A multi-region flexible neural interface for behavioral state decoding in freely moving mice

A multi-region flexible neural interface for behavioral state decoding in freely moving mice | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article A multi-region flexible neural interface for behavioral state decoding in freely moving mice Liuyang Sun, Ye Tian, Gen Li, Haoyang Su, Luyue Jiang, Yunfu Luo, and 9 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7980509/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 27 Apr, 2026 Read the published version in Microsystems & Nanoengineering → Version 1 posted 11 You are reading this latest preprint version Abstract High-density, long-term stable decoding of whole-brain function is crucial for advancing basic neuroscience research and developing neural disorder therapies. However, two major challenges remain: the lack of scalable interfaces capable of long-term, multi-regional recordings and the limited generalizability of existing decoding algorithms across days and individuals. Here, we developed an integrated platform that achieves accurate, stable, and generalizable decoding of behavioral states (resting, roaming, feeding and flash) with up to 89% accuracy. This platform combines multi-region flexible probes (MRFPs), enabling distributed recordings from 128 sites across eight brain regions over months, with a Conformer-based deep learning framework optimized for brain-wide neural dynamics. Comparative analyses demonstrate that distributed sampling, particularly from five or more regions, markedly enhances decoding performance over concentrated electrode configurations. Furthermore, the platform supports robust generalization across days and individuals without retraining, providing a practical solution for longitudinal and large-scale behavioral neuroscience studies. These results establish a foundation for stable, high-fidelity multi-region electrophysiology and offer a generalizable approach for decoding internal states from complex neural dynamics. Physical sciences/Engineering/Electrical and electronic engineering Physical sciences/Nanoscience and technology/Nanobiotechnology/Bionanoelectronics brain-computer interfaces multi-region flexible probes brain decoding behavioral decoding Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1. Introduction Brain-computer interfaces (BCIs) enable direct communication between neural circuits and external devices, providing insights into brain function and transformative applications for neurological disorders. BCIs have shown remarkable potential in decoding neural signals from single brain regions, such as the motor cortex for prosthetic control 1 – 5 , the visual cortex for vision restoration 6 – 10 , and Broca’s area for speech synthesis 11 – 15 . However, this focus on single regions limits their ability to capture the distributed neural dynamics underlying complex behavioral states, which are critical for both fundamental neuroscience and clinical translation. For example, abstract rule processing involves coordinated activity between thalamic nuclei and prefrontal cortex 16 , 17 . Likewise, neuropsychiatric disorders such as Alzheimer’s disease arise from distributed network dysfunction rather than isolated regional deficits 18 – 20 . These examples highlight the urgent need for neural interfaces that can decode interactions across multiple brain regions. A first challenge lies in neural signal acquisition. Decoding distributed neural interactions requires interfaces capable of stable, high-resolution recording across anatomically distinct brain regions in naturalistic settings. Optical and fMRI methods have shown promise for brain-wide decoding, but are limited to specific animal models 21 – 27 , such as zebrafish 28 – 30 , or constrained behavioral paradigms 31 , 32 , restricting their applicability to freely moving animals. Electrophysiology offers a direct strategy. However, conventional rigid probes, such as microwires or Utah Arrays, suffer from mechanical mismatch with brain tissue, leading to tissue damage and unstable recordings, while their bulky backends limit spatial coverage 33 , 34 . High-density probes like Neuropixels have expanded neural sampling capabilities, but their size still often requires head fixation in small animals, hindering naturalistic behavior 35 , 36 . Although flexible probes have improved surgical accessibility and recording stability, their potential for brain-wide recording remains largely unrealized 37 – 39 . A second challenge lies in neural signal decoding. Even with high-quality recordings, the ability of decoding methods to generalize across different days (cross-day) and individuals (cross-subject) is essential for widespread practical and clinical application 5 . However, achieving this goal faces obstacles due to the inherent and profound variability of neural signals 40 . Most current approaches remain confined to a “single-subject-single-model” paradigm, where a model is effective only for a specific individual within a specific timeframe, severely limiting its versatility and translational potential 11 , 13 . Deep learning provides a powerful framework for deciphering complex temporal dynamics, yet existing models face fundamental shortcomings. For example, convolutional neural networks (CNNs) are suitable for extracting local features but their fixed receptive fields struggle to capture relationships across extended time windows 41 – 43 . Recurrent neural networks (RNNs), while designed for sequential information, often fail to model long-range temporal dependencies due to issues like the vanishing gradient problem 44 – 46 . Hybrid models such as the EEG Conformer, which integrate CNN and Transformer architectures, have demonstrated state-of-the-art performance in EEG-based tasks 47 – 49 such as emotion recognition and motor imagery, yet their utility for decoding large-scale local field potentials (LFPs) in naturalistic behavior remains underexplored. In this work, we present an integrated platform that addresses both challenges. We designed a multi-region flexible probe (MRFP) with omega-shaped shanks for stable, long-term recordings across multiple brain regions in freely moving mice, and paired it with a modified Conformer-based AI model optimized for brain-wide dynamics (Fig. 1 ). This MRFP-AI framework achieved up to 89% decoding accuracy of behavioral states (resting, roaming, feeding and flash), with performance maintained at 85% across days and 70% across individuals. Optimal decoding performance was observed with a 4-second time window. Notably, distributed multi-region recordings consistently outperformed single-region configurations, underscoring the value of brain-wide sampling. By jointly enabling stable, scalable recordings and robust generalization in decoding, this platform reveals distributed neural contributions to behavior, providing a foundation for exploring brain-wide dynamics and advancing therapeutic strategies for neuropsychiatric disorders. 2. Results 2.1. Design and characterization of MRFP Conventional flexible neural probes are typically fabricated with needle-like, linear shanks to enable minimally invasive brain insertion 50 , 51 . However, their capacity to access multiple widely distributed regions is constrained by the low stretchability of common flexible materials such as polyimide (PI) or SU-8. This limitation poses two challenges: first, shanks are prone to mechanical failure under excessive tensile strain; second, after implantation, relative micromotion between the brain and skull may displace the shanks, resulting in the degradation of recording fidelity and signal stability 52 – 54 . To overcome these limitations, we designed an MRFP that introduces omega-shaped deformation elements into the probe structure, enabling substantial in-plane stretchability without compromising compactness or mechanical integrity. This design was developed as an optimization of our previously reported flexible electrode system, enhancing both its mechanical adaptability and spatial resolution 54 . The MRFP consists of eight shanks arranged linearly (Fig. 2 a), each integrated with 16 gold recording sites positioned in a staggered zig-zag layout (Fig. 2 b). Recording sites are spaced at 200 µm intervals, with an extended pitch of 266 µm incorporated on the two lateral shanks to enable access to deeper brain regions. Notably, the probe design is inherently customizable, and we developed a 16-shank variant, with eight electrodes per shank (Fig. S1 ), offering a scalable solution for applications in multi-region electrophysiology. The omega-shaped structure allows uniaxial stretch up to 50% without mechanical failure (Fig. S2). This stretchability is not derived from the material properties alone, but rather from the geometric reconfiguration of the structure under strain. By leveraging mechanical deformation in this manner, we achieved a compact and flexible design capable of conforming to the brain’s curved surface and accommodating micromotion over time. To demonstrate the MRFP’s implantation feasibility and anatomical coverage, we performed insertion tests in a mouse brain phantom composed of 0.8% agarose gel. In agarose brain phantoms, MRFP successfully achieved implantation across more than 1 cm of brain-mimicking tissue (Fig. 2 c), supporting its suitability for targeting distributed brain regions. Next, we evaluated the device’s mechanical durability and electrical performance through cyclic uniaxial stretching and electrochemical impedance spectroscopy. Individual shanks from three separate MRFPs were subjected to 100 and 300 stretch cycles under 50% strain. Impedance spectra showed minimal drift over time (Fig. 2 d), with mean impedances at 1 kHz of 1.34 ± 0.45 MΩ (pre-stretch), 1.43 ± 0.41 MΩ (100 cycles), and 1.34 ± 0.45 MΩ (300 cycles), consistent with stable gold-electrode performance. Long-term in vivo performance was assessed by tracking impedance over seven weeks post-implantation in Mice01, which still yielded 115 functional channels at the end of the session. As shown in Fig. 2 e, mean impedances remained stable over time (2.25 ± 0.17 MΩ on day 1; 2.26 ± 0.19 MΩ on day 7; 2.20 ± 0.19 MΩ on day 21; 2.14 ± 0.21 MΩ on day 49), and only five channel failures were observed (Fig. S3), confirming the MRFP’s suitability for chronic, high-density, multi-region recordings. 2.2. Implantation of MRFP The wide anatomical reach of the MRFP necessitates a more complex implantation procedure. While previous methods, such as pre-arranged tungsten wire array 55 , were effective for localized insertions but unsuitable for distributed architectures. We therefore developed a needle-threading-inspired strategy using tungsten wires to guide each shank (Fig. S4). Instead of full craniotomies, only microholes are drilled to minimize surgical damage. A tungsten wire with a sharp tip (via electrochemical etching) is threaded through a pinhole at the tip of the shank, allowing insertion with stereotaxic guidance. After implantation, the metal wire is retracted, leaving the shank in position. Figure 2 f illustrates a typical MRFP implantation targeting eight regions: M2 (AP + 1.69 mm, ML ± 1 mm), MnPO (AP + 0.25 mm, ML 0 mm), DLS (AP 0 mm, ML 3 mm), S1BF (AP − 1.31 mm, ML ± 2.5 mm), CA1 (AP − 2 mm, ML 1.8 mm), and V1 (AP − 2.69 mm, ML − 2.5 mm). 2.3. Recording environment and status identification A central goal of multi-region decoding is to capture the neural signatures representing both an animal’s internal states and its responses to external stimuli. We therefore designed a behavioral paradigm to explicitly span this internal-external spectrum. Mice engaged in three naturalistic, self-initiated behaviors (resting, roaming, and feeding), together with an externally triggered state induced by a visual flash. To enable the naturalistic expression and reliable labeling of these behaviors, we constructed a custom arena (30 × 30 × 30 cm 3 ) with a dedicated feeding location and an overhead flash lamp (Fig. 3 a). Behavioral tracking was performed on continuous video recordings using DeepLabCut 56 to extract positional data from key landmarks. Roaming was defined by coordinated movement of the head and tail across the arena, while resting was identified when both markers remained stationary. Feeding behavior was determined when the head entered a predefined zone around the food trough and remained there for a sustained period without concurrent tail movement. Flash status was triggered by activating the lamp at a frequency of 10 Hz, and lasted 5 minutes for each trial. Representative trajectories and movement distance of the head marker (Fig. 3 c) illustrate clear spatial and movement patterns distinguishing each state. This behavioral labeling strategy, based on self-initiated actions in an unconstrained setting, provided a reliable foundation for training and evaluating neural decoding models. Neural recordings were acquired from four mice, each continuously monitored over a 168-hour period to capture behavioral events used for subsequent model training. 2.4. LFP-Based neural network for status classification Transformer models, which use attention mechanisms to capture global dependencies, have demonstrated significant success in natural language processing and computer vision 57 . Their application has extended to EEG decoding, where they show promise in exploiting long-term temporal relationships. However, conventional Transformers often overlook local feature learning, which is essential for effective LFP decoding. To address this gap, we introduced an LFP Conformer (L-Conformer) architecture that originates from an open-source hybrid model, integrating the local feature extraction capability of CNNs with the global dependency modeling capability of Transformers 47 ; its architecture is illustrated in Fig. 3 d (see Section 4.6 for detailed information on the model’s specific structure and implementation). Optimized for multi-region LFP signals, this model can map neural activity to discriminative features corresponding to the four behavioral states, thereby laying a foundation for accurate decoding in subsequent analyses. 2.5. AI classification of behavioral states in mice Temporal resolution and model benchmarking for behavioral state decoding. We first evaluated whether combining MRFP-based multi-region neural recordings with the L-Conformer architecture could support accurate behavioral state classification. A key factor in training the decoder was the temporal window used to associate neural activity with behavioral labels. Short windows risked fragmenting sustained behaviors (e.g., feeding), while long windows increased the chance of including state transitions and label ambiguity. To determine the optimal resolution, we segmented the raw neural data into windows ranging from 0.1 s to 16 s and trained L-Conformer on each dataset. The results revealed a distinct inverted U-shaped relationship between decoding performance and window duration (Fig. 4 a). Specifically, validation accuracy (valid-acc) rose from 57.9% at a 0.1 s window to a peak of 89.49% at the 4 s window. The confusion matrix for a representative subject (Mice01) at this optimal 4 s window confirms high classification fidelity for all four states (Fig. 4 b). Conversely, extending the window beyond this peak led to a progressive decline in performance, with accuracy dropping to 75.46% at 16 s, likely due to the integration of non-relevant temporal information. This finding indicates that a 4 s time window most effectively captures the neural dynamics differentiating the “rest”, “roam”, “feed” and “flash” behavioral states in our case. Consequently, all subsequent analyses were conducted using a 4 s time window. To visualize the feature representations learned by the model, we projected the high-dimensional feature matrix into a three-dimensional space using t-distributed stochastic neighbor embedding (t-SNE) (Fig. 4 c). The resulting embedding reveals a distinct underlying structure. A majority of the data points form a continuous manifold, suggesting that the model has captured a smooth, progressive relationship among these samples. Notably, trajectories corresponding to the flash state formed a distinct and isolated cluster in the feature space, indicating that the model effectively captured the discriminative features of this state, resulting in a classification accuracy of 100%. The continuous color gradient is applied to the manifold simply to aid in visualizing its three-dimensional trajectory. We further evaluated several state-of-the-art neural decoding architectures, including Cebra, PatchTST, and EEGNet, to identify the most suitable model for our dataset. Among these, L-Conformer consistently achieved the highest classification accuracy (Fig. 4 d). This superior performance stems from its unique architectural design. Specifically, the Cebra model is primarily constructed for spike signals, whereas our study focuses on LFPs; the PatchTransformer model, while considering temporal information, only focuses on the relationships between signal patches, overlooking the fine-grained dynamics within them; and other models like EEGNet rely predominantly on convolutional modules, lacking effective modeling for long-range temporal dependencies. In contrast, our model integrates convolutional modules with a Transformer-based self-attention mechanism, enabling it to capture both local features and establish global temporal associations, thus achieving superior results. Spatial contributions of multi-region and distributed recordings. Neural decoding efforts over the past decades have mainly focused on motor control 58 – 61 , with recording sites typically concentrated within the M1. This concentration was largely constrained by the limited scalability of electrode technologies. As advances in electrode integration have substantially increased channel counts, a key question has emerged: whether additional recording sites should remain localized in a single brain region or be distributed across multiple anatomically distinct regions. After validation of L-Conformer, we firstly assessed the contribution of each brain regions to classification accuracy. As shown in Fig. 5 a, training the decoding model using data from a single brain region resulted in variable performance across regions. Incorporating data from all recorded brain regions markedly improved model performance, with averaged accuracy increasing to 87.77% (blue dotted line). For instance, when the decoder was trained only on signals from the left V1, accuracy for all behavioral states except flash was substantially reduced (Fig. 5 b), reflecting the limited capacity of single-region recordings. A leave-one-region-out analysis further confirmed that removing signals from any region decreased performance (Fig. 5 c), underscoring the importance of integrating multi-region signals for capturing distributed, behaviorally relevant dynamics. To compare concentrated and distributed decoding strategies, we constructed two datasets from the same recordings: one including only M1 signals, and another sampling from multiple distinct regions (pink bars). In both configurations, decoding accuracy improved as more channels were included (Fig. 5 d). At lower channel counts, the concentrated configuration achieved higher accuracy; however, once signals from more than five regions were included, the distributed configuration exceeded the concentrated one. A maximum accuracy of 84.37% was reached when signals from all available regions were utilized. Notably, the performance of the concentrated configuration plateaued, whereas the distributed configuration continued to yield incremental gains. These observations suggest that densely clustering electrodes within a single region may lead to signal redundancy with diminishing returns in decoding accuracy. To quantitatively support this notion, we fitted the decoding accuracies of both configurations using a Logistic growth model, a sigmoidal function commonly employed to characterize systems exhibiting saturation effects. The resulting curves revealed a saturating trend for the concentrated strategy when the number of recording sites exceeds six. In contrast, the distributed configuration maintained a steadily rising trajectory, suggesting that expanding coverage across additional brain regions may further enhance decoding accuracy. Generalization across days and individuals. Beyond the spatial configuration of electrode sites, a key limitation of conventional decoding algorithms is their dependence on session- and subject-specific calibration. For a given individual, neural recordings acquired on different days typically require recalibration to maintain decoding accuracy, limiting practical deployment across time. Moreover, due to substantial inter-individual variability in neural activity patterns, most models must be trained separately for each subject, hindering generalizability and posing a significant barrier to the development of scalable, cross-subject decoding frameworks. To evaluate the robustness of L-Conformer, we assessed the its ability to generalize across days and across individuals. In the cross-day paradigm, training was performed on neural recordings from all preceding days, and testing was conducted on a newly acquired session from the subsequent day without fine-tuning (Fig. 6 a). As shown in Fig. 6 b, accuracy started at 76% on day 2 and steadily improved with additional data, reaching 85% by day 5 and closely matching within-day performance (Fig. 6 c). These results demonstrate that L-Conformer supports robust cross-day decoding. Notably, high decoding accuracy of 85% was achieved using only four days of training data (approximately 1,500 samples), highlighting the model’s training efficiency. In the cross-individual paradigm, four mice were divided into a foundational group (n = 3) and a test group (n = 1). Neural recordings from the foundational group were used to train L-Conformer and generate a transferable core network containing key learned parameters. This foundational core enabled downstream construction of two testing strategies for new individuals: a zero-shot (ZS) approach and a conventional fine-tuning (FT) approach (Fig. 6 d). In the ZS strategy, the pretrained model was directly applied to data from the new subject without any re-calibration, while the FT strategy incorporated a proportion of the new subject’s data into the training set to improve accuracy. As shown in Fig. 6 e, the ZS approach achieved an average accuracy of ~ 70%, indicating baseline generalizability sufficient for practical use. FT further improved performance in most cases, with accuracy increasing in proportion to the amount of subject-specific data provided. These results highlight the model’s transferability and adaptability to previously unseen subjects. Notably, the ZS strategy required no retraining, reducing processing time by ~ 15 minutes under our current hardware setup and enabling faster deployment of cross-individual decoding. 3. Discussion In this study, we developed a MRFP-AI platform for decoding behavioral states in freely moving mice. The platform integrates MRFPs capable of large-scale, high-density electrophysiological recordings with a modified Conformer-based AI decoding framework. An omega-shaped structural design of MRFP provided in-plane stretchability without increasing device size or compromising implantability, enabling implantation across eight brain regions with high yield and chronic stability. Combined with the AI framework, this platform allows robust classification of behavioral states by decoding large-scale neural dynamics. Using this multi-region decoding platform, we addressed a key question in neural interface design: with limited electrode numbers, is it more effective to concentrate sampling within a single region or to distribute coverage across multiple regions? To answer this, we systematically compared the decoding performance using concentrated versus distributed electrode configurations. Our results show that multi-region recordings offer a clear advantage. While some regions, such as M1, achieved moderate classification performance alone, integrating signals across multiple regions markedly improved accuracy. When more than five brain areas were included, decoding accuracy using a distributed configuration surpassed that of concentrated strategies. These findings suggest that key features of behavioral states, including resting, roaming, feeding, and flash, are not localized but instead emerge from distributed network dynamics spanning cortical and subcortical structures. Furthermore, we demonstrated that performance gains plateaued in concentrated configurations after ~ 5 channels, highlighting redundancy within single regions and the importance of sampling diverse functional nodes. Beyond within-session decoding, we evaluated the platform’s robustness under cross-day and cross-individual conditions, two benchmarks essential for real-world deployment of neural decoders. In the cross-day setting, decoding accuracy improved steadily with accumulating data, reaching 85% by day 5. This suggests that while inter-day variability exists, it can be mitigated by accumulating neural context over time. In the cross-individual setting, our transfer learning framework achieved approximately 70% zero-shot accuracy, with further improvements through modest fine-tuning, demonstrating the model’s generalizability and potential for rapid deployment. Together, these findings represent both technical and conceptual advances. Technically, we demonstrate that high-density, spatially distributed recordings can be achieved using a flexible and scalable platform suitable for chronic behavioral studies. Conceptually, our results highlight the importance of accessing multiple functionally distinct brain regions to decode global behavioral states. This distributed sampling strategy enables a systems-level view of neural dynamics, laying the foundation for deeper mechanistic investigations into large-scale brain function. 4. Methods 4.1. Fabrication of MRFP The neural probe was fabricated by standard MEMS process (Fig. S4). First, the wafer was cleaned using RCA process. After photolithography, we adopted electron beam evaporating technique to deposit 100 nm thick Ni on the wafer, and the sacrificial layer was patterned via the lift-off process. Then 1 µm thick PI was spin coated and solidified in a 350°C nitrogen oven for 10 h as the bottom insulation layer. After that, we fabricated the recording sites and interconnects following photolithography, electron beam evaporation and lift-off, including a sandwich metal layer containing 5 nm Ti, 100 nm Au and 5 nm Ti. Backend pads for bonding were formed through the same process, including a 5 nm Ti layer, a 150 nm Ni layer and a 50 nm Au layer. Another 1 µm PI layer was cured on top of the wafer as top encapsulation. Before PI etching, a 100 nm Al layer was deposited by sputtering, patterned by photolithography and wet etching to form a hard mask. Finally, we employed RIE to expose the electrodes and pads, and etched out profile of the probe. 4.2. Surgical procedures All procedures were conducted in accordance with institutional animal care guidelines and approved protocols. Mice were initially anesthetized in an induction chamber using 4–5% isoflurane. Once anesthetized, the fur on the scalp was shaved, and animals were placed in a stereotaxic frame (RWD Life Science Co., Ltd) equipped with a temperature-regulated heating pad. To prevent corneal drying and injury, erythromycin ophthalmic ointment was applied to the eyes using a sterile cotton swab. Anesthesia was maintained with 2.5% isoflurane for the first 10 minutes and subsequently reduced to 1% for the remainder of the surgical procedure. Upon reaching a stable anesthetic plane, a trapezoid-shaped incision was made to expose the skull. The overlying fascia and periosteum were carefully removed using fine scissors, forceps, and sterile cotton swabs. Implantation sites were identified and marked using a stereotaxic coordinate system. Burr holes were drilled at designated locations with a dental drill. Flexible shanks of the MRFP were positioned above the burr holes using fine tweezers and inserted into the brain tissue alongside a tungsten wire for structural support. Once the desired implantation depth was reached, the tungsten wire was withdrawn, leaving the flexible shanks in place. A steel wire was inserted to serve as a reference electrode and soldered to the reference pad of the probe’s PCB using low-temperature solder. The probe assembly was secured to the skull with dental acrylic to ensure long-term mechanical stability. Postoperative analgesia was administered via subcutaneous injection of meloxicam (5 mg/kg) for pain management. Mice were allowed to recover for at least seven days prior to device testing and neural data acquisition. 4.3. Data preprocessing The raw LFP data is structured as a matrix with dimensions C×T, where C represents the number of electrode channels and T denotes the number of time samples. Initially, the raw data were bandpass filtered to isolate neural signals within the 0.1–250 Hz frequency band, coupled with notch filtering to specifically attenuate 50 Hz line noise contamination, and then z-score normalized to reduce variability and non-stationarity. The filtered data is normalized using the following equation: $$\:{\mathcal{X}}_{0}=\frac{{\mathcal{X}}_{\mathcal{i}}-\mathcal{u}}{\sqrt{{\sigma\:}^{2}}}$$ , where \(\:{\mathcal{X}}_{\mathcal{i}}\) is the bandpass-filtered data, and \(\:{\mathcal{X}}_{0}\) is the normalized output. The mean ( \(\:\mathcal{u}\) ) and standard deviation ( \(\:\sigma\:\) ) are calculated exclusively from the training dataset and subsequently applied to the test dataset. 4.4. Feature visualization through low-dimensional embedding and geodesic gradient mapping To investigate the internal feature representations learned by the model and intuitively illustrate the topological relationships between distinct neural states, we implemented a multi-step visualization pipeline. First, we extracted the output logits of each EEG window from the final hidden layer of the trained EEG Conformer model, just prior to classification. These high-dimensional feature vectors were then projected into a three-dimensional space using the t-distributed Stochastic Neighbor Embedding (t-SNE) algorithm, with the target dimensionality set to 3. This transformation yielded a set of 3D coordinates (x, y, z), with each point corresponding to a single EEG segment. Next, we constructed a k-nearest neighbor (k-NN) graph to encode the local structure of the embedded space. Each point was treated as a node, and edges were created between each node and its 15 nearest neighbors in the 3D Euclidean space. To establish an unbiased origin for color mapping, we first calculated the geometric centroid of the entire point cloud. The node farthest from this centroid was designated as the starting node of the manifold traversal. From this starting point, we applied Dijkstra’s algorithm to compute the shortest path lengths along the k-NN graph to all other nodes. These geodesic distances reflected the accumulated path length along the intrinsic data manifold, rather than direct Euclidean distances. For isolated nodes not connected to the graph (i.e., with infinite path lengths), we assigned the maximum finite distance in the dataset to preserve continuity in the colormap mapping. The resulting distance values were linearly normalized to the [0, 1] range using min–max scaling, and the normalized values were then mapped to a continuous colormap (rainbow gradient). This procedure allowed each point in the final 3D visualization to be color-coded based on its relative geodesic distance from the origin node. This color-gradient embedding provided strong visual evidence for how the model organized and separated distinct neural states within its learned representation space. 4.5. Animals In all the experiments, the mice we used were adult ICR mice (30–40g at start of experiments) provided by Phenotek Biotechnology Co., Ltd. All mice were aged between 8 and 10 weeks prior to commencement of experiments and were maintained on a 12-h lightdark cycle with ad libitum access to food and water. All our animal experiments obtained ethical approval from the Ethics Committee for Animal Management at the Phenotek Biotechnology (approval number 20240708-01). Our experimental conditions and handling procedures for mice complied with the ethical requirements for experimental animals. 4.6. L-Conformer Model Architecture L-Conformer integrates the local feature extraction capability of CNNs and the global temporal dependency modeling capability of Transformers. The overall framework of the L-Conformer, as illustrated in Fig. 3 d, comprises three primary components: a convolutional module, a self-attention module, and a fully-connected classifier. Convolutional module. To efficiently extract low-level spatiotemporal features, we employed a two-stage convolutional encoder. A temporal convolution layer with 40 kernels of size 1×25 and a stride of 1×1 captures time-dependent patterns, followed by a spatial convolution layer with 40 kernels of size C×1 and a stride of 1×1 to model inter-electrode interactions. Batch normalization is applied after these convolutions to accelerate training and mitigate overfitting. The Exponential Linear Unit (ELU) is used as the activation function. Following this, an average pooling layer with a kernel size of 1×75 and a stride of 1×15 is employed. This layer smooths the temporal features, reduces computational complexity, and further minimizes the risk of overfitting. Finally, the feature maps from the convolutional module are reshaped by compressing the electrode channel dimension and transposing the feature channel with the temporal dimension. This rearrangement treats all feature channels at a given timestep as a unified “token” for the subsequent attention module. This approach is based on the premise that leveraging context-dependent representations within low-level spatiotemporal features enhances EEG decoding by capitalizing on the coherence of neural activity. Self-Attention Module. To complement the limited receptive field of the convolutional module, a self-attention module is used to learn global temporal dependencies across the feature tokens. The tokens from the convolutional module are linearly projected into three matrices of equal dimension: Query (Q), Key (K), and Value (V). The core of the attention mechanism involves computing the dot product between the Query and Key matrices to determine inter-token correlations. A scaling factor, \(\:\frac{1}{\sqrt{{d}_{k}}}\) (where \(\:{d}_{k}\) is the dimension of the key), is applied to prevent vanishing gradients and ensure training stability. The resulting attention weights are obtained via a Softmax function and are then applied to the Value matrix to yield the final output, as described by the formula: \(\:Attention(Q,K,V)=Softmax\left(\frac{\text{Q}{\text{K}}^{\text{T}}}{\sqrt{{\text{d}}_{\text{k}}}}\right)\) V To further enhance the model’s representational power, a multi-head attention (MHA) mechanism is implemented. The input tokens are partitioned into h segments, each processed independently by a separate self-attention submodule. The outputs of these parallel heads are then concatenated to form the final representation: $$\:MHA(Q,\:K,\:V)=Concat(\:{\text{h}\text{e}\text{a}\text{d}}_{0},\:\dots\:\:,{\:\text{h}\text{e}\text{a}\text{d}}_{\text{h}-1})$$ , where \(\:{\:\text{h}\text{e}\text{a}\text{d}}_{\text{i}}=\text{A}\text{t}\text{t}\text{e}\text{n}\text{t}\text{i}\text{o}\text{n}({\text{Q}}_{\text{i}},{\text{K}}_{\text{i}},{\text{V}}_{\text{i}}\) ). Classifier. The final component is a classifier module, which consists of two fully-connected layers. This module receives the output from the self-attention block and generates an M-dimensional output vector, corresponding to the M possible classes. A Softmax transformation is applied to this vector to produce class probabilities. The entire framework is optimized end-to-end using the cross-entropy loss function. The entire L-Conformer framework is optimized end-to-end using the cross-entropy loss function, which minimizes the difference between the predicted class probabilities and the true behavioral state labels (one-hot encoded). This optimization strategy ensures that the model learns discriminative features directly aligned with the task of behavioral state decoding. Declarations Acknowledgments This work was supported by National Key R & D Program of China grant 2021ZD0201600 and Young Scientists Fund of the National Natural Science Foundation of China grant 62305368. Author contributions These authors contributed equally to this work: Ye Tian, Gen Li, Haoyang Su. Ye Tian led the experimental design, flexible electrode design and fabrication, and animal experiments. Gen Li led the data processing, neural network construction, training and testing, and data visualization. Haoyang Su led the data preprocessing and system design. Luyue Jiang, Yunfu Luo, Yingkang Yang, Lei Huang, Jiazhi Li, and Shuang Jin contributed to animal experiments and device design. Peijie Chen and Yiming Gao contributed to system fabrication and device fabrication. Yike Xiang and Yi Wei contributed to neural network construction. Yifei Ye and Liuyang Sun designed and supervised the project. All authors discussed the results and commented on the manuscript at all stages. Data availability The data that support the findings of this study are available from the corresponding author upon reasonable request. Competing Interests The authors declare that they have no conflict of interest. References Flesher, S. N. et al. 1-A brain-computer interface that evokes tactile sensations improves robotic arm control. SCIENCE 372, 831-+ (2021). Hochberg, L. R. et al. Neuronal ensemble control of prosthetic devices by a human with tetraplegia. Nature 442, 164–171 (2006). Hochberg, L. R. et al. Reach and grasp by people with tetraplegia using a neurally controlled robotic arm. NATURE 485, 372-U121 (2012). Handelman, D. A. et al. Shared Control of Bimanual Robotic Limbs With a Brain-Machine Interface for Self-Feeding. FRONTIERS IN NEUROROBOTICS 16, 918001 (2022). Wolpaw, J. R., Birbaumer, N., McFarland, D. J., Pfurtscheller, G. & Vaughan, T. M. Brain–computer interfaces for communication and control. Clinical Neurophysiology 113, 767–791 (2002). Beauchamp, M. S. et al. Dynamic Stimulation of Visual Cortex Produces Form Vision in Sighted and Blind Humans. CELL 181, 774-+ (2020). Nirenberg, S. & Pandarinath, C. Retinal prosthetic strategy with the capacity to restore normal vision. PROCEEDINGS OF THE NATIONAL ACADEMY OF SCIENCES OF THE UNITED STATES OF AMERICA 109, 15012–15017 (2012). Foroushani, A. N., Pack, C. C. & Sawan, M. Cortical visual prostheses: from microstimulation to functional percept. JOURNAL OF NEURAL ENGINEERING 15, 021005 (2018). Fernandez, E. et al. Visual percepts evoked with an intracortical 96-channel microelectrode array inserted in human occipital cortex. JOURNAL OF CLINICAL INVESTIGATION 131, 151331 (2021). Niketeghad, S. & Pouratian, N. Brain Machine Interfaces for Vision Restoration: The Current State of Cortical Visual Prosthetics. NEUROTHERAPEUTICS 16, 134–143 (2019). Willett, F. R. et al. A high-performance speech neuroprosthesis. NATURE 620, (2023). Metzger, S. L. et al. A high-performance neuroprosthesis for speech decoding and avatar control. NATURE 620, 1037-+ (2023). Card, N. S. et al. An Accurate and Rapidly Calibrating Speech Neuroprosthesis. NEW ENGLAND JOURNAL OF MEDICINE 391, 609–618 (2024). Moses, D. A., Leonard, M. K., Makin, J. G. & Chang, E. F. Real-time decoding of question-and-answer speech dialogue using human cortical activity. NATURE COMMUNICATIONS 10, 3096 (2019). Wahl, M. et al. The Human Thalamus Processes Syntactic and Semantic Language Violations. Neuron 59, 695–707 (2008). Baehner, F. et al. Abstract rule learning promotes cognitive flexibility in complex environments across species. NATURE COMMUNICATIONS 16, 5396 (2025). Phillips, J. M. et al. Primate thalamic nuclei select abstract rules and shape prefrontal dynamics. NEURON 113, 2014–2027 (2025). Tayebi, H. et al. Applications of brain-computer interfaces in neurodegenerative diseases. NEUROSURGICAL REVIEW 46, 131 (2023). Belkacem, A. N., Jamil, N., Palmer, J. A., Ouhbi, S. & Chen, C. Brain Computer Interfaces for Improving the Quality of Life of Older Adults and Elderly Patients. FRONTIERS IN NEUROSCIENCE 14, 692 (2020). Davenport, F. et al. Neurodegenerative disease of the brain: a survey of interdisciplinary approaches. JOURNAL OF THE ROYAL SOCIETY INTERFACE 20, 20220406 (2023). Cayce, J. M. et al. Calcium imaging of infrared-stimulated activity in rodent brain. CELL CALCIUM 55, 183–190 (2014). Boyden, E. S., Zhang, F., Bamberg, E., Nagel, G. & Deisseroth, K. Millisecond-timescale, genetically targeted optical control of neural activity. NATURE NEUROSCIENCE 8, 1263–1268 (2005). Fenno, L., Yizhar, O. & Deisseroth, K. The Development and Application of Optogenetics. in ANNUAL REVIEW OF NEUROSCIENCE, VOL 34 (eds. Hyman, S. E., Jessell, T. M., Shatz, C. J., Stevens, C. F. & Zoghbi, H. Y.) vol. 34 389–412 (2011). Bai, L. et al. Volumetric Imaging of Neural Activity by Light Field Microscopy. NEUROSCIENCE BULLETIN 38, 1559–1568 (2022). Dong, C. et al. Fluorescence Imaging of Neural Activity, Neurochemical Dynamics, and Drug-Specific Receptor Conformation with Genetically Encoded Sensors. ANNUAL REVIEW OF NEUROSCIENCE 45, 273–294 (2022). Logothetis, N. K., Pauls, J., Augath, M., Trinath, T. & Oeltermann, A. Neurophysiological investigation of the basis of the fMRI signal. NATURE 412, 150–157 (2001). Critchley, H. D., Wiens, S., Rotshtein, P., Öhman, A. & Dolan, R. J. Neural systems supporting interoceptive awareness. NATURE NEUROSCIENCE 7, 189–195 (2004). Yang, C. et al. A population code for spatial representation in the zebrafish telencephalon. NATURE 634, 397-+ (2024). Renninger, S. L. & Orger, M. B. Two-photon imaging of neural population activity in zebrafish. METHODS 62, 255–267 (2013). Marques, J. C., Li, M., Schaak, D., Robson, D. N. & Li, J. M. Internal state dynamics shape brainwide activity and foraging behaviour. NATURE 577, 239-+ (2020). Delafontaine-Martel, P. et al. Whole brain vascular imaging in a mouse model of Alzheimer’s disease with two-photon microscopy. JOURNAL OF BIOMEDICAL OPTICS 23, 076501 (2018). Johannssen, H. C. & Helmchen, F. Two-photon imaging of spinal cord cellular networks. EXPERIMENTAL NEUROLOGY 242, 18–26 (2013). Normann, R. A. & Fernandez, E. Clinical applications of penetrating neural interfaces and Utah Electrode Array technologies. JOURNAL OF NEURAL ENGINEERING 13, 061003 (2016). Maynard, E. M., Nordhausen, C. T. & Normann, R. A. The Utah Intracortical Electrode Array: A recording structure for potential brain-computer interfaces. ELECTROENCEPHALOGRAPHY AND CLINICAL NEUROPHYSIOLOGY 102, 228–239 (1997). Paulk, A. C. et al. Large-scale neural recordings with single neuron resolution using Neuropixels probes in human cortex. NATURE NEUROSCIENCE 25, 252-+ (2022). Chung, J. E. et al. High-density single-unit human cortical recordings using the Neuropixels probe. NEURON 110, 2409-+ (2022). Guan, S. et al. Elastocapillary self-assembled neurotassels for stable neural activity recordings. SCIENCE ADVANCES 5, eaav2842 (2019). Luan, L. et al. Ultraflexible nanoelectronic probes form reliable, glial scar-free neural integration. SCIENCE ADVANCES 3, e1601966 (2017). Zhou, T. et al. Syringe-injectable mesh electronics integrate seamlessly with minimal chronic immune response in the brain. PROCEEDINGS OF THE NATIONAL ACADEMY OF SCIENCES OF THE UNITED STATES OF AMERICA 114, 5894–5899 (2017). Pachitariu, M., Sridhar, S., Pennington, J. & Stringer, C. Spike sorting with Kilosort4. Nat Methods 21, 914–921 (2024). Acharya, U. R., Oh, S. L., Hagiwara, Y., Tan, J. H. & Adeli, H. Deep convolutional neural network for the automated detection and diagnosis of seizure using EEG signals. COMPUTERS IN BIOLOGY AND MEDICINE 100, 270–278 (2018). Schirrmeister, R. T. et al. Deep Learning With Convolutional Neural Networks for EEG Decoding and Visualization. HUMAN BRAIN MAPPING 38, 5391–5420 (2017). Lawhern, V. J. et al. EEGNet: a compact convolutional neural network for EEG-based brain-computer interfaces. JOURNAL OF NEURAL ENGINEERING 15, 056013 (2018). Tortora, S., Ghidoni, S., Chisari, C., Micera, S. & Artoni, F. Deep learning-based BCI for gait decoding from EEG with LSTM recurrent neural network. JOURNAL OF NEURAL ENGINEERING 17, 046011 (2020). Mante, V., Sussillo, D., Shenoy, K. V. & Newsome, W. T. Context-dependent computation by recurrent dynamics in prefrontal cortex. NATURE 503, 78-+ (2013). Güler, N. F., Übeyli, E. D. & Güler, I. Recurrent neural networks employing Lyapunov exponents for EEG signals classification. EXPERT SYSTEMS WITH APPLICATIONS 29, 506–514 (2005). Song, Y., Zheng, Q., Liu, B. & Gao, X. EEG Conformer: Convolutional Transformer for EEG Decoding and Visualization. IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING 31, 710–719 (2023). Polikov, V. S., Tresco, P. A. & Reichert, W. M. Response of brain tissue to chronically implanted neural electrodes. JOURNAL OF NEUROSCIENCE METHODS 148, 1–18 (2005). Abiri, R., Borhani, S., Sellers, E. W., Jiang, Y. & Zhao, X. A comprehensive review of EEG-based brain-computer interface paradigms. JOURNAL OF NEURAL ENGINEERING 16, 011001 (2019). Musk, E. An Integrated Brain-Machine Interface Platform With Thousands of Channels. JOURNAL OF MEDICAL INTERNET RESEARCH 21, e16194 (2019). Gu, C., Jiang, J., Tao, T. H., Wei, X. & Sun, L. Long-term flexible penetrating neural interfaces: materials, structures, and implantation. SCIENCE CHINA-INFORMATION SCIENCES 64, 221401 (2021). Kim, M. G., Yu, K., Niu, X. & He, B. Investigation of Displacement of Intracranial Electrode Induced by Focused Ultrasound Stimulation. IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT 70, 9600509 (2021). Peng, L. et al. Biomechanics characterization of an implantable ultrathin intracortical electrode through finite element method. SCIENTIFIC REPORTS 15, 19938 (2025). Cheng, Q. et al. High-Resolution Recording of Neural Activity in Epilepsy Using Flexible Neural Probes. ADVANCED MATERIALS TECHNOLOGIES (2023) doi: 10.1002/admt.202301210 . Zhao, Z. et al. Parallel, minimally-invasive implantation of ultra-flexible neural electrode arrays. JOURNAL OF NEURAL ENGINEERING 16, 035001 (2019). Mathis, A. et al. DeepLabCut: markerless pose estimation of user-defined body parts with deep learning. NATURE NEUROSCIENCE 21, 1281-+ (2018). Lee, Y.-E. & Lee, S.-H. EEG-Transformer: Self-attention from Transformer Architecture for Decoding EEG of Imagined Speech. 10TH INTERNATIONAL WINTER CONFERENCE ON BRAIN-COMPUTER INTERFACE (BCI2022) (2022) doi: 10.1109/BCI53720.2022.9735124 . Oh, S. et al. A stealthy neural recorder for the study of behaviour in primates. NATURE BIOMEDICAL ENGINEERING 9, (2025). Collinger, J. L. et al. High-performance neuroprosthetic control by an individual with tetraplegia. LANCET 381, 557–564 (2013). Kuiken, T. A. et al. Targeted Muscle Reinnervation for Real-time Myoelectric Control of Multifunction Artificial Arms. JAMA-JOURNAL OF THE AMERICAN MEDICAL ASSOCIATION 301, 619–628 (2009). Musallam, S., Corneil, B. D., Greger, B., Scherberger, H. & Andersen, R. A. Cognitive control signals for neural prosthetics. SCIENCE 305, 258–262 (2004). Additional Declarations There is no conflict of interest Supplementary Files 1028MINESI.docx A multi-region flexible neural interface for behavioral state decoding in freely moving mice Cite Share Download PDF Status: Published Journal Publication published 27 Apr, 2026 Read the published version in Microsystems & Nanoengineering → Version 1 posted Editorial decision: revise 04 Jan, 2026 Review # 3 received at journal 31 Dec, 2025 Review # 1 received at journal 28 Dec, 2025 Review # 2 received at journal 23 Dec, 2025 Reviewer # 3 agreed at journal 11 Dec, 2025 Reviewer # 2 agreed at journal 11 Dec, 2025 Reviewer # 1 agreed at journal 11 Dec, 2025 Reviewers invited by journal 11 Dec, 2025 Submission checks completed at journal 30 Oct, 2025 Editor assigned by journal 29 Oct, 2025 First submitted to journal 29 Oct, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7980509","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":558642080,"identity":"05239c54-e8d8-4dcc-a4e5-99a6d82635f9","order_by":0,"name":"Liuyang Sun","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABDklEQVRIiWNgGAWjYBACPmYEm/EBA4OEDJBuwKuFDUkLswFQCw9hLchsCSDBQ9BhbOw8ZhI/d9QyyM/IPVZ1o8aCh4H/cOsGhho7Bv7Z2G1jY+Yxk+w9c5zB4EZe2u2cY0CHSSS23WA4lswgcecATi0SvG3HGAwkcsxu57CBtDACtbAdAIok4LblL1CL/Iwcs+Kcf0At/AeBWv7h1yLN21bDwHAjx4w5tw0UYkCHMbbh08JWbC3bdoDH4MwbY+ncPgkeNpBfEvuSeSRuYNfCz3944823bXVy8u05hp9zvtXJ8fMff3bjwzc7Of4Z2LUAAQswOg4jogMcUwn4I4j5AwNDHR75UTAKRsEoGPEAAEGwTTwwHJq+AAAAAElFTkSuQmCC","orcid":"https://orcid.org/0000-0001-9896-4382","institution":"Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences","correspondingAuthor":true,"prefix":"","firstName":"Liuyang","middleName":"","lastName":"Sun","suffix":""},{"id":558642081,"identity":"3edc6d34-2173-4430-ac5c-1a6aebdd8427","order_by":1,"name":"Ye Tian","email":"","orcid":"","institution":"Shanghai Institute of Microsystem and Information Technology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Ye","middleName":"","lastName":"Tian","suffix":""},{"id":558642082,"identity":"ef9ec607-b6dd-4a01-8521-89c069bfd5d9","order_by":2,"name":"Gen Li","email":"","orcid":"","institution":"Shanghai Institute of Microsystem and Information Technology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Gen","middleName":"","lastName":"Li","suffix":""},{"id":558642083,"identity":"f8a7eeb9-20bf-4731-b38b-b710a1d47cf8","order_by":3,"name":"Haoyang Su","email":"","orcid":"","institution":"Shanghai Institute of Microsystem and Information Technology, CAS","correspondingAuthor":false,"prefix":"","firstName":"Haoyang","middleName":"","lastName":"Su","suffix":""},{"id":558642084,"identity":"ee578e4e-7c12-4b9d-959b-9be7ffc99177","order_by":4,"name":"Luyue Jiang","email":"","orcid":"","institution":"Shanghai Institute of Microsystem and Information Technology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Luyue","middleName":"","lastName":"Jiang","suffix":""},{"id":558642085,"identity":"beea37b5-2ee8-498b-8ec8-e8477aa9dded","order_by":5,"name":"Yunfu Luo","email":"","orcid":"","institution":"Shanghai Institute of Microsystem and Information Technology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Yunfu","middleName":"","lastName":"Luo","suffix":""},{"id":558642086,"identity":"fa6da665-e97f-43da-b81c-3a0643ced9d7","order_by":6,"name":"Yingkang Yang","email":"","orcid":"","institution":"Shanghai Institute of Microsystem and Information Technology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Yingkang","middleName":"","lastName":"Yang","suffix":""},{"id":558642087,"identity":"3524482d-da79-4649-9496-90bd7cc26e02","order_by":7,"name":"Lei Huang","email":"","orcid":"","institution":"Shanghai Institute of Microsystem and Information Technology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Lei","middleName":"","lastName":"Huang","suffix":""},{"id":558642088,"identity":"d11bc8e4-9e64-4de8-b9d1-06550474b680","order_by":8,"name":"Jiazhi Li","email":"","orcid":"","institution":"Shanghai Institute of Microsystem and Information Technology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Jiazhi","middleName":"","lastName":"Li","suffix":""},{"id":558642089,"identity":"5f8db4c2-951f-4c9e-9d1d-02291026c933","order_by":9,"name":"Shuang Jin","email":"","orcid":"","institution":"Shanghai Institute of Microsystem and Information Technology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Shuang","middleName":"","lastName":"Jin","suffix":""},{"id":558642090,"identity":"4b7820e6-ba49-4f51-af52-847ab9b51f01","order_by":10,"name":"Peijie Chen","email":"","orcid":"","institution":"Shanghai Institute of Microsystem and Information Technology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Peijie","middleName":"","lastName":"Chen","suffix":""},{"id":558642091,"identity":"5578f739-0355-4473-95aa-ad0c844a572a","order_by":11,"name":"Yiming Gao","email":"","orcid":"","institution":"Shanghai Institute of Microsystem and Information Technology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Yiming","middleName":"","lastName":"Gao","suffix":""},{"id":558642092,"identity":"5c8b7a78-fc3c-4947-bc68-ea0ee5ab2acf","order_by":12,"name":"Yike Xiang","email":"","orcid":"","institution":"Shanghai Institute of Microsystem and Information Technology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Yike","middleName":"","lastName":"Xiang","suffix":""},{"id":558642093,"identity":"c1d131f4-7511-4852-b740-2f6e5f5294a6","order_by":13,"name":"Yi Wei","email":"","orcid":"","institution":"Shanghai Institute of Microsystem and Information Technology Chinese Academy of Sciences","correspondingAuthor":false,"prefix":"","firstName":"Yi","middleName":"","lastName":"Wei","suffix":""},{"id":558642094,"identity":"6eb7632a-2681-4bb1-82a3-aad76bdc9329","order_by":14,"name":"Yifei Ye","email":"","orcid":"","institution":"Shanghai Institute of Microsystem and Information Technology, CAS","correspondingAuthor":false,"prefix":"","firstName":"Yifei","middleName":"","lastName":"Ye","suffix":""}],"badges":[],"createdAt":"2025-10-29 13:42:18","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7980509/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7980509/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41378-026-01258-5","type":"published","date":"2026-04-27T04:00:00+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":98300230,"identity":"7447bb72-4442-4fd6-922d-40672136b7dd","added_by":"auto","created_at":"2025-12-16 10:01:02","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":20876534,"visible":true,"origin":"","legend":"","description":"","filename":"1028MINEAmultiregionflexibleneuralinterfaceforbehavioralstatedecodinginfreelymovingmice.docx","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/9d753b98b90395eaf2cae78f.docx"},{"id":98436895,"identity":"8c2d082a-7a60-4b46-9c7f-8e7c3d8fc06e","added_by":"auto","created_at":"2025-12-17 16:56:25","extension":"png","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1762368,"visible":true,"origin":"","legend":"","description":"","filename":"figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/7a4d0f0a5e884b899278a608.png"},{"id":98436548,"identity":"83f48b28-ea0b-439c-8488-8cb610e37e54","added_by":"auto","created_at":"2025-12-17 16:55:51","extension":"png","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10206150,"visible":true,"origin":"","legend":"","description":"","filename":"figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/450107863af01dbe91a37d59.png"},{"id":98436532,"identity":"21753968-f168-478b-9f26-c2a938ed6f3a","added_by":"auto","created_at":"2025-12-17 16:55:49","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6522240,"visible":true,"origin":"","legend":"","description":"","filename":"figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/50ce4815d4ec24e99cc7e1ba.png"},{"id":98300221,"identity":"08588fde-b172-4973-afba-12f580887bf1","added_by":"auto","created_at":"2025-12-16 10:01:02","extension":"jpg","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":660585,"visible":true,"origin":"","legend":"","description":"","filename":"figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/c51787d968eb2b0ea0fbc9b2.jpg"},{"id":98300227,"identity":"72c906e8-3b97-4a3a-b0fe-52c693688ce4","added_by":"auto","created_at":"2025-12-16 10:01:02","extension":"jpg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":796747,"visible":true,"origin":"","legend":"","description":"","filename":"figure5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/f6eede71c972825770a5921c.jpg"},{"id":98300228,"identity":"16ee40df-d582-4bb2-9c33-26546bd7c62a","added_by":"auto","created_at":"2025-12-16 10:01:02","extension":"jpg","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":832004,"visible":true,"origin":"","legend":"","description":"","filename":"figure6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/cb6508916ce702372d536ee4.jpg"},{"id":98435865,"identity":"786c636a-26f2-46c5-8542-13877db4c52a","added_by":"auto","created_at":"2025-12-17 16:54:32","extension":"json","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":14684,"visible":true,"origin":"","legend":"","description":"","filename":"MICRONANO05083.json","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/be02615cb017dc01a6c1472d.json"},{"id":98300234,"identity":"0e4107eb-a906-41e9-8d8b-371a2ce40fba","added_by":"auto","created_at":"2025-12-16 10:01:02","extension":"docx","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4442569,"visible":true,"origin":"","legend":"","description":"","filename":"1028MINESI.docx","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/466c3144a4f4960153dba8fa.docx"},{"id":98300236,"identity":"e721cf7f-0f23-4c3c-a5e7-193b97588c25","added_by":"auto","created_at":"2025-12-16 10:01:02","extension":"xml","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":138718,"visible":true,"origin":"","legend":"","description":"","filename":"MICRONANO050830enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/9a966866f6b40fcd0e66aba2.xml"},{"id":98300232,"identity":"6f5396cf-233b-4496-9a33-f22541755b7f","added_by":"auto","created_at":"2025-12-16 10:01:02","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1762368,"visible":true,"origin":"","legend":"","description":"","filename":"figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/b01d69a02c2725ff44d00f7b.png"},{"id":98434632,"identity":"a1c8b00a-490d-4633-a6eb-c512ffdf070d","added_by":"auto","created_at":"2025-12-17 16:52:23","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10206150,"visible":true,"origin":"","legend":"","description":"","filename":"figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/04f52b0c59d7f68a2342bcc4.png"},{"id":98300244,"identity":"95542a7e-da17-42a4-9d9e-c09b49c01b33","added_by":"auto","created_at":"2025-12-16 10:01:02","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6522240,"visible":true,"origin":"","legend":"","description":"","filename":"figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/96b0c153294e89a149044480.png"},{"id":98300252,"identity":"a5a23929-06ec-4326-a5ed-f95bb9fbb56e","added_by":"auto","created_at":"2025-12-16 10:01:03","extension":"jpg","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":660585,"visible":true,"origin":"","legend":"","description":"","filename":"figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/29c127a19041304cb4968425.jpg"},{"id":98435233,"identity":"0585b9a0-1104-4d8f-9557-45501770e417","added_by":"auto","created_at":"2025-12-17 16:53:22","extension":"jpg","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":796747,"visible":true,"origin":"","legend":"","description":"","filename":"figure5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/891f7eedcf6047b3e6d737f2.jpg"},{"id":98436466,"identity":"9d3827f9-ae4d-45e9-8f1f-f495a6f96da1","added_by":"auto","created_at":"2025-12-17 16:55:42","extension":"jpg","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":832004,"visible":true,"origin":"","legend":"","description":"","filename":"figure6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/48efbae15b3a175afdd7ae4f.jpg"},{"id":98436526,"identity":"e09604bf-72c5-4965-ac38-dfbfbd25b861","added_by":"auto","created_at":"2025-12-17 16:55:49","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1762368,"visible":true,"origin":"","legend":"","description":"","filename":"figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/8a8d2ceb8ce3a212a8c10113.png"},{"id":98436911,"identity":"208791c3-b9db-42ce-a9fb-80190e1275d5","added_by":"auto","created_at":"2025-12-17 16:56:27","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10206150,"visible":true,"origin":"","legend":"","description":"","filename":"figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/507bd8927a75dd558276fb73.png"},{"id":98436574,"identity":"3f5b1f9c-01d1-4e2f-bb92-539c47bbdae3","added_by":"auto","created_at":"2025-12-17 16:55:54","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6522240,"visible":true,"origin":"","legend":"","description":"","filename":"figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/25193538c6126f57270e12a3.png"},{"id":98300250,"identity":"9c58478d-0f9c-4422-ac5f-e2c753219ae9","added_by":"auto","created_at":"2025-12-16 10:01:03","extension":"jpeg","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":660585,"visible":true,"origin":"","legend":"","description":"","filename":"figure4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/1edd7675cd70116e7bcee184.jpeg"},{"id":98300249,"identity":"d1687ec0-0be3-4db6-bd54-197bbb1fba89","added_by":"auto","created_at":"2025-12-16 10:01:03","extension":"jpeg","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":796747,"visible":true,"origin":"","legend":"","description":"","filename":"figure5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/6f07ae3f1e44af796c75fd68.jpeg"},{"id":98300238,"identity":"17e9b213-7d45-434f-b6dc-41287c4f206a","added_by":"auto","created_at":"2025-12-16 10:01:02","extension":"jpeg","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":832004,"visible":true,"origin":"","legend":"","description":"","filename":"figure6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/19792173b5f9b8a729557bec.jpeg"},{"id":98300246,"identity":"0587dea0-67e8-4692-965d-0b7a33f78b6a","added_by":"auto","created_at":"2025-12-16 10:01:03","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":289720,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/f93b1a46a416c1c99a8336ef.png"},{"id":98434655,"identity":"86dfcc81-4309-4f4f-9f1d-d5a46b04da8f","added_by":"auto","created_at":"2025-12-17 16:52:26","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1223721,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/ec9fa3e7960119b2e76cf8df.png"},{"id":98435837,"identity":"61b06611-fb10-44fd-82e4-ecdda4f9d3c3","added_by":"auto","created_at":"2025-12-17 16:54:30","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1080069,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefigure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/e32b2926ac28ef2b7fae7756.png"},{"id":98300245,"identity":"a36b6814-cfdb-4b2f-942c-95cfaab266e3","added_by":"auto","created_at":"2025-12-16 10:01:02","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":179833,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefigure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/98d1299371b74acb13827b76.png"},{"id":98436957,"identity":"5acda3f5-8305-4b69-b82d-0fbc6de1a18b","added_by":"auto","created_at":"2025-12-17 16:56:30","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":223566,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefigure5.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/68873648b89db66942edc307.png"},{"id":98300253,"identity":"e199e358-57c9-4c11-8003-8ce91e325dc6","added_by":"auto","created_at":"2025-12-16 10:01:03","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":284653,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefigure6.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/39980ac4abfa89d426df83db.png"},{"id":98436664,"identity":"1e8ce37b-d3e5-4782-9483-740b51be8e2e","added_by":"auto","created_at":"2025-12-17 16:56:02","extension":"png","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":289720,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/670bcaaf7b6a39d4e49fbca3.png"},{"id":98300251,"identity":"7b22d1eb-0e03-4413-a49a-4a5686245422","added_by":"auto","created_at":"2025-12-16 10:01:03","extension":"png","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1223721,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/f19cb88cf019befd8d245122.png"},{"id":98300260,"identity":"2f547cb7-817a-4db0-ad9d-fb044d47bbe6","added_by":"auto","created_at":"2025-12-16 10:01:03","extension":"png","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1080069,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/0d2d9579b4380c2af594db96.png"},{"id":98300248,"identity":"0927ee51-6889-4a84-90de-6d9da37274cb","added_by":"auto","created_at":"2025-12-16 10:01:03","extension":"png","order_by":31,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":179833,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/71e058b2253a3cc5670d81df.png"},{"id":98300243,"identity":"9711be07-3ad5-4541-b442-2149c3936804","added_by":"auto","created_at":"2025-12-16 10:01:02","extension":"png","order_by":32,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":223566,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/bc2f3299fc714fc8a8e80fab.png"},{"id":98300247,"identity":"bf02a1e0-11c7-47d1-a35b-86bce473b4c5","added_by":"auto","created_at":"2025-12-16 10:01:03","extension":"png","order_by":33,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":284653,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/d560c4e35480812acb803e13.png"},{"id":98300256,"identity":"c2120b62-cc77-4f76-9431-928708733cac","added_by":"auto","created_at":"2025-12-16 10:01:03","extension":"xml","order_by":34,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":137026,"visible":true,"origin":"","legend":"","description":"","filename":"MICRONANO050830structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/a3bbd1a4bfa1c054e2811a26.xml"},{"id":98436583,"identity":"32f22f49-8b98-42b6-bf86-15c7f484f9c5","added_by":"auto","created_at":"2025-12-17 16:55:56","extension":"html","order_by":35,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":152206,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/46baefd85173bb82906c1d0f.html"},{"id":98300218,"identity":"5a512979-33db-4cd8-ae23-b0315713974c","added_by":"auto","created_at":"2025-12-16 10:01:02","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1750752,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eMulti-region behavioral state neural decoding in mice using MRFP combined with customized algorithm.\u003c/strong\u003e \u003cstrong\u003ea\u003c/strong\u003eSchematic diagram of a MRFP \u003cem\u003ein vivo\u003c/em\u003e. The shanks are implanted in distinct brain regions to acquire neural activities. \u003cstrong\u003eb\u003c/strong\u003e Trajectories of the animal under different states are captured by a camera in a customized arena. \u003cstrong\u003ec\u003c/strong\u003e Conceptual flow of behavioral states (resting, roaming, feeding and flash) decoding. The neural signals and trajectories are engaged in fundamental core building of the model.\u003c/p\u003e","description":"","filename":"figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/ea02a0cfc9fdfe49581282e8.png"},{"id":98436570,"identity":"441480a7-cdc5-4983-b9f3-4539282dc94e","added_by":"auto","created_at":"2025-12-17 16:55:54","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":4132083,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDesign, characterization and implantation of MRFP. a\u003c/strong\u003e Optical images of MRFP of 8-shank configuration. \u003cstrong\u003eb\u003c/strong\u003e (Left) Zoom in view of a shank from MRFP demonstrating flexibility. (Right) Microscope image of electrodes on the shank. Each shank incorporated 16 electrodes arranged in a zig-zag configuration, with individual electrode sites measuring 30 × 30 μm and separated by 200 μm intervals. \u003cstrong\u003ec\u003c/strong\u003eA MRFP implanted in a mouse brain phantom composed of 0.8% agar. \u003cstrong\u003ed\u003c/strong\u003e Box chart of impedance 24 electrodes from three MRFPs before cyclic stretch and after 100 and 300 cycles. The mean impedances at 1 kHz were 1.34 ± 0.45 MΩ, 1.43 ± 0.41 MΩ and 1.34 ± 0.45 MΩ, respectively. \u003cstrong\u003ee\u003c/strong\u003e Box chart of impedance during seven-week recording session from Mice01. The mean impedances at 1 kHz were 2.25 ± 0.17 MΩ, 2.26 ± 0.19 MΩ, 2.20 ± 0.19 MΩ and 2.14 ± 0.21 MΩ, respectively. \u003cstrong\u003ef\u003c/strong\u003e A MRFP implanted through minimal invasive micro-holes in a mouse, and the eight shanks were inserted into eight brain regions.\u003c/p\u003e","description":"","filename":"figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/2c6c8d16da2a4ff2d34a5007.png"},{"id":98300223,"identity":"37d72612-e254-4d6f-b08e-938700974c1b","added_by":"auto","created_at":"2025-12-16 10:01:02","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":3723063,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStatus identification of mice and flow diagram of AI-based decoding algorithm. a\u003c/strong\u003e Photo of a freely moving mouse in the arena for multi-region recording. The arena measured 30 × 30 × 30 cm, with a food trough positioned along one side to provide a consistent feeding location. A camera was installed on top of the arena to capture video record, from which trajectories of the mouse was extracted. \u003cstrong\u003eb\u003c/strong\u003e A frame from the video footage, showing the mouse under the state of feeding. The food was stored in a food through installed on the wall. Both head and tail of the mouse was labeled for consequent DeepLabCut analysis. Marker positions were mapped into coordinates, where the coordinate space was defined by the pixel resolution of each video frame. \u003cstrong\u003ec\u003c/strong\u003e (Top) Displacement of the mouse’s head marker under three states during a representative recording session. (Middle) Trajectories of both head and tail markers corresponding to different states. The dotted square in the middle trajectory indicates the position of the food through. (Bottom) Photos of the mouse during roaming, flash, feeding and resting. \u003cstrong\u003ed\u003c/strong\u003eStructural diagram of the decoding model. Multi-channel neural signals undergo sequential processing through temporal convolution, depthwise spatial convolution, and pooling operations. The resulting features are transformed into Query (Q), Key (K), and Value (V) representations for multi-head attention computation. Finally, a classification head generates predictions of the mouse's behavioral states.\u003c/p\u003e","description":"","filename":"figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/eaa762c2a563ea6ac29614ad.png"},{"id":98435619,"identity":"55fa036e-a8d9-4dc2-ad58-e52baacbe84b","added_by":"auto","created_at":"2025-12-17 16:54:08","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1582536,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePerformance characterization and feature space analysis of the L-Conformer decoding model. \u003c/strong\u003e\u0026nbsp;\u003cstrong\u003ea \u003c/strong\u003eValidation accuracy (valid-acc) as a function of input time window length across four animals. An inverted U-shaped trend is observed, with optimal decoding performance achieved at a 4-second window for most mice. \u003cstrong\u003eb\u003c/strong\u003e Confusion matrix for Mice01 using the optimal 4-second window, yielding an average classification accuracy of 87.45% across four behavioral states. \u003cstrong\u003ec \u003c/strong\u003eVisualization of the high-dimensional features learned by the model, projected into a two-dimensional space using t-SNE. \u003cstrong\u003ed \u003c/strong\u003eComparison of decoding performance between the proposed model and several state-of-the-art neural decoding architectures using the same cross-regional dataset. The modified EEG Conformer exhibited the highest validation accuracy.\u003c/p\u003e","description":"","filename":"figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/657b92ec599b9108f170b371.png"},{"id":98300224,"identity":"284a8e7a-0a80-41df-b2d1-667965c72843","added_by":"auto","created_at":"2025-12-16 10:01:02","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":1920456,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eContribution of individual brain regions to decoding accuracy and comparison of decoding strategies. a\u003c/strong\u003eDecoding performance of models trained using data from single brain regions. The inset schematic shows the implantation locations of the 8 shanks, with colors corresponding to the x-axis labels. The blue dashed line indicates the decoding accuracy (87.77%) achieved when using signals from all brain regions. \u003cstrong\u003eb\u003c/strong\u003eConfusion matrix of Mice01 when the model was trained solely on signals from the left V1 region, showing reduced classification accuracy for all states except “flash”. \u003cstrong\u003ec\u003c/strong\u003e Comparison between the concentrated decoding strategy (all electrodes placed in a single brain region) and the distributed decoding strategy (electrodes implanted across multiple regions). When the number of participating brain regions exceeded five, the distributed strategy outperformed the concentrated one, and decoding accuracy continued to increase with the number of regions included. \u003cstrong\u003ed\u003c/strong\u003e Effect of removing a single brain region from the training dataset. Models trained on data from only seven regions exhibited reduced accuracy compared to those trained on all eight, underscoring the additive contribution of each region to multi-region decoding.\u003c/p\u003e","description":"","filename":"figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/7facb2b298ef01f2aa0e7e8e.png"},{"id":98435488,"identity":"db57066d-4694-41af-9649-5e7f4fe6ca28","added_by":"auto","created_at":"2025-12-17 16:53:56","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":2002538,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCross-day and cross-individual generalization performance of the decoding model. a\u003c/strong\u003e Schematic of the cross-day decoding paradigm. The model was trained on neural data from all days preceding the test day and evaluated on the newly acquired session. \u003cstrong\u003eb\u003c/strong\u003e Cross-day decoding accuracy across five consecutive days. Both training and testing accuracy improved with increasing training data. Notably, performance plateaued after day 4, indicating sufficient model generalization with a limited dataset. \u003cstrong\u003ec\u003c/strong\u003eConfusion matrix showing decoding performance for Mice01 on day 5, confirming high classification fidelity across all behavioral states. \u003cstrong\u003ed\u003c/strong\u003e Schematic of the cross-individual decoding paradigm. Data from three of four mice were used to train a shared model, which was then evaluated on the remaining subject using two approaches: zero-shot (direct inference without retraining) and fine-tuning (incrementally adding subject-specific data to the training set). \u003cstrong\u003ee\u003c/strong\u003eComparison between zero-shot and fine-tuning performance. The zero-shot model achieved an average accuracy of 70%, while fine-tuning progressively improved performance, reaching optimal accuracy when 80% of the new subject’s data was incorporated into training.\u003c/p\u003e","description":"","filename":"figure6.png","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/2e0d3efbdb70a619b09bb1c6.png"},{"id":107973481,"identity":"8cf0462b-67cd-4e86-a05a-42edd78b82a2","added_by":"auto","created_at":"2026-04-28 07:12:00","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":15481633,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/071ead6f-16cd-486c-9cec-8f68f42b068b.pdf"},{"id":98300219,"identity":"20e6920d-e3d7-42c5-a0a6-d468c045cd3e","added_by":"auto","created_at":"2025-12-16 10:01:02","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":4442569,"visible":true,"origin":"","legend":"A multi-region flexible neural interface for behavioral state decoding in freely moving mice","description":"","filename":"1028MINESI.docx","url":"https://assets-eu.researchsquare.com/files/rs-7980509/v1/e5996ffd4eb2d77dcf3524f3.docx"}],"financialInterests":"There is no conflict of interest","formattedTitle":"A multi-region flexible neural interface for behavioral state decoding in freely moving mice","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eBrain-computer interfaces (BCIs) enable direct communication between neural circuits and external devices, providing insights into brain function and transformative applications for neurological disorders. BCIs have shown remarkable potential in decoding neural signals from single brain regions, such as the motor cortex for prosthetic control\u003csup\u003e\u003cspan additionalcitationids=\"CR2 CR3 CR4\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e, the visual cortex for vision restoration\u003csup\u003e\u003cspan additionalcitationids=\"CR7 CR8 CR9\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e, and Broca\u0026rsquo;s area for speech synthesis\u003csup\u003e\u003cspan additionalcitationids=\"CR12 CR13 CR14\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. However, this focus on single regions limits their ability to capture the distributed neural dynamics underlying complex behavioral states, which are critical for both fundamental neuroscience and clinical translation. For example, abstract rule processing involves coordinated activity between thalamic nuclei and prefrontal cortex\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e,\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. Likewise, neuropsychiatric disorders such as Alzheimer\u0026rsquo;s disease arise from distributed network dysfunction rather than isolated regional deficits\u003csup\u003e\u003cspan additionalcitationids=\"CR19\" citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. These examples highlight the urgent need for neural interfaces that can decode interactions across multiple brain regions.\u003c/p\u003e\u003cp\u003eA first challenge lies in neural signal acquisition. Decoding distributed neural interactions requires interfaces capable of stable, high-resolution recording across anatomically distinct brain regions in naturalistic settings. Optical and fMRI methods have shown promise for brain-wide decoding, but are limited to specific animal models\u003csup\u003e\u003cspan additionalcitationids=\"CR22 CR23 CR24 CR25 CR26\" citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e, such as zebrafish\u003csup\u003e\u003cspan additionalcitationids=\"CR29\" citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e, or constrained behavioral paradigms\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e,\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e, restricting their applicability to freely moving animals. Electrophysiology offers a direct strategy. However, conventional rigid probes, such as microwires or Utah Arrays, suffer from mechanical mismatch with brain tissue, leading to tissue damage and unstable recordings, while their bulky backends limit spatial coverage\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e,\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e. High-density probes like Neuropixels have expanded neural sampling capabilities, but their size still often requires head fixation in small animals, hindering naturalistic behavior\u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e,\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e. Although flexible probes have improved surgical accessibility and recording stability, their potential for brain-wide recording remains largely unrealized\u003csup\u003e\u003cspan additionalcitationids=\"CR38\" citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eA second challenge lies in neural signal decoding. Even with high-quality recordings, the ability of decoding methods to generalize across different days (cross-day) and individuals (cross-subject) is essential for widespread practical and clinical application\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. However, achieving this goal faces obstacles due to the inherent and profound variability of neural signals\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e. Most current approaches remain confined to a \u0026ldquo;single-subject-single-model\u0026rdquo; paradigm, where a model is effective only for a specific individual within a specific timeframe, severely limiting its versatility and translational potential\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e,\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. Deep learning provides a powerful framework for deciphering complex temporal dynamics, yet existing models face fundamental shortcomings. For example, convolutional neural networks (CNNs) are suitable for extracting local features but their fixed receptive fields struggle to capture relationships across extended time windows\u003csup\u003e\u003cspan additionalcitationids=\"CR42\" citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u003c/sup\u003e. Recurrent neural networks (RNNs), while designed for sequential information, often fail to model long-range temporal dependencies due to issues like the vanishing gradient problem\u003csup\u003e\u003cspan additionalcitationids=\"CR45\" citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e. Hybrid models such as the EEG Conformer, which integrate CNN and Transformer architectures, have demonstrated state-of-the-art performance in EEG-based tasks\u003csup\u003e\u003cspan additionalcitationids=\"CR48\" citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e\u003c/sup\u003e such as emotion recognition and motor imagery, yet their utility for decoding large-scale local field potentials (LFPs) in naturalistic behavior remains underexplored.\u003c/p\u003e\u003cp\u003eIn this work, we present an integrated platform that addresses both challenges. We designed a multi-region flexible probe (MRFP) with omega-shaped shanks for stable, long-term recordings across multiple brain regions in freely moving mice, and paired it with a modified Conformer-based AI model optimized for brain-wide dynamics (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). This MRFP-AI framework achieved up to 89% decoding accuracy of behavioral states (resting, roaming, feeding and flash), with performance maintained at 85% across days and 70% across individuals. Optimal decoding performance was observed with a 4-second time window. Notably, distributed multi-region recordings consistently outperformed single-region configurations, underscoring the value of brain-wide sampling. By jointly enabling stable, scalable recordings and robust generalization in decoding, this platform reveals distributed neural contributions to behavior, providing a foundation for exploring brain-wide dynamics and advancing therapeutic strategies for neuropsychiatric disorders.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"2. Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1. Design and characterization of MRFP\u003c/h2\u003e\u003cp\u003eConventional flexible neural probes are typically fabricated with needle-like, linear shanks to enable minimally invasive brain insertion\u003csup\u003e\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e,\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e\u003c/sup\u003e. However, their capacity to access multiple widely distributed regions is constrained by the low stretchability of common flexible materials such as polyimide (PI) or SU-8. This limitation poses two challenges: first, shanks are prone to mechanical failure under excessive tensile strain; second, after implantation, relative micromotion between the brain and skull may displace the shanks, resulting in the degradation of recording fidelity and signal stability\u003csup\u003e\u003cspan additionalcitationids=\"CR53\" citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eTo overcome these limitations, we designed an MRFP that introduces omega-shaped deformation elements into the probe structure, enabling substantial in-plane stretchability without compromising compactness or mechanical integrity. This design was developed as an optimization of our previously reported flexible electrode system, enhancing both its mechanical adaptability and spatial resolution\u003csup\u003e\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e\u003c/sup\u003e. The MRFP consists of eight shanks arranged linearly (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea), each integrated with 16 gold recording sites positioned in a staggered zig-zag layout (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb). Recording sites are spaced at 200 \u0026micro;m intervals, with an extended pitch of 266 \u0026micro;m incorporated on the two lateral shanks to enable access to deeper brain regions. Notably, the probe design is inherently customizable, and we developed a 16-shank variant, with eight electrodes per shank (Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e), offering a scalable solution for applications in multi-region electrophysiology.\u003c/p\u003e\u003cp\u003eThe omega-shaped structure allows uniaxial stretch up to 50% without mechanical failure (Fig. S2). This stretchability is not derived from the material properties alone, but rather from the geometric reconfiguration of the structure under strain. By leveraging mechanical deformation in this manner, we achieved a compact and flexible design capable of conforming to the brain\u0026rsquo;s curved surface and accommodating micromotion over time. To demonstrate the MRFP\u0026rsquo;s implantation feasibility and anatomical coverage, we performed insertion tests in a mouse brain phantom composed of 0.8% agarose gel. In agarose brain phantoms, MRFP successfully achieved implantation across more than 1 cm of brain-mimicking tissue (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec), supporting its suitability for targeting distributed brain regions.\u003c/p\u003e\u003cp\u003eNext, we evaluated the device\u0026rsquo;s mechanical durability and electrical performance through cyclic uniaxial stretching and electrochemical impedance spectroscopy. Individual shanks from three separate MRFPs were subjected to 100 and 300 stretch cycles under 50% strain. Impedance spectra showed minimal drift over time (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ed), with mean impedances at 1 kHz of 1.34\u0026thinsp;\u0026plusmn;\u0026thinsp;0.45 MΩ (pre-stretch), 1.43\u0026thinsp;\u0026plusmn;\u0026thinsp;0.41 MΩ (100 cycles), and 1.34\u0026thinsp;\u0026plusmn;\u0026thinsp;0.45 MΩ (300 cycles), consistent with stable gold-electrode performance.\u003c/p\u003e\u003cp\u003eLong-term in vivo performance was assessed by tracking impedance over seven weeks post-implantation in Mice01, which still yielded 115 functional channels at the end of the session. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ee, mean impedances remained stable over time (2.25\u0026thinsp;\u0026plusmn;\u0026thinsp;0.17 MΩ on day 1; 2.26\u0026thinsp;\u0026plusmn;\u0026thinsp;0.19 MΩ on day 7; 2.20\u0026thinsp;\u0026plusmn;\u0026thinsp;0.19 MΩ on day 21; 2.14\u0026thinsp;\u0026plusmn;\u0026thinsp;0.21 MΩ on day 49), and only five channel failures were observed (Fig. S3), confirming the MRFP\u0026rsquo;s suitability for chronic, high-density, multi-region recordings.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.2. Implantation of MRFP\u003c/h2\u003e\u003cp\u003eThe wide anatomical reach of the MRFP necessitates a more complex implantation procedure. While previous methods, such as pre-arranged tungsten wire array\u003csup\u003e\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e\u003c/sup\u003e, were effective for localized insertions but unsuitable for distributed architectures. We therefore developed a needle-threading-inspired strategy using tungsten wires to guide each shank (Fig. S4). Instead of full craniotomies, only microholes are drilled to minimize surgical damage. A tungsten wire with a sharp tip (via electrochemical etching) is threaded through a pinhole at the tip of the shank, allowing insertion with stereotaxic guidance. After implantation, the metal wire is retracted, leaving the shank in position. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ef illustrates a typical MRFP implantation targeting eight regions: M2 (AP\u0026thinsp;+\u0026thinsp;1.69 mm, ML\u0026thinsp;\u0026plusmn;\u0026thinsp;1 mm), MnPO (AP\u0026thinsp;+\u0026thinsp;0.25 mm, ML 0 mm), DLS (AP 0 mm, ML 3 mm), S1BF (AP \u0026minus;\u0026thinsp;1.31 mm, ML\u0026thinsp;\u0026plusmn;\u0026thinsp;2.5 mm), CA1 (AP \u0026minus;\u0026thinsp;2 mm, ML 1.8 mm), and V1 (AP \u0026minus;\u0026thinsp;2.69 mm, ML \u0026minus;\u0026thinsp;2.5 mm).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.3. Recording environment and status identification\u003c/h2\u003e\u003cp\u003eA central goal of multi-region decoding is to capture the neural signatures representing both an animal\u0026rsquo;s internal states and its responses to external stimuli. We therefore designed a behavioral paradigm to explicitly span this internal-external spectrum. Mice engaged in three naturalistic, self-initiated behaviors (resting, roaming, and feeding), together with an externally triggered state induced by a visual flash. To enable the naturalistic expression and reliable labeling of these behaviors, we constructed a custom arena (30 \u0026times; 30 \u0026times; 30 cm\u003csup\u003e3\u003c/sup\u003e) with a dedicated feeding location and an overhead flash lamp (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea). Behavioral tracking was performed on continuous video recordings using DeepLabCut\u003csup\u003e\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e\u003c/sup\u003e to extract positional data from key landmarks.\u003c/p\u003e\u003cp\u003eRoaming was defined by coordinated movement of the head and tail across the arena, while resting was identified when both markers remained stationary. Feeding behavior was determined when the head entered a predefined zone around the food trough and remained there for a sustained period without concurrent tail movement. Flash status was triggered by activating the lamp at a frequency of 10 Hz, and lasted 5 minutes for each trial. Representative trajectories and movement distance of the head marker (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ec) illustrate clear spatial and movement patterns distinguishing each state. This behavioral labeling strategy, based on self-initiated actions in an unconstrained setting, provided a reliable foundation for training and evaluating neural decoding models. Neural recordings were acquired from four mice, each continuously monitored over a 168-hour period to capture behavioral events used for subsequent model training.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e2.4. LFP-Based neural network for status classification\u003c/h2\u003e\u003cp\u003eTransformer models, which use attention mechanisms to capture global dependencies, have demonstrated significant success in natural language processing and computer vision\u003csup\u003e\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e\u003c/sup\u003e. Their application has extended to EEG decoding, where they show promise in exploiting long-term temporal relationships. However, conventional Transformers often overlook local feature learning, which is essential for effective LFP decoding. To address this gap, we introduced an LFP Conformer (L-Conformer) architecture that originates from an open-source hybrid model, integrating the local feature extraction capability of CNNs with the global dependency modeling capability of Transformers\u003csup\u003e\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e\u003c/sup\u003e; its architecture is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ed (see Section 4.6 for detailed information on the model\u0026rsquo;s specific structure and implementation). Optimized for multi-region LFP signals, this model can map neural activity to discriminative features corresponding to the four behavioral states, thereby laying a foundation for accurate decoding in subsequent analyses.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e2.5. AI classification of behavioral states in mice\u003c/h2\u003e\u003cp\u003e\u003cb\u003eTemporal resolution and model benchmarking for behavioral state decoding.\u003c/b\u003e We first evaluated whether combining MRFP-based multi-region neural recordings with the L-Conformer architecture could support accurate behavioral state classification. A key factor in training the decoder was the temporal window used to associate neural activity with behavioral labels. Short windows risked fragmenting sustained behaviors (e.g., feeding), while long windows increased the chance of including state transitions and label ambiguity. To determine the optimal resolution, we segmented the raw neural data into windows ranging from 0.1 s to 16 s and trained L-Conformer on each dataset. The results revealed a distinct inverted U-shaped relationship between decoding performance and window duration (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea). Specifically, validation accuracy (valid-acc) rose from 57.9% at a 0.1 s window to a peak of 89.49% at the 4 s window. The confusion matrix for a representative subject (Mice01) at this optimal 4 s window confirms high classification fidelity for all four states (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb). Conversely, extending the window beyond this peak led to a progressive decline in performance, with accuracy dropping to 75.46% at 16 s, likely due to the integration of non-relevant temporal information. This finding indicates that a 4 s time window most effectively captures the neural dynamics differentiating the \u0026ldquo;rest\u0026rdquo;, \u0026ldquo;roam\u0026rdquo;, \u0026ldquo;feed\u0026rdquo; and \u0026ldquo;flash\u0026rdquo; behavioral states in our case. Consequently, all subsequent analyses were conducted using a 4 s time window.\u003c/p\u003e\u003cp\u003eTo visualize the feature representations learned by the model, we projected the high-dimensional feature matrix into a three-dimensional space using t-distributed stochastic neighbor embedding (t-SNE) (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ec). The resulting embedding reveals a distinct underlying structure. A majority of the data points form a continuous manifold, suggesting that the model has captured a smooth, progressive relationship among these samples. Notably, trajectories corresponding to the flash state formed a distinct and isolated cluster in the feature space, indicating that the model effectively captured the discriminative features of this state, resulting in a classification accuracy of 100%. The continuous color gradient is applied to the manifold simply to aid in visualizing its three-dimensional trajectory. We further evaluated several state-of-the-art neural decoding architectures, including Cebra, PatchTST, and EEGNet, to identify the most suitable model for our dataset. Among these, L-Conformer consistently achieved the highest classification accuracy (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ed).\u003c/p\u003e\u003cp\u003eThis superior performance stems from its unique architectural design. Specifically, the Cebra model is primarily constructed for spike signals, whereas our study focuses on LFPs; the PatchTransformer model, while considering temporal information, only focuses on the relationships between signal patches, overlooking the fine-grained dynamics within them; and other models like EEGNet rely predominantly on convolutional modules, lacking effective modeling for long-range temporal dependencies. In contrast, our model integrates convolutional modules with a Transformer-based self-attention mechanism, enabling it to capture both local features and establish global temporal associations, thus achieving superior results.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eSpatial contributions of multi-region and distributed recordings.\u003c/b\u003e Neural decoding efforts over the past decades have mainly focused on motor control\u003csup\u003e\u003cspan additionalcitationids=\"CR59 CR60\" citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e\u003c/sup\u003e, with recording sites typically concentrated within the M1. This concentration was largely constrained by the limited scalability of electrode technologies. As advances in electrode integration have substantially increased channel counts, a key question has emerged: whether additional recording sites should remain localized in a single brain region or be distributed across multiple anatomically distinct regions.\u003c/p\u003e\u003cp\u003eAfter validation of L-Conformer, we firstly assessed the contribution of each brain regions to classification accuracy. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea, training the decoding model using data from a single brain region resulted in variable performance across regions. Incorporating data from all recorded brain regions markedly improved model performance, with averaged accuracy increasing to 87.77% (blue dotted line). For instance, when the decoder was trained only on signals from the left V1, accuracy for all behavioral states except flash was substantially reduced (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eb), reflecting the limited capacity of single-region recordings. A leave-one-region-out analysis further confirmed that removing signals from any region decreased performance (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec), underscoring the importance of integrating multi-region signals for capturing distributed, behaviorally relevant dynamics.\u003c/p\u003e\u003cp\u003eTo compare concentrated and distributed decoding strategies, we constructed two datasets from the same recordings: one including only M1 signals, and another sampling from multiple distinct regions (pink bars). In both configurations, decoding accuracy improved as more channels were included (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ed). At lower channel counts, the concentrated configuration achieved higher accuracy; however, once signals from more than five regions were included, the distributed configuration exceeded the concentrated one. A maximum accuracy of 84.37% was reached when signals from all available regions were utilized. Notably, the performance of the concentrated configuration plateaued, whereas the distributed configuration continued to yield incremental gains. These observations suggest that densely clustering electrodes within a single region may lead to signal redundancy with diminishing returns in decoding accuracy. To quantitatively support this notion, we fitted the decoding accuracies of both configurations using a Logistic growth model, a sigmoidal function commonly employed to characterize systems exhibiting saturation effects. The resulting curves revealed a saturating trend for the concentrated strategy when the number of recording sites exceeds six. In contrast, the distributed configuration maintained a steadily rising trajectory, suggesting that expanding coverage across additional brain regions may further enhance decoding accuracy.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eGeneralization across days and individuals.\u003c/b\u003e Beyond the spatial configuration of electrode sites, a key limitation of conventional decoding algorithms is their dependence on session- and subject-specific calibration. For a given individual, neural recordings acquired on different days typically require recalibration to maintain decoding accuracy, limiting practical deployment across time. Moreover, due to substantial inter-individual variability in neural activity patterns, most models must be trained separately for each subject, hindering generalizability and posing a significant barrier to the development of scalable, cross-subject decoding frameworks.\u003c/p\u003e\u003cp\u003eTo evaluate the robustness of L-Conformer, we assessed the its ability to generalize across days and across individuals. In the cross-day paradigm, training was performed on neural recordings from all preceding days, and testing was conducted on a newly acquired session from the subsequent day without fine-tuning (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ea). As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eb, accuracy started at 76% on day 2 and steadily improved with additional data, reaching 85% by day 5 and closely matching within-day performance (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ec). These results demonstrate that L-Conformer supports robust cross-day decoding. Notably, high decoding accuracy of 85% was achieved using only four days of training data (approximately 1,500 samples), highlighting the model\u0026rsquo;s training efficiency.\u003c/p\u003e\u003cp\u003eIn the cross-individual paradigm, four mice were divided into a foundational group (n\u0026thinsp;=\u0026thinsp;3) and a test group (n\u0026thinsp;=\u0026thinsp;1). Neural recordings from the foundational group were used to train L-Conformer and generate a transferable core network containing key learned parameters. This foundational core enabled downstream construction of two testing strategies for new individuals: a zero-shot (ZS) approach and a conventional fine-tuning (FT) approach (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ed). In the ZS strategy, the pretrained model was directly applied to data from the new subject without any re-calibration, while the FT strategy incorporated a proportion of the new subject\u0026rsquo;s data into the training set to improve accuracy. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ee, the ZS approach achieved an average accuracy of ~\u0026thinsp;70%, indicating baseline generalizability sufficient for practical use. FT further improved performance in most cases, with accuracy increasing in proportion to the amount of subject-specific data provided. These results highlight the model\u0026rsquo;s transferability and adaptability to previously unseen subjects. Notably, the ZS strategy required no retraining, reducing processing time by ~\u0026thinsp;15 minutes under our current hardware setup and enabling faster deployment of cross-individual decoding.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"3. Discussion","content":"\u003cp\u003eIn this study, we developed a MRFP-AI platform for decoding behavioral states in freely moving mice. The platform integrates MRFPs capable of large-scale, high-density electrophysiological recordings with a modified Conformer-based AI decoding framework. An omega-shaped structural design of MRFP provided in-plane stretchability without increasing device size or compromising implantability, enabling implantation across eight brain regions with high yield and chronic stability. Combined with the AI framework, this platform allows robust classification of behavioral states by decoding large-scale neural dynamics.\u003c/p\u003e\u003cp\u003eUsing this multi-region decoding platform, we addressed a key question in neural interface design: with limited electrode numbers, is it more effective to concentrate sampling within a single region or to distribute coverage across multiple regions? To answer this, we systematically compared the decoding performance using concentrated versus distributed electrode configurations. Our results show that multi-region recordings offer a clear advantage. While some regions, such as M1, achieved moderate classification performance alone, integrating signals across multiple regions markedly improved accuracy. When more than five brain areas were included, decoding accuracy using a distributed configuration surpassed that of concentrated strategies. These findings suggest that key features of behavioral states, including resting, roaming, feeding, and flash, are not localized but instead emerge from distributed network dynamics spanning cortical and subcortical structures. Furthermore, we demonstrated that performance gains plateaued in concentrated configurations after ~\u0026thinsp;5 channels, highlighting redundancy within single regions and the importance of sampling diverse functional nodes.\u003c/p\u003e\u003cp\u003eBeyond within-session decoding, we evaluated the platform\u0026rsquo;s robustness under cross-day and cross-individual conditions, two benchmarks essential for real-world deployment of neural decoders. In the cross-day setting, decoding accuracy improved steadily with accumulating data, reaching 85% by day 5. This suggests that while inter-day variability exists, it can be mitigated by accumulating neural context over time. In the cross-individual setting, our transfer learning framework achieved approximately 70% zero-shot accuracy, with further improvements through modest fine-tuning, demonstrating the model\u0026rsquo;s generalizability and potential for rapid deployment.\u003c/p\u003e\u003cp\u003eTogether, these findings represent both technical and conceptual advances. Technically, we demonstrate that high-density, spatially distributed recordings can be achieved using a flexible and scalable platform suitable for chronic behavioral studies. Conceptually, our results highlight the importance of accessing multiple functionally distinct brain regions to decode global behavioral states. This distributed sampling strategy enables a systems-level view of neural dynamics, laying the foundation for deeper mechanistic investigations into large-scale brain function.\u003c/p\u003e"},{"header":"4. Methods","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e4.1. Fabrication of MRFP\u003c/h2\u003e\u003cp\u003eThe neural probe was fabricated by standard MEMS process (Fig. S4). First, the wafer was cleaned using RCA process. After photolithography, we adopted electron beam evaporating technique to deposit 100 nm thick Ni on the wafer, and the sacrificial layer was patterned via the lift-off process. Then 1 \u0026micro;m thick PI was spin coated and solidified in a 350\u0026deg;C nitrogen oven for 10 h as the bottom insulation layer. After that, we fabricated the recording sites and interconnects following photolithography, electron beam evaporation and lift-off, including a sandwich metal layer containing 5 nm Ti, 100 nm Au and 5 nm Ti. Backend pads for bonding were formed through the same process, including a 5 nm Ti layer, a 150 nm Ni layer and a 50 nm Au layer. Another 1 \u0026micro;m PI layer was cured on top of the wafer as top encapsulation. Before PI etching, a 100 nm Al layer was deposited by sputtering, patterned by photolithography and wet etching to form a hard mask. Finally, we employed RIE to expose the electrodes and pads, and etched out profile of the probe.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003e4.2. Surgical procedures\u003c/h2\u003e\u003cp\u003e All procedures were conducted in accordance with institutional animal care guidelines and approved protocols. Mice were initially anesthetized in an induction chamber using 4\u0026ndash;5% isoflurane. Once anesthetized, the fur on the scalp was shaved, and animals were placed in a stereotaxic frame (RWD Life Science Co., Ltd) equipped with a temperature-regulated heating pad. To prevent corneal drying and injury, erythromycin ophthalmic ointment was applied to the eyes using a sterile cotton swab. Anesthesia was maintained with 2.5% isoflurane for the first 10 minutes and subsequently reduced to 1% for the remainder of the surgical procedure. Upon reaching a stable anesthetic plane, a trapezoid-shaped incision was made to expose the skull. The overlying fascia and periosteum were carefully removed using fine scissors, forceps, and sterile cotton swabs. Implantation sites were identified and marked using a stereotaxic coordinate system. Burr holes were drilled at designated locations with a dental drill. Flexible shanks of the MRFP were positioned above the burr holes using fine tweezers and inserted into the brain tissue alongside a tungsten wire for structural support. Once the desired implantation depth was reached, the tungsten wire was withdrawn, leaving the flexible shanks in place. A steel wire was inserted to serve as a reference electrode and soldered to the reference pad of the probe\u0026rsquo;s PCB using low-temperature solder. The probe assembly was secured to the skull with dental acrylic to ensure long-term mechanical stability. Postoperative analgesia was administered via subcutaneous injection of meloxicam (5 mg/kg) for pain management. Mice were allowed to recover for at least seven days prior to device testing and neural data acquisition.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003e4.3. Data preprocessing\u003c/h2\u003e\u003cp\u003eThe raw LFP data is structured as a matrix with dimensions C\u0026times;T, where C represents the number of electrode channels and T denotes the number of time samples. Initially, the raw data were bandpass filtered to isolate neural signals within the 0.1\u0026ndash;250 Hz frequency band, coupled with notch filtering to specifically attenuate 50 Hz line noise contamination, and then z-score normalized to reduce variability and non-stationarity. The filtered data is normalized using the following equation:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:{\\mathcal{X}}_{0}=\\frac{{\\mathcal{X}}_{\\mathcal{i}}-\\mathcal{u}}{\\sqrt{{\\sigma\\:}^{2}}}$$\u003c/div\u003e\u003c/div\u003e,\u003c/p\u003e\u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\mathcal{X}}_{\\mathcal{i}}\\)\u003c/span\u003e\u003c/span\u003eis the bandpass-filtered data, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\mathcal{X}}_{0}\\)\u003c/span\u003e\u003c/span\u003eis the normalized output. The mean (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mathcal{u}\\)\u003c/span\u003e\u003c/span\u003e) and standard deviation (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\sigma\\:\\)\u003c/span\u003e\u003c/span\u003e) are calculated exclusively from the training dataset and subsequently applied to the test dataset.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003e4.4. Feature visualization through low-dimensional embedding and geodesic gradient mapping\u003c/h2\u003e\u003cp\u003eTo investigate the internal feature representations learned by the model and intuitively illustrate the topological relationships between distinct neural states, we implemented a multi-step visualization pipeline. First, we extracted the output logits of each EEG window from the final hidden layer of the trained EEG Conformer model, just prior to classification. These high-dimensional feature vectors were then projected into a three-dimensional space using the t-distributed Stochastic Neighbor Embedding (t-SNE) algorithm, with the target dimensionality set to 3. This transformation yielded a set of 3D coordinates (x, y, z), with each point corresponding to a single EEG segment. Next, we constructed a k-nearest neighbor (k-NN) graph to encode the local structure of the embedded space. Each point was treated as a node, and edges were created between each node and its 15 nearest neighbors in the 3D Euclidean space. To establish an unbiased origin for color mapping, we first calculated the geometric centroid of the entire point cloud. The node farthest from this centroid was designated as the starting node of the manifold traversal. From this starting point, we applied Dijkstra\u0026rsquo;s algorithm to compute the shortest path lengths along the k-NN graph to all other nodes. These geodesic distances reflected the accumulated path length along the intrinsic data manifold, rather than direct Euclidean distances. For isolated nodes not connected to the graph (i.e., with infinite path lengths), we assigned the maximum finite distance in the dataset to preserve continuity in the colormap mapping. The resulting distance values were linearly normalized to the [0, 1] range using min\u0026ndash;max scaling, and the normalized values were then mapped to a continuous colormap (rainbow gradient). This procedure allowed each point in the final 3D visualization to be color-coded based on its relative geodesic distance from the origin node. This color-gradient embedding provided strong visual evidence for how the model organized and separated distinct neural states within its learned representation space.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003e4.5. Animals\u003c/h2\u003e\u003cp\u003eIn all the experiments, the mice we used were adult ICR mice (30\u0026ndash;40g at start of experiments) provided by Phenotek Biotechnology Co., Ltd. All mice were aged between 8 and 10 weeks prior to commencement of experiments and were maintained on a 12-h lightdark cycle with ad libitum access to food and water. All our animal experiments obtained ethical approval from the Ethics Committee for Animal Management at the Phenotek Biotechnology (approval number 20240708-01). Our experimental conditions and handling procedures for mice complied with the ethical requirements for experimental animals.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003e4.6. L-Conformer Model Architecture\u003c/h2\u003e\u003cp\u003eL-Conformer integrates the local feature extraction capability of CNNs and the global temporal dependency modeling capability of Transformers. The overall framework of the L-Conformer, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ed, comprises three primary components: a convolutional module, a self-attention module, and a fully-connected classifier.\u003c/p\u003e\u003cp\u003e\u003cb\u003eConvolutional module.\u003c/b\u003e To efficiently extract low-level spatiotemporal features, we employed a two-stage convolutional encoder. A temporal convolution layer with 40 kernels of size 1\u0026times;25 and a stride of 1\u0026times;1 captures time-dependent patterns, followed by a spatial convolution layer with 40 kernels of size C\u0026times;1 and a stride of 1\u0026times;1 to model inter-electrode interactions. Batch normalization is applied after these convolutions to accelerate training and mitigate overfitting. The Exponential Linear Unit (ELU) is used as the activation function. Following this, an average pooling layer with a kernel size of 1\u0026times;75 and a stride of 1\u0026times;15 is employed. This layer smooths the temporal features, reduces computational complexity, and further minimizes the risk of overfitting.\u003c/p\u003e\u003cp\u003eFinally, the feature maps from the convolutional module are reshaped by compressing the electrode channel dimension and transposing the feature channel with the temporal dimension. This rearrangement treats all feature channels at a given timestep as a unified \u0026ldquo;token\u0026rdquo; for the subsequent attention module. This approach is based on the premise that leveraging context-dependent representations within low-level spatiotemporal features enhances EEG decoding by capitalizing on the coherence of neural activity.\u003c/p\u003e\u003cp\u003e\u003cb\u003eSelf-Attention Module.\u003c/b\u003e To complement the limited receptive field of the convolutional module, a self-attention module is used to learn global temporal dependencies across the feature tokens. The tokens from the convolutional module are linearly projected into three matrices of equal dimension: Query (Q), Key (K), and Value (V). The core of the attention mechanism involves computing the dot product between the Query and Key matrices to determine inter-token correlations. A scaling factor, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\frac{1}{\\sqrt{{d}_{k}}}\\)\u003c/span\u003e\u003c/span\u003e (where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{d}_{k}\\)\u003c/span\u003e\u003c/span\u003e is the dimension of the key), is applied to prevent vanishing gradients and ensure training stability. The resulting attention weights are obtained via a Softmax function and are then applied to the Value matrix to yield the final output, as described by the formula:\u003c/p\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Attention(Q,K,V)=Softmax\\left(\\frac{\\text{Q}{\\text{K}}^{\\text{T}}}{\\sqrt{{\\text{d}}_{\\text{k}}}}\\right)\\)\u003c/span\u003e\u003c/span\u003eV\u003c/p\u003e\u003cp\u003eTo further enhance the model\u0026rsquo;s representational power, a multi-head attention (MHA) mechanism is implemented. The input tokens are partitioned into h segments, each processed independently by a separate self-attention submodule. The outputs of these parallel heads are then concatenated to form the final representation:\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:MHA(Q,\\:K,\\:V)=Concat(\\:{\\text{h}\\text{e}\\text{a}\\text{d}}_{0},\\:\\dots\\:\\:,{\\:\\text{h}\\text{e}\\text{a}\\text{d}}_{\\text{h}-1})$$\u003c/div\u003e\u003c/div\u003e,\u003c/p\u003e\u003cp\u003ewhere\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\:\\text{h}\\text{e}\\text{a}\\text{d}}_{\\text{i}}=\\text{A}\\text{t}\\text{t}\\text{e}\\text{n}\\text{t}\\text{i}\\text{o}\\text{n}({\\text{Q}}_{\\text{i}},{\\text{K}}_{\\text{i}},{\\text{V}}_{\\text{i}}\\)\u003c/span\u003e\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003cb\u003eClassifier.\u003c/b\u003e The final component is a classifier module, which consists of two fully-connected layers. This module receives the output from the self-attention block and generates an M-dimensional output vector, corresponding to the M possible classes. A Softmax transformation is applied to this vector to produce class probabilities. The entire framework is optimized end-to-end using the cross-entropy loss function.\u003c/p\u003e\u003cp\u003eThe entire L-Conformer framework is optimized end-to-end using the cross-entropy loss function, which minimizes the difference between the predicted class probabilities and the true behavioral state labels (one-hot encoded). This optimization strategy ensures that the model learns discriminative features directly aligned with the task of behavioral state decoding.\u003c/p\u003e\u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003eAcknowledgments\u003c/p\u003e\n\u003cp\u003eThis work was supported by National Key R \u0026amp; D Program of China grant 2021ZD0201600 and Young Scientists Fund of the National Natural Science Foundation of China grant 62305368.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;Author contributions\u003c/p\u003e\n\u003cp\u003eThese authors contributed equally to this work: Ye Tian, Gen Li, Haoyang Su. Ye Tian led the experimental design, flexible electrode design and fabrication, and animal experiments. Gen Li led the data processing, neural network construction, training and testing, and data visualization. Haoyang Su led the data preprocessing and system design. Luyue Jiang, Yunfu Luo, Yingkang Yang, Lei Huang, Jiazhi Li, and Shuang Jin contributed to animal experiments and device design. Peijie Chen and Yiming Gao contributed to system fabrication and device fabrication. Yike Xiang and Yi Wei contributed to neural network construction. Yifei Ye and Liuyang Sun designed and supervised the project. All authors discussed the results and commented on the manuscript at all stages.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eData availability\u003c/p\u003e\n\u003cp\u003eThe data that support the findings of this study are available from the corresponding author upon reasonable request.\u003c/p\u003e\n\n\u003cp\u003eCompeting Interests\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no conflict of interest.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eFlesher, S. N. \u003cem\u003eet al.\u003c/em\u003e 1-A brain-computer interface that evokes tactile sensations improves robotic arm control. \u003cem\u003eSCIENCE\u003c/em\u003e 372, 831-+ (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHochberg, L. R. \u003cem\u003eet al.\u003c/em\u003e Neuronal ensemble control of prosthetic devices by a human with tetraplegia. \u003cem\u003eNature\u003c/em\u003e 442, 164\u0026ndash;171 (2006).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHochberg, L. R. \u003cem\u003eet al.\u003c/em\u003e Reach and grasp by people with tetraplegia using a neurally controlled robotic arm. \u003cem\u003eNATURE\u003c/em\u003e 485, 372-U121 (2012).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHandelman, D. A. \u003cem\u003eet al.\u003c/em\u003e Shared Control of Bimanual Robotic Limbs With a Brain-Machine Interface for Self-Feeding. \u003cem\u003eFRONTIERS IN NEUROROBOTICS\u003c/em\u003e 16, 918001 (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWolpaw, J. R., Birbaumer, N., McFarland, D. J., Pfurtscheller, G. \u0026amp; Vaughan, T. M. Brain\u0026ndash;computer interfaces for communication and control. \u003cem\u003eClinical Neurophysiology\u003c/em\u003e 113, 767\u0026ndash;791 (2002).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBeauchamp, M. S. \u003cem\u003eet al.\u003c/em\u003e Dynamic Stimulation of Visual Cortex Produces Form Vision in Sighted and Blind Humans. \u003cem\u003eCELL\u003c/em\u003e 181, 774-+ (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNirenberg, S. \u0026amp; Pandarinath, C. Retinal prosthetic strategy with the capacity to restore normal vision. \u003cem\u003ePROCEEDINGS OF THE NATIONAL ACADEMY OF SCIENCES OF THE UNITED STATES OF AMERICA\u003c/em\u003e 109, 15012\u0026ndash;15017 (2012).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eForoushani, A. N., Pack, C. C. \u0026amp; Sawan, M. Cortical visual prostheses: from microstimulation to functional percept. \u003cem\u003eJOURNAL OF NEURAL ENGINEERING\u003c/em\u003e 15, 021005 (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFernandez, E. \u003cem\u003eet al.\u003c/em\u003e Visual percepts evoked with an intracortical 96-channel microelectrode array inserted in human occipital cortex. \u003cem\u003eJOURNAL OF CLINICAL INVESTIGATION\u003c/em\u003e 131, 151331 (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNiketeghad, S. \u0026amp; Pouratian, N. Brain Machine Interfaces for Vision Restoration: The Current State of Cortical Visual Prosthetics. \u003cem\u003eNEUROTHERAPEUTICS\u003c/em\u003e 16, 134\u0026ndash;143 (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWillett, F. R. \u003cem\u003eet al.\u003c/em\u003e A high-performance speech neuroprosthesis. \u003cem\u003eNATURE\u003c/em\u003e 620, (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMetzger, S. L. \u003cem\u003eet al.\u003c/em\u003e A high-performance neuroprosthesis for speech decoding and avatar control. \u003cem\u003eNATURE\u003c/em\u003e 620, 1037-+ (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCard, N. S. \u003cem\u003eet al.\u003c/em\u003e An Accurate and Rapidly Calibrating Speech Neuroprosthesis. \u003cem\u003eNEW ENGLAND JOURNAL OF MEDICINE\u003c/em\u003e 391, 609\u0026ndash;618 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMoses, D. A., Leonard, M. K., Makin, J. G. \u0026amp; Chang, E. F. Real-time decoding of question-and-answer speech dialogue using human cortical activity. \u003cem\u003eNATURE COMMUNICATIONS\u003c/em\u003e 10, 3096 (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWahl, M. \u003cem\u003eet al.\u003c/em\u003e The Human Thalamus Processes Syntactic and Semantic Language Violations. \u003cem\u003eNeuron\u003c/em\u003e 59, 695\u0026ndash;707 (2008).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBaehner, F. \u003cem\u003eet al.\u003c/em\u003e Abstract rule learning promotes cognitive flexibility in complex environments across species. \u003cem\u003eNATURE COMMUNICATIONS\u003c/em\u003e 16, 5396 (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePhillips, J. M. \u003cem\u003eet al.\u003c/em\u003e Primate thalamic nuclei select abstract rules and shape prefrontal dynamics. \u003cem\u003eNEURON\u003c/em\u003e 113, 2014\u0026ndash;2027 (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTayebi, H. \u003cem\u003eet al.\u003c/em\u003e Applications of brain-computer interfaces in neurodegenerative diseases. \u003cem\u003eNEUROSURGICAL REVIEW\u003c/em\u003e 46, 131 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBelkacem, A. N., Jamil, N., Palmer, J. A., Ouhbi, S. \u0026amp; Chen, C. Brain Computer Interfaces for Improving the Quality of Life of Older Adults and Elderly Patients. \u003cem\u003eFRONTIERS IN NEUROSCIENCE\u003c/em\u003e 14, 692 (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDavenport, F. \u003cem\u003eet al.\u003c/em\u003e Neurodegenerative disease of the brain: a survey of interdisciplinary approaches. \u003cem\u003eJOURNAL OF THE ROYAL SOCIETY INTERFACE\u003c/em\u003e 20, 20220406 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCayce, J. M. \u003cem\u003eet al.\u003c/em\u003e Calcium imaging of infrared-stimulated activity in rodent brain. \u003cem\u003eCELL CALCIUM\u003c/em\u003e 55, 183\u0026ndash;190 (2014).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBoyden, E. S., Zhang, F., Bamberg, E., Nagel, G. \u0026amp; Deisseroth, K. Millisecond-timescale, genetically targeted optical control of neural activity. \u003cem\u003eNATURE NEUROSCIENCE\u003c/em\u003e 8, 1263\u0026ndash;1268 (2005).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFenno, L., Yizhar, O. \u0026amp; Deisseroth, K. The Development and Application of Optogenetics. in ANNUAL REVIEW OF NEUROSCIENCE, VOL 34 (eds. Hyman, S. E., Jessell, T. M., Shatz, C. J., Stevens, C. F. \u0026amp; Zoghbi, H. Y.) vol. 34 389\u0026ndash;412 (2011).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBai, L. \u003cem\u003eet al.\u003c/em\u003e Volumetric Imaging of Neural Activity by Light Field Microscopy. \u003cem\u003eNEUROSCIENCE BULLETIN\u003c/em\u003e 38, 1559\u0026ndash;1568 (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDong, C. \u003cem\u003eet al.\u003c/em\u003e Fluorescence Imaging of Neural Activity, Neurochemical Dynamics, and Drug-Specific Receptor Conformation with Genetically Encoded Sensors. \u003cem\u003eANNUAL REVIEW OF NEUROSCIENCE\u003c/em\u003e 45, 273\u0026ndash;294 (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLogothetis, N. K., Pauls, J., Augath, M., Trinath, T. \u0026amp; Oeltermann, A. Neurophysiological investigation of the basis of the fMRI signal. \u003cem\u003eNATURE\u003c/em\u003e 412, 150\u0026ndash;157 (2001).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCritchley, H. D., Wiens, S., Rotshtein, P., \u0026Ouml;hman, A. \u0026amp; Dolan, R. J. Neural systems supporting interoceptive awareness. \u003cem\u003eNATURE NEUROSCIENCE\u003c/em\u003e 7, 189\u0026ndash;195 (2004).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYang, C. \u003cem\u003eet al.\u003c/em\u003e A population code for spatial representation in the zebrafish telencephalon. \u003cem\u003eNATURE\u003c/em\u003e 634, 397-+ (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRenninger, S. L. \u0026amp; Orger, M. B. Two-photon imaging of neural population activity in zebrafish. \u003cem\u003eMETHODS\u003c/em\u003e 62, 255\u0026ndash;267 (2013).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMarques, J. C., Li, M., Schaak, D., Robson, D. N. \u0026amp; Li, J. M. Internal state dynamics shape brainwide activity and foraging behaviour. \u003cem\u003eNATURE\u003c/em\u003e 577, 239-+ (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDelafontaine-Martel, P. \u003cem\u003eet al.\u003c/em\u003e Whole brain vascular imaging in a mouse model of Alzheimer\u0026rsquo;s disease with two-photon microscopy. \u003cem\u003eJOURNAL OF BIOMEDICAL OPTICS\u003c/em\u003e 23, 076501 (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJohannssen, H. C. \u0026amp; Helmchen, F. Two-photon imaging of spinal cord cellular networks. \u003cem\u003eEXPERIMENTAL NEUROLOGY\u003c/em\u003e 242, 18\u0026ndash;26 (2013).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNormann, R. A. \u0026amp; Fernandez, E. Clinical applications of penetrating neural interfaces and Utah Electrode Array technologies. \u003cem\u003eJOURNAL OF NEURAL ENGINEERING\u003c/em\u003e 13, 061003 (2016).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMaynard, E. M., Nordhausen, C. T. \u0026amp; Normann, R. A. The Utah Intracortical Electrode Array: A recording structure for potential brain-computer interfaces. \u003cem\u003eELECTROENCEPHALOGRAPHY AND CLINICAL NEUROPHYSIOLOGY\u003c/em\u003e 102, 228\u0026ndash;239 (1997).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePaulk, A. C. \u003cem\u003eet al.\u003c/em\u003e Large-scale neural recordings with single neuron resolution using Neuropixels probes in human cortex. \u003cem\u003eNATURE NEUROSCIENCE\u003c/em\u003e 25, 252-+ (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChung, J. E. \u003cem\u003eet al.\u003c/em\u003e High-density single-unit human cortical recordings using the Neuropixels probe. \u003cem\u003eNEURON\u003c/em\u003e 110, 2409-+ (2022).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGuan, S. \u003cem\u003eet al.\u003c/em\u003e Elastocapillary self-assembled neurotassels for stable neural activity recordings. \u003cem\u003eSCIENCE ADVANCES\u003c/em\u003e 5, eaav2842 (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLuan, L. \u003cem\u003eet al.\u003c/em\u003e Ultraflexible nanoelectronic probes form reliable, glial scar-free neural integration. \u003cem\u003eSCIENCE ADVANCES\u003c/em\u003e 3, e1601966 (2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhou, T. \u003cem\u003eet al.\u003c/em\u003e Syringe-injectable mesh electronics integrate seamlessly with minimal chronic immune response in the brain. \u003cem\u003ePROCEEDINGS OF THE NATIONAL ACADEMY OF SCIENCES OF THE UNITED STATES OF AMERICA\u003c/em\u003e 114, 5894\u0026ndash;5899 (2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePachitariu, M., Sridhar, S., Pennington, J. \u0026amp; Stringer, C. Spike sorting with Kilosort4. \u003cem\u003eNat Methods\u003c/em\u003e 21, 914\u0026ndash;921 (2024).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAcharya, U. R., Oh, S. L., Hagiwara, Y., Tan, J. H. \u0026amp; Adeli, H. Deep convolutional neural network for the automated detection and diagnosis of seizure using EEG signals. \u003cem\u003eCOMPUTERS IN BIOLOGY AND MEDICINE\u003c/em\u003e 100, 270\u0026ndash;278 (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSchirrmeister, R. T. \u003cem\u003eet al.\u003c/em\u003e Deep Learning With Convolutional Neural Networks for EEG Decoding and Visualization. \u003cem\u003eHUMAN BRAIN MAPPING\u003c/em\u003e 38, 5391\u0026ndash;5420 (2017).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLawhern, V. J. \u003cem\u003eet al.\u003c/em\u003e EEGNet: a compact convolutional neural network for EEG-based brain-computer interfaces. \u003cem\u003eJOURNAL OF NEURAL ENGINEERING\u003c/em\u003e 15, 056013 (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTortora, S., Ghidoni, S., Chisari, C., Micera, S. \u0026amp; Artoni, F. Deep learning-based BCI for gait decoding from EEG with LSTM recurrent neural network. \u003cem\u003eJOURNAL OF NEURAL ENGINEERING\u003c/em\u003e 17, 046011 (2020).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMante, V., Sussillo, D., Shenoy, K. V. \u0026amp; Newsome, W. T. Context-dependent computation by recurrent dynamics in prefrontal cortex. \u003cem\u003eNATURE\u003c/em\u003e 503, 78-+ (2013).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eG\u0026uuml;ler, N. F., \u0026Uuml;beyli, E. D. \u0026amp; G\u0026uuml;ler, I. Recurrent neural networks employing Lyapunov exponents for EEG signals classification. \u003cem\u003eEXPERT SYSTEMS WITH APPLICATIONS\u003c/em\u003e 29, 506\u0026ndash;514 (2005).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSong, Y., Zheng, Q., Liu, B. \u0026amp; Gao, X. EEG Conformer: Convolutional Transformer for EEG Decoding and Visualization. \u003cem\u003eIEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING\u003c/em\u003e 31, 710\u0026ndash;719 (2023).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePolikov, V. S., Tresco, P. A. \u0026amp; Reichert, W. M. Response of brain tissue to chronically implanted neural electrodes. \u003cem\u003eJOURNAL OF NEUROSCIENCE METHODS\u003c/em\u003e 148, 1\u0026ndash;18 (2005).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAbiri, R., Borhani, S., Sellers, E. W., Jiang, Y. \u0026amp; Zhao, X. A comprehensive review of EEG-based brain-computer interface paradigms. \u003cem\u003eJOURNAL OF NEURAL ENGINEERING\u003c/em\u003e 16, 011001 (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMusk, E. An Integrated Brain-Machine Interface Platform With Thousands of Channels. \u003cem\u003eJOURNAL OF MEDICAL INTERNET RESEARCH\u003c/em\u003e 21, e16194 (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGu, C., Jiang, J., Tao, T. H., Wei, X. \u0026amp; Sun, L. Long-term flexible penetrating neural interfaces: materials, structures, and implantation. \u003cem\u003eSCIENCE CHINA-INFORMATION SCIENCES\u003c/em\u003e 64, 221401 (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKim, M. G., Yu, K., Niu, X. \u0026amp; He, B. Investigation of Displacement of Intracranial Electrode Induced by Focused Ultrasound Stimulation. \u003cem\u003eIEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT\u003c/em\u003e 70, 9600509 (2021).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePeng, L. \u003cem\u003eet al.\u003c/em\u003e Biomechanics characterization of an implantable ultrathin intracortical electrode through finite element method. \u003cem\u003eSCIENTIFIC REPORTS\u003c/em\u003e 15, 19938 (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCheng, Q. et al. High-Resolution Recording of Neural Activity in Epilepsy Using Flexible Neural Probes. ADVANCED MATERIALS TECHNOLOGIES (2023) doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/admt.202301210\u003c/span\u003e\u003cspan address=\"10.1002/admt.202301210\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhao, Z. \u003cem\u003eet al.\u003c/em\u003e Parallel, minimally-invasive implantation of ultra-flexible neural electrode arrays. \u003cem\u003eJOURNAL OF NEURAL ENGINEERING\u003c/em\u003e 16, 035001 (2019).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMathis, A. \u003cem\u003eet al.\u003c/em\u003e DeepLabCut: markerless pose estimation of user-defined body parts with deep learning. \u003cem\u003eNATURE NEUROSCIENCE\u003c/em\u003e 21, 1281-+ (2018).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLee, Y.-E. \u0026amp; Lee, S.-H. EEG-Transformer: Self-attention from Transformer Architecture for Decoding EEG of Imagined Speech. 10TH INTERNATIONAL WINTER CONFERENCE ON BRAIN-COMPUTER INTERFACE (BCI2022) (2022) doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/BCI53720.2022.9735124\u003c/span\u003e\u003cspan address=\"10.1109/BCI53720.2022.9735124\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOh, S. \u003cem\u003eet al.\u003c/em\u003e A stealthy neural recorder for the study of behaviour in primates. \u003cem\u003eNATURE BIOMEDICAL ENGINEERING\u003c/em\u003e 9, (2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCollinger, J. L. \u003cem\u003eet al.\u003c/em\u003e High-performance neuroprosthetic control by an individual with tetraplegia. \u003cem\u003eLANCET\u003c/em\u003e 381, 557\u0026ndash;564 (2013).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKuiken, T. A. \u003cem\u003eet al.\u003c/em\u003e Targeted Muscle Reinnervation for Real-time Myoelectric Control of Multifunction Artificial Arms. \u003cem\u003eJAMA-JOURNAL OF THE AMERICAN MEDICAL ASSOCIATION\u003c/em\u003e 301, 619\u0026ndash;628 (2009).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMusallam, S., Corneil, B. D., Greger, B., Scherberger, H. \u0026amp; Andersen, R. A. Cognitive control signals for neural prosthetics. \u003cem\u003eSCIENCE\u003c/em\u003e 305, 258\u0026ndash;262 (2004).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"microsystems-and-nanoengineering","isNatureJournal":false,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"micronano","sideBox":"Learn more about [Microsystems \u0026 Nanoengineering](http://www.nature.com/micronano/)","snPcode":"41378","submissionUrl":"https://mts-micronano.nature.com/","title":"Microsystems \u0026 Nanoengineering","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Nature AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"brain-computer interfaces, multi-region flexible probes, brain decoding, behavioral decoding","lastPublishedDoi":"10.21203/rs.3.rs-7980509/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7980509/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eHigh-density, long-term stable decoding of whole-brain function is crucial for advancing basic neuroscience research and developing neural disorder therapies. However, two major challenges remain: the lack of scalable interfaces capable of long-term, multi-regional recordings and the limited generalizability of existing decoding algorithms across days and individuals. Here, we developed an integrated platform that achieves accurate, stable, and generalizable decoding of behavioral states (resting, roaming, feeding and flash) with up to 89% accuracy. This platform combines multi-region flexible probes (MRFPs), enabling distributed recordings from 128 sites across eight brain regions over months, with a Conformer-based deep learning framework optimized for brain-wide neural dynamics. Comparative analyses demonstrate that distributed sampling, particularly from five or more regions, markedly enhances decoding performance over concentrated electrode configurations. Furthermore, the platform supports robust generalization across days and individuals without retraining, providing a practical solution for longitudinal and large-scale behavioral neuroscience studies. These results establish a foundation for stable, high-fidelity multi-region electrophysiology and offer a generalizable approach for decoding internal states from complex neural dynamics.\u003c/p\u003e","manuscriptTitle":"A multi-region flexible neural interface for behavioral state decoding in freely moving mice","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-16 10:00:57","doi":"10.21203/rs.3.rs-7980509/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"revise","date":"2026-01-04T08:14:43+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"This content is not available.","date":"2025-12-31T06:18:57+00:00","index":3,"fulltext":"This content is not available."},{"type":"editorInvitedReview","content":"This content is not available.","date":"2025-12-28T18:04:25+00:00","index":1,"fulltext":"This content is not available."},{"type":"editorInvitedReview","content":"This content is not available.","date":"2025-12-23T13:33:05+00:00","index":2,"fulltext":"This content is not available."},{"type":"reviewerAgreed","content":"This content is not available.","date":"2025-12-11T08:15:29+00:00","index":3,"fulltext":"This content is not available."},{"type":"reviewerAgreed","content":"This content is not available.","date":"2025-12-11T07:44:19+00:00","index":2,"fulltext":"This content is not available."},{"type":"reviewerAgreed","content":"This content is not available.","date":"2025-12-11T07:20:34+00:00","index":1,"fulltext":"This content is not available."},{"type":"reviewersInvited","content":"","date":"2025-12-11T07:16:01+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-10-30T06:06:25+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-10-29T13:37:04+00:00","index":"","fulltext":""},{"type":"submitted","content":"Microsystems \u0026 Nanoengineering","date":"2025-10-29T13:37:03+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"microsystems-and-nanoengineering","isNatureJournal":false,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"micronano","sideBox":"Learn more about [Microsystems \u0026 Nanoengineering](http://www.nature.com/micronano/)","snPcode":"41378","submissionUrl":"https://mts-micronano.nature.com/","title":"Microsystems \u0026 Nanoengineering","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Nature AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"f211eb11-6f63-458c-a295-b35f10f94b57","owner":[],"postedDate":"December 16th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":59467624,"name":"Physical sciences/Engineering/Electrical and electronic engineering"},{"id":59467625,"name":"Physical sciences/Nanoscience and technology/Nanobiotechnology/Bionanoelectronics"}],"tags":[],"updatedAt":"2026-04-28T07:11:45+00:00","versionOfRecord":{"articleIdentity":"rs-7980509","link":"https://doi.org/10.1038/s41378-026-01258-5","journal":{"identity":"microsystems-and-nanoengineering","isVorOnly":false,"title":"Microsystems \u0026 Nanoengineering"},"publishedOn":"2026-04-27 04:00:00","publishedOnDateReadable":"April 27th, 2026"},"versionCreatedAt":"2025-12-16 10:00:57","video":"","vorDoi":"10.1038/s41378-026-01258-5","vorDoiUrl":"https://doi.org/10.1038/s41378-026-01258-5","workflowStages":[]},"version":"v1","identity":"rs-7980509","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7980509","identity":"rs-7980509","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}