Journal of Neural Engineering

ObjectiveIntraneural nerve interfaces often operate in a monopolar configuration with a common and distant ground electrode. This configuration leads to a wide spreading of the electric field. Therefore, this approach is suboptimal for intraneural nerve interfaces when selective stimulation is required. ApproachWe designed a multilayer electrode array embedding three-dimensional concentric bipolar electrodes. First, we validated the higher stimulation selectivity of this new electrode array compared to classical monopolar stimulation using simulations. Next, we compared them in-vivo by intraneural stimulation of the rabbit optic nerve and recording evoked potentials in the primary visual cortex. Main resultsSimulations showed that three-dimensional concentric bipolar electrodes provide a high localisation of the electric field in the tissue so that electrodes are electrically independent even for high electrode density. Experiments in-vivo highlighted that this configuration leads to evoked responses with lower amplitude and more localised cortical patterns due to the fewer fibres activated by the electric stimulus in the nerve. SignificanceHighly focused electric stimulation is crucial to achieving high selectivity in fibre activation. The multilayer array embedding three-dimensional concentric bipolar electrodes improves selectivity in optic nerve stimulation. This approach is suitable for other neural applications, including bioelectronic medicine.

Brain-computer interfaces (BCIs) can restore function for individuals with neuro-logical disorders and have the potential to transform the way people interact with digital systems. However, the development of advanced BCI applications, such as fluent speech synthesis, is dependent on the underlying information transfer capacity of the physical neural interface employed. A significant barrier to progress has been the lack of standardized, application-agnostic methods for benchmarking BCI system performance prior to clinical trials. Here, we introduce SONIC, a novel preclinical benchmarking paradigm designed to evaluate the information transfer rate (ITR) of a BCI system. This paradigm treats the brain and BCI as a noisy communication channel, where information is sent into the brain via precisely controlled sensory stimuli and read out by the neural interface. We implemented this paradigm in an ovine model by presenting rapid sequences of pure tones while recording neural activity from the primary auditory cortex with the Paradromics Connexus(C) BCI, a fully implanted system utilizing high-density intracortical micro-electrode arrays with wireless power and data transmission. A convolutional neural network was used to decode tones based on neural features. Our results demonstrate an achieved ITR of over 200 bits per second (bps), which is the highest reported BCI ITR to date. For reference, this rate exceeds the linguistic information content of human speech. This ITR is achieved with a total neural interface, filtering, and data aggregation delay of 56 milliseconds. Further analysis demonstrated that ITR remains high (> 100 bps) for the lowest total delay tested (11 ms), supporting the needs of latency-sensitive applications (e.g., direct speech synthesis). This work establishes a new benchmark for BCI performance and demonstrates that the Connexus BCI possesses the bandwidth necessary to support highly advanced applications. This benchmark provides a robust framework for preclinical BCI evaluation, enabling principled system design optimization to accelerate the translation of next-generation neurotechnology.

ObjectiveThis study aims to reveal longitudinal changes in functional network connectivity within and across different brain structures near the chronically implanted microelectrode. While it is well established that the foreign-body response (FBR) contributes to the gradual decline of the signals recorded from brain implants over time, how does the FBR impact affect the functional stability of neural circuits near implanted Brain-Computer Interfaces (BCIs) remains unknown. This research aims to illuminate how the chronic FBR can alter local neural circuit function and the implications for BCI decoders. ApproachThis study utilized multisite Michigan-style microelectrodes that span all cortical layers and the hippocampal CA1 region to collect spontaneous and visually-evoked electrophysiological activity. Alterations in neuronal activity near the microelectrode were tested assessing cross-frequency synchronization of LFP and spike entrainment to LFP oscillatory activity throughout 16 weeks after microelectrode implantation. Main ResultsThe study found that cortical layer 4, the input-receiving layer, maintained activity over the implantation time. However, layers 2/3 rapidly experienced severe impairment, leading to a loss of proper intralaminar connectivity in the downstream output layers 5/6. Furthermore, the impairment of interlaminar connectivity near the microelectrode was unidirectional, showing decreased connectivity from Layers 2/3 to Layers 5/6 but not the reverse direction. In the hippocampus, CA1 neurons gradually became unable to properly entrain to the surrounding LFP oscillations. SignificanceThis study provides a detailed characterization of network connectivity dysfunction over long-term microelectrode implantation periods. This new knowledge could contribute to the development of targeted therapeutic strategies aimed at improving the health of the tissue surrounding brain implants and potentially inform engineering of adaptive decoders as the FBR progresses. Our studys understanding of the dynamic changes in the functional network over time opens the door to developing interventions for improving the long-term stability and performance of intracortical microelectrodes.

ObjectiveOver the past decade, neural electrodes have played a crucial role in bridging biological tissues with electronic and robotic devices. This study focuses on evaluating the optimal tip profile and insertion speed for effectively implanting Paradromics high-density Fine Microwire Arrays (FA) prototypes into the primary visual cortex (V1) of mice and rats, addressing the challenges associated with the "bed-of-nails" effect and tissue dimpling. ApproachTissue response was assessed by investigating the impact of electrodes on the blood-brain barrier (BBB) and cellular damage, with a specific emphasis on tailored insertion strategies to minimize tissue disruption during electrode implantation. Main ResultsElectro-sharpened arrays demonstrated a marked reduction in cellular damage within 50 m of the electrode tip compared to blunt and angled arrays. Histological analysis revealed that slow insertion speeds led to greater BBB compromise than fast and pneumatic methods. Successful single-unit recordings validated the efficacy of the optimized electro-sharpened arrays in capturing neural activity. SignificanceThese findings underscore the critical role of tailored insertion strategies in minimizing tissue damage during electrode implantation, highlighting the suitability of electro-sharpened arrays for long-term implant applications. This research contributes to a deeper understanding of the complexities associated with high-channel-count microelectrode array implantation, emphasizing the importance of meticulous assessment and optimization of key parameters for effective integration and minimal tissue disruption. By elucidating the interplay between insertion parameters and tissue response, our study lays a strong foundation for the development of advanced implantable devices with a reduction in reactive gliosis and improved performance in neural recording applications.

Speech-based brain-computer interfaces (BCIs) can offer an intuitive means of communication for those who have lost the ability to speak due to paralysis. Significant progress has been made in classifying individual words from high numbers of electrocorticographic (ECoG) electrodes on the sensorimotor cortex (SMC). As implantations of larger grids with more ECoG electrodes are associated with higher surgical risk, we here examined whether confined electrode configurations can match the classification accuracy of larger grids. To this end, we analyzed data from eight able-bodied participants with high-density ECoG grids (64 to 128 electrodes) who performed a task that involved speaking 12 Dutch words. Word pronunciation was associated with changes in high frequency band activity in two SMC foci, one in the ventral SMC and another in the dorsal SMC. Using a combinatorics approach, we found that a smaller, rectangular, configuration with a surface area of 325 mm2 to 561 mm2 (32 electrodes) could achieve a word classification accuracy similar to that of the larger grids: 76{+/-}16% versus 75{+/-}17% across participants, respectively (practical chance level 16.7%). The best configurations were oriented vertically and centered on the central sulcus. These findings indicate that a 32-electrode ECoG grid placed optimally can be sufficient for achieving high word classification accuracy on a closed set of words. We conclude that targeted placement of small ECoG grids can reduce surgical demands on end users and justify energy- and complexity-efficient designs of fully implantable BCI devices for individuals with severe paralysis.

Retinal prostheses must be able to activate cells in a selective way in order to restore high-fidelity vision. However, inadvertent activation of far-away retinal ganglion cells (RGCs) through electrical stimulation of axon bundles can produce irregular and poorly controlled percepts, limiting artificial vision. Therefore, the problem of axon bundle activation can be defined as the axonal stimulation of RGCs with unknown soma and receptive field locations, typically outside the electrode array. Here, a new algorithm is presented that utilizes electrical recordings to determine the stimulation current amplitudes above which bundle activation occurs. The method exploits several spatiotemporal characteristics of electrically-evoked spikes to overcome the challenge of detecting small axonal spikes in extracellular recordings. The algorithm was validated using large-scale ex vivo stimulation and recording experiments in macaque retina, by comparing algorithmically and manually identified bundle activation thresholds. The algorithm could be used in a closed-loop manner by a future epiretinal prosthesis to reduce poorly controlled visual percepts associated with bundle activation. The method may also be applicable to other types of retinal implants and to cortical implants. ContributionsPT developed the algorithm and analyzed the data, with input from SMi and EJC. NB and NS helped with the analysis. SMa and LG performed dissections and collected the data. PT and VFH performed manual identification. PH, AS and AML developed and supported recording hardware and software. PT, EJC and SMi wrote the manuscript. NS and SMa edited it. EJC and SMi supervised the project.

Brain-machine interfaces (BMIs) for vision restoration require models that accurately simulate the anatomy and electrical properties of visual pathways. However, current models focus only on isolated structures, such as the retina or brain, and overlook surrounding tissues. Here, we present a comprehensive computational model of the human head, incorporating the entire visual pathway--including the eye, optic nerve, and brain--along with critical neighboring tissues such as the orbit, paranasal sinuses, enabling precise simulations. Validation using human and large animal data demonstrated a strong correlation between the simulated and measured electrical potentials. Component elimination analysis revealed that the optimized comprehensive model outperformed simplified versions. The models utility was demonstrated through multiple applications: (1) comparative analysis of electrical neuromodulation technologies for optic neuropathy, revealing the filed intensity limitations of noninvasive approaches and the safety concerns of invasive intraorbital approach; (2) identification of optimal stimulation site, revealing that transnasal stimulation at the optic chiasm outperformed traditional approaches; and (3) in silico design of electrode arrays for optic nerve prosthetics, demonstrating theoretical advantages in invasiveness and visual field coverage compared to existing retinal and cortical prosthetics. This validated and versatile computational resource supports the development of neuromodulation strategies and visual BMI technologies.

ObjectiveUtah arrays are widely used in both humans and non-human primates (NHPs) for intracortical brain-computer interfaces (BCIs), primarily for detecting electrical signals from cortical tissue to decode motor commands. Recently, these arrays have also been applied to deliver electrical stimulation aimed at restoring sensory functions. A key challenge limiting their longevity is the micromotion between the array and cortical tissue, which may induce mechanical strain in surrounding tissue and contribute to performance decline. This strain, due to mechanical mismatch, can exacerbate glial scarring around the implant, reducing the efficacy of Utah arrays in recording neuronal activity and delivering electrical stimulation. ApproachTo investigate this, we employed a finite element model (FEM) to predict tissue strains resulting from micromotion. Main ResultsOur findings indicated that strain profiles around edge and corner electrodes were greater than those around interior shanks, affecting both maximum and average strains within 50 {micro}m of the electrode tip. We then correlated these predicted tissue strains with in-vivo electrode performance metrics. We found negative correlations between 1 kHz impedance and tissue strains in human motor arrays and NHP area V4 arrays at 1-mo, 1-yr, and 2-yrs post-implantation. In human motor arrays, the peak-to-peak waveform voltage (PTPV) and signal-to-noise ratio (SNR) of spontaneous activity were also negatively correlated with strain. Conversely, we observed a positive correlation between the evoked SNR of multi-unit activity and strain in NHP area V4 arrays. SignificanceThis study establishes a spatial dependence of electrode performance in Utah arrays that correlates with tissue strain.

ObjectiveThis study investigates speech decoding from neural signals captured by intracranial electrodes. Most prior works can only work with electrodes on a 2D grid (i.e., Electrocorticographic or ECoG array) and data from a single patient. We aim to design a deep-learning model architecture that can accommodate both surface (ECoG) and depth (stereotactic EEG or sEEG) electrodes. The architecture should allow training on data from multiple participants with large variability in electrode placements and the trained model should perform well on participants unseen during training. ApproachWe propose a novel transformer-based model architecture named SwinTW that can work with arbitrarily positioned electrodes by leveraging their 3D locations on the cortex rather than their positions on a 2D grid. We train subject-specific models using data from a single participant and multi-patient models exploiting data from multiple participants. Main ResultsThe subject-specific models using only low-density 8x8 ECoG data achieved high decoding Pearson Correlation Coefficient with ground truth spectrogram (PCC=0.817), over N=43 participants, outperforming our prior convolutional ResNet model and the 3D Swin transformer model. Incorporating additional strip, depth, and grid electrodes available in each participant (N=39) led to further improvement (PCC=0.838). For participants with only sEEG electrodes (N=9), subject-specific models still enjoy comparable performance with an average PCC=0.798. The multi-subject models achieved high performance on unseen participants, with an average PCC=0.765 in leave-one-out cross-validation. SignificanceThe proposed SwinTW decoder enables future speech neuropros-theses to utilize any electrode placement that is clinically optimal or feasible for a particular participant, including using only depth electrodes, which are more routinely implanted in chronic neurosurgical procedures. Importantly, the generalizability of the multi-patient models suggests that such a model can be applied to new patients that do not have paired acoustic and neural data, providing an advance in neuroprostheses for people with speech disability, where acoustic-neural training data is not feasible.

Visual cortical prostheses offer a promising path to sight restoration, but current systems elicit crude, variable percepts and rely on manual electrode-by-electrode calibration that does not scale. This work introduces an automated data-driven neural control method for a visual neuroprosthesis using a deep learning framework to generate optimal multi-electrode stimulation patterns that evoke targeted neural responses. Using a 96-channel Utah electrode array implanted in the occipital cortex of a blind participant, we trained a deep neural network to predict single-trial evoked responses. The network was used in two complementary control strategies: a learned inverse network for real-time stimulation synthesis and a gradient-based optimizer for precise targeting of desired neural responses. Both approaches significantly outperformed conventional methods in controlling neural activity, required lower stimulation currents, and adapted stimulation parameters to resting state data, reliably evoking more stable percepts. Crucially, recorded neural responses better predicted perceptual outcomes than stimulation parameters alone, underscoring the value of our neural population control framework. This work demonstrates the feasibility of data-driven neural control in a human implant and offers a foundation for next-generation, model-driven neuroprosthetic systems, capable of enhancing sensory restoration across a range of clinical applications.

BackgroundPeripheral nerve stimulation (PNS) is most effective when specific nerve fiber subpopulations are activated, while minimizing off-target activation, which may cause undesirable side effects. This selectivity depends primarily on electrode design and charge delivery. We hypothesized that selective PNS could be achieved through electrode placement and intrafascicular electric field steering using Longitudinal Intrafascicular Electrodes (LIFEs). MethodsLIFEs were implanted into the tibial fascicle of the sciatic nerve of 17 anesthetized adult rats. We tested whether electrodes positioned at different cross-sectional and longitudinal locations within the same fascicle, together with different electric field-steering approaches produced distinct activation patterns in the gastrocnemius lateralis muscle. Muscle responses were measured using high-density epimysial electromyography (HD-eEMG). ResultsElectrodes placed at different locations within the same fascicle activated distinct muscle regions, demonstrating intrafascicular selectivity. Bipolar stimulation recruited nerve fibers differently than monopolar stimulation, showing that electric field steering can further shape the selective recruitment. In both configurations, increasing the stimulation amplitude produced a graded increase in muscle activation. Furthermore, our findings demonstrated that HD-eEMG is an effective tool for evaluating intrafascicular selectivity. ConclusionThese findings suggest that improving on-target selectivity may support next-generation bioelectronic therapies with better outcomes and fewer side effects, potentially enabling more precise, organ-specific neuromodulation. Using multiple intrafascicular electrodes may provide two complementary strategies for enhancing selectivity: strategic intrafascicular placement to access different fiber subpopulations and bipolar configurations to steer recruitment beyond what a single electrode can achieve.

Recent advances in brain-machine interfaces (BMIs) using neural network decoders and increased channel count have improved the restoration of speech and motor function, but at the cost of higher power consumption. For wireless, implantable BMIs to be clinically viable, power consumption must be limited to prevent thermal tissue damage and enable long use without frequent charging. Here, we show how neural network "pruning" creates sparse decoders that require fewer computations and active channels for reduced power consumption. Across multiple movement decoding tasks using brain and muscle signals, recurrent neural network decoders can be compressed by over 100x while maintaining strong performance, enabling decoding on the implant with <1% power increase compared to decoding externally. Pruning also allows for deactivating up to 89% of channels, reducing BMI power by up to 5x. Counterintuitively, our findings suggest that BMIs employing a subset of a larger number of channels may achieve lower power consumption than BMIs with fewer channels, for a given performance level. These results suggest a path toward power-efficient, implantable BMIs suitable for long-term clinical use.

Brain-machine interfaces have shown promise in extracting upper extremity movement intention from the thoughts of nonhuman primates and people with tetraplegia. Attempts to restore a users own hand and arm function have employed functional electrical stimulation (FES), but most work has restored discrete grasps. Little is known about how well FES can control continuous finger movements. Here, we use a low-power brain-controlled functional electrical stimulation (BCFES) system to restore continuous volitional control of finger positions to a monkey with a temporarily paralyzed hand. In a one-dimensional, continuous, finger-related target acquisition task, the monkey improved his success rate to 83% (1.5s median acquisition time) when using the BCFES system during temporary paralysis from 8.8% (9.5s median acquisition, equivalent to chance) when attempting to use his temporarily paralyzed hand. With two monkeys under general anesthesia, we found FES alone could control the monkeys fingers to rapidly reach targets in a median 1.1s but caused oscillation about the target. Finally, when attempting to perform a virtual two-finger continuous target acquisition task in brain-control mode following temporary hand paralysis, we found performance could be completely recovered by executing recalibrated feedback-intention training one time following temporary paralysis. These results suggest that BCFES can restore continuous finger function during temporary paralysis using existing low-power technologies and brain-control may not be the limiting performance factor in a BCFES neuroprosthesis.

BackgroundIntracortical electrical stimulation has emerged as a promising approach for sensory restoration, such as a cortical visual prosthesis, yet its effectiveness is limited by current spread and electrode density constraints. ObjectiveTo determine whether intracortical bipolar current steering--via modulation of the return electrode position--can enhance neural activation selectivity compared to traditional monopolar stimulation, with the aim of improving spatial precision in sensory restoration. MethodsWe applied intracortical stimulation and used two-photon calcium imaging on acute brain slices to directly visualize neural responses to bipolar stimulation. Biophysical computational modeling was used to complement the experimental results. The analysis included both cellular and population-level assessments to evaluate the impact of several stimulation patterns, such as current direction, electrode spacing and current amplitude, on recruitment patterns. ResultsBipolar stimulation selectively activated distinct neural populations based on the direction of the current flow. This approach decreased the overlap between activated groups and increased the number of independently addressable neural clusters by up to 9-fold relative to monopolar stimulation. Moreover, the electrode configuration and spacing critically influenced the spatial spread of activation. ConclusionsIntracortical bipolar current steering enhances neural activation selectivity by engaging independent neural populations through current directionality. These findings suggest that this strategy may improve the spatial precision of neural prosthetics and sensory restoration without the need for an increased electrode density.

To provide appropriate levels of stimulation, retinal prostheses must be calibrated to an individuals perceptual thresholds ( system fitting). Nonfunctional electrodes may then be deactivated to reduce power consumption and improve visual outcomes. However, thresholds vary drastically not just across electrodes but also over time, thus calling for a more flexible electrode deactivation strategy. Here we present an explainable artificial intelligence (XAI) model fit on a large longitudinal dataset that can 1) predict at which point in time the manufacturer chose to deactivate an electrode as a function of routine clinical measures ( predictors) and 2) reveal which of these predictors were most important. The model predicted electrode deactivation from clinical data with 60.8% accuracy. Performance increased to 75.3% with system fitting data, and to 84% when thresholds from follow-up examinations were available. The model further identified subject age and time since blindness onset as important predictors of electrode deactivation. An accurate XAI model of electrode deactivation that relies on routine clinical measures may benefit both the retinal implant and wider neuroprosthetics communities.

Chronically implanted brain-computer interfaces (BCIs) provide amazing opportunities to those living with disability and for the treatment of chronic disorders of the nervous system. However, this potential has yet to be fully realized in part due to the lack of stability in measured signals over time. Signal disruption stems from multiple sources including mechanical failure of the interface, changes in neuron health, and glial encapsulation of the electrodes that alter the impedance. In this study we present an algorithmic solution to the problem of long-term signal disruption in chronically implanted neural interfaces. Our approach utilizes a generative adversarial network (GAN), based on the original Unsupervised Image to Image Translation (UNIT) algorithm, which learns how to recover degraded signals back to their analogous non-disrupted ("clean") exemplars measured at the time of implant. We demonstrate that this approach can reliably recover simulated signals in two types of commonly used neural interfaces: multi-electrode arrays (MEA), and electrocorticography (ECoG). To test the accuracy of signal recovery we employ a common BCI paradigm wherein a classification algorithm (neural decoder) is trained on the starting (non-disrupted) set of signals. Performance of the decoder demonstrates expected failure over time as the signal disruption accumulates. In simulated MEA experiments, our approach recovers decoder accuracy to >90% when as many as 13/ 32 channels are lost, or as many as 28/32 channels have their neural responses altered. In simulated ECoG experiments, our approach shows stabilization of the neural decoder indefinitely with decoder accuracies >95% over simulated lifetimes of over 1 year. Our results suggest that these types of neural networks can provide a useful tool to improve the long-term utility of chronically implanted neural interfaces.

Detecting attempted movements of a paralyzed limb is a key step for neural interfaces for motor rehabilitation and restoration after a stroke. In this paper, we present a systematic evaluation of electroencephalographic (EEG) and electromyographic (EMG) activity to decode when stroke patients with severe upper-limb paralysis attempt to move their affected arm. EEG and EMG recordings of 35 chronic stroke patients were analyzed. We trained classifiers to discriminate between rest and movement attempt states relying on brain, muscle, or both types of features combined. Our results reveal that: i) EEG and residual EMG features provide complementary information to detect attempted movements, obtaining significantly higher decoding accuracy when both sources of activity are combined; ii) EMG-based, but not EEG-based, decoding accuracy correlates with the degrees of impairment of the patient; and iii) the percentage of patients that achieve decoding accuracy above the chance level strongly depends on the type of features considered, and can be as low as 50% of them if only ipsilesional EEG is used. These results offer new perspectives to develop improved neurotechnologies that establish a more accurate contingent link between the central and peripheral nervous system after a stroke, leveraging Hebbian learning and facilitating functional plasticity and recovery.

Intracortical brain-computer interfaces (BCIs) that decode complex movements, such as handwriting and speech, can require substantial training data to achieve high performance. We investigated whether leveraging the neural activity recordings of previous users could reduce this initial data collection burden for new BCI users (an approach we call "cross-brain transfer"). Using intracortical recordings from five BrainGate2 clinical trial participants, we tested cross-brain transfer for both speech and handwriting neural decoders trained and evaluated on general, unconstrained corpora of spoken and written English. We found that cross-brain transfer improved decoding performance when training data from the target user was limited (< 200 sentences), and that dataset-specific input layers to the decoder were critical for combining data across users. Without trainable input layers, transfer failed and performed worse than training from scratch on target user data only. Finally, we measured the effectiveness of cross-brain transfer relative to training with (1) more data from the same user and (2) more electrode-permuted data from the same user, which simulates sampling from another brain with identical neural latent structure. In some cases (T16 speech, T12 handwriting), cross-brain transfer appeared as effective as additional permuted data from the same user, while in others (T12 speech, T15 speech) electrode-permuted data was more beneficial. Our results successfully demonstrate and characterize cross-brain transfer learning between multiple intracortical BCI users, for both speech and handwriting, using a general open-ended dataset not restricted to small sets of words or phrases. This work highlights a promising path towards addressing a key barrier to the clinical translation of BCIs, while clarifying when cross-brain transfer may be most beneficial and the decoder design choices needed to realize those gains.

Practical application of brain-computer interfaces (BCIs) requires stable mapping between neuronal activity and behavior through various behavioral contexts and for different individuals. Due to neural activity instability, BCIs require frequent recalibration to maintain robust performance. Early approaches to addressing BCI stability issues mainly focused on tackling the challenge of neural activity changes over time. However, future BCI applications involve diversified scenarios and subjects, requiring solutions that address neural variability across time, subjects, and tasks. This study proposes a meta-learning-based algorithm to achieve BCI stability, named "Meta-AlignNN." By capitalizing on the consistency of neural population dynamics, it provides a unified solution for maintaining BCI stability, robustness, and scalability across subjects, time, and tasks. Tested over two years on four tasks with three monkeys, as well as on public datasets, the approach has achieved significantly excellent performance in both offline decoding and real-time brain-control, outperforming existing methods. Our findings provides a foundation for meeting the clinical demands of longterm, efficient, and stable usability across patients and tasks, offering a compelling solution for practical BCI applications.

Reconstructing speech envelopes from electroen-cephalography (EEG) signals is a challenging but valuable task for brain-computer interfaces (BCIs), with applications in assistive communication for individuals with speech impairments. While deep learning has improved reconstruction accuracy, most existing approaches are restricted to single-layer architectures such as convolutional neural networks (CNNs). This limits their ability to capture the full complexity of spatio-temporal and structural EEG patterns. In this work, we systematically extend the VLAAI framework by evaluating 26 architectures that integrate CNNs, long short-term memory networks (LSTMs), and graph convolutional networks (GCNs) in both single-layer and hybrid configurations. Experiments on the 64-channel Spar-rKULee dataset demonstrate that CNNs remain the strongest standalone models, but hybrid designs--particularly CNN-LSTM and CNN-GCN-LSTM--achieve competitive or superior performance. These results highlight the importance of combining spatial, temporal, and graph-based processing, and provide practical guidelines for hybrid architecture design. Our study offers the first large-scale comparative analysis of hybrid models for EEG-based speech envelope reconstruction, advancing robust BCI systems for non-invasive speech decoding.