Objective: Disorders of consciousness (DoC) diagnosis critically depends on accurate state discrimination to guide treatment and prognosis. Current EEG-based techniques face challenges of incomplete electrode coverage and manual feature reliance, due to the complex nature of DoC conditions. Methods: This study proposes DoC-Informer, a CNN-Transformer framework for automated DoC discrimination under adaptive EEG settings. By integrating a channel-independent architecture-enabled by electrode position encoding and spatial transformers-with channel masking training, the framework employs: 1) Shallow Temporal Feature Encoding with parallel temporal convolutions to extract channel-independent temporal features; 2) Spatiotemporal Representation Modeling using a Spatial Transformer (with 3D electrode position encoding) to infer spatial dependencies and a Temporal Transformer for long-range dynamics. A Channel Masking Training Strategy enhances robustness to incomplete data. Results: Extensive experiments on two real-world DoC datasets (including UWS and MCS patients) demonstrate DoC-Informer's superiority over the cutting-edge deep learning counterparts and a machine learning baseline, with results showing: 1) State-of-the-art performance, 2) Robustness to channel loss, and 3) Validated module efficacy via ablation studies. Conclusion and Significance: DoC-Informer bridges brain science and clinical needs by integrating anatomical priors (electrode coordinates) with adaptive deep learning. Its resilience to variable EEG configurations offers a practical solution for real-world DoC diagnosis, particularly in settings with sparse or incomplete recordings. The source code of the framework is available at https://github.com/pilonglin/docinformer.

Low-rank tensor decomposition has attracted growing attention in multivariate data analysis owning to its potentials of capturing latent representations of multivariate data correlating various data attributes. However, existing studies still have limited capabilities to represent the nonlinear relationships within multivariate data. Moreover, their performance tends to degrade under the circumstance of intensive noise embedded in the source data as well as improper setting of initial tensor rank. To tackle the challenges, this study proposes a Deep Bayesian Low-rank Tensor Decomposition (DBLTD) framework to handle the multivariate data: 1) DBLTD first establishes a probabilistic tensor decomposition model and imposes a Gaussian-Gamma sparsity prior over factor matrices and noise tensor to induce the sparse components of multidomain factors and separate the noise from the source data, respectively, 2) A generative convolution network is designed on the basis of multiple convolution blocks to produce the distributions of deep factors along with deep parameters in a self-training manner, 3) The framework develops a deep variational Bayesian inference algorithm with guaranteed convergence for model training, where the loss function is constituted of reconstruction error and sparse constraint term over deep factors and parameters. Experimental results on diverse multivariate datasets indicate that DBLTD is effective to learn the tensor rank and reduce the impact of various noises during multivariate data decomposition. Furthermore, DBLTD has demonstrated remarkable advantages in decomposition based medical applications like blind EEG denoising, structural EEG feature construction and MRI image restoration.

Decision-making on clinical Electroencephalogram (EEG) holds significant promise but faces critical reliability challenges due to signal complexity, noise, and inter-subject variability, which can lead to overconfident or untrustworthy predictions in safety-critical scenarios. This study presents TRUEE (Trusted Reliable Uncertainty Estimation framework for EEG), integrating multi-view learning and subjective logic to deliver clinically trustworthy uncertainty quantification. TRUEE advances reliability through three core innovations: 1) Multi-View Representation, decomposing EEG into temporally, spectrally, and time-frequency aligned views to capture complementary neurophysiological patterns; 2) View-specific Opinion Generation, employing evidential deep learning to transform each view's features into subjective opinions (belief masses and uncertainty); and 3) Reliable Opinion Aggregation, dynamically aggregating multi-view opinions via weighted belief fusion and reliability-aware trust discounting operation. Extensive experiments on four EEG datasets (diagnosis of Autism Spectrum Disorders (ASD), mental arithmetic, seizure detection, and motor imagery) demonstrate TRUEE's superiority over conventional uncertainty methods and multi-view baselines, achieving 86.5% classification accuracy with 0.175 expected calibration error on ASD detection. The framework shows particular strength in robustness, it maintains 74.1 %/67.1 % accuracy under cross-age generalization tasks and successfully gains highest improvement (6 %similar to 7 %) after rejecting unreliable predictions with lowest rejection rate (11.1 %similar to 23.1 %). By bridging evidential learning with multi-view neurodynamic analysis, TRUEE provides clinicians with uncertainty-aware predictions that align with real-world diagnostic challenges, offering a critical step toward deployable, safety-conscious EEG decision-making tools.

Comprehensive cognitive-affective profiling for precision ASD assessment is hindered by inadequate computational modeling of emotion-behavior dynamics. This gap persists due to limited methods for reconciling EEG, eye-tracking, and facial expression modalities-critical for capturing ASD heterogeneity drivers. This study proposes the Cognitive-Affective Ability Assessment (CA(3)) framework, focusing on attention control in conjunction with valence-arousal, to address these limitations through: (1) Ability-Specific Neural Disentanglement extracting attention- and emotion-specific representations from EEG using ET and facial expression recordings as anchors; (2) Symmetrical Cross-Ability Alignment modeling contextual dependencies via a symmetric cross-attention mechanism; and (3) Uncertainty-Aware Ability Integration fusing per-ability predictions using Dirichlet modeling and Dempster-Shafer theory while quantifying confidence. Extensive experiments have been conducted to evaluate CA(3) vs. state-of-the-arts counterparts on the multimodel datasets (EEG, ET and facial expression recordings) with CCNU (27 ASD vs. 30 typically developing (TD) children) and BNU (49 ASD vs. 48 TD), and the results demonstrate that: 1) In the ASD cognitive-affective abilities assessment task, accuracy is improved by 2.7% for attention control and 2.5% for emotion perception, and by 3.5% for cognitive-affective abilities assessment, compared with current state-of-the-art methods. 2) For attention control and valence-arousal, composite ability gaps are measured between ASD and typically developing children: 15.6% and 19.9% for mild ASD, 28.2% and 39.1% for moderate ASD, and 44.4% and 61.3% for severe ASD, respectively. Overall, the framework effectively bridges computational assessment with clinically actionable insights, enabling robust decision-making through personalized cognitive-affective ASD profiles. The source code and preprocessed two datasets for the CA(3) framework are available at https://github.com/51cloud/CA3.

Real-time decoding of brain signals is critical for applications ranging from Brain-Computer Interface (BCI) to clinical neurodiagnostics. While various learning methods are used for predicting EEG streams, Spiking Neural Networks (SNNs) offer a compelling approach due to their sparse, event-driven computation, which naturally aligns with the temporal dynamics of neural signals. However, applying SNNs to high-dimensional, evolving EEG data reveals critical limitations: deep architectures are prone to convergence failure from excessive signal diffusion, and adapting to EEG stream requires computationally expensive global retraining. To address these limitations, this study proposes a Brain-inspired Broad Learning framework (B-2 L-SNN). Mimicking hippocampal synaptic mechanisms, the framework involves: 1) a dual-layer, parallel architecture with Random Vector Functional-Link (RVFL) mappings that enables robust pattern interpretation while addressing the diffusion issue; and 2) a synaptic growth-inspired optimization that achieves efficient incremental learning by dynamically extending the network, eliminating the need for iterative global updates. Experiments on five benchmark datasets demonstrate that B-2 L-SNN significantly outperforms competing SNN variants and other methods. For instance, on a motor imagery task, it achieves 94.45% accuracy while reducing inference time by 64.9%. In dynamic tests, it maintained 92.37% accuracy with only an additional 2.74 seconds of training, proving its efficient incremental learning capability. Overall, this study provides an effective alternative SNN design, supporting the advancement of sophisticated neuroengineering applications.

Accurate detection of K-complex events in sleep electroencephalography (EEG), particularly with respect to their precise onsets and durations, is critical for research on sleep processes and clinical applications. However, the automated localization of K-complexes has long faced performance bottlenecks due to the significant individuality in their morphology across different subjects.This study develops a MUlti-Scale learning framework for K-complex localization (MUSK), which assumes that the individuality of K-complexes can be better characterized at multiple spatial/spectrum scales, with temporal context simultaneously integrated. K-complex localization operates in three phases: (1) Multi-scale feature learning: local features with multiple time and frequency scales are efficiently learned and fused by Two-Dimensional Convolutional Neural Network (2D-CNN); (2) Temporal context modeling: Long Short-Term Memory (LSTM) layers model the long-term temporal context of K-complex-related EEG evolution; and (3) By-sample decision making: CNN then provides a variety of instances of point-wise prediction to pinpoint the exact onset and duration of each K-complex event.Experiments have been carried out on the Montreal Archive of Sleep Studies (MASS) dataset against the state-of-the-art counterparts, the results indicate that: MUSK performs the best in both identification and localization of K-complexes, with F1-score, mean Intersection over Union (IoU) up to 83.9%, 87.3%. Ablation experiments validate MUSK’s effective design of introducing multi-scale features and temporal context. Overall, MUSK enables advancements driven by Artificial Intelligence (AI), including multi-scale learning, temporal context integration, and the handling of individual variability, to support more personalized and efficient neuro-engineering solutions for sleep-related problems.

Objective.Raw electroencephalography (EEG) requires robust rejection of inevitable bad EEG epochs to ensure data reliability. While automated methods reduce manual inspection burdens, existing approaches struggle with parameter optimization, scenario adaptation, and label dependency. This study presentsEEGEpochNet, an end-to-end model for accurate bad EEG epoch rejection.Approach.EEGEpochNet is developed through three modules: (1)multi-level morphological representation: A multi-branch 1D-convolutional neural network (CNN) with U-Net-encoded multi-level features captures scale-invariant patterns mimicking expert visual analysis, eliminating handcrafted feature engineering; (2)temporal evolution modeling: bidirectional gated recurrent unit decode electrophysiological dynamics to distinguish artifacts from normal activity; and 3)self-supervised contrastive learning: a symmetric loss leverages unlabeled data to learn domain-invariant EEG representations, reducing reliance on labeled examples.Main results.Extensive experiments have been performed to compareEEGEpochNetto five state-of- the-art counterparts (e.g. Autoreject and BRCNN) on a semi-simulated dataset and two real datasets (the EEG recordings from children and adults): (1) EEGEpochNet performs the best withF1-scores of 93.05%, 95.33%, and 84.41%, and (2) the capability of self-supervised learning makesEEGEpochNetfar superior to supervised methods when labeled data are limited.Significance.Overall,EEGEpochNetprovides a parameter-efficient framework to deploy reliable EEG analysis toward clinical-grade automation.

Objective.EEG-based assessment of motor response capability is critical for understanding the heterogeneous cognitive profiles of children with autism spectrum disorder (ASD), e.g. typically P300 event assessment. However, mainstream methods generally struggle to address the entangled trial-to-trial variability and temporal instability of P300 signatures in ASD, fixed time windows or trial-averaging primarily effective only for typically developing subjects.Approach.To overcome this limitation, this paper introducesDynaSPARC(DynamicSingle-trialP300Assessment ofResponseCapability), a framework based on the premise that behavioral reaction time () provides a latent regulation for inferring trial-specific cognitive dynamics.DynaSPARCenables reliable assessment of motor response capability in ASD through: (1)Dynamic Temporal Windowing, which uses a nonlinear mapping of standardizeddifferences to adaptively parameterize the start and length of the P300 window for each trial within a physiologically constrained range (e.g. 300-800 ms); (2)Collaborative Spatio-temporal Attention, which employs a learnable temporal filter to pinpoint P300 latency and a knowledge-guided channel weighting scheme for a reliable reconstruction of the P300 signature.Main results.Evaluation on an EEG dataset from 67 children (29 with ASD vs 38 typically developing) through a motor-cognitive task demonstrates thatDynaSPARCachieves superior performance over fixed-window methods: (1) A more temporally localized and spatially plausible single-trial P300 was obtained, and a substantially stronger-peak-latency coupling was preserved (Spearman= 0.61,= 1.33vs= 0.11, p = 0.391), consistent with more stable peak-latency estimates (reduced jitter); (2) The peak scalp topography showed a more canonical, spatially focused centro-parietal positivity; (3) The classification accuracy,F1score, recall, and precision reached 84.77%, 78.43%, 84.03%, and 76.14%, respectively.Significance.This work establishes a new paradigm for EEG-based assessment, moving beyond static averaging to model the dynamic interplay between neural latency and behavioral output.

Alzheimer’s disease (AD) is a common central neurodegenerative disease disorder characterized primarily by cognitive impairment and non-cognitive neuropsychiatric symptoms that significantly impact patients’ daily lives and behavioral functioning. The pathogenesis of AD remains unclear and current Western medicines treatment are purely symptomatic, with a singular pathway, limited efficacy, and substantial toxicity and side effects. In recent years, as research into AD has deepened, there has been a gradual increase in the exploration and application of medicinal plants for the treatment of AD. Numerous studies have shown that medicinal plants and their active ingredients can potentially mitigate AD by regulating various molecular mechanisms, including the production and aggregation of pathological proteins, oxidative stress, neuroinflammation, apoptosis, mitochondrial dysfunction, neurogenesis, neurotransmission, and the brain-gut microbiota axis. In this review, we analyzed the pathogenesis of AD and comprehensively summarized recent advancements in research on medicinal plants for the treatment of AD, along with their underlying mechanisms and clinical evidence. Ultimately, we aimed to provide a reference for further investigation into the specific mechanisms through which medicinal plants prevent and treat AD, as well as for the identification of efficacious active ingredients derived from medicinal plants.

Discovering biomarkers for Autism Spectrum Disorder (ASD) is essential for elucidating its etiology, enabling early diagnosis, and refining treatment strategies. Electroencephalogram (EEG) microstates reflect the brain's overall dynamic changes, aiding in exploring differences in brain function patterns between ASD and Typically Developing (TD) groups. To this end, this study proposes an adaptive EEG microstate analysis approach based on Hidden Markov Models (HMMs) for the discovery of ASD biomarkers. Specifically, the proposed method, within the HMM framework, adaptively extracts millisecondscale transient brain microstate patterns that recur over time and models microstates using a multivariate Gaussian distribution rather than static topological structures. Resting-state EEG from 178 children aged 3 to 6 are used to validate the proposed approach. The analysis of the four microstates (#1, #2, #3, and #4) reveals significant differences between ASD and TD groups. Temporally, the ASD group shows difficulty in microstate transitions, primarily between microstates #2 and #3. In the frequency and spatial domains, TD individuals exhibit stronger brain region activation and interaction in microstates #2 and #4, whereas the ASD group shows reduced activity. Notably, during microstate #3, the ASD group demonstrates higher spectral power and channel coherence. Additionally, Ttests on intergroup feature differences and the results of the ASD discrimination task (accuracy: 88.89%) further confirm the potential of microstate features in assessing ASD tendencies. Overall, this study holds the potential to reveal novel insights into the neural mechanisms underlying ASD and identify valuable biomarkers for clinical assessment and diagnosis.

Electroencephalogram (EEG) analysis has heavily relied on sophisticated machine learning methods. However, the limited availability of diverse and extensive EEG datasets often underscores the need for reliable data augmentation approaches. This study introduces a new Generative Adversarial Network framework, HiFi-EEG-GAN, consisting of a supervisor, generator, and discriminator, aiming at generating artificial EEG that closely mimics real-world counterparts with “high fidelity” (Hi-Fi). The framework emphasizes two core tasks: 1) Global Dynamics Supervision: The supervisor model distills the global dynamics of real EEG into a Gaussian-like representation. This representation regulates the subsequent EEG generation using Kullback-Leibler (KL) divergence, focusing on macroscopic dynamics; and 2) Hi-Fi EEG Generation: EEG generator replicates real EEG’s time, frequency, and spatial characteristics using a composite architecture. This process is further regulated by another discriminator, focusing on microscopic details. The HiFi-EEG-GAN framework (design validated through ablation study) outperforms state-of-the-art counterparts (e.g., FT-Surrogate, EEG-GAN, BWGAN-GP) in terms of fidelity and diversity in data augmentation. Notable performance metrics include r1NNC (0.88), FID (13.97), and MMD (0.09). In two test cases, classification accuracy improves by 3.2% 6.8% in ASD and 2.5% 8.2% in mental arithmetic tasks, surpassing its counterparts with EEG augmentation by HiFi-EEG-GAN.

Detecting sleep spindles in Electroencephalography (EEG) is crucial for assessing cognitive function and neurological disorders. How to provide robust detection across diverse datasets and clinical settings has long remained an open problem. Aiming at this challenge, this study introduces BiLSTM-Attention Spindle neT (BLAST), a deep learning framework focusing on the localization of individual spindle events operating in three phases: (1) Representation: BLAST first enhances morphological and temporal representations through the U-Net to learn hierarchical features, followed by BiLSTM (Bidirectional Long Short-Term Memory) to model the intricate temporal dynamics within EEG; (2) Highlighting: It then enforces attention to spindle bodies with an attention mechanism, which diminishes irrelevant background signals and magnifies salient features that discriminate spindles from other activities; and (3) Detection: These refined features are efficiently integrated through skip connections, preserving both abstract representations and fine-grained details for precise localization of each spindle event based on predictions at each sampling point. BLAST has been extensively evaluated against seven state-of-the-art counterparts (e.g., SEED and CDTSD) using three distinct datasets. The results demonstrate the following: (1) BLAST achieves the best overall performance, with precision, recall, and F1-score of 86.4 %, 86.9 %, and 85.9 %, respectively, on the publicly available MASS dataset; (2) It exhibits the highest consistency with expert evaluation on downstream tasks, such as spindle density and average duration, with correlation coefficients of 0.828 and 0.821, respectively; and (3) Transfer learning on the open-access DREAMS and private BCH-Children datasets further highlights BLAST's strong generalization potential across various clinical settings.

Spiking Neural Network (SNN) provides an irreplaceable mechanism for series prediction, particularly in scenarios where time and computing resources are both critical, exhibiting superiority over other deep learning models with sparse and spike-based communication inherent to SNN methods. However, as the depth of the network increases, a critical challenge arises: excessive diffusion of signals in the deeper network can seriously affect its accuracy and efficiency. Inspired by the organizational principles of hippocampal circuits, this study proposes a broad learning framework inspired by the brain (B2L), offering an alternative SNN architecture to mitigate the problem of excessive diffusion. It adopts a broad, incrementally extendable structure while adopting sparse coding for data representation, dedicated to: (1) Construction of the B2L SNN: Combining preliminary feature extraction with enhancement processing, a dual-layer broad structure establishes random vector functional-link (RVFL) mapping between rate-coded spike sequences and output labels, effectively preventing excessive diffusion to upstream neurons; (2) Extendable Optimization of B2L SNN: Drawing from the concept of incremental learning, B2L SNN employs a weight optimization method to search for optimal configurations by expanding RVFL mappings from initial architecture to static datasets through incremental neuron integration. Experiments across six benchmarks (MNIST, CIFAR-10, CIFAR-100, NORB, N-MNIST, and BCI IV 2b) demonstrate: (1) B2L SNN achieves 20.3xfaster training time than four SNN variants (4058 seconds -> 197 s) via random vector functional-link mapping; (2) 96.97% NORB accuracy (+3.64% over SNN2ANN) with 2.79 ms inference time (3.06xspeedup), also achieving SOTA 89.52% accuracy on BCI IV 2b and leading N-MNIST inference speed (3.85ms); (3) Scalability: With 1.5xexpansion in feature layers (PF: 60 -> 100, EF: 7k -> 11k), B2L SNN achieves 1.04% accuracy improvement on MNIST alongside 12.3% faster training than BLS baselines.

Deep learning for Electroencephalography (EEG) has become dominant in the tasks of discrimination and evaluation of brain disorders. However, despite its significant successes, this approach has long been facing challenges due to the limited availability of labeled samples and the individuality of subjects, particularly in complex scenarios such as Autism Spectrum Disorders (ASD). To facilitate the efficient optimization of EEG discrimination models in the face of these limitations, this study has developed a framework called STEM (Self-Training EEG Model). STEM accomplishes this by self-training the model, which involves initializing it with limited labeled samples and optimizing it with self-constructed samples. (1) Model initialization with multi-task learning: A multi-task model (MAC) comprising an AutoEncoder and a classifier offers guidance for subsequent pseudo-labeling. This guidance includes task-related latent EEG representations and prediction probabilities of unlabeled samples. The AutoEncoder, which consists of depth-separable convolutions and BiGRUs, is responsible for learning comprehensive EEG representations through the EEG reconstruction task. Meanwhile, the classifier, trained using limited labeled samples through supervised learning, directs the model's attention towards capturing task-related features. (2) Model optimization aided by pseudo-labeled samples construction: Next, trustworthy pseudo-labels are assigned to the unlabeled samples, and this approach (PLASC) combines the sample's distance relationship in the feature space mapped by the encoder with the sample's predicted probability, using the initial MAC model as a reference. The constructed pseudo-labeled samples then support the self-training of MAC to learn individual information from new subjects, potentially enhancing the adaptation of the optimized model to samples from new subjects. The STEM framework has undergone an extensive evaluation, comparing it to state-of-the-art counterparts, using resting-state EEG data collected from 175 ASD-suspicious children spanning different age groups. The observed results indicate the following: (1) STEM achieves the best performance, with an accuracy of 88.33% and an F1-score of 87.24%, and (2) STEM's multi-task learning capability outperforms supervised methods when labeled data is limited. More importantly, the use of PLASC improves the model's performance in ASD discrimination across different age groups, resulting in an increase in accuracy (3%-8%) and F1-scores (4%-10%). These increments are approximately 6% higher than those achieved by the comparison methods.

The risk of adverse effects in Electroconvulsive Therapy (ECT), such as cognitive impairment, can be high if an excessive stimulus is applied to induce the necessary generalized seizure (GS); Conversely, inadequate stimulus results in failure. Recent efforts to automate this task can facilitate statistical analyses on individual parameters or qualitative predictions. However, this automation still significantly lags behind the requirements in clinical practices. This study addresses this issue by predicting the probability of GS induction under the joint restriction of a patient's EEG (electroencephalogram) and the stimulus parameters, sustained by a two-stage learning model (namely ECTnet): 1) Temporal-Spatial Feature Learning . Channel-wise convolution via multiple convolution kernels first learns the deep features of the EEG, followed by a “ConvLSTM” constructing the temporal-spatial features aided with the enforced convolution operations at the LSTM gates; 2) GS Prediction . The probability of seizure induction is predicted based on the EEG features fused with stimulus parameters, through which the optimal parameter setting(s) may be obtained by minimizing the stimulus charge while ensuring the probability above a threshold. Experiments have been conducted on EEG data from 96 subjects with mental disorders to examine the performance and design of ECTnet. These experiments indicate that ECTnet can effectively automate the selection of optimal stimulus parameters: 1) an AUC of 0.746, F1-score of 0.90, a precision of 89% and a recall of 93% in the prediction of seizure induction have been achieved, outperforming the state-of-the-art counterpart, and 2) inclusion of parameter features increases the F1-score by 0.054.

Electroencephalogram (EEG) artifact removal has been investigated for decades with the goal of reconstructing the clean signals for the subsequent EEG analysis. However, existing denoising methods still have limited capabilities to handle the highly mixed artifacts and the fine-grained temporal dependency of artifact-free EEG without a priori knowledge of the artifacts. To address the challenges, this study proposes a CNN-Transformer-based dual-stage collaborative ensemble learning framework (namely CT-DCENet) in the form of three modules: 1) randomized collaboration module initially utilizes four individual learners to reveal multi-group morphological characteristics of the denoised EEG, 2) linear ensemble module integrates the outputs of four individual learners via weighted linear combination to preliminarily estimate the denoised EEG, 3) information complementation module takes in the residual between the contaminated EEG and the above estimated EEG, and critically applies CNN-Transformer-based feature extractor and denoising head to learn the detailed characteristics of the denoised EEG. CT-DCENet is conducted in a dual-stage training manner to derive the morphological characteristics & the detailed characteristics of the artifact-free EEG successively. The experimental results on the public EEG datasets indicate that 1) CT-DCENet significantly outperforms the state-of-the-art counterparts (e.g., DuoCL, GCTNet) under the conditions of various artifacts and noise intensities, where the increases of SNR & PCC are 0.79 dB, 0.6% and the decrease of RRMSE is 1.9% for the removal of EMG, ECG, EOG mixed artifacts, 2) the reconstructed EEG by CT-DCENet can well fit the clean EEG with a low error achieved, especially for the peak amplitude, the high-frequency area and the boundary area of the EEG waveform, providing promising EEG data for the downstream task-oriented EEG analysis.

Facial expression recognition plays a key role in promoting the development of comprehensive intelligence and building friendly human-computer interaction. Due to the interference of feature noise in expression data, the lightweight facial expression recognition model with fewer parameters is difficult to learn more expression features through simple training, which limits the improvement of its recognition performance. An efficient facial expression recognition network based on Spot-adaptive Knowledge Distillation is proposed in this paper. Inspired by VoVNetV2, the network designed in this paper is lightweightly improved using Depthwise Separable Convolution and the parameter-free SimAM attention mechanism, reducing the number of parameters to 0.21 M. To further improve the recognition accuracy of the model, Spot-adaptive Knowledge Distillation is employed to improve the characterization ability of the model. The recognition accuracies of the student network designed in this paper on the KDEF and RAF-DB datasets are 93.05% and 81.17% respectively after spot-adaptive distillation.