IEEE Transactions on Neural Systems and Rehabilitation Engineering

People with lower limb amputation lack somatosensory feedback from their prosthesis, and this loss contributes to functional deficits, including balance and gait impairments. Recent advances in neuroprostheses have demonstrated that electrical stimulation of sensory nerves in the residual limb and spinal cord can restore lost sensations in the lower limb. To maximize the efficacy of these somatosensory neuroprostheses, the restored sensations should be intuitive, seamlessly integrating into the sensorimotor network. However, it is challenging to quantify the intuitiveness of these evoked sensations. Recent studies have proposed using crossmodal congruency effect (CCE) tasks for this purpose in people with upper-limb amputation. The current study tests the feasibility of the CCE task for assessing the intuitiveness of sensory feedback in the lower limb. We hypothesized that CCE score would reliably differentiate between a more natural (pneumatic) sensation and a less natural (electric) sensation at two locations: the knee and the foot. Across fifteen able-bodied individuals, we observed that the CCE task differentiates sensory modalities at the knee, but not at the foot. Identification of external factors affecting the CCE is needed before it can be implemented to measure intuitiveness of sensory feedback in lower-limb amputees.

Non-invasive human-machine interfaces (HMIs) are critical in developing prosthetic systems that offer intuitive, simultaneous, and proportional control over multiple degrees of freedom (DOFs). This study introduces a novel system for intuitive concurrent control of hand and wrist movements using sonomyography based imaging of muscle activity. Our method uses a sparse set of ultrasound scanlines to reduce computational complexity while enhancing usability. We evaluated four regression techniques for wrist and hand angle prediction, focusing on performance with a reduced sonomyographic feature set. We also explored the feasibility of a sonomyography-based system by simulating various factors that could affect prediction, including feature selection and scanline count. Our findings demonstrate that Gaussian process regression excels in predicting wrist and hand angles with just eight equispaced transducers in offline settings. Real-time evaluations with 10 non-disabled participants showed a 93 % success rate for two-DOF tasks using linear regression. The system was tested with an individual with amputation, achieving a 46 % success rate for two-DOF control in a 2D space, even though the ground truth data for model training was collected from the contralateral limb. This study validates our sonomyography-based approach for accurate wrist and hand angle estimation, reducing complexity and demonstrating potential in real-world scenarios.

Age-related decline in postural control is an important factor that increases the fall risk of older adults. Fingertip vibrotactile stimulation has been developed to provide light touch-like somatosensory input. However, evidence regarding differences among older age groups is limited. This study examined the effects of fingertip vibrotactile stimulation on postural control in 348 community-dwelling older adults classified as young-old (age 65-74 years), old-old (age 75-84 years), and oldest-old (age 85 years or older). Participants stood with eyes closed and feet together under stimulation and no stimulation conditions. The center of pressure (COP) velocity and COP area were measured using a force plate. The natural log-transformed COP area was used for the analysis. Linear mixed models were used to examine the effects of age group, stimulation conditions, and measurement segments. The COP velocity under the stimulation condition was significantly lower than that under the no stimulation condition; however, the COP area did not change significantly. Significant main effects of age group were observed for both COP indices, but no interaction between age group and stimulation condition was observed. Fingertip vibrotactile stimulation may reduce the COP velocity across older age groups, thus reflecting the effects on postural adjustment frequency.

ObjectiveElectroencephalography (EEG) is a key tool for studying auditory processing in cochlear implant (CI) users. In particular, EEG recordings obtained during continuous speech are becoming increasingly important for assessing speech and language processing in CI users, and may be utilized for neurofeedback. However, CIs also induce strong stimulation artifacts that are time-locked to the stimulus and mask the neural responses that have smaller magnitudes. Existing artifact reduction methods are typically based on event-related potentials (ERPs) or require manual component selection, making them unsuitable for naturalistic listening conditions or large datasets. ApproachWe develop CORSICA (CORrelation-baSed ICA artifact rejection), a reproducible, parameter-efficient method for CI artifact reduction in EEG responses to continuous speech. CORSICA operates on independent components (ICs) obtained through Infomax ICA and requires no manual component labelling, with performance governed by a single tunable threshold. It exploits the observation that CI artifacts temporally follow the audio signal without delay, whereas neural responses have an inherent lag due to auditory pathway latencies. For each IC, CORSICA computes the cross-correlation with the speech stimulus. Artifacts are identified by a high signal-to-noise ratio (SNR) of the correlation peak near zero lag, and the component is rejected if this SNR exceeds a threshold. To benchmark CORSICA, we evaluate two alternatives: a TRF-based SNR method, in which temporal response functions are fitted to each IC and artifact-driven peaks near zero lag are used for rejection, and a variant replacing ICA with second-order blind identification (SOBI) as the source separation step. Main resultsCORSICA effectively suppressed CI artifacts while preserving neural activity, enabling recovery of physiologically plausible TRFs with only 2% of ICs rejected. Both benchmark methods confirmed the validity of the SNR-based rejection framework, but CORSICA outperformed the TRF-based alternative in artifact suppression quality. Replacing ICA with SOBI as the source separation step required more ICs to be rejected, further supporting ICA as the preferred backbone for CORSICA. SignificanceCORSICA provides a fully objective, label-free approach to identifying CI artifacts in speech-evoked EEG data, with no manual intervention required. By centering artifact rejection on a single interpretable threshold, it offers a reproducible preprocessing standard for future EEG studies on speech processing in CI users. ConclusionOur findings demonstrate that objective CI artifact suppression in speech-evoked EEG data is feasible on the basis of the ICs temporal response patterns.

Background: To examine the association between ultrasound-based muscle activity detection, or sonomyography (SMG) derived metrics, and clinical measures of upper-extremity function in individuals with cervical spinal cord injury (cSCI), and to evaluate SMG-based trajectories derived directly from muscle activity as a muscle-level assessment compared to conventional kinematic approaches. Methods: Eight individuals with cSCI (n = 8; American Spinal Injury Association Impairment Scale grades A to C; injury levels C5 to C6) participated. Participants performed a wrist tenodesis based target achievement task while SMG data were collected. SMG derived metrics were correlated with performance-based upper extremity function assessed using the Jebsen Taylor Hand Function Test (JTHFT) and self-reported function assessed using the Capabilities of Upper Extremity Questionnaire (CUE-Q). Associations were quantified using distance correlation (dCorr). Results: Strong associations between SMG-derived metrics and clinical measures were observed. Movement Arrest Period Ratio (MAPR) showed the strongest association with JTHFT performance (dCorr = 0.75), while Time to Peak Velocity (TTPV) demonstrated a moderate association (dCorr = 0.62). Rate of Change of Acceleration (ROCAcc) showed a strong correlation with CUE-Q scores (dCorr {approx} 0.70), and spectral arc length (SAL) showed moderate correlations (dCorr {approx} 0.66). Conclusions: SMG-derived metrics show meaningful associations with both performance-based and self-reported measures of upper-extremity function in individuals with cSCI. These findings suggest that SMG metrics can serve as objective tools to complement clinical assessments for tracking functional status and recovery. Larger studies are needed to confirm these observations.

Surface electromyography (sEMG) is a robust non-invasive modality for human-machine interaction, yet its application remains largely limited to coarse motor tasks such as grasping or rotation. The decoding of fine motor skills, specifically handwriting, remains a challenging problem with potential relevance for prosthetic control and natural communication interfaces. In this work, we explore a Transformer-based alternative to classical signal-processing pipelines that treats multi-channel sEMG signals as complex time series. We introduce DualMyo, a specialized model integrating Patch Embeddings and Rotary Positional Embeddings (RoPE) to capture the intricate spatio-temporal dynamics of myoelectric activity. Our experimental results show strong intra-session performance. Furthermore, we address the inherent challenges of signal drift and sensor displacement in cross-session applications. We show that a lightweight fine-tuning strategy of 10 epochs enables DualMyo to effectively adapt to session variability, achieving approximately 91\% accuracy with two examples per digit. These findings provide a promising step toward adaptive sEMG-based handwriting interfaces, although further validation is required for real-time and multi-subject deployment and neuromuscular control.

Physiological mirror activity (pMA) is the increase in involuntary muscle activity observed on the contralateral side during unilateral voluntary movement in neurologically healthy participants. This cross-sectional study aimed to explore the relationship between pMA and corticomuscular coherence (CMC) during finger dexterity tasks in young and older adults. Thirty-one right-handed young adults and 24 older adults performed a left-hand finger dexterity task. Electroencephalogram (EEG) signals were recorded from C3 and C4, and electromyogram (EMG) signals were collected from bilateral finger flexors and extensors. pMA was quantified as the change in right-hand EMG from rest to task. Gamma-band CMC was calculated from task-related EEG-EMG pairs, and its association with pMA was analyzed. In young adults, greater pMA was associated with lower CMC (C3- and C4-right flexors), whereas in older adults, greater pMA was associated with higher CMC (C3-left flexor). Young adults may suppress pMA emergence by appropriately monitoring and inhibiting activity, in the hand not performing the task. Conversely, in older adults, the mobilization of the ipsilateral motor cortex may have contributed to pMA emergence. This study suggests that the neuromuscular mechanisms involved in pMA during finger dexterity tasks differ between young and older adults.

Static force plates (FP) are the gold standard for measuring ground reaction forces (GRF) and computing joint moments through inverse dynamics in gait analysis. However, they are restricted to controlled environments, and the number of steps analyzed is limited by the plates embedded in the floor. To address these limitations, portable solutions such as sensorized insoles, socks, or shoes have emerged. Yet, creating wearable systems capable of measuring three-dimensional GRF in real-world conditions remains challenging. Current sensorized shoes often incorporate thick sensors (up to 2 cm), reducing usability and limiting their application in pathological populations or dynamic tasks like running. This study evaluates the usability of ShokacShoes, a novel sensorized shoe integrating three thin, three-dimensional force sensors, and explores its potential as a Wearable Force Plate (WFP). Eight healthy participants performed slow, natural, and fast walking using two insole configurations. Force and temporal metrics were derived from WFP and FP data. Results indicate that WFP enables accurate step segmentation and detects significant effects of speed and insole type on temporal and force metrics, confirming its reliability under different walking conditions. Comparisons with FP revealed differences in force metrics and signal morphology, though temporal parameters remained consistent. These results are likely due to sensor quantity and positioning. Thereby, ShokacShoes represent a valid solution capable of measuring three-dimensional forces within commercial footwear. Future work will focus on validating the applicability of a new version of ShokacShoes against gold-standard FP in a comprehensive validation study involving diverse real-world scenarios and pathological conditions.

Back-ground. Beneficial rehabilitation interventions for severely impaired stroke patients are limited. Owing to practical constraints in the routine clinical use of electroencephalogram (EEG)-based brain-computer interfaces, this study investigates the feasibility of using a more practical electromyography (EMG) to detect movement intention in severe stroke subjects without visible movement. Currently, no large-scale studies provide strong evidence in favour of EMG-based human-machine interaction for closed-loop control of robotic assistance for severe stroke. Objective. To screen severely impaired stroke subjects without active wrist extension for the presence of residual EMG activity. Methods. High-density surface EMG was recorded from the wrist extensor muscles of 100 severely impaired stroke survivors while they repeatedly attempted wrist extension. EMG activity during "Rest" and "Move" states was compared, and subjects showing statistically greater muscle activity during Move than Rest were classified as having residual EMG. A modified Hodges detector combined with the probability difference-sum ratio (PDSR) was used for classification, with a threshold of 0.73 identifying subjects with residual EMG. Results. Of the 100 subjects without active wrist extension (Muscle power: MRC < 2), 64 exhibited residual EMG activity, supporting the feasibility of EMG for movement intention detection. Among these, 35 demonstrated consistent muscle activity (Detection probability > 0.2); representing suitable candidates for EMG-driven robot-assisted therapy. Conclusions. A substantial proportion of severely impaired stroke subjects without active movement could benefit from a simpler EMG-based interface for robot-assisted therapy. Distinct neural mechanisms (intact voluntary drive or abnormal co-activation) may contribute to the residual muscle activity, which should be considered while designing control strategies.

Clinical gait retraining typically relies on multi-sensor arrays and high-dimensional feedback displays, imposing setup and interpretation burdens that limit routine clinical deployment. We developed a single-IMU visual biofeedback system that delivers real-time feedback of Lower Limb Trajectory Error (LLTE), a composite kinematic error metric integrating knee position and shank angle across the stance phase. Twenty able-bodied adults walked on a treadmill under two visual biofeedback targets (flexed-knee, extended-knee) while receiving either corrected (n=10) or uncorrected (n=8) feedback, where the correction accounted for limb orientation at initial contact. LLTE and stance-phase knee kinematics adapted consistently under the flexed-knee target for both feedback groups, with feedback formulation moderating the temporal trajectory of change. Adaptation toward the extended-knee target was limited, likely because participants were already operating near terminal knee extension and because the scalar error metric provided limited directional information for correction. Ankle range of motion (ROM) changed significantly across the stance phase under both target conditions, while hip ROM did not. Multiscale multivariate sample entropy (MSMVSE) increased monotonically with time scale across all conditions, with no statistically distinguishable difference between corrected and uncorrected feedback. These results suggest that single-IMU LLTE biofeedback can modify gait mechanics and that adaptation was expressed across multiple lower-limb segments rather than through changes at a single joint.

Combining complementary neurophysiological modalities offers a promising strategy for improving motor imagery (MI) brain-computer interfaces (BCIs), but learning shared representations across modalities remains largely unexplored. Here, we propose a two-phase deep learning framework for multimodal EEG-MEG decoding that explicitly decouples representation learning from downstream classification. In the first phase, a convolutional encoder-decoder learns a shared latent representation by predicting the power spectral density (PSD) of EEG and MEG signals directly from time-domain activity, rather than using the conventional objective of reconstructing the input signal. In the second phase, the encoder is frozen and its learned representations are reused, without further adaptation, to perform the classification of the downstream MI-BCI task. The framework was evaluated on simultaneous EEG and MEG recordings from 20 participants. The learned representations consistently outperformed conventional handcrafted spectral features, increasing median classification accuracy from 0.734 to 0.794. The multimodal framework also improved performance over MEG alone (median accuracy from 0.680 to 0.794) and yielded a modest increase over EEG alone (median accuracy from 0.765 to 0.794), providing a more robust decoding strategy than single-modality approaches. Furthermore, the learned latent representations were transferable across participants, with more than half of cross-subject models performing within 0.01 accuracy of their subject-specific counterparts. These findings demonstrate that task-agnostic representation learning can capture physiologically meaningful multimodal neural representations that remain transferable across individuals, offering a promising foundation for more robust and reusable BCI pipelines.

This paper presents a two-stage pipeline for implicit feature engineering in time series-based physiological stress detection using electrodermal activity (EDA) signals. In the first stage, we forecast three descriptive statistics of future EDA signals over short horizons (3, 5, and 10 seconds) based on a 60-second context window. In the second stage, a lightweight linear classifier detects stress from these predicted statistics. We evaluate three forecasting architectures spanning the domain expertise spectrum: a domain-specific bidirectional long short-term memory (BiLSTM) recurrent neural network, zero-shot and fine-tuned variants of Amazon Chronos T5 time series foundation model, and the Tabular Prior-data Fitted Network (TabPFN) applied to engineered physiological features. Experiments on the publicly available Wearable Stress and Affect Detection (WESAD) dataset, comprising chest-worn multimodal physiological signals from 15 subjects under baseline and stress conditions, demonstrate that the domain-specific BiLSTM achieves the highest classification performance, with area under the receiver operating characteristic curve (AUC) values ranging from 0.913 to 0.962. TabPFN follows with AUC values of 0.853-0.869, while Chronos variants yield 0.528-0.744. Notably, models using predicted features consistently outperform those using oracle features derived from the true future signals--the theoretical upper bound--suggesting effective noise filtering through learned sequence representations. Chronos models quickly reach performance saturation regardless of training depth, highlighting challenges in tokenizing continuous physiological time series. The proposed approach advances implicit feature engineering for wearable stress monitoring by leveraging forecasting as a powerful inductive bias, thereby improving robustness and providing insights into the limitations of the foundation model for physiological signals.

Breast cancer survivors frequently experience upper-limb impairments, making continuous monitoring essential for effective rehabilitation. We propose REINA (Recognize-Then-Infer Wearable-to-App AI Framework), a two-stage deep-learning approach for remote monitoring of motor function during breast cancer rehabilitation using wearable-device data. Inertial measurement unit (IMU) signals from wearable devices are first used to recognize physical activities via supervised learning, followed by an activity-specific recurrent neural network (RNN) to infer corresponding electromyography (EMG) signals. REINA establishes reliable inference of neuromuscular activity from wearable IMU data, enabling real-time, cost-effective assessment of motor function recovery in real-world settings.

Surface electromyography (sEMG) offers a silent and wearable input modality, but its practical use is limited by variability across users and recording sessions. This study presents a compact CNN- Transformer model for decoding isolated handwritten digits from eight-channel sEMG signals. The model combines trainable signal preprocessing, convolutional feature extraction, and Transformerbased temporal modeling. It was evaluated on ten recordings from five participants using recordingseen classification, leave-one-recording-out (LORO) generalization, and few-shot adaptation. The model achieved a mean macro F1 score of 0.924 {+/-} 0.059 in the recording-seen setting and 0.619 {+/-} 0.252 under zero-shot LORO evaluation. Adaptation using two labeled trials per digit increased macro F1 to 0.828 {+/-} 0.112, while ten trials per digit achieved 0.925 {+/-} 0.053. The proposed architecture also outperformed classical and neural baselines in the controlled LORO benchmark. These results indicate that compact CNN-Transformer models, combined with lightweight target-recording calibration, provide a promising basis for adaptive sEMG-based input systems.

Balance is fundamental to mobility and independence, yet it remains difficult to measure in a way that is both clinically meaningful and feasible for widespread adoption. There is a need for more quantitative, objective, and accessible balance assessments. To this end, we created the Stanford Balance Level Evaluation (StaBLE), a battery of 18 balance-challenging tasks and accompanying digital performance metrics computed from video that aimed to capture widely ranging balance abilities across 180 participants. We validated the StaBLE score against existing clinical scales and found that it correlated with age ({rho} = -0.76, p<0.001), the Activities-Specific Balance Confidence Scale ({rho} = 0.60, p<0.001), a Mini-Balance Evaluation Systems Test task ({tau}=0.55, p<0.001), 4-Stage Balance Test ({tau} = 0.55, p<0.001), and Short Physical Performance Battery ({tau} =0.57, p<0.001). StaBLE overcame ceiling effects in these existing scales and had a different underlying distribution for those identified at risk of falling and not (KS: 0.67, p<0.001). To reduce the time it takes to perform the test, we identified subsets of balance tasks that best predicted the StaBLE score across different populations. Together, our dataset, balance assessment protocol, and performance-based score demonstrate the validity of digital measures of balance.

The primary goal of a brain-computer interface (BCI) is to enable interactions with the environment by generating control signals directly from brain activity, independent of physical movement including gaze shifts. Attentional selection of peripheral objects has been used to control a BCI but requires sufficient visual acuity to reliably recognize the target stimulus. Here, we systematically investigated the effect of different stages of visual impairment on event-related potentials (ERPs) and BCI performance while healthy participants directed covert attention to peripheral visual stimuli. We hypothesized that the effect of visual impairments such as blurred or monocular vision might alter attention-related visual ERPs, but has only minor or no impact at all on BCI performance when color is the target feature. Early attention sensitive visual components were altered in monocular vision compared to binocular vision. Simulated vision degradation resulted in reduction of ERPs in the P300 range. The differences in ERPs across vision conditions did not affect the decoding accuracy of the BCI. The study not only revealed differences of visual ERPs in different viewing conditions but primarily demonstrated that a BCI can be reliably controlled by focusing attention on a visual stimulus whose distinguishing features such as color, can be recognized even with visual impairments. This provides a promising approach to enable paralyzed persons to communicate using a non-invasive, gaze-independent BCI.

Traditional clinical gait assessments focus on lower-limb kinematics and overall walking speed, often overlooking the upper-limb dynamics and inter-limb coordination that matter for real-world ambulation. Measuring these movements outside the laboratory is difficult, and this study presents and validates a wearable sensor algorithm to quantify arm swing kinematics and arm-leg coordination (Phase Locking Value, PLV) during overground walking in the home. Validated against optical motion capture, the algorithm detected swing events reliably and with negligible temporal bias. In home-based gait recordings from nearly 1500 community-dwelling older adults, arm-leg coordination was the strongest arm swing predictor of rhythmic gait stability once walking speed was accounted for. Exploratory factor analysis separated upper-limb function into distinct "coordination and stability'" and "pace and capacity'' axes, identifying arm swing as an independent dimension of the gait profile. Analysis of dynamic resilience during turns showed that frail older adults have slower recovery of arm-leg coordination, pointing to a loss of motor automaticity. Across a 30-year age span, arm swing amplitude declined with age while arm-leg coordination did not change detectably. In a separate laboratory cohort, coordination was also reduced in Parkinson's disease, indicating that the measure responds to neurological impairment as well as to frailty. This algorithm offers a scalable way to assess upper-limb gait dynamics in daily life. Shifting the clinical focus from walking speed alone to full-body movement may help detect instability early.

Pathological tremor is a neurological condition that impairs fine motor tasks, affecting 1% of the general population and 4% of the elderly. Tremors arise when muscles micro-oscillations synchronize and phase lock, typically within a 4-12 Hz frequency range. Administering beta-blockers can reduce tremor severity, but doses are hard to personalize, with heavy doses of propranolol correlating with low blood pressure, dizziness, and nausea. In this project, we aimed to model tremor and create a closed-loop control framework to suppress tremor amplitude while minimizing pharmacological dependence. Because side effects constrain the use of pharmacological suppression alone, we investigated noninvasive neuromodulation. We used vibrotactile stimulation (VTS) to disrupt pathological tremor synchronization and reduce oscillatory amplitude. We hypothesized that tremor suppression involving VTS followed a nonmonotonic relationship, tested by determining whether maximum relief requires an adaptable framework. The procedure consisted of constructing a propranolol-reduction simulation by implementing a Hill curve, where we calculated and utilized tremor reduction, heart rate (HR) drop, and blood pressure (BP) drop. We then built a device to capture tremor-related data and create vibration using two linear resonant actuator (LRA) coin motors. We connected it to a microcontroller, where we determined optimal vibration frequencies through a feedback loop. Across 50 trials, VTS alone reduced tremor amplitude by an average of 37.3%, reducing the propranolol dose needed to reach 50% total tremor reduction by 71.9%, lowering the modeled blood pressure drop from 38.1 to 18.9 mmHg. This device demonstrates proof-of-concept for a nonmonotonic tremor-vibration relationship to reduce dependency on propranolol in the treatment of pathological tremor. These propranolol dose-reduction estimates are derived from computational simulation and have not been clinically validated; they are not intended as a recommendation to alter prescribed medication.

Beta oscillations (13-30 Hz) are a prominent sensorimotor neural rhythm and an important biomarker in neurorehabilitation. These oscillations propagate along the corticospinal pathway and are expressed in the discharge patterns of spinal motor neurons, enabling the assessment of corticomuscular coupling. Moreover, peripheral beta band oscillations have recently emerged as a potential control signal for motor augmentation interfaces. However, it remains unclear whether peripheral beta activity simply reflects cortical oscillations or is partly shaped by peripheral mechanisms, and to what extent it can be voluntarily controlled. To address these questions, we developed a 10-day neurofeedback protocol in which participants learned to up-regulate peripheral beta band activity. Subjects were trained to exploit movement cancellation, a behaviour naturally associated with increased cortical and muscle beta band activity, as a two-state strategy to voluntarily modulate peripheral beta band power. Each session included a guided familiarization phase based on a GO/NO-GO task, in which participants familiarized with movement cancellation through guided visual cues, followed by an asynchronous control phase in which they self-initiated the same strategy without external guidance to increase peripheral beta band activity in a target window. Participants progressively improved their ability to voluntarily modulate peripheral beta band activity. Peripheral beta band power during movement cancellation increased significantly across training days in both the familiarization and asynchronous control phases. Intramuscular coherence in the beta band also increased, indicating enhanced common synaptic input to the motor neuron pool in this band. In contrast, cortical beta power and corticomuscular coherence remained unchanged. Together, these findings demonstrate that peripheral beta band activity is a dynamic neural feature that can be voluntarily shaped through training, supporting its potential as a non-invasive control signal for future neurorehabilitation and motor augmentation technologies.

EEG microstates are a distinct number of quasi-stable spatial distributions of brain activity. Microstate trajectories are strongly suspected to reflect the underlying neural mechanisms during information processing and are therefore also called the "building blocks" of human thought. In this study, we examined, if EEG microstate sequences can serve as potential triggers for a Brain-Computer Interface (BCI). To this end, a semi-supervised deep learning model architecture consisting of an LSTM-based autoencoder and a dense neural network was utilized to classify between left- and right-hand motor imagery EEG data, with the resulting classification output serving as the BCI trigger. On the one hand, this was done in a 2-step approach, in which the autoencoder and classifer have been trained separately. On the other hand, an end-to-end approach was employed, where training was performed by combining reconstruction and classification losses. Results show that the proposed model architecture was able to extract relevant features from microstate sequences and exploit them for within subjects and sessions classification. Applying transfer learning to session-to-session or across-subject transfer resulted in peak classification accuracies around 89%. We also investigated to what extent transfer learning has to be applied to reach considerable classification accuracies serving as the calibration time representative. We found that on average around 400s are needed for BCI calibration when emplyoing our approach to reach 80% classification accuracy. The present study signifies that the investigation of EEG microstate trajectories can be a promising approach for extracting BCI triggers, as it reduces the dimensionality of multi-channel recorded EEG signals to a distinct number of brain states over time. Deep learning methods, especially transfer learning, applied to EEG microstate trajectories seem promising regarding user-convenient and calibration-free BCIs in real-world applications.