Intracortical Microstimulation in Brain–Computer Interfaces: Evoking Perception and Plasticity Authors: Pengfei Hu, Chong Chen, Yunliang Zang, Xiaohong Li, Dong Ming Published: 2026 ABSTRACT Intracortical microstimulation (ICMS), the activation of specific neuronal populations via microelectrodes implanted in the cortex, has emerged as a key technology for invasive brain–computer interfaces (BCIs). This article reviews the technical foundations and functional applications of ICMS within the BCI field, highlighting its diverse capabilities in both perception construction and targeted neuromodulation, while emphasizing the interface and parameter constraints that shape its long-term use. At the technical level, we analyze the evolution of ICMS microelectrode interfaces from rigid arrays to flexible, biomimetic, and biohybrid strategies, alongside key pulse-train parameters relevant to neural recruitment and application safety. Functionally, we discuss how biomimetic and spatiotemporally patterned ICMS generates high-resolution artificial tactile and visual perception, and how ICMS can serve as learnable information channels to guide behavior. We further consider temporally contingent and closed-loop ICMS as plasticity-based approaches for modulating cortical functional connectivity and pathological network activity in selected experimental models, while noting translational challenges related to stability, scalability, safety, and patient variability. Finally, we extend the discussion to novel biohybrid neural interfaces, including cell-seeded interface modifications, axon-guidance strategies, and stem-cell- and brain-organoid-integrated platforms, which provide a theoretical reference for next-generation biointegrated neuromodulation technologies in BCIs. Together, this review evaluates ICMS in BCIs along 2 central lines: evoking artificial perception and engaging plasticity-based modulation, while highlighting that long-term translation will require coordinated advances in interface reliability, stimulation encoding, closed-loop calibration, safety evaluation, and biohybrid integration. Introduction Since the turn of the 21st century, brain–computer interfaces (BCIs) have continuously evolved, offering new hope for functional recovery in patients with neurological injuries, developmental disorders, and related conditions [1]. Neuromodulation technologies artificially activate the nervous system to inject information or modulate local cortical activity and serve as a fundamental technical pillar of BCIs [2,3]. Current neuromodulation technologies are broadly classified into noninvasive techniques, such as transcranial magnetic stimulation (TMS) [4] and transcranial direct current stimulation (tDCS) [5], and invasive techniques, such as epidural electrical stimulation [6], cortical electrical stimulation [7], intracortical microstimulation (ICMS) [8], and deep brain stimulation (DBS) [9]. Among these, ICMS has emerged as a prominent invasive modality capable of activating intracortical neuronal populations with high precision, enabling the fine manipulation of local neural circuits [10], and is developing into an effective tool for replacing or supplementing natural sensory inputs, probing and modulating dysfunctional neural circuits, and influencing behavior in experimental settings [11]. In early research, ICMS primarily served as an electrophysiological tool to map the functional topology of the cerebral cortex [12]. With the maturation of neural encoding and decoding technologies, ICMS has successfully induced stable and discriminable tactile [13] and visual [14] sensations in both animal models and human participants. Furthermore, it has been utilized as an auxiliary channel for information injection to guide the behavioral responses of animals [11]. This capability to inject information directly into the cortex makes ICMS an important foundation for sensory feedback in BCIs and indicates that its perceptual quality and behavioral utility are closely related to stimulation parameters, perceptual stability, and user-specific neural responses. In recent years, studies have also begun to examine whether ICMS can bias neural connectivity over longer time scales. Specific temporal patterns or activity-dependent ICMS sequences can induce plasticity-like changes within cortical networks in selected paradigms, suggesting that ICMS may modulate synaptic connection strength and cortical representations and may be beneficial for selected neurological diseases [15]. This direction extends ICMS from transient cortical activation toward circuit-level neuromodulation and brings timing rules, activity-specific biomarkers, closed-loop control, and the durability and safety of induced plasticity into closer focus. The above ICMS applications also depend on the stability of the implanted interface. To overcome physical limitations of conventional electrodes, including foreign body responses (FBRs), biohybrid neural interfaces (BNIs) that integrate living cellular components with electrodes provide new possibilities for improving tissue–device compatibility and supporting graft–host interactions, indicating that interface safety, controllability, and long-term functional stability require focused evaluation in future ICMS development [16]. Here, we propose an integrative framework for ICMS applications in BCIs, organized around hardware interfaces, stimulation parameters, artificial perception/information input, plasticity-based modulation, biohybrid integration, and translational challenges. This framework is not limited to a single device or application. Rather, it explains how ICMS has progressed from transient cortical activation toward parameterized stimulation design, artificial information input, closed-loop plasticity modulation, and biointegrated interface exploration, while identifying limitations in long-term stability, mechanistic interpretation, safety, and scalability for its application in BCIs (Fig. 1). Conceptual framework for ICMS in BCIs. The schematic provides an overview of the major domains covered in this review: ICMS-evoked artificial perception, the learned use and integration of ICMS patterns, plasticity-related circuit remodeling, and the evolution of rigid, flexible, and biohybrid interfaces, centered on ICMS delivery and pulse parameters. Technical Foundations: Microelectrode Interfaces and ICMS Parameters ICMS operates primarily by directly delivering electrical pulses to brain tissue via implantable microelectrode arrays (MEAs) [17]. The stimulation sequence generally consists of μA-level, biphasic, cathodic-first pulses, with a charge-balanced design to minimize tissue damage while enabling the sparse activation of nearby neurons at relatively low amplitudes [10]. To meet the demands for chronic in vivo application, invasive MEAs are currently undergoing a paradigm shift: from rigid to flexible [18], from single-function to multi-function, from standard array structures to biomimetic architectures, and ultimately toward BNIs [19]. These advancements aim to provide chronic, stable, and minimally invasive electrode interfaces (Fig. 2A). Evolution of microelectrode interfaces, FBR, and fundamental parameters of ICMS. (A) Schematic of the transition from rigid to flexible electrodes and biohybrid neural interfaces. Left to right: microwire electrode, Michigan probe, Utah array, thin-film electrode, flexible filament electrode, flexible mesh electrode, and biohybrid neural interfaces. (B) Rigid electrodes typically trigger FBR and glial scar encapsulation. During operation, ICMS activates adjacent neuronal populations via the exposed electrode contacts. (C) Typical parameters of an ICMS pulse train, including stimulation amplitude, frequency, pulse duration, interphase interval, and train duration. Microelectrode interfaces Rigid electrodes Current ICMS research predominantly employs microwire or silicon-based electrodes. Microwire electrodes consist of micrometer-diameter filaments typically fabricated from materials such as platinum, iridium, platinum/iridium alloys, gold, tungsten, or stainless steel. These filaments are coated with an insulating layer, leaving only the tips exposed as recording or stimulation sites [20]. While these electrodes feature a straightforward design and low cost, their limited coating compatibility and durability render them unsuitable for the rigorous demands of chronic, high-frequency stimulation [21]. Silicon-based MEAs emerged from the development of microelectromechanical systems (MEMS) technology and provide an alternative to microwire-based MEAs [22–24]. Structurally, these are primarily classified as Michigan probes and Utah arrays. Michigan probes generally use thin shank-like structures with contacts arranged along the electrode length, allowing recording and stimulation across cortical layers, whereas Utah arrays use needle-like 3-dimensional (3D) structures with contacts near the tips to cover broader cortical regions [25,26]. These designs achieve multicontact integration, enabling simultaneous signal recording and precise stimulation across multiple cortical regions; when combined with micrometer/nanometer-scale precision surface etching and high-charge-capacity coatings such as sputtered iridium oxide film, they can substantially enhance durability, signal fidelity, and stimulation capabilities [22]. However, metal microwires (~100 to 300 GPa) and silicon (~180 GPa) possess high Young’s moduli [18], resulting in a profound mechanical mismatch with the exceptionally soft brain tissue (~1 to 30 kPa) [27]. This discrepancy makes them prone to causing micromotion-induced tissue damage and FBR, triggering neuroinflammation and glial scar encapsulation [28] (Fig. 2B). Consequently, this hinders the establishment of stable, chronic interactions between the electrode interface and local neurons [29]. As the dominant tools for in vivo experiments, these electrodes have undergone extensive chronic performance evaluations in animal models and a subset of human patients [30,31]. Although compensatory strategies such as ultraflexible cable interconnects [32] and bioactive surface modification [33] have been proposed, such strategies can partially improve interface performance but cannot eliminate the bulk mechanical mismatch between stiff implants and soft brain tissue; the interactions between microelectrodes and brain tissue remain exceedingly complex, and the challenges surrounding their chronic applicability still demand further resolution. Flexible electrodes To meet the demands for high-quality modulation and chronic implantation, microelectrode technology is advancing toward flexible platforms. Some flexible electrodes can be discussed within a broader biomimetic electronics framework, which aims to reduce micromotion-induced mechanical mismatch and FBR through material mechanics, device geometry, and tissue-compliant design [19,34]. Structurally, current flexible electrodes generally fall into 3 categories: flexible thin-film electrodes [35], flexible filament electrodes [36,37], and flexible mesh electrodes [38]. Utilizing insulating polymers (such as polydimethylsiloxane [PDMS], Parylene C, SU-8, polyimide, and polyethylene terephthalate) as flexible substrates provides a Young’s modulus that closely matches that of brain tissue. This effectively reduces mechanical damage and neuroinflammatory responses [37,39], enhances the long-term stability of the electrodes, and may reduce glial encapsulation and improve tissue integration in selected experimental settings. To overcome the high interfacial impedance caused by electrode miniaturization, conducting polymers such as poly(3,4-ethylenedioxythiophene):polystyrene sulfonate and polypyrrole (PPy) can be introduced. These polymers enhance the electrochemical performance of the electrodes and reduce the ICMS charge injection threshold, thereby providing a favorable microenvironment for cell adhesion and synaptic growth [40]. To further improve interfacial biocompatibility, flexible electrodes can also incorporate bioactive surface modification strategies. By incorporating specific extracellular matrix proteins, such as laminin, anti-inflammatory peptides, or defined cell adhesion molecules onto flexible material surfaces or within hydrogel coatings, these strategies create a highly biomimetic local microenvironment that supports host neuronal adhesion and synaptic growth [41]. This approach can complement the mechanical advantages of flexible electrodes, but current progress should still be interpreted cautiously. Flexibility itself entails engineering constraints in chronic insulation and encapsulation stability, precise implantation, stimulation safety windows, and consistency in high-throughput manufacturing. Therefore, reduced impedance, improved histology, or enhanced cell adhesion alone is insufficient to demonstrate long-term, stable, and scalable ICMS application across conditions and species [18,42]. For flexible electrodes to become a complete solution for long-term ICMS, further clinical and functional validation in implanted systems, such as the Neuralink platform, is still needed [36]. These limitations also suggest that subsequent interface design should move beyond simple mechanical matching toward regulation of tissue–device interactions. Biohybrid neural interfaces While flexible substrates and bioactive surface modification strategies have improved electrode biocompatibility, they still cannot fully bridge the fundamental barrier between artificial devices and biological tissues [41,43]. To address this limitation, researchers are attempting to construct new BNIs based on biohybrid electronics. This concept extends bioactive surface modification by introducing living biological components into the tissue–device interface to mediate interactions between electronic devices and the nervous system and to enhance biocompatibility and tissue integration [19,44,45]. Such interfaces have been shown to effectively reduce electrode-implantation-induced FBR and promote closer integration between the electrode interface and the cerebral cortex [46]. Further, biohybrid regenerative electronics emphasizes regenerative-medicine and tissue-repair goals. BNIs constructed from biological components such as neural stem cells (NSCs) or brain organoids may allow ICMS to participate in graft development and integration, and may benefit the repair or reconstruction of injured tissue by monitoring and modulating graft function [47]. This strategy may provide new application routes for ICMS in traumatic brain injury (TBI) and other neurological diseases, but its ethical and safety constraints require further consideration [48]. The above ICMS-oriented microelectrode-interface progress has been conceptualized as an evolution from “biomimetic and bioactive electronics” to “biohybrid electronics” and ultimately to “living electronics and interfaces”, and may provide a more advanced hardware basis for chronic ICMS applications while expanding its application scope [19]. ICMS pulse-train parameters and neuronal recruitment ICMS is not simply the delivery of electrical pulses to the cortex. Electrode structure, contact location, and pulse-train parameters all influence which neural elements are recruited and what functional effects follow. Microampere-level extracellular pulses can recruit axons, somata, dendrites, and trans-synaptic pathways. Axonal activation is central to spatial spread [10,49–51]. Early studies proposed that ICMS forms an approximately spherical effective activation zone around the electrode tip and can be described by the I = kr2 current–distance relationship, where I is current, r is distance from the electrode, and k is the current–distance constant [52]. Subsequent imaging and modeling studies further indicate that ICMS is better understood as a sparse, distributed, and parameter-dependent recruitment process: increasing stimulation amplitude, pulse width, or delivered charge can enhance axonal recruitment and trans-synaptic effects, but the spatial extent of activation is also shaped by axonal orientation, cell type, and local circuit organization [10,53]. Common ICMS protocols use charge-balanced biphasic pulses, often cathodic-first, to limit net charge accumulation and tissue damage while enabling neural activation [54,55]. The key variables are stimulation amplitude, frequency, pulse duration, interphase interval, and pulse-train duration [56] (Fig. 2C). These variables jointly influence neural recruitment, perceptual effects, adaptation, and safety boundaries [10,57,58]. These effects are not simply linear: stimulation amplitude and pulse width mainly influence recruitment strength and spatial spread, whereas frequency and interpulse interval further shape short-latency excitation, prolonged inhibition, and pulse-to-pulse interactions. Therefore, the same parameters may produce different recruitment outcomes across electrode sites, cortical layers, or tissue states [59]. Thus, a single optimal ICMS parameter combination is unlikely to generalize across applications. Parameter differences reflect electrode-contact location, tissue state, stimulation target, perceptual modality, and task demands, as well as application-specific requirements for recruitment range, temporal structure, and safety boundaries [11,60]. For example, tactile studies emphasize perceived intensity, naturalness, and localization [13,61–65]; visual studies focus on phosphene location, brightness, discriminability, and shape perception [14,66,67]; learnable-information studies test whether temporal, spatial, or sensor-coupled ICMS patterns can be learned and integrated with natural sensory information [68–74]; and plasticity-induction studies rely more strongly on millisecond-scale timing between spontaneous neural activity and stimulation, with spike-triggered activity-dependent ICMS often using low-frequency, single-pulse triggering [15,75–78]. Table 1 summarizes representative studies by application category, species range, electrode type, major pulse-train parameter ranges, and experimental duration range. Table 1.Representative ICMS pulse-train parameters across major application categoriesCategorySpecies rangeElectrode typeAmplitudeFrequencyPulse widthInterphase intervalTrain durationExperimental duration rangeTactile perception [13,61–64,88,89]HumanUtah arrays2–100 μA0–300 Hz~200 μs/phase53–100 μs0.05–3 s (or real-time driven)1 d–84 mo (NR in some)Visual perception [14,66,67,91,204]Human/Nonhuman primatesMicrowire electrodes/Utah arrays1–300 μA50–300 Hz100–800 μs/phase0–100 μs0.125–3 s4–24 mo (NR in some)Learnable information input [68–74,94,95]Nonhuman primates/RodentsMicrowire electrodes/Utah arrays1–300 μA0–400 Hz or event-triggered100–200 μs/phase25–250 μs50 ms–5 s (or real-time driven)3 d–4 mo (NR in some)Plasticity modulation [15,75–78,115,117–121]Nonhuman primates/RodentsMicrowire electrodes/Michigan probes/Utah arrays20–330 μASpike-triggered single pulse (~7–19 Hz) or 300–1,000 Hz (paired stimulation)192–200 μs/phaseNRSingle-pulse stimulation or ~2.8–6.7 ms (paired stimulation)0.5 h–21 dNR, not reported; h, hours; d, days; mo, months Representative ICMS pulse-train parameters across major application categories NR, not reported; h, hours; d, days; mo, months In parameter selection, stronger, longer, or higher-frequency stimulation can enhance neural recruitment and behavioral effects, but may also increase threshold drift, tissue responses, chronic interface deterioration, neuronal loss, or myelinated fiber atrophy [79–81]. Therefore, Table 1 should be understood as summarizing representative choices rather than universal “safe ranges”. Species differences, application category, electrode interface, and experimental duration can all alter the trade-off between safety and stimulation effects, and parameter translation across studies should be calibrated with cross-species functional equivalence and computational models, with attention to the recruited cortical layers, neuronal populations, and functional outcomes [60,82]. Together, these findings suggest that future ICMS parameter optimization should move beyond empirical tuning toward task-specific, individualized, and potentially closed-loop strategies that account for circuit recruitment, perceptual stability, adaptation, and long-term tissue responses [60]. Based on diverse electrode platforms and parameter combinations, ICMS has made important progress, particularly in artificial tactile and visual perception. Artificial Perception and Neural Information Injection Initially, ICMS was primarily used to determine the functional representations of different cortical areas [12]. For example, activating the motor cortex elicits repetitive muscle movements [83], enabling the generation of precise functional cortical maps [84]. Recently, this approach has transitioned from fixed, simple stimulation sequences to high-frequency, long-duration, and complex patterns, revealing more intricate motor information encoding [85]. Furthermore, it can define effective intracortical connectivity networks via cortico-cortical evoked potentials (CCEPs) and stimulation-evoked effective connectivity (SEEC) [86,87], holding promise as a novel method for intraoperative functional cortical mapping. As research has progressed, ICMS has been increasingly applied to the development of sensory neuroprostheses in BCIs. In this context, the clinical goal of sensory restoration is essentially achieved by evoking artificial perception. Delivering ICMS to precise locations within the primary somatosensory cortex (S1) or primary visual cortex (V1) can induce diverse, transient perceptual experiences in primates and humans [13,14]. Establishing tactile or visual feedback via ICMS in BCIs may not only improve the psychological well-being of patients but also enhance motor task performance [11,88]. With advancements in neural encoding research, ICMS-evoked responses have become more defined and controllable, making ICMS-based artificial perception promising [14,61–64,66]. Next, we discuss how ICMS can evoke artificial perception and provide learnable information input, while noting limitations such as parameter dependence and perceptual stability. Evoking artificial tactile and visual perception ICMS-evoked artificial perception is not equivalent to a complete reconstruction of natural sensory activity, but is better understood as a cortical input that can be localized, modulated, and learned. Receptive fields (RFs) refer to body or skin regions where natural peripheral stimuli can evoke responses from a specific neuronal population or cortical site [12]. Based on these S1 RFs, ICMS can evoke tactile sensations (such as numbness, tingling, or pressure) in the corresponding areas [8,12], with perceptual intensity generally correlating positively with stimulation amplitude and frequency, which provides the basis for linear encoding that maps a single stimulation parameter onto perceptual intensity [13,65]. In this process, projected fields (PFs) refer to the body regions where participants report the sensation after stimulation of a specific cortical site, and usually correspond to or fall within the spatial extent of that site’s RFs. Based on this RF–PF relationship, ICMS can evoke spatially adjustable localized tactile percepts and support parameterized tactile encoding [89] (Fig. 3A). However, interactions among amplitude, frequency, electrode combination, and other parameters may be nonlinear, increasing variability in stimulation effects and reducing the feasibility of multiparameter linear encoding and parameter transfer [65,90]. Artificial perception paradigms in the primary somatosensory and visual cortices via ICMS. (A) In S1, ICMS evokes tactile perception based on the spatial correspondence between RFs and PFs. Spatiotemporally patterned ICMS coordinates stimulation amplitude and timing across channels to convey tactile edge curvature and apparent motion. (B) In V1, ICMS leverages the retinotopic organization of RFs to evoke phosphenes, constructing complex visual percepts, such as spots, shapes, and directional apparent motion, through patterned multielectrode stimulation. Overall, early S1 studies established the feasibility of localized artificial touch, whereas later human studies have shifted toward multielectrode combinations and more complex tactile encoding. Multielectrode ICMS is a spatial delivery strategy that distributes stimulation across multiple electrode contacts, which can reduce charge input at individual contacts and may improve perceptual spatial resolution or PF stability [8]. Subsequent human BCI work showed that multielectrode ICMS-evoked tactile percepts can remain relatively stable, that their intensity can be modulated by stimulation amplitude, and that they can be linked to prosthetic-hand sensor signals to provide tactile feedback in bidirectional BCIs. This feedback, however, still mainly relies on approaches such as linear amplitude modulation and provides basic tactile feedback without fully resolving perceptual naturalness [13,61,88]. Beyond spatial-distribution strategies, biomimetic ICMS places greater emphasis on stimulation timing and dynamic parameter design, aiming to approximate population-level neural responses evoked by natural sensory inputs by adjusting pulse amplitude, frequency, or spatiotemporal patterns rather than encoding a single sensory variable with a single parameter [8]. In natural touch, when making or breaking contact with an object, mechanoreceptors in the skin generate a high-frequency transient firing known as a “contact transient”, whereas the firing frequency decreases during sustained pressure. The biomimetic patterning of ICMS leverages this phenomenon by delivering a brief, high-amplitude or high-frequency burst at the onset of stimulation before reducing it to a lower baseline level. Compared to conventional linearly encoded ICMS, where the amplitude increases linearly with pressure, biomimetically encoded ICMS significantly improves the subjects’ tactile resolution, resulting in a lower just-noticeable difference (JND) and enhancing the naturalness and dynamic intensity range of the touch [61,63]. Spatiotemporally patterned ICMS further builds on multielectrode ICMS by coordinating electrode combinations in space with stimulation sequence, frequency, and amplitude changes in time. By using voltage field shaping (VFS) to dynamically adjust the amplitude ratios across electrodes, it can evoke clear perceptions of edges, curvature, and continuous motion [62]. More recent personalized and interactive ICMS paradigms allow participants to actively adjust amplitude, frequency, the biomimetic degree of stimulation encoding, and temporal overlap during sequential electrode activation, and can establish highly specific virtual-object perceptual models to convey complex tactile features such as object texture and compliance [64]. Together, these results support the extension of ICMS from linear intensity encoding to multielectrode, biomimetic, and more complex spatiotemporal pattern design. These strategies improve spatial delivery, temporal dynamics, or subjective controllability in different ways, but they still mainly enhance naturalness and controllability at the perceptual-outcome level and do not yet approximate the population-level activity patterns of natural tactile input [8]. When applied to V1, ICMS can induce phosphenes (such as spots or lines of light), with brightness, shape, spatial location, and discriminability closely related to current amplitude, pulse timing, electrode combination, and cortical location [14,67,91,92]. Utilizing the specific spatial arrangements of V1 RFs and the retinotopic map, spatiotemporally patterned ICMS can construct recognizable static letters (e.g., T and L) and directional apparent motion within the visual field of monkeys. This recognition ability can be directly acquired during visual tasks and transferred to ICMS-evoked phosphene patterns, demonstrating that high-channel-count ICMS can generate spatially structured phosphene patterns across visual cortex [14]. In similar experiments involving completely blind human participants, synchronized patterned multielectrode stimulation successfully evoked simple 2-dimensional (2D) shapes and letters distinguishable by the subjects, and a closed-loop system incorporating a bio-inspired retinal encoder allowed subjects to locate external objects via head scanning [66]. These findings indicate that visual ICMS has progressed from single-phosphene induction toward structured visual-cue construction (Fig. 3B). Current evidence, however, still mainly supports simple shapes, letters, and object localization. Further development remains limited by phosphene predictability, multielectrode combination rules, retinotopic mapping, and chronic stability, and long-term stable, generalizable image perception has not yet been achieved [8,14,66,67]. In summary, multiple complex stimulation strategies have advanced ICMS-based artificial perception from single-site evocation toward greater controllability, but their advantages and limitations differ. Multielectrode ICMS expands the spatial distribution of stimulation and can improve spatial controllability, while remaining influenced by electrode layout, interchannel interactions, and cortical layers. Biomimetic ICMS currently has application evidence mainly from tactile studies and can improve tactile discrimination and subjective naturalness, but its stability and transferability remain to be validated. Spatiotemporally patterned ICMS further integrates electrode combinations with stimulation timing and can convey more complex structured artificial perceptual cues in touch and vision, but it remains limited by stimulation-combination rules and chronic stability [13,14,61–64,66,67,88,93]. Therefore, the translational limitations of ICMS arise not only from stimulation-pattern design but also from its mode of neural recruitment. Electrical stimulation can bypass normal synaptic integration and generate neuronal activation patterns that differ from natural sensory input. Continuous stimulation can also cause rapid perceptual fading during stimulation and longer-lasting stimulation-induced depression of neuronal excitability. In human tactile ICMS, perceptual effects may fade in less than 1 min during continuous stimulation, which is an important obstacle to sustained feedback under clinical conditions [8]. Representative parameters for the tactile and visual ICMS studies discussed here are summarized in Table 2. Table 2.Representative stimulation parameters and electrode configurations in tactile and visual ICMS studiesStudySpeciesTarget areaElectrodeAmplitudeFrequencyPulse widthInterphase intervalTrain durationExperimental durationMain findings[13]Human (N = 1)S1Utah10–100 μA25–300 Hz200 μs cathodic, 400 μs anodic100 μs0.5–1 s6 moS1 ICMS evoked stable localized touch[89]Human (N = 1)S1Utah20–100 μA50–300 Hz200 μs per phase53 μs1 s3–56 dS1 ICMS evoked cutaneous/proprioceptive percepts[88]Human (N = 1)S1; M1 (recording)Utah14–64 μAAverage 100 Hz200 μs cathodic, 400 μs anodic100 μsNR23 moS1 ICMS tactile feedback improved BCI control[61]Human (N = 3)S1; M1 (recording)Utah20–90 μA50–200 Hz200 μs cathodic, 400 μs anodic100 μs0.1/1 s24–84 moMultielectrode/biomimetic ICMS supported stable precise PFs[63]Human (N = 3)S1Utah2–80 μA0–250 Hz200 μs cathodic, 400 μs anodic100 μs1 s1–3 dBiomimetic ICMS enhanced natural touch at lower charge[64]Human (N = 3)S1Utah10–100 μA20–150 Hz200 μs cathodic, 400 μs anodic100 μsNRNRCustomized ICMS evoked object-specific touch[62]Human (N = 2)S1; M1 (recording)Utah40–90 μA100 Hz200 μs cathodic, 400 μs anodic100 μs0.05–3 sNRPatterned ICMS evoked edges, curvature, and motion[67]Human (N = 1)V1Microwire1.9–80 μA50–200 Hz100–800 μs per phase0–100 μs0.125–3 s4 moV1 ICMS evoked stable phosphenes[204]Macaque monkey (N = 1)V1Microwire12–25 μA200 Hz200 μs per phaseNR0.2–1 s16 moV1 ICMS supported phosphene localization[91]Macaque monkey (N = 1)V1Utah18–300 μA200 Hz200 μs per phase100 μs200 ms19–24 moChronic V1 microstimulation supported spatial discrimination[14]Macaque monkey (N = 2)V1; V4 (recording)Utah1–210 μA300 Hz170 μs per phase60 μs167/450 msNRMultielectrode V1 ICMS supported simple pattern recognition[66]Human (N = 1)V1/V2 boundaryUtah1–128 μA300 Hz100–800 μs per phase60 μs166 ms6 moOccipital ICMS evoked phosphenes and simple patternsNR, not reported; M1, primary motor cortex; V2/V4, visual cortical areas V2/V4; h, hours; d, days; mo, months Representative stimulation parameters and electrode configurations in tactile and visual ICMS studies NR, not reported; M1, primary motor cortex; V2/V4, visual cortical areas V2/V4; h, hours; d, days; mo, months ICMS pattern learning and information integration Distinct from the percept-evocation strategies discussed above, another direction of research treats ICMS as meaningful, learnable signals. A classical paradigm involves training primates to learn and discriminate different spatial or temporal patterns of ICMS sequences delivered to the sensorimotor cortex, prompting them to perform corresponding behaviors to obtain rewards (Fig. 4A). For instance, owl monkeys successfully learned to discriminate 4 distinct spatial or temporal ICMS patterns applied to the hand representation area of S1, using these cues to locate food hidden behind either a left or right door [68]. Rhesus macaques were able to complete 3 distinct visual tasks: center-out reaching, continuous target tracking, and target selection, which were cued either by mechanical vibration on the hand or by ICMS within S1 [69]. Furthermore, during active exploration, delivering specific, reward-associated ICMS patterns to S1 enabled monkeys to learn to distinguish between 3 visually identical targets [70], or to discriminate texture coarseness conveyed by different invisible virtual gratings [94]. These studies show that, after training, ICMS can become an actionable cue for discrimination, reaching, and active exploration, rather than a direct reproduction of natural sensory encoding or task-related neural activity. Notably, in a control experiment, posterior parietal cortex stimulation failed to form an effective behavioral cue, suggesting that this form of information input is target dependent and may fundamentally rely on distinguishable somatic sensations evoked by stimulation [70]. ICMS pattern learning and multisensory information integration. (A) Following training, primates learn to interpret ICMS sequences with diverse spatiotemporal patterns to execute target selection, active exploration, and reaching toward invisible targets. (B) Rodents learn to use sensor- or behavior-coupled ICMS to localize and explore invisible targets. Infrared signals can be mapped onto S1 stimulation frequency, whereas patterned multichannel ICMS can distribute directional and distance-related information across electrodes. In addition, animals have also been shown to learn to integrate ICMS with visual cues to explore invisible targets (Fig. 4B). In rhesus macaques, delivering continuous spatial information via 8-channel ICMS to S1 to gradually replace visual signal (VIS) cues allowed the monkeys to learn a center-out reaching task toward an invisible target [74]. In rats, using S1 ICMS frequency as a single coding variable that changed continuously with behaviorally relevant variables such as the angle or distance to a target could also replace VIS cues. Through active exploration of invisible locations, rats learned to obtain water rewards or complete water-maze search tasks, while still distinguishing artificial input from native whisker deflections [71,72]. Subsequent work extended this coding approach to 2 dimensions: multielectrode ICMS encoded target angle and distance in real time through stimulation frequency and electrode channel, allowing mice to replace VIS cues, integrate the artificial input with natural vision, and complete a 3 × 3 grid-navigation task with high accuracy [73]. Additionally, gradually substituting visual cues with ICMS applied to the dorsal premotor cortex trained rhesus macaques to discriminate ICMS patterns across different electrode contact combinations, enabling highly accurate judgments in a 4-class reach-and-grasp task [95]. Interestingly, in some studies, combined VIS+ICMS cueing produced higher accuracy than VIS-only or ICMS-only cueing [73,74], and native tactile representations were not disrupted [71,72]. This suggests that ICMS-carried information can remain partly independent from existing tactile representations while being incorporated into other sensory modalities in a naturalistic manner. These studies suggest that ICMS does not necessarily have to perfectly replicate natural neural encoding to be effective. Instead, the brain can treat it as a novel sensory modality or an abstract instruction and, through learning, naturally comprehend these task cues or integrate them with other natural sensory modalities, thereby expanding or augmenting natural perception [11]. This is informative for expanding the roles and contexts of ICMS in BCI feedback, but these effects mainly arise in structured tasks with explicit training, reward mappings, and constrained behavioral goals. Learnable ICMS is therefore better interpreted as a task-dependent information channel than as a universally transferable sensory code. Recently, ICMS informed by the S1-to-primary motor cortex (M1) pathway significantly modulated grip force without prior associative training, suggesting that ICMS designed according to cortical information pathways may be more efficient [96]. Future work should compare cortical targets and training requirements, standardize training metrics, and test whether learned mappings remain stable across training stages and task modalities. Representative parameters for the learnable ICMS information-input paradigms discussed here are summarized in Table 3. Table 3.Representative stimulation parameters and electrode configurations in studies of ICMS pattern learning and information integrationStudySpeciesTarget areaElectrodeAmplitudeFrequencyPulse widthInterphase intervalTrain durationExperimental durationMain findings[68]Owl monkey (N = 2)S1Microwire50–200 μA100 Hz100 μs100 μs4 s3–4 moMultichannel S1 ICMS supported reaching cues[69]Rhesus macaque (N = 1)S1/PP; M1/PMd/SMA (recording)Microwire25–60 μA30–60 Hz150–200 μs per phase100 μs0.5–2 sNRS1 ICMS supported cue learning; PP ineffective[70]Rhesus macaque (N = 2)S1; M1 (recording)Microwire80–200 μA200 Hz or 400 Hz105–200 μs per phase25 μs50 msNRS1 ICMS supported active texture discrimination[71]Rat (N = 6)S1Microwire1–300 μA0–400 Hz, updated every 50 ms100 μs50 μs0.6–1.5 sApproximately 51 dS1 ICMS served as an infrared sensory channel[74]Rhesus macaque (N = 2)S1Utah30–80 μA100–300 Hz200 μs per phase250 μs0.2–1.5 sNRS1 ICMS supported artificial sensory-visual integration[95]Rhesus macaque (N = 2)PM/S1; M1 (control)MicrowirePM/S1:2–28/20–60 μA20–100 Hz200 μs per phaseNR50–1,000 msNRPM/S1 ICMS supported action instruction learning[72]Rat (N = 6)S1; A1 (control)Concentric bipolar electrode15–75 μA100 Hz200 μs per phaseNR500 ms3–7 dS1 ICMS guided active sensing[94]Rhesus macaque (N = 2)S1; M1 (recording)Microwire130 μAEvent-triggered, no fixed frequency100 μs per phaseNRNRNRS1 ICMS supported virtual texture discrimination[73]Mouse (N = 10)S1Microwire10–20 μA0–200 Hz, real-time updating200 μs per phase250 μs5 s≥7 dS1 ICMS supported visual integration during navigationNR, not reported; A1, primary auditory cortex; PM, premotor cortex; PMd, dorsal premotor cortex; PP, posterior parietal cortex; SMA, supplementary motor area; h, hours; d, days; mo, months Representative stimulation parameters and electrode configurations in studies of ICMS pattern learning and information integration NR, not reported; A1, primary auditory cortex; PM, premotor cortex; PMd, dorsal premotor cortex; PP, posterior parietal cortex; SMA, supplementary motor area; h, hours; d, days; mo, months Cortical representation reorganization by chronic ICMS During the process of mapping cortical functions or evoking tactile perception using ICMS, researchers observed a critical phenomenon: in rodents and primates, chronic, low-frequency, repetitive ICMS can induce sensorimotor representation areas to undergo sustained and reversible boundary expansion. This expansion is predominantly characterized by complete substitution, with only a small fraction overlapping with other functional areas, and it gradually recovers after the cessation of stimulation [97,98]. In seizure-kindling models, motor-map enlargement is also associated with enhanced synaptic and field-potential responses, suggesting that abnormal repetitive activity can reshape cortical maps through potentiation-like mechanisms [99]. This chronic stimulation also sensitized the brain’s perception of ICMS, causing the stimulation threshold required to evoke the same response to decrease over time [49]. This heightened sensitivity is likely related to the plastic adaptation of the cortex to ICMS [93]. The continuous delivery of paired ICMS sequences with a fixed temporal delay between 2 electrode contacts in the forelimb area of the rat S1 significantly reduced the rat’s perceptual threshold and gradually enhanced the inferred functional connectivity between the 2 stimulation sites, further supporting this hypothesis [75]. These studies reveal an important transition in the interpretation of ICMS, indicating that the brain can integrate ICMS into ongoing cortical activity and respond in a targeted manner, thereby going beyond mere perceptual evocation [11]. Repeated ICMS does not merely “passively activate” neural tissue to evoke transient percepts; rather, it may be incorporated into ongoing cortical dynamics, actively inducing alterations in cortical functional representations and dynamically adjusting neural connection strengths. The underlying mechanism may be consistent with the synaptic plasticity engaged during the brain’s motor learning processes [100]. These phenomena support the view that ICMS may move beyond transient artificial perception, but whether it can reliably produce neural connectivity remodeling still requires further discussion. Neural Connectivity Remodeling Based on Synaptic Plasticity Beyond artificial perception, another key question is whether temporally structured electrical pulses can continue to influence cortical connectivity and network responses after stimulation ends. Synaptic plasticity, the ability of synaptic connections to change strength in response to neural activity, serves as the foundation for brain development [101] and postinjury functional reorganization [102], providing a key mechanistic framework for understanding how ICMS may induce longer-lasting network changes (Fig. 5A). Based on classic Hebbian plasticity [103], neurons that fire together tend to wire together, whereas poorly coordinated connections are weakened or become unstable; at the network level, these 2 effects are expressed as long-term potentiation (LTP) and long-term depression (LTD) [104]. LTP is generally induced by high-frequency, high-intensity stimulation, whereas LTD is generally induced by low-frequency, weak stimulation; both may be accompanied by physiological changes such as dendritic spine remodeling and altered local circuit excitability [105,106]. Spike-timing-dependent plasticity (STDP) further introduces a millisecond-scale temporal window and reveals a rule for synaptic weight change: a presynaptic pulse preceding a postsynaptic pulse tends to induce LTP, whereas the reverse order tends to induce LTD [107] (Fig. 5B). Burst-timing-dependent plasticity (BTDP) further extends this timing view by emphasizing that brief postsynaptic bursts paired with presynaptic input may produce stronger or more durable synaptic modification, although direct in vivo evidence for BTDP-like mechanisms in ICMS remains limited [108,109]. For ICMS, these plasticity rules are useful not because electrical stimulation reproduces natural physiological processes in full, but because they define timing constraints under which artificial pulses may bias synaptic efficacy. Synaptic mechanisms relevant to ICMS-induced plasticity. (A) Basic organization of chemical synaptic transmission. (B) STDP is a timing-dependent Hebbian mechanism: presynaptic activity preceding postsynaptic activity induces LTP, whereas the reverse sequence induces LTD. (C) Homeostatic synaptic scaling readjusts synaptic strength toward baseline after sustained increases or decreases in neuronal activity. Traditionally, reorganizing cortical representations and synaptic connections relies on chronic natural inputs, such as physical tactile [110] or visual [111] stimulation, or continuous motor skill learning [112]. ICMS, however, because of its intracortical nature, may “bypass” peripheral receptors. By delivering electrical pulses with specific timing, ICMS can simulate the neural activity required for synaptic enhancement and induce STDP-like connectivity changes within cortical networks [15,76,77]. This feature helps distinguish ICMS from neuromodulation approaches that primarily act on peripheral nerves, cortical surfaces, or deep nuclei: ICMS does not primarily alter network excitability at a larger scale, but, under appropriate experimental conditions, introduces precisely timed activity into defined intracortical microcircuits [9,11,113]. Targeted connectivity modulation and damaged circuit repair Current ICMS research is shifting from open-loop stimulation (i.e., paired stimulation sequences with a fixed temporal delay between microelectrode contacts) [75] toward plasticity-based closed-loop modulation. This approach, commonly implemented through activity-dependent stimulation (ADS), involves the real-time detection of spontaneous neural activity in a recording area, such as spikes or specific band power of local field potentials (LFPs), followed by the delivery of ICMS to a target area after a millisecond-scale delay, thereby artificially strengthening the functional connection between the 2 regions [114] (Fig. 6A). Compared with conventional paired stimulation, ADS is more informative mechanistically because the pulse is contingent on ongoing circuit activity rather than imposed independently of the circuit state. Closed-loop ICMS paradigms and circuit remodeling. (A) In closed-loop ICMS, spike events or LFP features detected from ongoing recordings trigger stimulation according to predefined thresholds, delays, and pulse settings. Neural activity after stimulation is then recorded as feedback for the next detection cycle. (B) An ADS paradigm detects spikes at one cortical site and delivers ICMS at another after a predefined delay. Repeated pairing can guide a directional shift in cortical motor output representations. (C) In rodents with M1 lesions, a closed-loop bridge uses activity recorded from RFA to trigger S1 ICMS, strengthening the RFA-S1 pathway and improving motor recovery. Leveraging this ADS concept, ICMS can induce measurable neurofunctional changes between recording and target areas. In healthy primates and rodents, precisely controlling the delay between the trigger spike and the target-site ICMS to 5 to 50 ms significantly induces a directional shift of M1 motor representations from the recording area toward the stimulation area [15] (Fig. 6B). Furthermore, this paradigm significantly reduces the threshold for evoking tactile perception in S1 [75] and enhances functional connectivity between the recording and target regions [76,115]. These phenomena operate independently of the physical distance between the recording and target sites [15,78] and are broadly consistent with the classical STDP time window [107,116]. In contrast, long-delay stimulation exceeding 50 ms [15,76] or continuous open-loop paired stimulation uncoupled from spontaneous activity [75,78,115] fails to induce stable or persistent STDP-like connectivity enhancements in vivo. Together, these comparisons indicate that timing contingency is a key determinant of ICMS-induced plasticity, and artificially coupling presynaptic neuronal firing with ICMS-induced postsynaptic depolarization provides a possible method for neural connectivity remodeling. However, most demonstrations remain tied to selected cortical targets, animal models, and short observation windows, so their scalability and relevance to durable therapeutic remodeling require further validation. They should not yet be interpreted as equivalent to natural experience-dependent sensory or motor plasticity [8]. In TBI models, ICMS also demonstrates potential for supporting circuit-level compensation after injury. For example, in rats with M1 lesions, researchers developed a closed-loop neural prosthesis for functional recovery by bridging spike activity in the rostral forelimb area (RFA) with ICMS in S1, effectively enhancing the functional connectivity between these 2 regions with a 7.5 to 10 ms delay. This paradigm significantly enhanced intercortical input–output coupling and increased stimulus-spike synchronization. At the same time, fine motor function of the rats’ forelimb recovered rapidly within 14 days and significantly outperformed open-loop stimulation and no-stimulation controls [77,117] (Fig. 6C). The researchers suggested that this behavioral benefit may arise from ICMS-induced enhancement of neural circuit connectivity. Chronic in vivo monitoring further confirmed that ADS can induce local up-regulation of synaptophysin around the electrode, providing anatomical support for enhanced synaptic connectivity induced by ICMS [118]. Additionally, multiple studies demonstrate that ADS-patterned ICMS rapidly modulates neuronal excitability [119–121]. Taken together, these findings support a repair-oriented interpretation in which temporally contingent ICMS first reshapes excitability and synchrony and may then promote pathway-level coupling. Notably, the functional neural connectivity sculpted by ICMS is not permanent. It typically persists robustly for a period before slowly returning to baseline over approximately 1 week [15,75–77]. This transient nature is primarily governed by the influence of homeostatic plasticity, which introduces another dimension of ICMS-mediated plasticity modulation and will be discussed in the following section. Overall, ICMS may help restore damaged functional connectivity, but current evidence cannot yet fully distinguish this effect from physiological functional compensation; more refined pathway-level monitoring and causal validation will be needed to address this question. Representative parameters for the ICMS plasticity-induction paradigms discussed here are summarized in Table 4. Table 4.Representative stimulation parameters and electrode configurations in ICMS-based plasticity modulation studiesStudySpeciesTarget areaElectrodeAmplitudeFrequencyPulse widthInterphase intervalTrain durationExperimental durationMain findings[15]Macaque monkey (N = 2)M1Microwire25–80 μAADS (~9–19 Hz)200 μs per phaseNRSingle-pulse stimulation1–4 dM1 ADS induced lasting motor-output remapping[76]Rat (N = 3)Forelimb sensorimotor cortexMicrowire30 μAADS (≤10 Hz)200 μs per phaseNRSingle-pulse stimulation40–72 hShort-delay ADS reshaped functional connectivity[75]Rat (N = 5)Forelimb sensorimotor cortexMicrowire20 μA1,000 Hz200 μs per phaseNRApproximately 6 ms48–72 hPaired stimulation lowered ICMS detection thresholds[115]Macaque monkey (N = 2)S1Utah50 μAADS/random stimulation (<10 Hz)200 μs per phaseNR2-pulse stimulation, ~2.8 ms0.5 hS1 spike-triggered stimulation enhanced tactile plasticity[77]Rat (N = 6)S1; RFA (recording)Michigan60 μAADS (~8 Hz)192 μs per phaseNRSingle-pulse stimulation≤21 dADS improved reaching and RFA-S1 connectivity[78]Macaque monkey (N = 2)S1/M1Microwire80–330 μA300–330 HzNRNR6–6.7 ms1–3 hPaired stimulation induced STDP-like site strengthening[117]Rat (N = 11)S1FL/S1BF; RFA (recording)Michigan60 μAADS (~7 Hz)200 μs per phaseNRSingle-pulse stimulation3 hADS enhanced distant coupling more than random stimulation[118]Rat (N = 5)S1; RFA (recording)Michigan60 μAADS (~7 Hz)200 μs per phaseNRSingle-pulse stimulation21 dChronic ADS modulated RFA activity and S1 synaptic markers[121]Rat (N = 6)S1; RFA (recording)Michigan60 μAADS (≤35 Hz)200 μs per phaseNRSingle-pulse stimulation1 hPostinjury ADS increased firing and synchronized burstsNR, not reported; RFA, rostral forelimb area; S1BF, S1 barrel-field region; S1FL, S1 forelimb region; h, hours, d, days Representative stimulation parameters and electrode configurations in ICMS-based plasticity modulation studies NR, not reported; RFA, rostral forelimb area; S1BF, S1 barrel-field region; S1FL, S1 forelimb region; h, hours, d, days Achieving chronic ICMS modulation in freely behaving animals requires uninterrupted neural monitoring coupled with conditional stimulation delivery. Various integrated stimulation devices, including the Neurochip series [122–124], NeuroRighter [125,126], the wireless artifact-free neuromodulation device (WAND) [127], and the smart neurostimulation system (SNS) [128], have provided a technical basis for low-latency triggering, real-time artifact control, and programmable precise ICMS delivery; their technical parameters are summarized in Table 5. Recently, wireless distributed neuromodulation systems based on the “neurograin” concept [129] have emerged and may extend complex ICMS paradigms toward the whole-brain level [130]. Looking forward, miniaturized, low-power closed-loop devices integrated with on-chip machine learning will provide a reliable hardware basis for chronic ICMS-induced neural plasticity and for testing ADS-like plasticity rules, but their capacity for connectivity modulation should be jointly validated through long-term behavioral outcomes and safety assessment while considering activity-specific biomarkers, delay parameters, stimulation patterns, and safety boundaries, so that useful Hebbian strengthening, maladaptive synchronization, and compensatory downscaling can be distinguished [128,131,132]. Table 5.Comparison of technical parameters of closed-loop neuromodulation devices relevant to activity-dependent ICMSFeatureNeurochip [15,122]NeuroRighter [125,126]Neurochip-2 [123]WAND [127]Neurochip-3 [124]SNS [128]Dimensions44 mm × 20 mm × 15 mm15.2 cm × 8.9 cm (not head-fixed)55 mm × 35 mm× 20 mm36 mm × 33 mm × 15 mm56 mm × 40 mm × 50 mmNR (coin-sized cranial IPG)Wireless linkInfraredNR (tethered)InfraredNordic BLEInfraredInductive IPG-EP link; EP USB/BLERecording channels1 unipolar, 2 bipolar16–64312816 bipolar/32 unipolar60Sample rate11.7/2 kS·s−125 kS·s−12.6/5.3/10.5 kS·s−11 kS·s−15/10/20 kS·s−10.5 kS·s−1ADC resolution8 bits16 bits8 bits15 bits16 bitsNRStimulation channels1Up to 643128660Maximum current100 μANR10–200 μA; up to 5 mA5 mA10 mA5 mABiomarkerAction potentialAction potentialAction potential/spectral powerSpectral powerAction potential/spectral powerSpectral powerClosed-loop controlEvent detectionEvent detectionEvent detection/feature thresholdingFeature thresholdingEvent detection/feature thresholdingSpectral-feature classificationArtifact cancellationNo active cancellationSALPA-compatibleTransient gain reductionLinear interpolationFast-settle/blanking windowPoststimulation analysis windowOperation time≤60 hNR≤48 h~11.3 h≤24 h~8 hProcessing platformOn-board PSoCHost PC/APIOn-board PSoCOn-board FPGA/MCUOn-board MCU/FPGAImplanted processorClosed-loop timing0.5 ms example7.1–46.9 ms5 ms example8–1,024 ms<50 ms supported<25 msDemonstrated applicationCortical plasticityNetwork controlChronic recording/stimulationMovement modulationCircuit bridgingSpectral power modulationAnimal modelNonhuman primateRodent/cultureNonhuman primateNonhuman primateRodent/nonhuman primateOvineNR, not reported; ADC, analog-to-digital converter; BLE, Bluetooth low energy; EP, external processor; FPGA, field-programmable gate array; IPG, implantable pulse generator; MCU, microcontroller unit; PSoC, programmable system-on-chip; SALPA, subtraction of artifacts by local polynomial approximation; WAND, wireless artifact-free neuromodulation device; SNS, smart neurostimulation system Comparison of technical parameters of closed-loop neuromodulation devices relevant to activity-dependent ICMS NR, not reported; ADC, analog-to-digital converter; BLE, Bluetooth low energy; EP, external processor; FPGA, field-programmable gate array; IPG, implantable pulse generator; MCU, microcontroller unit; PSoC, programmable system-on-chip; SALPA, subtraction of artifacts by local polynomial approximation; WAND, wireless artifact-free neuromodulation device; SNS, smart neurostimulation system Modulating homeostatic plasticity for pathological network intervention Homeostatic plasticity refers to a negative-feedback plasticity mechanism by which neurons and neural circuits counteract activity perturbations caused by learning, development, or injury by regulating synaptic strength, intrinsic excitability, and network excitation/inhibition (E/I) balance, thereby maintaining their activity within a stable functional range [133] (Fig. 5C). This concept originated from the “sliding threshold” in the Bienenstock–Cooper–Munro theory and later developed into the theories of metaplasticity and synaptic scaling, emphasizing how prior neuronal activity regulates subsequent synaptic changes to restrain excessive Hebbian positive-feedback loops in the form of LTP or LTD [134–136]. Consequently, cortical changes induced by tetanic high-frequency ICMS are generally difficult to sustain [75,78]. In contrast, sparse stimulation patterns triggered by spontaneous neural activity (such as spikes) more closely resemble physiological firing, making them less likely to trigger homeostatic down-regulation and potentially induce more robust and persistent remodeling of neural connectivity [15,77,118]. These phenomena indirectly suggest that stimulation paradigms consistent with physiological activity patterns may be more favorable for balancing Hebbian reinforcement with homeostatic regulation. In diseases characterized by abnormal neural discharge, homeostatic plasticity can lead to an overcompensation of intrinsic excitability, thereby propelling disease progression [137]. ICMS may interfere with maladaptive coupling in these circuits, offering a potential strategy to modulate pathological homeostatic imbalances in such conditions. For instance, in focal epilepsy, the failure of inhibitory homeostatic mechanisms within the epileptogenic zone has been considered to increase α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptor density in lesioned areas, thereby leading to the strengthening of pathological excitatory synaptic connections, suggesting that abnormal homeostatic plasticity involves both synaptic efficacy and network-level regulation [138]. In addition to directly interrupting epileptiform activity in the cortex [139], ICMS can also precisely modulate abnormal homeostatic plasticity through activity-dependent closed-loop paradigms. Interictal epileptiform discharges (IEDs) represent a meaningful neural activity target; under pathological conditions, they often form cross-regional abnormal coupling with normal brain oscillations (such as sleep spindles), which can, through chronic plasticity mechanisms, continuously expand the epileptic network and impair cognitive function [140–142]. Modulating interictal patterns may gradually reshape neural networks, support physiological brain function, and suppress epileptic activity, as preliminarily shown in cortical surface stimulation studies [143]. Previous studies have shown that, in a rat model of hippocampal kindling, the real-time detection of hippocampal IEDs followed by the application of Gaussian-waveform ICMS to the medial prefrontal cortex (mPFC) with a millisecond-level delay prevented pathological coupling between the 2 regions. This intervention could effectively delay the network-wide propagation of the epileptic focus and improve cognitive memory function [142,144]. These findings support the mechanistic feasibility of closed-loop ICMS for intervening in epilepsy-related pathological network coupling, but the current evidence still comes mainly from animal experiments and depends on selected biomarkers, stimulation targets, and timing rules. Therefore, compared with more macroscopic or clinically mature neuromodulation therapies such as vagus nerve stimulation (VNS), DBS, and responsive neurostimulation (RNS), ICMS is currently better defined as a cortical-circuit-level proof of concept [145]. Similarly, Parkinson’s disease (PD) is associated with chronic abnormalities in cortical homeostatic plasticity. Progressive dopamine depletion disrupts plasticity in the motor cortex and cortico-striatal circuits [146], leading to motor deficits. Levodopa and other dopaminergic therapies may partially restore this balance but may also induce aberrant plasticity [147]. This dual effect suggests that more precise and controllable plasticity-modulating approaches are needed to compensate for the limitations of dopaminergic therapy alone [148]. Current PD-related studies primarily use ICMS as a high-resolution functional detection tool to assess motor cortical excitability, motor representations, and cortico-basal ganglia loop changes in 6-hydroxydopamine or 1-methyl-4-phenyl-1,2,3,6-tetrahydropyridine models [149,150]. In addition, studies in PD models indicate that ICMS can suppress pathological electrical activity in mouse brain slices and transiently improve motor capacity when applied to rat M1 [151,152]. Meanwhile, theta-burst patterned cortical stimulation can modulate motor cortical excitability in patients with PD, induce LTP/LTD-like changes, and improve motor performance [153], supporting the broader potential of plasticity-based cortical modulation. Overall, current PD-related ICMS studies still lack clinical validation and evidence of durable functional benefit; ICMS is therefore better understood as a high-resolution experimental tool for probing and modulating cortical microcircuits [9,154]. With further integration of closed-loop paradigms, ICMS may expand its role in PD through more precise plasticity modulation [155]. In summary, ICMS has the potential to interfere with abnormal homeostatic plasticity, but its translational relevance should be interpreted in the context of established clinical neuromodulation therapies. Among invasive neuromodulation technologies, DBS is relatively similar to ICMS in intervention form but typically uses larger electrodes to deliver milliampere-level currents to deep nuclei such as the subthalamic nucleus, recruiting heterogeneous neuronal populations and fiber tracts within a several-millimeter region [9,156]. By contrast, ICMS delivers microampere-level pulses to cortical microcircuits and recruits neural elements over a range of tens of micrometers, making it better suited for targeting plasticity-related nodes within local cortical circuits. Recent studies further indicate that ICMS recruits excitatory and inhibitory neurons in an activity-state-dependent manner and can induce postactivation neural response depression, suggesting that patterned ICMS may influence long-range plastic reorganization by regulating local E/I balance [10,157]. In the future, closed-loop or specifically patterned ICMS may provide a high-precision cortical-level complementary strategy for selected disorders characterized by abnormal neural discharge. Before clinical translation, however, reliable biomarkers, reproducible parameters, implantation safety, and durable functional benefits require systematic validation [158,159]. Although ICMS-based electrical neuromodulation shows important potential in reshaping existing circuits and intervening in abnormal plasticity, its efficacy is fundamentally constrained by the long-term stability of the electrode interface and the requirement of intact neural substrates. To address these physical and biological limitations, research has begun to explore whether BNIs can improve electrode–tissue interface stability and signal transmission while extending the potential applications of ICMS in injury repair. This is the emerging frontier we explore next. BNIs for ICMS Applications In chronic ICMS applications, BNIs can be understood as experimental strategies for interface stabilization, living signal bridging, and tissue repair. Whether rigid or flexible, MEAs remain constrained in long-term applications by the mismatch between device materials and brain tissue, which drives the need to integrate living biological components into neuroelectronic systems [43]. Compared with conventional electrodes that activate neurons through direct physical stimulation, BNIs introduce living neural components, creating possibilities for biointegration, synaptic bridging, and regenerative modulation, and may help overcome the limitations of purely physical stimulation [46,48], thereby offering new possible solutions for chronic ICMS. Current strategies can be read as a problem-oriented continuum: one group aims to improve the tissue–device interface or provide biologically mediated signal bridging, whereas another group, based on biohybrid regenerative electronics, further integrates stem cells, progenitors, or brain organoids to explore graft maturation, integration, and tissue reconstruction [19]. We provide a preliminary summary of representative studies of cortical BNIs in Table 6. Table 6.Summary of representative studies of biohybrid neural interfacesStudyBiological componentElectrode structureExperimental modelLimitations[44]Rat sciatic nerve segment; host cortical neuritesHollow glass cone electrodeRat S1Tissue-ingrowth dependent; limited channels[160]Human host cortical neuritesHollow neurotrophic glass coneHuman motor speech cortexComplex implantation; limited scalability[162]E14 mouse cortical NSCsParylene probe with microwell for NSCsRat cortexChronic benefit not maintained[164]Adult mouse subependymal NPCsSilicon probe seeded with NPCsRat cortex and in vitro testingChronic performance not tested[163]E18 rat hippocampal neurons; rat astrocytesFibrin-coated microelectrodeRat cortex and in vitro testingLiving-cell contribution unclear[184]E18 rat cortical neural cellsSilk film with patterned gold electrodesMouse cortex and in vitro testingIn vivo fully cell-loaded electrode function not tested[16]Rat neuroprogenitors/glial cellsPlatinum electrode with conductive hydrogelIn vitro testingIn vitro only; integration untested[161]Human host neuritesHollow neurotrophic glass coneHuman motor speech cortexSingle case; low throughput[185]Human hESC-derived NPCsPPy/PDMS conductive polymerRat cortical stroke modelLong-term integration unproven[187]Human hiPSC-derived cortical organoidsTransparent graphene MEAMouse cortexStimulation/repair not tested[165]E18 mouse cortical neurons4 × 4 biohybrid transition arrayRat cadaver cortex and in vitro testingProof of concept only[188]Human hESC-derived brain organoidsOrganoid-mediated OBCIMouse S1 injury modelLong-term safety unvalidated[166]E18 rat neural spheroids/axon bundlesStretchable array with spheroid wellsMouse cortex and in vitro testingIn vivo synapses not achieved[186]E13.5 mouse NSCsLiving artificial dura seeded with NSCsMouse cortical TBI modelIntegration and safety unvalidatedE, embryonic day; hESC, human embryonic stem cell; hiPSC, human induced pluripotent stem cell; NPCs, neural progenitor cells; NSCs, neural stem cells; OBCI/OBCIs, organoid–brain–computer interface(s) Summary of representative studies of biohybrid neural interfaces E, embryonic day; hESC, human embryonic stem cell; hiPSC, human induced pluripotent stem cell; NPCs, neural progenitor cells; NSCs, neural stem cells; OBCI/OBCIs, organoid–brain–computer interface(s) BNI strategies for interface integration For BNIs, the earliest and still instructive question is whether the host tissue can be induced to couple with an implanted structure more actively than it does with a conventional inert electrode. A classic strategy was the cone electrode, which used a sciatic nerve graft to guide host cortical neurites into the tip of a glass cone electrode and could support stable recordings for months [44]. Subsequent work replaced the sciatic nerve graft with neurotrophic factors and refined this strategy into the neurotrophic electrode, which supported approximately a decade of stable recording from human motor speech cortex, with histological evidence showing myelinated neural filaments inside the electrode and no obvious gliotic encapsulation [160,161]. Although these electrodes remain complex to implant, low in channel count, limited in scalability, and mainly supported by structural and recording evidence, they demonstrate the feasibility of incorporating living tissue into electrode structures and provide an early conceptual basis for the use of BNIs in chronic ICMS and bidirectional BCIs. Another strategy uses cell-seeded electrode interfaces, in which living cells are directly loaded onto probe surfaces to mitigate electrode-implantation-induced FBR. Representative strategies include direct seeding of NSCs, neural progenitor cells (NPCs), and other neural cells on polymer or rigid probes through hydrogel- or protein-modified coatings. In selected experiments, these methods improved early peri-electrode neuronal density, cell adhesion, or electrochemical performance [162–164]. Compared with cone and neurotrophic electrodes, these designs more directly target the tissue–device interface, but their effects remain time-dependent. For example, early benefits of probes loaded with NSCs were not maintained and even reversed by 6 weeks, whereas other studies were mainly evaluated within implantation windows shorter than 1 month. Therefore, their long-term development still requires further resolution of cell survival and coating stability so that durable biological buffering can be maintained under chronic ICMS conditions. Further emerging strategies move beyond passive interface modification by using engineered microstructures to guide cell growth and living signal bridging. The original “living electrode” proposed a layered hydrogel structure that combines conductivity with cell support [16]. Subsequently, one prototype study integrated neuron-loaded microwells, gelatin methacrylate-filled polymer microneedles, and axon-guiding microchannels into a biohybrid transition MEA, achieving directed axonal growth and synaptic connection formation, although functional signal transmission has not yet been validated in vivo [165]. Spheroid-based interfaces further brought this idea in vivo by combining neural spheroids, stretchable electrode arrays, and microfluidic channels; they supported axon-mediated signal conduction within 5 days after implantation, but subsequent axonal degeneration and neuroinflammation were observed [166]. These studies mark a shift in BNIs from passive compatibility toward living signal transfer, suggesting a possible route for maintaining chronic ICMS interfaces and supporting biologically coupled signal transmission. However, current evidence still comes mainly from in vitro or short-term animal validation. For such systems to develop into long-term stable functional interfaces, cell viability, immune compatibility, and chronic ICMS tolerance must be demonstrated together. BNIs based on biohybrid regenerative electronics ICMS can induce synaptic plasticity within cortical neural networks. However, when TBI, stroke, or other causes result in the loss of neural tissue, neuromodulation loses its viable targets. Relying solely on the plasticity of surrounding intact tissues is often insufficient to achieve complete functional recovery. In this context, BNIs based on biohybrid regenerative electronics differ from the interface-integration or signal-bridging strategies discussed above. By combining cell transplantation with electrode interfaces, they may further extend ICMS applications toward damaged-network repair, host–graft connections, and functional reconstruction [46,48]. This direction is supported by the convergence of 2 technical approaches: in vitro MEA platforms [167,168], which enable long-term recording and stimulation of NSCs, NPCs, and brain organoids, and in vivo brain interfaces, which provide chronic recording, stimulation, and potential closed-loop modulation after graft implantation [45,47]. This in vitro-to-in vivo continuity not only supports the development of BNIs based on biohybrid regenerative electronics but also imposes higher requirements for graft survival, integration, and control over graft fate. Electrical regulation of stem cell and organoid fates In the damaged cortex, transplanted NSCs [169], neurons derived from embryonic stem cells (ESCs) or induced pluripotent stem cells (iPSCs) [170,171], and brain organoids [172] can fill lesion cavities, chronically integrate with the host tissue, and actively participate in neural network activity, representing a potential neurorestorative strategy (Fig. 7A). However, whether in cell transplantation alone or biohybrid electronics-based cell implantation, grafts require posttransplantation guidance, and the growth of implanted cells must be controlled to support integration, reduce cell loss, and lower the risk of abnormal growth [173,174], highlighting the importance of ICMS-based regulation of stem cells transplanted in vivo. In vitro graft modulation and in vivo integration in BNIs. (A) NSCs, ESC- or iPSC-derived neurons, and brain organoids can be delivered as 2D or 3D grafts into cortical lesions to support host–graft integration and tissue repair. (B) Microelectrode-based electrical stimulation can modulate the proliferation, differentiation, and maturation of 2D or 3D neural cultures and may be used for graft preconditioning. (C) Cell-seeded electrode interfaces use hydrogel- or surface-modified rigid electrodes or flexible substrates primarily to support peri-electrode integration and mitigate FBR, with selected flexible systems also enabling electrical modulation of transplanted NPCs or NSCs. (D) Brain organoids are transplanted into the mouse S1 and coupled with flexible electrode arrays to enable recording and ICMS delivery during selected posttransplantation windows. This paradigm may support host–graft synchrony and integration. In vitro studies demonstrate that microelectrode-based electrical stimulation can modulate stem cell fate [174], effectively inducing the proliferation [175], differentiation, and maturation [176] of 2D NSC or iPSC cultures. As these cultures evolve from 2D to 3D structures, the advantages of electrical modulation become even more pronounced. 3D brain organoids can closely simulate neurogenesis, cortical stratification, cell diversity, and neural circuits of the human brain [177]. By utilizing 3D microelectrode cantilevers that penetrate into the tissue interior [178,179], ICMS-style electrical pulses not only promote organoid maturation and differentiation [180] but also support the modeling of neurological diseases using these organoids as platforms [181]. Furthermore, preconditioning with microelectrode-based electrical stimulation can also improve the viability of 2D or 3D cultures as transplant donors and enhance the quality of posttransplant integration [182,183]. These findings support the view that in vivo ICMS may regulate the proliferation, differentiation, and maturation of stem cells, progenitors, or brain organoids, thereby potentially compensating for low graft survival and nondirectional circuit integration [173] and providing preliminary support for constructing BNIs based on biohybrid regenerative electronics [47] (Fig. 7B). In vivo integration and regenerative circuit repair In biohybrid regenerative electronics, combining NPCs or NSCs with surface-modified flexible interfaces and electrical modulation may extend the role of cell-seeded electrode interfaces beyond local interface support toward host–graft coupling (Fig. 7C). An early study provided initial support for this interface strategy: rat cortical neurons seeded on silk films patterned with gold electrodes were associated with a limited inflammatory response after implantation in the mouse cortex [184]. In a rat model of M1 stroke, PPy-modified PDMS flexible microchambers enabled delivery and electrical modulation of human NPCs, significantly up-regulated stanniocalcin-2 expression in transplanted NPCs, accelerated sensorimotor recovery, and promoted endogenous stem cell proliferation [185]. In another study, a living artificial dura mater constructed by seeding NSCs onto a magnetically driven Au-nanostrip patch promoted the differentiation of transplanted NSCs and activated local neuroprotective A2 astrocytes, thereby remodeling the post-TBI immune-inflammatory microenvironment [186]. These studies suggest that ICMS-based BNIs can move beyond passive interface modification by regulating transplanted cell fate and gradually approach active repair. However, long-term stimulation/recording control and stable graft–host circuit integration have not yet been achieved, and further development toward closed-loop modulation systems remains necessary. Beyond preseeding cells onto electrode surfaces, an alternative effective strategy involves allowing the living graft to initially colonize the host tissue in vivo before applying modulation via MEAs. Brain organoid transplantation provides a more complex 3D substrate for BNIs: organoids can integrate with injured host cortex, form bidirectional synaptic connections, participate in host neural activity, and be monitored over time using MEAs [187]. On this basis, recent work on organoid–brain–computer interfaces (OBCIs) has advanced by introducing active ICMS modulation. By transplanting brain organoids into the S1 cortical lesion area of mice and implanting flexible electrode probes, researchers constructed an OBCI with both signal acquisition and ICMS delivery functions (Fig. 7D). Applying 10 days of ICMS at 30 and 60 days after transplantation could promote organoid maturation and integration while significantly enhancing electrophysiological coupling between the organoids and the host and improving mouse von Frey nociceptive behavioral performance, suggesting that ICMS could promote functional integration between the organoids and host cortex [188]. By establishing a bidirectional channel for stimulation and recording, this paradigm provides a novel perspective on utilizing BNIs to integrate with the host network for brain-injury repair. Compared with 2D stem-cell grafts, organoid-based BNIs have the advantages of being 3D and amenable to monitoring, but they still require longer in vivo validation of posttransplantation survival, vascularization, projection specificity, and electrophysiological function, while tumorigenicity should be evaluated as an independent safety dimension [187,188]. In summary, ICMS-oriented BNIs combine the high spatiotemporal resolution of ICMS with the bioactivity and regenerative capacity of biological tissues, and may become an emerging framework for chronic neural interfaces and graft–host network modulation. However, current evidence should still be interpreted cautiously: existing studies differ in cell source, interface structure, implantation site, and final outcomes. Their near-term value is not that they already solve the limitations of chronic ICMS, but that they provide testable routes along the same continuum: improving the interface, bridging signals, and guiding repair-oriented host–graft coupling. Future work should compare and evaluate these strategies in terms of cell fate, immune response, signal fidelity, stimulation controllability, and ethical governance, and should establish evaluation boundaries for BNIs with different research goals [19,43,46–48]. Challenges and Future Perspectives Taken together, the future translation of ICMS can be understood within the integrative framework established in this review: hardware interfaces influence chronic usability, stimulation parameters delimit neural recruitment and safety boundaries, artificial-perception encoding shapes sensory-feedback quality and task support, plasticity modulation determines whether ICMS can extend into a circuit-intervention tool, and BNIs expand its potential for in vivo applications. Accordingly, future challenges mainly concentrate on 5 aspects: chronic reliability, mechanistic encoding and calibration, translational positioning, safety and ethics, and governance of BNIs. 1. Chronic interface reliability and hardware scalability Currently, hardware-interface reliability remains a major bottleneck for chronic ICMS applications. Long-term Utah-array studies in humans and nonhuman primates support the feasibility of chronic ICMS, yet they also reveal heterogeneity in recording quality, perceptual thresholds, electrode stability, and tissue responses [30,31,66,93]. Therefore, ICMS hardware development should not be understood simply as replacing rigid electrodes with flexible electrodes. Chronic reliability and scalability should instead be treated as core engineering goals, with mature and emerging routes integrated in stages. Current studies support this staged view. Flexible electrodes are moving toward nanoscale and ultraflexible designs with advanced manufacturing technologies, but their long-term insulation integrity and electrochemical stability in complex in vivo environments still require rigorous validation [19]. Conventional rigid electrodes remain important for near-term clinical translation, and their chronic use can be optimized at the levels of interface design, stimulation delivery, and system packaging: bioactive surface modification may improve local tissue compatibility [32,33], multielectrode stimulation may reduce the burden on individual contacts [61,66], and wireless or fully implantable packaging may help support long-term use [127–129,189]. In the future, the hardware foundation for ICMS should be able to maintain stable performance over years. 2. Mechanistic encoding and calibration for artificial perception ICMS can now evoke distinguishable tactile and visual artificial percepts. With training, these percepts or stimulation patterns can also serve as task-related cues, but their translational value still depends on whether such artificial inputs can remain stable, interpretable, and usable over the long term. Current work has produced 2 complementary strategies. Spatially, multielectrode synchronous or spatiotemporally patterned stimulation can adjust activation range, distribute channel burden, and generate patterned percepts. Temporally, biomimetic dynamic amplitude or frequency modulation can improve tactile naturalness and reduce perceptual fading caused by sustained activation [62–64,190–192]. However, perceptual encoding and stimulation-recruitment mechanisms remain insufficiently defined. Future studies should clarify how stimulation amplitude, frequency, pulse width, pulse shape, multielectrode recruitment, and cortical layers affect perceptual quality, safety boundaries, and long-term stability, while integrating electrode design, encoding algorithms, subjective reports, and closed-loop calibration to promote more scalable sensory neuroprostheses. At the same time, a bidirectional translational framework could align human perceptual performance, animal mechanistic recordings, and computational models, so that perceptual quality, threshold drift, or perceptual fading observed in humans can become mechanistic clues for animal experiments, cross-species parameter calibration, and human ICMS parameter optimization [8,60]. 3. Translational positioning of ICMS within neuromodulation As ICMS moves from transient percept evocation toward plasticity modulation, it should be positioned within the broader neuromodulation landscape. DBS, VNS, RNS, TMS, and tDCS already occupy distinct roles across clinical and experimental settings, with different trade-offs in invasiveness, stimulation scale, spatial specificity, safety, and maturity [4–6,9,145]. In this context, ICMS primarily acts on local cortical microcircuits and is better positioned as a complement rather than a replacement for existing clinical neuromodulation methods. More feasible translational directions include local microcircuit probing, cortical information input, and closed-loop plasticity induction. ICMS may also provide a mechanistic reference for larger-scale neuromodulation such as DBS [60]. Current closed-loop and distributed stimulation systems already reflect this trend by linking recording, stimulation, timing rules, and behavioral or network outcomes [15,76,77,127–129,142,144,189,193]. For diseases characterized by homeostatic imbalances, such as epilepsy and PD, the effects of ICMS should be evaluated in terms of target scale, threshold reproducibility, decoding reliability, functional benefit, surgical risk, regulatory feasibility, and patient variability. Future studies should also include sham or active-stimulation controls, longer observation windows, and cross-species parameter mapping to determine whether ICMS can induce durable and beneficial plastic changes, intervene in disease progression, and avoid disrupting physiological circuit function [194,195]. 4. Safety and ethical constraints of ICMS Building on the interface reliability, stimulation strategy, and chronic-use constraints discussed above, the safety and ethical limitations of clinical ICMS require more comprehensive consideration. As an invasive technology, electrode implantation involves surgical risks, including hemorrhage, infection, and acute tissue injury, and long-term use also involves chronic tissue responses, threshold drift, and electrode failure [8,159]. During application, ICMS may also affect surrounding glial cells and pericytes, potentially supporting long-term use by modulating local vascular function and tissue responses, but this direction still lacks chronic human evidence [157,196]. Therefore, safety evaluation should run throughout ICMS application rather than being handled only at the final regulatory stage. Technical efforts such as standardized array implantation roadmaps and robot-assisted array implantation are being developed to reduce placement variability and procedural burden [3,36]. At the same time, because ICMS directly delivers signals to the brain as an information-input strategy in invasive BCIs, its applications should also consider informed consent, patient psychology, long-term device management, and posttrial responsibility [197], and should establish comprehensive and stringent standards for identity authentication, information encryption, and system protection to safeguard neural-data privacy [198]. 5. Controllability and governance of BNIs BNIs should be developed into controllable and governable platforms rather than remain conceptually attractive extensions of electrode technology. The development of BNIs reflects multiple technical routes: cell-seeded coating strategies aim to improve interface integration, axon-guidance strategies aim to support living signal bridging, and stem-cell- or organoid-integrated strategies aim to guide regenerative coupling [16,19,43,45–47,162,163,166,188]. Feasible supporting technologies include microfluidic control of the local environment, on-site oxygen generation and immunoprotection, wireless multifunctional living interfaces, vascularized organoids, and 3D flexible electrodes for organoid electrophysiology [199–202]. At the same time, more exploratory cell-based BCI concepts have started to appear in industry-oriented proposals. For example, a Neuralink-related patent proposes transplanting genetically modified neurons above cortical layer 1 (L1) to form an artificial cortical layer zero (L0), potentially enabling bidirectional communication through thin-film microelectrodes or related formats, making this layer an input/output layer for BCIs [203]. Furthermore, stringent safety evaluation systems must be established, including assessment of long-term survival of living components, immune compatibility, standardized manufacturing, regulatory classification, and ethical acceptability, with the goal of ensuring the safe and controllable integration of living grafts with the host cortex while guaranteeing the long-term quality and reliability of stimulation [47,48]. Further maturation and translation of ICMS require systematic validation across hardware interfaces, stimulation encoding, closed-loop modulation, disease-specific applications, and biointegration. Future studies should combine cross-species experiments, long-term follow-up, standardized safety evaluation, and reproducible behavioral and network outcomes, with criteria tailored to microelectrode materials, stimulation parameters, perceptual encoding, closed-loop algorithms, and BNIs. These studies should assess long-term usability, mechanistic interpretability, functional or therapeutic benefit, and safety to define the applicable settings and translational value of each technical route and build a continuous evidence chain from experimental microstimulation to chronically usable BCI information interfaces and precision neuromodulation.