribution peaks around 7 Hz, and the Poisson-process autocorrelation peaks at 0 ms, consistent with statistical independence of spikes.Source: S2 Spike encoding properties · pp. 26–27
  • At θ=50, firing rates fluctuate between approximately 2 and 30 Hz across 128 channels, producing denser trains than at θ=10 while preserving input temporal structure.The firing rate tracks the modulation signal.Source: S2 Spike encoding properties · p. 27

S2.1 Pseudo-code of the spike encoding

  • The encoder accumulates amplitude-scaled normalized RMS input and emits a binary spike when a channel’s threshold is reached, then resets the accumulator and draws a new threshold.The input is a normalized RMS envelope and the output is a spike train.Source: S2.1 Pseudo-code of the spike encoding · p. 27

S3 Baseline methods configurations

The study compares the proposed method with six baseline methods. It uses literature-described MLP and CNN architectures and selects tuned configurations for four other baselines using R2.

S3 Baseline methods configurations

S4 Hyperparameter tuning

Hyperparameter searches were conducted separately for the one-finger pinch, two-finger pinch, and full-grasp tasks. The selected configuration for each task was then used across participants and experimental conditions.

S4 Hyperparameter tuning

S5 Full performance comparison

SNNPhysio significantly outperformed HGB across many task–joint combinations, with the strongest gains in full-hand grasping and under pronated wrist posture. Improvements varied across tasks, joints, and wrist positions.

S5 Full performance comparison

  • SNNPhysio significantly outperformed HGB in 45% of 450 task–joint combinations and achieved higher median R2 in 81%.Significant gains occurred in most full-hand grasp conditions.Source: S5 Full performance comparison · p. 30
  • Full-hand grasp improved significantly for 67–93% of joints across wrist postures, while one-finger pinch ranged from 0–33%.Two-finger pinch showed a more variable 13–87% improvement range across wrist postures; gains were concentrated in central-finger joints, especially during grasping.Source: S5 Full performance comparison · p. 30
  • Pronated wrist posture (180°) had the highest proportion of significantly improved joints (59%), versus 30% at supinated posture (0°).The strongest and most consistent pronated-posture improvements occurred during full-hand grasping.Source: S5 Full performance comparison · p. 30
  • Elbow posture showed no consistent effect: 45% of joints improved significantly with an extended elbow and 44% with a flexed elbow.The authors note that finer analyses are needed to characterize individual wrist and elbow effects.Source: S5 Full performance comparison · p. 30

S6 Complementary neural analysis across tasks

The SNN’s population measures varied across decoded tasks, including synchrony, coherence, and gain between layers. Coherence also increased with pool size, while short activity segments make task differences more informative than absolute coherence values.

S6 Complementary neural analysis across tasks

  • The analysis measured synchrony, coherence, and inter-layer gain from test-set spike trains under extended-elbow, 90°-wrist conditions.Coherence was estimated from about 3 seconds of activity, so task differences are more informative than absolute values.Source: S6 Complementary neural analysis… · pp. 30–31
  • Synchrony and sCST coherence differed across tasks in both layers, with low synchrony overall (median χ2 ≤0.02).For layer 2, one-finger pinch had higher synchrony than two-finger pinch and full-hand grasp; coherence contrasts differed by layer and task.Source: S6 Complementary neural analysis… · p. 30, Figure S4 · p. 32
  • Gain differed across tasks: one-finger pinch was higher (4.07 [3.94, 4.45]) than two-finger pinch (1.57 [1.46, 1.71]) and full-hand grasp (1.75 [1.67, 2.05]).Full-hand grasp gain was also higher than two-finger pinch gain; the figure reports participant-level measures for 23 participants.Source: S6 Complementary neural analysis… · p. 31, Figure S4 · p. 32
  • Coherence increased with the number of pooled neurons in all tasks and layers and was concentrated below 3 Hz.The authors relate this pool-size pattern to common synaptic input observed in motoneuron pools; in the SNN, it arises because neurons receive the same encoded sEMG channels.Source: S6 Complementary neural analysis… · p. 31

Related papers

Loading 2610.33979v1…