A Synergistic Rehabilitation Approach for Post-Stroke Patients with a Hand Exoskeleton: A Feasibility Study with Healthy Subjects

Abstract

Hand exoskeletons are increasingly used to support post-stroke reach-to-grasp, yet most intention-detection strategies trigger assistance from local hand events without considering the synergy between proximal arm transport and distal hand shaping. We evaluated whether proximal arm kinematics, alone or fused with EMG, can predict flexor and extensor digitorum activity for synergy-aligned hand assistance. We trained nine models per participant: linear regression (LINEAR), feedforward neural network (NONLINEAR), and LSTM, each under EMG-only, kinematics-only (KIN), and EMG+KIN inputs. Performance was assessed by RMSE on test trials and by a synergy-retention analysis, comparing synergy weights from original EMG versus a hybrid EMG in which extensor and flexor digitorum measure signals were replaced by model predictions. Results have shown that kinematic information can predict muscle activity even with a simple linear model (average RMSE around 30% of signal amplitude peak during go-to-grasp contractions), and synergy analysis indicated high cosine similarity between original and hybrid synergy weights (on average 0.87 for the LINEAR model). Furthermore, the LINEAR model with kinematics input has been tested in a real-time go-to-grasp motion, developing a high-level control strategy for a hand exoskeleton, to better simulate post-stroke rehabilitation scenarios. These results suggest the intrinsic synergistic motion of go-to-grasp actions, offering a practical path, in hand rehabilitation contexts, for timing hand assistance in synergy with arm transport and with minimal setup burden.

1. Introduction

Impairments of hand function are among the most disabling consequences of stroke and critically limit independence in activities of daily living (ADLs) [1,2]. Over the past two decades, robotic technologies for upper-limb rehabilitation have proliferated in both research and clinical environments with the aim of delivering highly repetitive, task-oriented, and measurable therapy [3,4,5]. Large meta-analyses and reviews indicate that electromechanical and robot-assisted training can improve arm function and strength and, in some cases, ADLs, although clinical effect sizes and the translation to meaningful functional gains remain debated and appear sensitive to device design and dosing parameters [6,7,8,9].

A key lesson emerging from rehabilitation robotics and human–robot interaction is that assistance must be intuitive and patient-cooperative to sustain engagement and improve usability [5,10]. In wearable orthoses, the bottleneck is often the intention detection strategy (IDS), the algorithmic interface that maps user signals to device actions, which strongly affects robustness, timing, workload, and acceptance in daily use [10]. IDSs used in practice span surface electromyography (sEMG), kinematic sensing, manual triggers, and brain–computer interfaces (BCIs), each with distinct trade-offs in setup burden, reliability, and responsiveness [10,11,12,13].

Hand-specific robotic devices include end-effector systems and wearable exoskeletons/gloves designed for therapy and, increasingly, for assistance [3,12,14,15]. These devices can be either rigid or soft, depending on the clinical case and the rehabilitation scenario. Soft hand exoskeletons are very-well adaptable and versatile but they may struggle in achieving high forces, especially on hands with a high spasticity, and they usually feature more complex systems for closed-loop control [16,17]. Rigid exoskeletons may be less adaptable but they can convey high forces and they are, in most of the cases, more bulky but with a more precise and simple control schemes [18]. Control strategies for these devices can be grouped as follows. (i) EMG-driven control: widely used to trigger hand opening/clos ing via onset thresholds or pattern-recognition models; however, EMG onset detection is non-trivial in noisy, post-stroke signals, and performance can degrade with fatigue and spasticity [19,20,21]. (ii) Kinematics-driven control: local finger/wrist angles or inertial measurements are used to infer grasp states; in bilateral modes, sensors on the unaffected hand command the paretic-hand exoskeleton (master–slave mirroring) [22]. (iii) BCI-driven control: motor-imagery BCIs coupled to hand robots have shown promising clinical signals in controlled trials and systematic reviews, yet deployment remains constrained by robustness and setup-time issues [10,23]. (iv) Mirrored/bimanual strategies: robot-mediated mirror therapy and bilateral training replicate the movement of the unaffected hand on the affected side; these approaches can modulate cortical activation but, by design, do not infer intention from the paretic limb [22,24,25].

Neuroscience and motor-control studies emphasize that reach-to-grasp is a coordinated behavior in which proximal arm transport and distal hand shaping are tightly coupled via task-dependent synergies [1]. Post-stroke, interjoint coordination and the timing of grasp relative to arm transport are often disrupted [1,2]. These insights motivate control policies that fuse proximal and distal information to assist the hand in synergy with the arm. While there are demonstrations of integrated reach-to-grasp assistance combining a multi-joint arm exoskeleton with neuromuscular stimulation of the hand [26], recent reviews of hand exoskeletons and wearable soft devices primarily catalogue EMG-, kinematics-, BCI-, and mirrored-control schemes, with little emphasis on controllers that explicitly exploit shoulder–elbow kinematics to synchronize hand opening/closing with the phases of reaching [10,11,12,13]. In many clinical and research systems, hand assistance is still triggered by discrete events (e.g., EMG onset, decoded MI, button press, or mirrored command) that are largely independent of concurrent arm motion.

Natural prehension is a coordinated behaviour in which proximal arm transport and distal hand shaping are temporally coupled; for example, hand closure typically begins during the deceleration phase of the transport [27,28]. After stroke, this coupling is disrupted, with altered timing and longer deceleration phases during reach-to-grasp [2,29]. These observations motivate assistance strategies that respect arm–hand synergies rather than treating the hand in isolation.

A second, complementary pillar is the role of temporal contingency in driving neuroplasticity. Paired associative stimulation (PAS) shows that precisely timed pairing of afferent input with cortical stimulation induces Hebbian, timing-dependent changes in corticospinal excitability in humans [30,31,32]. Related work demonstrates rapid, training-induced reorganization of motor representations with short bouts of practice [33]. Together, these studies support the principle that the closer the temporal coupling between intention, movement, and afferent feedback, the stronger the plasticity response.

Clinical neurorehabilitation technologies that make feedback contingent on intention align with these principles. In chronic stroke, brain–computer interface (BCI) control coupled to functional electrical stimulation (FES) produced significantly greater and lasting upper-limb recovery than sham FES, with concomitant neuroplastic signatures [34]. By contrast, a multicentre randomized trial comparing cyclic NMES, EMG-triggered NMES, and sensory stimulation for the paretic hand found no superiority of EMG-triggered over cyclic stimulation; notably, participants in the EMG-triggered arm were instructed to relax during stimulation, effectively unpairing intention and afferent consequences [35]. These findings underscore that mere event triggering is insufficient unless intention and feedback are closely time-locked.

Beyond local hand feedback, somatosensory inputs and movements from adjacent limb segments can modulate hand motor pathways, providing a rationale for synergy-level contingency. Heteronymous reflex pathways link muscles across the upper limb [36,37], and cutaneous/proprioceptive inputs alter hand corticospinal excitability [38,39,40]. Forearm/hand posture and ongoing movement further shape cortical drive to intrinsic and extrinsic hand muscles [41,42]. In parallel, bilateral/bimanual paradigms that exploit interlimb coupling show beneficial effects on post-stroke motor outcomes [43]. Collectively, these data indicate that pairing hand assistance with phase-specific signals from the moving arm (and, where appropriate, from the contralateral limb) is biologically plausible and clinically relevant.

Building on these principles, we frame and evaluate intention-prediction strategies that use EMG, proximal arm kinematics, or their fusion to trigger hand exoskeleton open/close in synergy with ongoing arm transport during functional reach-to-grasp. This design places the controller within a contingency-informed framework, time-locking assistance to volitional state and to intersegmental context, which prior experimental and clinical evidence identifies as favourable for inducing and consolidating motor plasticity [30,33,34].

The state of the art thus reveals a relative lack of intention-decoding approaches that leverage proximal arm kinematics, alone or fused with EMG, to coordinate hand assistance with ongoing arm transport during goal-directed tasks. Addressing this gap is consistent with patient-cooperative control principles and could improve timing and naturalness of assistance during functional reach-to-grasp [5,10].

The objective of this study is to develop and evaluate an intention–prediction strategy for a hand exoskeleton that can support post-stroke opening and closing in synergy with arm motion during goal-directed actions (e.g., grasping a bottle). Specifically, we compare multiple machine-learning (ML) techniques trained to predict hand open/close from three input conditions: (1) sEMG alone, (2) kinematics from a sensorized 3-DoF arm carrying the exoskeleton at the end of the kinematic chain, and (3) a multimodal fusion of EMG and arm kinematics. Our research questions are does fusing arm-kinematic context with EMG improve the accuracy and timing of open/close predictions, compared to single-modality baselines, for controlling a hand exoskeleton during functional tasks? Are such predictions suitable to feed a hand exoskeleton control in an actual rehabilitation setting? Can such a control retain the original arm/hand patterns during grasping?

The contributions of this work are (i) a comparative evaluation of several ML intention-decoding pipelines across EMG-only, arm-kinematics–only, and multimodal inputs for dynamic and synergistic hand exoskeleton control; (ii) an experimentally validated online control strategy that uses the best predictor (in terms of trade-off between performance, simplicity of setup, and complexity of the model) to drive timely, task-consistent assistance of hand opening/closing while the arm moves.

2. Materials and Methods

2.1. Participants

Eight adult participants took part in the study (N = 8; height $182.0\pm 7.2$ cm; arm length $54.8\pm 4.6$ cm; forearm length $27.8\pm 3.3$ cm; hand length $19.7\pm 1.1$ cm; hand width $8.3\pm 0.7$ cm; age $31.6\pm 2.6$ years). Participants have been selected to cover a broad range of hand and arm sizes, rather than on a broader age range, which are the variables that are most related to the wearability of the hand exoskeleton. Hand length, for instance, goes from a minimum of 17.4 cm to a maximum of 20.5 cm. Arm length and forearm length as well go respectively from a minimum of 45.0 cm and 20.0 cm, to a maximum of 60.0 cm and 30.1 cm. All subjects reported to be right-handed.

2.2. System Description

Rehabilitation Platform with the Hand Exoskeleton

The system used in this study is a wheeled rehabilitation platform that has been designed to accommodate different exoskeletons on the end-effector. A passive, sensorized 3-DoF arm served as the end-effector support, constrained to move on the horizontal plane (see Figure 1). The dominant-arm forearm was secured to the end-effector via a rigid forearm support. On the end-effector of the sensorized arm, the HandExos hand exoskeleton is mounted: it features four under-actuated finger modules (with a parallel kinematics that constrain their motion on the horizontal plane) with an independent thumb module, five linear electromagnetic actuators (Actuonix Miniature Linear Motor P16-50, Victoria, BC, Canada), a control unit, and it is mainly composed by 3D printed mechanical parts. Control electronics was embedded in the structure of the device at the hand dorsum, including amplification and digital conversion of the force sensors signals, signal processing and control on a microcontroller unit (Teensy 4.1 carrier board with ARM Cortex-M7, Sherwood, OR, USA), native Ethernet connection with the system platform, and drivers of the actuators with H-Bridge ICs (Texas Instruments DRV8833, Dallas, TX, USA). The user can wear the hand exoskeleton inserting, for each of their fingers excluding the thumb, their proximal and middle phalanxes in two C-shaped components that can be removable, thanks to its sliding insertion in the aforementioned finger kinematic structure, and also with different sizes. These mechanical user–machine interfaces grant the orthogonal transmission of forces from the human to the hand exoskeleton and vice versa.

2.3. Hand Exoskeleton Low-Level Control Architecture

The low-level loop is a proportional position controller: the error $e=P_{\mathrm{ref}}-P_{\mathrm{curr}}$ is amplified by a gain $K_p$ to form a drive command $V_{\mathrm{ref}}$ (see Figure 2). A static nonlinearity models actuator dead-zone and saturation, and the command is applied to an H-bridge PWM stage that powers a linear actuator. The actuator–mechanism is represented as two cascaded integrators. Position feedback $P_{\mathrm{curr}}$ is measured by a linear potentiometer mounted on the actuator, scaled by a gain k and low-pass filtered to suppress sensor and PWM ripple before being subtracted from $P_{\mathrm{ref}}$. The chosen filtering and control loop adds a non-relevant delay to the actuation system, being quantified overall to be 30–50 ms. Let $P_{\mathrm{ref}}\left(t\right)$ be the position setpoint and $P\left(t\right)$ the measured position. The feedback is low-pass filtered and scaled (filter $F\left(s\right)$ with impulse response $f\left(t\right)$, scalar k). The drive-to-position dynamics of the H-bridge–actuator–mechanism are denoted $G\left(s\right)$ (e.g., two integrators with gain). Linearizing the static nonlinearity after the proportional stage, the closed-loop transfer is

$${\displaystyle \frac{P\left(s\right)}{P_{\mathrm{ref}}\left(s\right)}}={\displaystyle \frac{K_{p}G\left(s\right)}{1+K_{p}G\left(s\right)kF\left(s\right)}}.$$

The actual command sent to the drive includes the dead-zone/saturation map $\varphi (·)$ and reads

$$V_{\mathrm{ref}}\left(t\right)=\varphi \left(K_p\left[P_{\mathrm{ref}}\left(t\right)-k\left(fP\right)\left(t\right)\right]\right),$$

where $K_p$ is the proportional gain, $fP$ denotes convolution, and $V_{\mathrm{ref}}\left(t\right)$ is applied to the PWM H-bridge. This cascaded structure provides closed-loop position tracking with bounded velocity and current via the drive saturation.

Setup Sensors

Three bottle locations were defined on the workspace: one central position, one at $-{35}^{\circ }$ (right of center), and one at $+{35}^{\circ }$ (left of center), all at the same radial distance from the subject. Surface EMG (sEMG) was acquired at 2222 Hz using a Delsys wireless system equipped with Quattro and Avanti sensors (Delsys Inc., Boston, MA, USA) and Ag/Cl electrodes. Twelve muscles were monitored: anterior (AD), middle (MD), and posterior deltoids (PD); pectoralis major (PM); infraspinatus (INF); teres major (TM); biceps brachii (short head, BS); triceps brachii (long head, TL); brachioradialis (BR); extensor digitorum (ED); flexor digitorum (FD); and abductor pollicis brevis (APB). A representation of the electrodes placement is shown in Figure 3. Joint angles and the end-effector position of the 3-DoF passive arm were sampled at 100 Hz using RLS Orbis absolute encoders mounted at each joint.

2.4. Offline Experimental Protocol

Participants were seated on a chair positioned in front of the HandExo rehabilitation platform. After electrode placement, each participant performed maximum voluntary contraction (MVC) trials used for subsequent normalization of EMG amplitudes: these trials include flexion/extension of the shoulder, internal/external rotation of the shoulder, flexion/extension of the elbow, and open/close of fingers. Participants then donned the wrist splint and were guided to a standardized initial posture: the wrist center located 70 cm from the central bottle, with the elbow configured so that the central bottle could be reached at full elbow extension (see Figure 1). For each of the three bottle locations (center, $-{35}^{\circ }$, $+{35}^{\circ }$), participants completed 7 reach-to-grasp-and-return trials, starting from the initial position, reaching to grasp the bottle, and returning to the initial position. In total, each participant performed $3\times 7=21$ trials.

2.5. Data Acquisition and Synchronization

EMG and kinematic data streams were synchronized and temporally aligned offline to a common time base. All data processing and modeling were conducted in MATLAB R2023b (MathWorks, Natick, MA, USA). Raw EMG signals underwent the following steps: (i) normalization to each muscle’s MVC, (ii) full-wave rectification, (iii) band-pass filtering in the 3–250 Hz range, and (iv) envelope extraction via low-pass filtering at 4 Hz. Unless otherwise specified in subsequent sections, all muscles followed the same preprocessing pipeline. Joint angles from each of the three revolute joints and the end-effector Cartesian position of the passive arm were sampled at 100 Hz from the RLS Orbis absolute encoders. No additional filtering was applied beyond that required for synchronization and model input formatting.

2.6. Prediction Target and Modeling Overview

The modeling objective was to predict the (preprocessed, MVC-normalized) amplitudes of ED and FD during reach-to-grasp movements. To emulate assistive use, predictions are intended to drive a hand exoskeleton during reach-to-grasp tasks within a serious-game environment that mirrors the physical setup (virtual bottle positions in the game replicate the three recorded physical locations i.e., center, $-{35}^{\circ }$, $+{35}^{\circ }$). We trained three algorithmic families:

1.

Linear regression (LR),

2.

Nonlinear regression using a feedforward neural network (NN),

3.

Long short-term memory network (LSTM).

Each algorithm was trained under three input conditions:

1.

EMG-only,

2.

Kinematics-only,

3.

EMG+Kinematics.

This yielded a total of $3\times 3=9$ models per participant. The EMG input set consisted of the electromyographic signals from all channels except the target ED and FD channels, resulting in 10 EMG features. The Kinematics input set included 10 kinematic channels describing the movement trajectories: end-effector position and velocity (4 channels) and joint angles and velocity (6 channels, having 3 DoF on the passive arm). The combined EMG+Kinematics input set was formed by concatenating the 10 EMG features (excluding channels 6 and 7) with the 10 kinematic features, yielding a total of 20 input features. In all cases, the models were trained to predict the EMG activity of channels 6 and 7. Model hyperparameters and feature construction are detailed in the next subsection. For each participant and each bottle location, the 7 trials were split into 3 for training, 2 for validation, and 2 for testing. Splits were performed at the trial level to prevent temporal leakage between sets. Unless otherwise noted, models were trained and evaluated in a within-subject manner using these subject-specific splits.

2.7. Architectures and Settings

In this work, three classes of models were compared for the estimation task: linear regression, nonlinear regression, and recurrent neural networks of the Long Short-Term Memory (LSTM) type. These models have been chosen to preliminary investigate the feasibility of the aforementioned rehabilitation approach, selecting the most common ones in literature, in the EMG signals estimation tasks, with increasing complexity. Thus, the comparison of the algorithms is either not intended to be extensive or to aim at finding the best model for the proposed task, but rather to give to the reader an objective evaluation of the performance of the synergistic go-to-grasp rehabilitation approach, on which specific future optimization can be run.

Three regression models were trained to estimate electromyographic activity in channels 6 and 7 from different sets of input features. The first model was a linear regression with ridge regularization, implemented as two independent least-squares regressors (one for each channel). The second model was a non-linear regression approach based on gradient boosting of regression trees (LSBoost), which aggregates the predictions of multiple shallow regression trees to capture non-linear input–output relationships. The third model was a recurrent neural network using a long short-term memory (LSTM) architecture to model the temporal dynamics of the signals. The LSTM network consisted of a sequence input layer with dimensionality equal to the number of features, a first LSTM hidden layer with 100 units, a dropout layer with 20% probability, a second LSTM hidden layer with 50 units, a fully connected layer with two output units corresponding to the target channels, and a regression layer with a mean squared error loss function. The network was trained with the Adam optimizer for a maximum of 100 epochs using a mini-batch size of 8, a learning rate of ${10}^{-3}$, and gradient thresholding set to 1. Validation data were used to monitor performance and prevent overfitting. These specific settings have been set to build a moderate-capacity, regularized sequence model appropriate for smooth EMG envelopes: two LSTM layers provide enough capacity to model temporal dependencies without moving into an over-parameterized regime; dropout equals to 0.2 is a standard regularizer for small datasets, while mini-batch size increases the number of weight updates per epoch when training data are limited. Finally the gradient thresholding set to 1 is included specifically to prevent rare exploding-gradient events in recurrent optimization, which is a common setting in the state of the art.

2.8. Performance Metrics

Given the time-continuous nature of the target signals, models accuracy was quantified using the root mean squared error (RMSE) between the predicted and measured EMG envelopes of FD and ED. For a given muscle m and trial with N time samples, let $y_m\left(t\right)$ denote the measured (MVC-normalized, enveloped) EMG and ${\widehat{y}}_m\left(t\right)$ the model estimate. The RMSE is

$${\mathrm{RMSE}}_m=\sqrt{{\displaystyle \frac{1}{N}}\sum _{t=1}^N{\left({\widehat{y}}_m\left(t\right)-y_m\left(t\right)\right)}^2}.$$

(1)

We report ${\mathrm{RMSE}}_{\mathrm{FD}}$ (flexor digitorum) and ${\mathrm{RMSE}}_{\mathrm{ED}}$ (extensor digitorum) separately; summary values within a condition are obtained by averaging across trials (and, where indicated, across the two target muscles).

2.9. Statistical Analysis

To assess differences in estimation accuracy, we compared the root mean squared error (RMSE) across factors Model (Linear, Nonlinear, LSTM) and Input type (EMG, Kinematics, EMG+Kinematics). Because of the limited sample size and potential violations of normality, we employed a non-parametric two-way analysis based on the Scheirer–Ray–Hare test, a rank-based extension of the Kruskal–Wallis test suitable for factorial designs. RMSE values were first converted to ranks, and the sums of ranks were partitioned into main effects for Model, Input type, and their interaction. For each term we computed the corresponding H statistic, degrees of freedom, and p-value using the chi-square approximation. To quantify effect sizes in a way robust to ties, we reported rank-based eta squared (${\eta }_H^2$) and epsilon squared (${ϵ}^2$). These measures are defined as the proportion of variance in ranks attributable to each factor, with ${ϵ}^2$ providing a less biased estimate. Interaction effects were carefully checked: in cases where cross-cell rank dispersion was essentially zero (for example, nearly identical medians across cells), the H statistic approached zero and effect sizes were reported as zero rather than undefined.

2.10. Muscle Synergy Analysis and Rationale

Because the controller is intended to operate in synergy with arm transport during reach-to-grasp, we verified that replacing the two missing channels (ED, FD) with model predictions does not alter the subject’s original low-dimensional muscle coordination. To this end, we compared muscle synergies extracted from the original EMG against those from a hybrid EMG in which channels 6–7 were replaced by the model estimates. Preserving synergy structure supports the physiological plausibility of the predictions and their suitability for synergy-informed assistance. For each participant, we concatenated the EMG envelopes from the test set only (all test trials and repetitions) into a nonnegative matrix $X\in {\mathbb{R}}_{\ge 0}^{M\times T}$ with M = 12 channels and T time samples at 2222 Hz. Two versions were formed: (i) $X_{\mathrm{real}}$: original EMG (all real channels), and (ii) $X_{\mathrm{hyb}}$: the same matrix where channels 6 and 7 (ED and FD) were replaced by the corresponding model predictions (per model and input condition). All signals were those used for evaluation (MVC-normalized, rectified, 3–200 Hz band-pass, and 4 Hz envelope). Muscle synergies were computed by non-negative matrix factorization (NMF), seeking

$$X\approx WH,W\in {\mathbb{R}}_{\ge 0}^{M\times r},H\in {\mathbb{R}}_{\ge 0}^{r\times T},$$

(2)

where r is the number of synergies, W contains time-invariant synergy weights, and H their time courses. For each $r\in \{1,\dots ,8\}$ we ran NMF with 20 random initializations (MATLAB R2023b, nnmf, multiplicative updates, max 500 iterations) and retained the factorization with the smallest reconstruction error. We computed the global variance accounted for (VAF) as

$$\mathrm{VAF}\left(r\right)=1-{\displaystyle \frac{{\parallel X-{\widehat{X}}_r\parallel }_F^2}{{\parallel X\parallel }_F^2}},{\widehat{X}}_r=W_r{H}_r,$$

(3)

and selected the dimensionality by the 90% rule,

$$r^{★}=min\{r:\mathrm{VAF}\left(r\right)\ge 0.90\}.$$

Dimensionality was determined separately for $X_{\mathrm{real}}$ and $X_{\mathrm{hyb}}$. To assess preservation, we (i) compared the VAF curves ${\mathrm{VAF}}_{\mathrm{real}}\left(r\right)$ vs. ${\mathrm{VAF}}_{\mathrm{hyb}}\left(r\right)$ and (ii) verified whether $r_{\mathrm{hyb}}^{★}=r_{\mathrm{real}}^{★}$. Because NMF solutions are column-permuted, we matched synergies between $W_{\mathrm{real}}$ and $W_{\mathrm{hyb}}$ using a permutation that maximized overall column-wise similarity. For similarity we used cosine similarity:

$$s(W_i,{\widehat{w}}_j)={\displaystyle \frac{w_i^{\top }{\widehat{w}}_j}{{\parallel w_i\parallel }_2{\parallel {\widehat{w}}_j\parallel }_2}},w_i,{\widehat{w}}_j\in {\mathbb{R}}_{\ge 0}^M.$$

(4)

The optimal one-to-one pairing was found by maximizing the cumulative cosine similarity. We report per-subject mean cosine similarity $\overline{s}$ across the $r^{★}$ matched pairs, and aggregate $\overline{s}$ across subjects, models, and input conditions.

2.11. Online Test of the Hand Exoskeleton: High-Level Control Strategy

The control strategy can be designed trying to mimic fingers extensor and flexor antagonists muscles behaviour. Considering a typical position control (a control output between 0 and 1, where 0 stands for fully-closed hand exoskeleton kinematics, and 1 stands for fully-open hand exoskeleton kinematics) the control strategy can be designed as follows: when both are at rest, the fingers closing amount is at mid-way (i.e., $0.5$), when flexors are active close the hand ($0.5-k_{f}u\left(1\right))$ and when extensors are active open the hand ($0.5+k_{e}u\left(2\right)$). Finally the reference position for the control law can be written as follows:

$$P_{ref}=0.5+k_e{e}_{est}-k_f{f}_{est}$$

(5)

where $e_{est}$ and $f_{est}$ are low-bounded to zero and $k_e$ and $k_f$ are tuned considering the maximum amplitude of ED and FD in healthy subjects, so that $k_e{e}_{est}<=0.5$ and $k_f{f}_{est}<=0.5$. The $k_e$ and $k_f$ coefficients are subject-specific and depend on the normalized values of EMG signals of ED and FD. They can be re-tuned for lower residual activations for post-stroke patients. Basically, the control strategy can be imagined to be in such a way that each finger is pulled by two antagonist springs, one pulling towards the full opening (i.e., the extensor) and one pulling towards the full closing (i.e., the flexor). To summarize, the model predicts the extensor and flexor muscle activity, which are $e_{est}$ and $f_{est}$ respectively, and it feeds the computation of the reference position for each of the linear actuator of the hand exoskeleton, i.e., $P_{ref}$, that is eventually used in the control scheme.

Online Experimental Protocol

After implementing the control law on the hand exoskeleton (see Section 2.11), 3 out of 8 subjects repeated the go-to-grasp experiments while wearing the exoskeleton: in this test, the subject was instructed to simply get close to the bottle and passively follow the assistance of the hand exoskeleton during pre-shaping and closing actions. In an actual setup with post-stroke patients, EMG placement timing cannot be ignored, thus, in testing the online control not only performance but also usability has been considered. For this reason, to simplify the setup, to make implementation simpler and given the extracted performance of the offline tests (see Figure 4), only the LINEAR model with kinematics inputs was used. Right after the end of the offline experiment and consequent training of the model on the subject’s data, coefficients and bias of the linear model were used in the online tests. Before testing the online control, an estimation of the prediction delay, in the case of the LINEAR model with KIN inputs, has been performed, in order to ensure that the actuation can be delivered on time to the user while approaching the object to grasp. The delay has computed first detecting the index where a 10% variation of the initial value of the channel occurs both on the measured signal and the prediction and then calculating their difference, dividing it by the sampling frequency.

3. Results

3.1. Offline Prediction Accuracy Across Models and Inputs

Across eight participants and 168 trial-level observations, descriptive RMSE for the combined target (ED+FD) showed that LINEAR and NONLINEAR models produced low errors for all inputs, while LSTM errors were larger and more input-sensitive (see Figure 4). Channel-wise patterns (reported for completeness) were approximately consistent with the combined metric and they have been reported in

Table 1. RMSE (extensor digitorum, ch. 6) across subjects: mean ± SD.

Model	EMG	EMG_KIN	KIN
LINEAR	0.0035 ± 0.0023	0.0032 ± 0.0017	0.0032 ± 0.0017
LSTM	0.0063 ± 0.0025	0.0090 ± 0.0020	0.0042 ± 0.0018
NONLINEAR	0.0043 ± 0.0027	0.0040 ± 0.0018	0.0044 ± 0.0023

and

Table 2. RMSE (flexor digitorum, ch. 7) across subjects: mean ± SD.

Model	EMG	EMG_KIN	KIN
LINEAR	0.0017 ± 0.0013	0.0017 ± 0.0013	0.0019 ± 0.0017
LSTM	0.0057 ± 0.0017	0.0103 ± 0.0080	0.0033 ± 0.0009
NONLINEAR	0.0020 ± 0.0016	0.0020 ± 0.0016	0.0026 ± 0.0024

.

A Scheirer–Ray–Hare (rank two-way) test on combined RMSE (factors: Model, Input) yielded a highly significant main effect of Model $(H=123.76,df=2,$ $p<1\times {10}^{-20},$ ${\eta }_H^2=0.287,{ϵ}^2=0.283)$, a significant main effect of Input $(H=7.39,df=2,$ $p=0.0249,{\eta }_H^2=0.017,{ϵ}^2=0.012)$, and a significant Model × Input interaction $(H=30.96,df=4,p=3.12\times {10}^{-6},{\eta }_H^2=0.072,{ϵ}^2=0.063)$. The interaction reflects the disproportionate degradation of LSTM with EMG-rich inputs (EMG, EMG_KIN) relative to KIN, whereas LINEAR and NONLINEAR varied little across inputs (see Figure 5).

3.2. Synergy Retention with Predicted Channels

Using NMF on test-set EMG and replacing channels 6–7 with model predictions to form a hybrid EMG, we compared synergy weight matrices W via cosine similarity. Across {model, input} cells (Figure 6), median similarities clustered high for LINEAR and NONLINEAR (typically 0.85–0.89) with low-to-moderate dispersion. LSTM showed input sensitivity: KIN yielded the highest similarities (mean $0.873\pm 0.023$, median $0.868$), while EMG_KIN was lowest (mean $0.825\pm 0.020$, median $0.845$), with wider interquartile ranges. Aggregated over inputs, mean similarities were LINEAR $0.874\pm 0.010$, NONLINEAR $0.876\pm 0.012$, LSTM $0.846\pm 0.013$. Aggregated over models, KIN achieved the highest mean similarity $0.869\pm 0.0117$, versus EMG $0.864\pm 0.0132$ and EMG_KIN $0.863\pm 0.0108$.

Mean VAF curves for original vs. hybrid EMG largely overlapped across the synergy range (Figure 7); the 90% VAF criterion was reached at $r\approx 6$ on average (overall $r=6.06\pm 0.75$, median 6). At the selected dimensionality r, the average VAF difference $\Delta \mathrm{VAF}={\mathrm{VAF}}_{\mathrm{hyb}}-{\mathrm{VAF}}_{\mathrm{true}}$ was small overall ($0.00456\pm 0.0159$), with model/input breakdowns indicating minor, condition-dependent shifts: LINEAR showed higher VAF in the hybrid (EMG $+0.0267\pm 0.0116$, KIN $+0.0177\pm 0.00746$, EMG_KIN $+0.0169\pm 0.00926$); NONLINEAR yielded small negative deltas (KIN $-0.00703\pm 0.00508$, EMG $-0.00819\pm 0.00347$, EMG_KIN $-0.00545\pm 0.00556$); LSTM deltas were near zero to mildly negative for KIN/EMG_KIN (KIN $-0.00589\pm 0.0150$, EMG_KIN $-0.000312\pm 0.0146$) and mildly positive for EMG ($+0.00673\pm 0.0151$). Taken together, the synergy dimensionality was preserved (no systematic shift in r), and synergy composition was largely retained, with the highest retention observed for KIN inputs and for LINEAR/NONLINEAR models.

3.3. Preliminary Real-Time Control of the Hand Exoskeleton

Before evaluating the online control performance, an estimation of the actuation delay has been computed, without including any lag deriving from the electromechanical system itself, but only considering the prediction delay. On the LINEAR model with KIN inputs, the delay has been calculated to be $279\pm 397$ ms with a median value of the distribution that includes all trials and all subjects equal to 147 ms. Furthermore, for reproducibility purposes, the values used for $k_e$ and $k_f$ are here reported, as mean and std values on the tested population: $k_e$ was set to $0.0022\pm 0.0027$ and $k_f$ was set to $0.0049\pm 0.0072$. Preliminary results on the real-time hand exoskeleton synergistic go-to-grasp control can be found in Figure 8. Differently from what expected, only on a single case out of 6 (2 channels and 3 trials) the amplitude of EMG signals were lower in the online test (i.e., wearing the hand exoskeleton) compared to the offline test. Across the three bottle-placement trials, the offline condition showed extensor digitorum RMS values between 0.0036 and 0.0055 with standard deviations between 0.0015 and 0.0043, and flexor digitorum RMS values between 0.0035 and 0.0061 with standard deviations between 0.0026 and 0.0046. In the online condition, the extensor digitorum RMS values ranged from 0.0035 to 0.0079 with standard deviations from 0.00086 to 0.0045, while the flexor digitorum RMS values ranged from 0.0093 to 0.0274 with standard deviations from 0.0073 to 0.0221.

These numerical differences indicate higher RMS amplitudes in the online condition for both muscles, with the most pronounced increase observed in the flexor digitorum channel. Concerning muscle synergies, a comparison has ben performed between the synergies extracted during the offline task (without the exo) and the online task (with the exo), on the three subjects (see Figure 9). For the bottle–left trial, the offline (NoExo) centroids showed: C1 dominated by DELT P (weight $\approx 0.9$), C2 by DELT M (≈0.8–$0.9$), C3 by DELT A (≈0.85–$0.9$), C4 with larger loading on TRI LAT and BIC SH (both ≈0.55–$0.7$, with INFRASP $\approx 0.6$), and C5 by BRACH (≈0.9). Online (Exo) retained the deltoid-dominated clusters (C1 DELT P ≈ 0.9, C2 DELT M ≈ 0.9, C3 DELT A ≈ 0.7) but showed a redistribution in the remaining modules: C4 emphasized TERES MAJ/BIC SH (≈0.7–$0.9$), whereas C5 became pectoral-dominant (PECT M ≈ 0.9). For the bottle–center trial, offline synergies were organized as C1 DELT P (≈0.9), C2 DELT M (≈0.9), C3 TRI LAT (≈0.9–$0.95$), C4 DELT A (≈0.9), and C5 BRACH (≈0.7–$0.8$); online preserved the deltoid structure (C1/C2/C3 with DELT P/DELT M/DELT A at ≈0.7–$0.9$) and increased the contribution of scapular/rotator muscles, with C4–C5 highlighting TERES MAJ/INFRASP (both approaching $\approx 0.9$). For the bottle–right trial, offline clusters again emphasized DELT P (C1 and C5, both ≈0.9–$0.95$), DELT M (C2 ≈ 0.6–$0.7$), TRI LAT (C3 ≈ 0.95), and DELT A (C4 ≈ 0.9); online retained the deltoid pattern (C1/C2/C3/C4) and exhibited stronger pectoral/teres-major loadings in C5/C4 (PECT M and TERES MAJ each ≈0.9). Overall, across targets the deltoid-dominated modules were retained from offline to online, while the remaining clusters showed moderate reweighting toward shoulder–scapular muscles (pectoral, teres major, infraspinatus) under the hand exoskeleton assistance.

4. Discussions

4.1. Models and Flexor/Extensor Activity Prediction

We compared three intention-prediction families (LINEAR, NONLINEAR, LSTM) trained on three input configurations (EMG, KIN, EMG_KIN) to estimate the EMG envelopes of ED and FD for synergy-aligned control of a hand exoskeleton during reach-to-grasp. The rationale was twofold: first, to align assistance with the natural coordination between arm transport and hand shaping described in motor-control studies of prehension and in post-stroke kinematic evidence of disrupted coupling [2,27,28,29]; second, to exploit temporal contingency between intention, movement, and sensory consequences, a known driver of plasticity in humans and a feature of assistive paradigms that show stronger gains when feedback is intention-locked [30,31,32,33,34,35].

Two consistent trends emerged. First, model choice strongly influenced performance: LINEAR and NONLINEAR produced similarly low errors across inputs, whereas LSTM errors were larger and markedly input-dependent (rank two-way on combined RMSE: Model, $H=123.76$, $df=2$, $p<{10}^{-20}$; Model × Input, $H=30.96$, $df=4$, $p=3.12\times {10}^{-6}$). Second, input type also mattered: across models, KIN yielded the lowest average RMSE, and adding EMG (EMG or EMG_KIN) did not systematically improve accuracy (Input, $H=7.39$, $df=2$, $p=0.0249$). The interaction reflects the disproportionate degradation of LSTM with EMG-rich inputs relative to KIN, whereas LINEAR and NONLINEAR varied little across inputs (see Figure 5). These findings are compatible with intention–detection literature for wearables: EMG-driven control can be responsive but is sensitive to electrode placement, skin impedance, fatigue, and calibration drift, with non-trivial setup burden and dedicated hardware [10,19,21]; kinematic signals from encoders, IMUs, or camera tracking are faster to set up and typically more stable over repeated practice [11,12,13,22]. In this light, the best predictor for clinical deployment may balance accuracy with setup time, cost, and reliability. Our results suggest that a simple linear or shallow nonlinear model on arm kinematics can achieve low errors suitable for driving open/close assistance, consistent with reports with previous works that lightweight regressors, not specifically using kinematic inputs, suffice when features are informative and stable [44,45]. Deep or recurrent architectures are prominent in EMG decoding when large training sets and high-density arrays are available [21], but performance can degrade with limited data and non-stationarity, helping to explain the LSTM pattern observed here. A plausible reason the LSTM underperformed in our setting is the mismatch between model capacity and the effective amount of training data available per condition, combined with optimization difficulties induced by long, high-resolution sequences. Although recurrent networks can, in principle, exploit temporal context, in our protocol each participant contributes only a limited number of independent trials per target to learn the mapping, while the LSTM architecture (two recurrent layers with a relatively large number of units) has enough flexibility to fit trial-specific idiosyncrasies rather than stable task-invariant structure. In small-data regimes like this, the variance of the estimator becomes dominant: the LSTM can inadvertently learn noise and non-stationarities (e.g., envelope fluctuations or trial-to-trial changes) that do not generalize to test trials, whereas simpler linear models are implicitly regularized and therefore remain more robust. In addition, the LSTM is trained on time series derived from signals acquired at very different native sampling rates (high-rate EMG then smoothed into an envelope versus lower-rate kinematics), which can yield long effective sequences once synchronized. Long sequences make recurrent optimization more sensitive to vanishing/exploding gradients; consequently, stabilizing choices such as gradient thresholding/clipping and dropout, while necessary, can bias learning toward overly smooth dynamics and reduce the ability to capture sharper transitions in muscle envelopes. In such a scenario, the LSTM may converge to conservative, smoothed predictions that appear qualitatively reasonable but incur higher pointwise error, particularly when the relationship between inputs and outputs is largely instantaneous (or adequately captured by shallow nonlinearities), limiting the advantage of a memory-based model. In any case, the goal of this study was not to perform an optimization of hyperparameters of the LSTM, but rather investigate and compare plausible different models, and to test the one with a good trade-off between complexity of models, setup simplicity, and predicting performance.

Existing control strategies for hand exoskeletons include EMG triggers or pattern recognition, kinematic thresholds, mirrored or bilateral strategies, and BCI-driven intents [11,12,13,15,22,25]. Most of the hand exoskeletons, both rigid and soft, found in literature employ an EMG-driven control, relying on residual muscle activity at a distal level, making the actuation of the robotic device not possible in case of severe impairments or high spasticity. Of course the availability of hand muscles activity grant higher motion prediction accuracy and suitable actuation delays, but it requires bulky, uncomfortable, and time-consuming EMG setups. BCI–robot or BCI–FES paradigms produce clinically meaningful gains versus control conditions in randomized trials and meta-analyses [23,34]; mirror or bilateral schemes modulate cortical activation and can be implemented with simple sensors [24]. Across the forearm/hand digitorum literature, EMG estimation has been tackled with markedly different inputs and modeling assumptions: for isometric gripping, Mogk and Keir derived regression equations that predict normalized forearm activation levels (including extensor digitorum communis) primarily from grip force and wrist/forearm posture, reporting RMSEmodel approximately 8.9–11% MVE (with EDC around 11% MVE) and variance explained that was higher for flexors (>70%) than extensors (up to 60%), without any explicit treatment of prediction delay, synergy constraints, or reliable prediction error metric in non-maximal voluntary contraction [46]. Furthermore, the usage of either grip force or wrist posture is not a feasible rehabilitation approach for a patient with a severe lesion on the distal segment. In contrast, synergy-based approaches treat EMG structure itself as low-dimensional: Ajiboye and Weir used NMF-derived muscle synergies learned from a subset of static ASL hand postures and showed that as few as 11 postures could define an 8-dimensional synergy framework achieving $>=$90% prediction (variance explained) of EMG patterns across 33 held postures, explicitly preserving synergy structure [47]. Extending synergies to dynamic/non-isometric hand function, Cole and Ajiboye extracted time-invariant synergies from static and dynamic datasets and evaluated reconstruction by variance accounted for (VAF), finding that three functional synergies achieved $66.0\pm 4.9$% VAF (vs $42.5\pm 4.4$% chance), again inherently synergy-preserving [48]. None of these works performed an actual real-time control and latency estimation of the actuation. Furthermore, the real-time usage of muscle synergies assume that EMG setup are included in the protocol which is not practical in daily clinical scenarios. Musculoskeletal inverse-dynamics studies instead estimate model-based muscle activity/activation proxies from measured kinematics: Melzner/Engelhardt et al. drove a detailed hand model with task kinematics from seven dynamic hand movements and evaluated agreement to EMG via timing-based metrics—computing on-/off-set time differences (mean 0.58 s) and weighted kappa—thereby providing an explicit timing discrepancy measure that functions as a practical “delay” indicator between simulated and measured activity [49]. Nevertheless, this approach assume that hand motion is preserved, since motion capture data is used in the inverse-dynamics based model, which makes it unsuitable for post-stroke rehabilitation scenarios. Standard pipelines typically trigger hand assistance from local hand or wrist events or mirrored commands without explicitly synchronizing open/close with ongoing shoulder–elbow transport. The present strategy contributes by leveraging proximal kinematics to time hand assistance within a reach-to-grasp phase structure, a design aligned with the neurophysiology of prehension and contingency principles [2,27,28,29,30], and related to hybrid arm–hand systems that assist transport and grasp together [26]. The proposed model achieved 30% RMSE on average compared to the typical contraction level during pre-shaping extension and prehension (corresponding to less than 1% of the MVC). To the best of authors’ knowledge, there are no studies in literature that performs EMG estimation of flexor and extensor digitorum during dynamic tasks.

4.2. Muscle Synergies Retention Analysis on the Offline Test

We examined whether replacing the two target channels with model estimates preserves each subject’s low-dimensional coordination. Cosine similarity between synergy weights (W) from original versus hybrid EMG remained high: for LINEAR and NONLINEAR, median values across inputs typically fell in the $0.83$–$0.91$ range. LSTM was more input-sensitive, with KIN yielding the highest similarity (median $\approx 0.91$) and EMG_KIN among the lowest (medians $\approx 0.83$–$0.85$), accompanied by wider dispersion. Thus, even when pointwise estimation errors differ, the predicted channels tend to integrate coherently with the subject’s synergy structure, particularly when only kinematics are used as inputs. This coherence accords with reports that upper-limb activations during reach-to-grasp can be captured by a few task-related synergies coupling proximal and distal groups [26,27].

VAF curves for original and hybrid EMG overlapped and crossed the 90% threshold at approximately six synergies on average, with negligible differences in VAF at the selected dimensionality. The low-dimensional control space therefore was neither inflated nor collapsed by channel replacement, an important property for synergy-informed assistance and for interpretable monitoring of recovery. Together with known intersegmental interactions (heteronymous reflex links and posture-dependent changes of corticospinal excitability) [36,37,38,39,40,41,42], these findings support the feasibility of hybrid synergies in which hand channels are inferred from proximal context without corrupting global coordination. In practice, this enables two complementary selection criteria for an online controller: choose the model with the lowest RMSE, or, when preserving natural coordination is prioritized (for training transfer or decoding robustness), choose the model and input that best retain synergy structure even if pointwise error is slightly higher. In our data, KIN-based inputs offered an attractive trade-off on both fronts.

For a synergy-aligned hand exoskeleton designed for frequent outpatient use, a pragmatic recipe follows from these results. First, prioritize proximal kinematics (encoders or IMUs, or camera-based tracking) for minimal setup burden and stable signals [11,12,13,22], adding EMG only when a clear incremental benefit is demonstrable for a given user and task [10,19,20,21]. Second, begin with linear or shallow nonlinear regression and escalate complexity only if data volume and stability justify it. Third, gate open/close by transport phase estimates (for example, onset of deceleration) to respect prehension timing [27,28] and to maximize intention–feedback contingency, a factor linked to stronger plastic changes [33,34,35,36,37,38]. Finally, monitor not only task scores and RMSE but also synergy changes (for example, cosine similarity of W, VAF and the selected number of synergies) as a proxy of physiological plausibility during training.

4.3. Real-Time Hand Exoskeleton Go-to-Grasp Control

The kinematics-driven linear model ran stably in real time across all targets (left/center/right), and the clustered synergy maps indicate that assistance did not disrupt the proximal coordination underpinning reach-to-grasp. Using the elbow of the k-means sum-of-distances curve, the slope flattened at about five clusters, adopted as the common dimensionality for comparison. With that choice, offline (NoExo) and online (Exo) centroids showed closely corresponding patterns across trials, deltoid groupings and pectoralis–biceps–triceps couplings remained salient, and brachioradialis and scapular rotators loaded similarly in homologous clusters, despite exclusion of the two hand-muscle channels (see Figure 9). This qualitative preservation tracks well with prior reports that upper-limb robotic assistance tends to exert limited effects on synergy structure in healthy subjects and during aided reaching, with most spatial weights conserved and only modest adjustments in specific muscles or amplitudes [50,51,52]. Similar “module preservation with modulation” has also been observed in lower-limb robot-aided locomotion, supporting the broader view that exoskeletons often leave the fundamental modular organization intact while altering timing or gain [53].

In contrast, the RMS boxplots for flexor and extensor digitorum during the same online sessions showed higher medians and markedly wider dispersion than offline for all targets, with several large outliers. A likely contributor is timing misalignment between the healthy user’s internally learned reach-to-grasp synergy, a well-consolidated cortico-spinal pattern, and the position-controlled exoskeleton. If a mismatch happens between what the subject would do and what the exoskeleton actually does, the subject basically opens or closes the hand against the exoskeleton (equivalent to an isometric contraction); this causes an increase of the EMG activity of the aforementioned muscles: even small early/late triggers can make the user generate forces that are not perfectly aligned with the imposed motion, promoting co-contraction and elevated EMG. Consistent with this interpretation, the estimated controller features a prediction delay for the LINEAR+KIN model, computed prior to online testing (see Section 2.11). Delays of this magnitude can span a meaningful fraction of the transport–hand-shaping sequence, particularly near closure onset, increasing the risk of user–robot opposition and explaining the broader RMS dispersion. Mitigations follow directly: reduce algorithmic latency (shorter windows, lighter filtering, pipelined processing), incorporate explicit movement-phase estimation to anticipate open/close timing, and add online adaptation of assistance gain and compliance (e.g., impedance control) to absorb small timing errors. A brief per-subject calibration of timing and gain before training could further temper peaks while preserving the synergy coherence observed in the heatmaps. Nevertheless, it represents a limited criticality, since it is not likely to happen with a post-stroke patient with a low residual muscle activity, as it would be used as a fully-passive exercise.

In rehabilitation scenarios, the proposed kinematics-only intention predictor is attractive because it avoids EMG setup (skin prep, electrode placement, recalibration) and remains usable when distal EMG is weak or unreliable, common in patients with severe finger paresis or spasticity. For users who contribute little voluntary hand drive and largely allow the exoskeleton to transport the arm, opposition forces against a position-controlled device are expected to be lower, which should mitigate the timing-mismatch effects we observed in healthy subjects (elevated, more variable finger-muscle RMS). In other words, a more passive distal strategy can reduce user–robot “tug-of-war,” while proximal kinematics still provide a stable, high-SNR cue for synchronizing hand opening/closing with transport. Clinically, this could translate into simpler sessions (faster donning, fewer sensors), easier deployment in outpatient and home settings (IMUs/encoders or camera tracking suffice), and broader eligibility, including patients with minimal hand activation who cannot tolerate EMG-driven triggers. At the same time, maintaining engagement remains essential: progressive assistance, impedance/compliance tuning, and brief per-patient timing calibration can help personalize the controller so that it supports movement without overriding it, preserving the arm–hand coordination we saw in the synergy maps while limiting unintended co-contraction.

4.4. Applicability of the Proposed Method in Post-Stroke Rehabilitation Contexts

One interesting discussion point regards the applicability of the proposed methodology on the target population, i.e post-stroke patients. There are many examples in the literature showing that muscle synergies, which reflect co-contraction patterns on coordinated movements, are more likely to be disrupted after stroke [45,54,55]. This element may prevent the proposed methodology to be applied, since models cannot be trained after the stroke onset. If the coordinated pattern is disrupted on post-stroke patients but it existed prior stroke it means that it is reasonable, from a rehabilitation point of view, trying to restore it to give back to patients the previous coordinated movement. In this case one strategy could be to involve the healthy limb: in an actual rehabilitation setup, this coordinated pattern can be extracted from the healthy hand and arm and used to train a model and finally to actuate the hand exoskeleton worn on the affected hand. The approach of using the healthy limb to rehabilitate the impaired one is a well-established practice in neuro-rehabilitation, for example in mirroring setups. This approach assumes that the impaired side, which can be dominant or non-dominant, shares common patterns with the healthy one: Flindall and Gonzalez in a 2013 study [56] have shown no significant differences in kinematic measurements, such as peak velocity, movement time, and deceleration phase time, between left and right hand during pick-and-place tasks; in another study by Grosskopf et al. [57] comparable kinematic features resulted from dominant and non-dominant arms during prehension tasks. The maximum hand aperture during pre-shaping seems to be the only relevant difference, that is a tunable parameter on the presented hand exoskeleton device. These results suggest that the proposed approach could be applicable to the target patient population, although it is an hypothesis that needs to be validated in a pilot study. Nevertheless, results shown in this work mainly represent an advancement on the hand rehabilitation technology, without any clinically relevant outcome.

4.5. Limitations

This study has a couple of limitations. First, the sample size is small (N = 8), which limits generalizability; nevertheless, the between-subject standard deviations of the performance metrics were modest, suggesting relatively low variation across subjects in this dataset. Secondly, although the presented methodology was designed to be used in rehabilitation setups with post-stroke patients, the current study is a feasibility study with healthy subjects only, to be intended as a pre-clinical evaluation of the synergistic control strategy on a hand exoskeleton. Finally, the presented methodology is not suitable in those cases in which the hand opening and closing is decoupled from the arm transport action, for example in static hand operations or object manipulations. Robotic rehabilitation setups with hand exoskeletons can, thus, combine different control strategies to target goal-directed reach-to-grasp actions and manipulation ones separately in different sessions.

5. Conclusions

This study compared linear, shallow nonlinear, and recurrent predictors trained on EMG, arm kinematics, and their fusion to estimate flexor and extensor digitorum activity for synergy-aligned hand exoskeleton control during reach-to-grasp. The main findings are (i) model choice strongly affected accuracy, with linear and shallow nonlinear models yielding the lowest RMSE and LSTM showing larger, input-sensitive errors; (ii) kinematics alone provided the most favorable accuracy–setup trade-off, with the lowest average RMSE and no additional sensor burden; and (iii) in the offline analysis, substituting the two target channels with model predictions largely preserved each subject’s muscle-synergy composition and dimensionality, supporting the physiological plausibility of the inferred signals.

Beyond the offline benchmark, we demonstrated real-time feasibility: a kinematics-driven LINEAR model controlled the hand exoskeleton during go-to-grasp with three subjects and three target locations. Control was stable across trials, but the distal EMG (flexor/extensor digitorum) showed higher and more variable RMS online than offline, consistent with occasional timing mismatches between the user’s reach-to-grasp pattern and the position-controlled device. These observations motivate delay-aware scheduling of open/close events, brief per-subject timing calibration, and adaptive compliance/assistance gains to absorb small phase errors while maintaining coordination with arm transport.

From an implementation standpoint, these results continue to favor kinematics-based sensing with simple regressors as a practical baseline for intention detection when the goal is to synchronize hand opening and closing with arm transport phases. For rehabilitation scenarios, the approach is appealing in patients with limited hand mobility, where distal drive is weak and interaction forces with a position-controlled device are expected to be smaller; in such cases, the model’s reliance on proximal kinematics can still coordinate assistance while reducing interference from involuntary or fatigued finger activity.

Future steps include expanding closed-loop tests to larger healthy cohorts, then piloting on post-stroke patients, testing robustness across sessions and users, benchmarking against established controllers (e.g., EMG-threshold or mirrored control), and adding neural endpoints to probe whether synergy-aligned timing enhances training-related plasticity.