Combined Study of Behavior and Spike Discharges Associated with Negative Emotions in Mice

Abstract

In modern society, people are increasingly exposed to chronic stress, leading to various mental disorders. However, the activities of brain regions, especially neural firing patterns related to specific behaviors, remain unclear. In this study, we introduce a novel approach, NeuroSync, which integrates open-field behavioral testing with electrophysiological recordings from emotion-related brain regions, specifically the central amygdala and the paraventricular nucleus of the hypothalamus, to explore the mechanisms of negative emotions induced by chronic stress in mice. By applying machine vision techniques, we quantified behaviors in the open field, and signal processing algorithms elucidated the neural underpinnings of the observed behaviors. Synchronizing behavioral and electrophysiological data revealed significant correlations between neural firing patterns and stress-related behaviors, providing insights into real-time brain activity underlying behavioral responses. This research combines deep learning and machine learning to synchronize high-resolution video and electrophysiological data, offering new insights into neural-behavioral dynamics under chronic stress conditions.

Introduction

In modern society, individuals are increasingly subjected to various forms of chronic stress, including work-related pressures, financial insecurity, and social demands [1]. These persistent stressors are known to contribute to the development of several psychiatric disorders, such as anxiety, depression, and post-traumatic stress disorder [2]. Despite extensive research, the precise neural mechanisms underlying these conditions, particularly depression, remain elusive. Although there’s no definitive and consistently effective treatment for depression in clinical practice, deep brain stimulation (DBS) has emerged as a promising therapeutic approach [3]. Nevertheless, the broad application is limited by an incomplete understanding of the neural firing patterns in key brain regions during depressive states [4]. Elucidating these neural firing patterns is essential for improving the efficacy of DBS and developing more targeted interventions for the treatment of depression and related disorders [5].

Currently, depression-like behaviors can be evaluated in different rodent models [6]. The chronic stress model induces anhedonia through unpredictable, mild stressors and is commonly used to study the loss of pleasure associated with depression [7]. The lipopolysaccharide (LPS) model mimics neuroinflammation-driven depression, exploring the relationship between neuroinflammation and depressive behaviors [8]. The social defeat model exposes rodents to repeated aggression, leading to social withdrawal and despair, making it an important tool for studying the effects of social stress on depression [9]. In terms of behavioral assays, the open-field test evaluates anxiety-like behaviors by assessing the activity levels of mice in an open space [10]. The social defeat test measures aggressive responses and submissive behaviors when mice are confronted by an intruder [11]. The forced swim test [12] and tail suspension test [13] assess despair and depression-like behaviors by measuring the immobility time in inescapable situations. Despite the usefulness of these models and assays, current methods often rely on broad observations and lack the precision required to detect more subtle behavioral changes, highlighting the need for more refined and quantitative approaches [14].

Multichannel electrode recordings enable detailed analysis of neural activity in depression research by simultaneously capturing electrical signals from multiple brain regions [15]. This technique provides high-resolution data on neural oscillations and connectivity patterns and is used to investigate abnormal neural circuit dynamics in animal models of depression, revealing disruptions in brain network synchronization associated with depressive behaviors [16]. In addition, multichannel recordings assess the impact of antidepressant treatments on neural activity, offering insights into how these therapies alter brain function at a precise level [17]. This technique is widely used to study the roles of the central amygdala (CEA) and the paraventricular nucleus of the hypothalamus (PVN) in negative emotional disorders. Prolonged exposure to negative emotions enhances the excitability of the CEA [18], which, through α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptor-mediated synaptic transmission, induces depressive-like behaviors. As the primary regulatory center of the hypothalamic-pituitary-adrenal (HPA) axis, the PVN leads to excessive activation of this axis, causing persistent secretion of corticotropin-releasing hormone and vasopressin, resulting in physiological and emotional dysregulation [19]. However, there have been no reported studies on multichannel electrode recordings that are synchronized with behavioral assessments.

To address these challenges, we established a negative emotion model to examine the impact of negative emotions on mouse behavior and the underlying neural mechanisms. During data collection, we simultaneously recorded both behavioral videos and electrophysiological waveforms from the mice. The analysis of the behavioral data used deep learning models for emotion recognition, identifying various behavioral patterns during the open-field experiment. For the electrophysiological data, signal processing techniques, including filters and machine learning classification algorithms, were applied to extract neuronal spike discharges. This integrated analytical approach provides valuable insights into the relationship between negative emotions and neural activity.

Materials and Methods

Animals

All experiments involving mice were approved by the Animal Care and Use Committee of the University of Science and Technology of China. A total of 24 male C57BL/6 mice, aged 6 to 8 weeks, were used in this study [20]. The mice were housed in groups of 3 to 4 per cage in a controlled environment, with the temperature maintained at ~22°C and humidity between 35% and 55%.

Mouse Model

To induce negative emotions, restraint stress, and LPS injection were used. Mice in the restraint stress model were confined in perforated plastic tubes for 6–8 h daily over 14 days [21]. In the LPS model, LPS (3 mg/mL in sterile saline) was administered intraperitoneally (300 µg/kg) once daily for 28 days [22].

The experiment followed a structured timeline (Fig. 1). After a 7-day habituation, mice underwent electrode implantation and a 7-day recovery. Stress induction was then initiated, with simultaneous electrophysiological recordings from the CEA and PVN. This was followed by behavioral tests, including the open field test (OFT), elevated plus maze (EPM), and sucrose preference test (SPT), conducted at 24-h intervals.

Fig. 1.

Detailed Experimental Timeline of the Study

Construction of Behavioral Testing Platforms

To systematically evaluate the behavioral and emotional changes in mice with induced negative emotions, we established a series of behavioral testing platforms. These platforms quantitatively assessed key behavioral parameters, such as activity levels, anxiety-related behaviors, and reward preferences, providing objective measures of the modeling outcomes. The following sections detail the methods used to construct these platforms, which served as the foundation for the subsequent behavioral analyses.

Elevated Plus Maze Testing Platform

The EPM was used to assess anxiety levels and exploratory behavior in mice [23]. The apparatus consisted of a central platform, two open arms, and two enclosed arms, representing environments with varying levels of perceived safety. The open arms, measuring 5 cm in width and 30 cm in length, were designed without walls to simulate an exposed environment, while the enclosed arms, of the same dimensions, included walls ~15 cm in height to create a secure, enclosed environment. At the intersection of all four arms was the central platform, measuring 5 cm × 5 cm. Each mouse was carefully positioned on the central platform, oriented toward one of the open arms, and allowed to freely explore the maze for 5 min. During the experiment, the number of entries into the open arms was recorded to evaluate anxiety-related behavior. Increased exploration of the open arms typically indicates reduced anxiety levels, while decreased exploration reflects heightened anxiety.

Sucrose Preference Testing Platform

To evaluate preferential behavior toward reward stimuli, the SPT was applied. This test measures the hedonic capacity of mice and serves as an indicator of depressive-like or anhedonic behavior in preclinical models [24]. The experimental setup included two identical drinking bottles—one filled with 1% (w/v) sucrose solution and the other with pure water—fixed inside the cage to ensure free access by the mice. Conducted in a quiet environment with evenly distributed lighting, the test consisted of two phases: adaptation and testing. During the 48-h adaptation phase, both bottles were filled with pure water, and their positions were exchanged daily to eliminate positional bias. In the subsequent 24-h testing phase, one bottle was replaced with 1% sucrose solution, while the other remained filled with pure water. The remaining volume in each bottle was recorded every 12 h, and the sucrose preference index (SPI) was calculated to analyze the hedonic behavior of the mice. The SPI was determined using the following formula:

$$SPI=\frac{\text{Sucrose consumption}}{\text{Total liquid consumption}}\times 100\%$$ 1

Open Field Testing Platform

The OFT was conducted to evaluate activity levels and anxiety-related behaviors, key indicators of emotional state, and exploratory tendencies in mice [25]. The apparatus consisted of a 50 cm × 50 cm square arena enclosed by 30 cm-high walls. The field was divided into three zones: a central region (four innermost squares of a 4 × 4 grid), a peripheral region (all non-corner outer squares), and a corner region (four corner squares). Each mouse was placed in the center at the start and allowed to explore freely for 5 min. Movement trajectories, central region entries, and time spent in each zone were recorded to assess activity and spatial preferences. In addition, the arena was used to synchronize electrophysiological recordings with behavioral video data, both being recorded for the 5-min duration of the test, enabling an integrated analysis of neural activity and behavior.

Multi-tetrode Electrophysiological Recording

The electrodes used in the study consisted of a 4×4 16-channel microelectrode array (model BTAME04-I, manufactured by Nanjing Changxiang Medical Technology Co., Ltd., China), constructed from nichrome alloy and insulated with an external polytetrafluoroethylene layer. These electrodes were evenly spaced, with an inter-electrode distance of 250 μm, a total diameter of 35 μm, and an impedance of < 2 MΩ [26]. The electrodes were implanted into the CEA and PVN regions, following the mouse brain stereotaxic atlas, and were left in place for a 14-day postoperative recovery period. Afterward, brain signals were recorded from the mice using the Cereplex Direct system from Blackrock Neurotech, at a sampling rate of 30,000 Hz.

Hardware and Software Equipment

Both Windows and Linux systems are compatible for key-point annotation, pose clustering recognition, and neuronal spike detection and classification. We used a digital camera to record behavioral videos, ensuring accurate posture recognition by setting the video resolution to at least 1920 × 1080 pixels, with a bit rate of 1 Mbps and a frame rate of 30 Hz. To train the tracking model, a desktop or server equipped with a GPU was required. The hardware configuration in our setup included a 64 GB 2666 MHz DDR4 RAM, Intel Core i5-12400 CPU, and a NVIDIA GeForce GTX 3080 GPU. For stable program operation, the following versions of the software were required: Pytorch ≥ 1.13.1, CUDA ≥ 11.8, MatLab ≥ R2022a, Ultralytics ≥ 8.2.32, and Nvidia Driver ≥ 520.06.

Research Workflow

To investigate the effects of negative emotions, we established the NeuroSync workflow (Fig. 2A). First, we established a negative emotion mouse model using restraint stress and LPS injection. Electrodes were implanted in both the CEA and PVN regions to enable the simultaneous recording of behavioral videos and electrophysiological waveforms, forming the foundational dataset for subsequent analyses. For emotion recognition in the open field behavioral videos, we developed and trained the OF-Tracking Net based on a deep learning model, which addressed challenges related to diverse posture recognition and the occlusion caused by head-mounted electrodes. For multichannel electrode recordings across different brain regions, we applied a Butterworth high-pass filter and used a random forest classification algorithm to accurately extract neuronal spike discharges. Through the NeuroSync approach, we uncovered relationships between specific neural circuits and negative emotional behaviors by synchronously analyzing both behavioral data and neuronal activity.

Fig. 2.

Integrated workflow of behavioral analysis, neural tracking, and signal processing in depression model mice. A The NeuroSync method synchronizes behavioral video recordings with neural signals following depression induction via restraint stress and LPS injection. B Enhanced feature extraction and detection using the OF-Tracking Net, incorporating linear attention and dynamic deformable convolution. C Electrical signal processing involving filtering, feature extraction, spike detection, and clustering to identify distinct neuronal populations.

Mouse Posture Classification Based on OF-Tracking Net

The method for mouse posture classification was as follows. First, the open-field behavioral videos of the mice were processed through the OF-Tracking Net to obtain bounding boxes and key-point coordinates for each frame. Next, the distance and angle feature information of the key points was calculated to extract the features representing the current posture of the mouse. Finally, the Ward hierarchical clustering algorithm was applied to group similar postures across frames, producing the mouse posture sequence as the output.

OF-Tracking Net Key-point Detection

To recognize and analyze mouse postures in the open field experiment, we developed the OF-Tracking Net, based on the YOLO v8 backbone [27] (Fig. 2B). This model incorporated a linear attention mechanism to enhance its ability to focus on key areas, thereby improving key-point detection. In addition, a dynamic snake convolution module was implemented to capture finer details and edge information, significantly improving the model's robustness under occlusion conditions. The structure of the MousePoseKey dataset and each component of the OF-Tracking Net are described as follows:

1. MousePoseKey Dataset and Hyperparameter Settings

To achieve mouse pose recognition and synchronization analysis with electrophysiological signals, we established a dataset to comprehensively evaluate the proposed model and other state-of-the-art baselines. The dataset consisted of 1,000 randomly-extracted frames from 20 open-field behavior videos of 10 mice wearing head-mounted electrodes. Using the LabelMe software, the coordinates of the mouse's nose, ears, tail, and bounding box were annotated to create the mouse key-point dataset. Key points were marked at the center of the respective parts when clearly visible and recorded as “visible”. For occluded key points, annotators used their expertise to reasonably estimate the locations, which were then recorded as invisible. Details of the dataset are provided in Table 1.

Table 1.

Summary of the MousePoseKey Dataset

Features	Occluded points	Visible points
Snout	852	148
Left ear	793	207
Right ear	826	174
Tail	917	83

For the training and validation process, 80% of the data was allocated to model training, while the remaining 20% was reserved for validation. The training was set to run for 300 epochs, with early stopping applied if the model's validation loss did not improve for 20 consecutive epochs, to prevent overfitting.

• (2)YOLO v8 Basic Structure

The basic network model of YOLO v8 consists of a feature extraction network, a feature enhancement network, and a regression output network.

The feature extraction network forms the foundation of the model, responsible for extracting features from the input images. It is composed of a series of convolutional blocks, the C2f module, and the Spatial Pyramid Pooling - Fast (SPPF) module. The convolutional block consists of a 2D convolution layer with a kernel size of 3×3, a stride of 2, and a padding width of 1, as well as a BatchNorm2d layer for batch normalization and a Sigmoid Linear Unit activation function. This block is used for the initial extraction of low-level features from the image while reducing the size of the feature map. The C2f module comprises two 2D convolution layers with a kernel size of 1×1, a stride of 1, and a padding width of 0, along with two Bottleneck blocks and a concatenation module. It is designed to capture basic image features while gathering multi-scale information, providing richer feature representations for subsequent tasks. The SPPF module includes two 1×1 kernel size 2D convolution layers with a stride of 1 and a padding width of 0, a concat block, and three MaxPool2d layers. This reduces the computational load and allows the model to extract features at different scales, thereby capturing detailed information about the target object across various scales.

The feature enhancement network is responsible for merging feature maps of different scales to enhance the model's multi-scale detection capability. YOLO v8 employs PANet as the neck, which connects the backbone and head to enhance multi-scale feature fusion. PANet utilizes both a bottom-up path and a top-down path, where up-sampling restores low-resolution feature maps to high resolution, and down-sampling reduces high-resolution feature maps. This process facilitates the fusion and transfer of information across different scales. In addition, PANet uses lateral concatenation operations to merge feature maps from different levels, enhancing the richness and diversity of feature representations. The fused feature maps contain multi-scale information ranging from low-level to high-level, thereby improving the accuracy of object detection and pose recognition.

The regression output network consists of three detection modules and a bidirectional coupled head with a regression branch and a prediction branch. Each detection module comprises two convolutional blocks with a kernel size of 3 × 3, a stride of 1, and a padding width of 1, along with a 2D convolutional layer with a kernel size of 1 × 1, a stride of 1, and a padding width of 0. The bidirectionally coupled head is used for both bounding box prediction and class prediction. Each branch contains two convolutional blocks, with a separate Conv2d layer at the end to provide the bounding box loss and class loss.

• (3)Linear Attention Mechanism

To enhance the OF-Tracking Net's ability to handle diverse postures, we introduced a linear attention mechanism (LA) module after the SPPF layer [28]. This module has lower computational complexity and higher processing efficiency, allowing it to focus on key features while reducing interference from irrelevant background information.

The linear attention mechanism adds a Depthwise Convolution (DWC) module to the attention matrix, enhancing the model's ability to express feature diversity. Here, we define $O$ as computational complexity, and the output can be represented as:

$$O=\varphi (Q)\varphi {(K)}^{T}V+\text{DWC}\left(V\right)$$ 2

In each query process, the DWC module focuses only on a few adjacent features $V$ in space rather than all features. This locality ensures that even if the linear attention values for two queries are the same, different outputs can still be obtained from different local features, thereby preserving feature diversity. This can be expressed as:

$$O=\left(\begin{array}{c}\varphi (Q)\varphi {(K)}^T+M_{\text{DWC}}\end{array}\right)V=M_{\mathit{eq}}V$$ 3

Subsequently, a new mapping function is used to mimic the sharp distribution of the original softmax attention, allowing the module to benefit from both linear complexity and high expressiveness. Finally, the module can be formulated as:

$$O=\text{Sim}(Q,K)V={\varphi }_p(Q){\varphi }_p{(K)}^{T}V+\text{DWC}(V)$$ 4

The linear attention mechanism, through cross-layer connections and feature fusion with the SPPF layer, enables the model to identify and enhance important features for more effective feature extraction.

• (4)Dynamic Snake Convolution

To enhance the model's performance when handling occluded parts, a dynamic snake convolution (DSC) module was introduced and embedded within the C2F module [29]. This module improves the model's ability to capture features in complex scenes by adaptively altering the shape of the convolution kernels.

DSC flattens the standard convolution kernel. Taking a standard 2D convolution as an example, with the central coordinate $K_i=(x_i,y_i)$, the specific position of each grid in K is represented as: $K_{i\pm c}=(x_{i\pm c},y_{i\pm c})$, where $c=\{0,1,2,3,4\}$ indicates the horizontal distance from the central grid. The stretched convolution kernel is represented along the x-axis and y-axis as:

$$K_{i\pm c}=\left\{\begin{array}{c}(x_{i+c},y_{i+c})=(x_i+c,y_i+{\mathrm{\Sigma }}_i^{i+c}\mathrm{\Delta }y)\hfill \\ (x_{i-c},y_{i-c})=(x_i-c,y_i+{\mathrm{\Sigma }}_{i-c}^i\mathrm{\Delta }y)\hfill \end{array}\right)$$ 5

$$K_{j\pm c}=\left\{\begin{array}{c}(x_{j+c},y_{j+c})=(x_j+{\mathrm{\Sigma }}_j^{j+c}\mathrm{\Delta }x,y_j+c)\hfill \\ (x_{j-c},y_{j-c})=(x_j+{\mathrm{\Sigma }}_{j-c}^j\mathrm{\Delta }x,y_j-c)\hfill \end{array}\right)$$ 6

Therefore, the bilinear interpolation kernel will ultimately be decomposed into two one-dimensional kernels as:

$$B(K,K^{\prime })=b(K_x,K_x^{\prime })·b(K_y,K_y^{\prime })$$ 7

When key-points are occluded, DSC allows the convolution kernels to adaptively adjust their shape during the learning process based on the input feature map. This enables the model to better capture elongated and curved structures near the occluded points, further improving the precision of key-point detection when handling complex and irregular targets.

The final model outputs the coordinates of four key points on the mouse's body, recorded as:

• $p_1(x_1,y_1)$, the center point of the nose tip.

• $p_2(x_2,y_2)$, the center point of the left ear.

• $p_3(x_3,y_3)$, the center point of the right ear.

• $p_4(x_4,y_4)$, the base point of the tail.

Posture Feature Extraction

To accurately evaluate the current posture of the mouse, we calculated five static features, i.e., features extracted from each frame, and two dynamic features, i.e., features extracted from two consecutive frames.

Based on the key-point coordinates, distance features can be obtained using the Euclidean distance formula: nose-to-left ear, nose-to-right ear, and nose-to-tail, recorded as $\rho$.

$${\rho }_{\mathit{ab}}=\sqrt{{\left(x_a-x_b\right)}^2+{\left(y_a-y_b\right)}^2},|X|=\sqrt{x_a^2+y_b^2}(1\le a,b\le 4)$$ 8

The cosine theorem can be used to calculate angle features: right ear-nose-tail and left ear-nose-tail, recorded as $cos\alpha$.

$$cos{\alpha }_a=\frac{{\rho }_{\mathit{ab}}^2+{\rho }_{\mathit{ac}}^2-{\rho }_{\mathit{bc}}^2}{2{\rho }_{\mathit{ab}}{\rho }_{\mathit{ac}}}(1\le a,b,c\le 4)$$ 9

By calculating the sum of the distances between each key point in two consecutive frames and then averaging them, the movement distance of the mouse between adjacent frames can be determined and recorded as ${\rho }_{\mathit{frame}}$.

$${\rho }_{\mathit{frame}}=\frac{\sqrt{\sum _{i=1}^4{\left(x_{i(n)}-x_{i(n-1)}\right)}^2}}{4}(n\ge 2)$$ 10

By aligning the starting points of the central axis (the line connecting the nose and the tail) through translation, and then applying the cosine theorem to calculate the angle between the central axes, the movement angle of the mouse between adjacent frames can be obtained, recorded as $cos{\alpha }_{\mathit{frame}}$.

$$cos{\alpha }_{\mathit{frame}}=\frac{{\rho }_{14(n)}^2+{\rho }_{14(n-1)}^2-{\rho }_{4(n)(n-1)}^2}{2{\rho }_{14(n)}{\rho }_{14(n-1)}}(n\ge 2)$$ 11

Posture Clustering Recognition

We weighted and normalized the calculated features, then applied the Ward hierarchical clustering algorithm to merge the posture information of each frame, completing the recognition of the mouse's current behavior [30].

When two adjacent clusters, $A$ and $B$, are merged, the within-cluster sum of squared variances before and after merging needs to be calculated as:

$$I_{\mathit{AB}}=SS{E}_{\mathit{AB}}-(SS{E}_A+SS{E}_B)$$ 12

In the formula, $I_{\mathit{ab}}$ is used to represent the inter-cluster distance. The within-cluster sum of squared variances before and after merging are, respectively:

$$SS{E}_A=\sum _{i=1}^{n_A}{\left(y_i-{\overline{y}}_A\right)}^{\prime }\left(y_i-{\overline{y}}_A\right)\text{for}{y}_i\in A$$ 13

$$SS{E}_B=\sum _{i=1}^{n_B}{\left(y_i-{\overline{y}}_B\right)}^{\prime }\left(y_i-{\overline{y}}_B\right)\text{for}{y}_i\in B$$ 14

$$SS{E}_{\mathit{AB}}=\sum _{i=1}^{n_{\mathit{AB}}}{\left(y_i-{\overline{y}}_{\mathit{AB}}\right)}^{\prime }\left(y_i-{\overline{y}}_{\mathit{AB}}\right)\text{for}{y}_i\in AB$$ 15

The value ${\overline{y}}_{\mathit{AB}}$ in the formula can be calculated using the following equation:

$${\overline{y}}_{\mathit{AB}}=\frac{\sum _{i=1}^{n_A}{y}_i+\sum _{j=1}^{n_B}{y}_j}{n_A+n_B}=\frac{n_A{\overline{y}}_A+n_B{\overline{y}}_B}{n_A+n_B}$$ 16

Therefore, according to the Ward method, the two clusters $A$ and $B$ should be merged such that the increase in the $\mathit{SSE}$ before and after merging is minimized. Substituting the above expressions and simplifying yields:

$$\begin{array}{c}{I}_{\mathit{AB}}=SS{E}_{\mathit{AB}}-(SS{E}_A+SS{E}_B)\\ =\frac{n_A{n}_B}{n_A+n_B}{({\overline{y}}_A-{\overline{y}}_B)}^{\prime }({\overline{y}}_A-{\overline{y}}_B)\\ \end{array}$$ 17

In the formula, $y_A$ and $y_B$ are the cluster centers of categories $A$ and $B$, respectively, $n_A$ and $n_A$ are the total number of samples in categories clusters $A$ and $B$.

Using the Ward hierarchical clustering method, we effectively recognized and classified mouse postures in the open-field experiment. Five main action categories were identified:

1. Stationary: The mouse is in a completely stationary state with no significant body movement, primarily used to assess its resting and alertness states.

2. Sniffing: The mouse lowers its head to sniff the ground or the surrounding environment, exhibiting exploratory behavior related to curiosity and adaptability to a new environment.

3. Climbing: The mouse uses its forelimbs to grasp objects for climbing or attempts to overcome obstacles, reflecting its activity capability and environmental exploration behavior.

4. Walking: The mouse freely moves around the field, displaying relatively continuous movement trajectories, which helps evaluate its locomotor abilities.

5. Stepping: The mouse moves at a slower pace, typically showing cautious exploration, which can be used to assess its anxiety levels and adaptability to the environment.

Spike Discharge Extraction

The raw signals recorded by the electrodes are continuous time-series data [31]. To quantify the activity of the two brain regions for subsequent analysis, we developed a comprehensive spike discharge extraction process. This process is divided into three stages: signal preprocessing for multielectrode recordings, spike waveform feature calculation, and classification using a random forest algorithm (Fig. 2C).

Preprocessing of Multielectrode Recording Signals

We used a 4th-order Butterworth high-pass filter with a cutoff frequency of 250 Hz to filter the raw signals. This filtering process effectively removes electrode drift and motion artifacts caused by poor contact between the electrodes and the skin, ensuring the clarity of neuronal spike signals. The 4th-order filter provides a sufficiently steep filtering effect while minimizing excessive phase distortion, preserving the temporal characteristics of the filtered signals. Following this, baseline correction is applied to the filtered signals to stabilize the baseline around zero, ensuring the accuracy of subsequent spike detection and classification.

Calculation of Spike Waveform Features

After obtaining the potential spike discharge segments, we calculated a total of eight waveform features: minimum spike value, left-side maximum value, right-side maximum value, total waveform width, left-side width, right-side width, left-side slope, and right-side slope. These features provide valuable information reflecting various functional characteristics of a neuronal discharge, such as discharge intensity, duration, and the degree and rate of depolarization and repolarization.

Random Forest Classification Algorithm

We applied a random forest algorithm to extract spike discharges by constructing and integrating multiple decision trees [32]. This algorithm creates a set of independent decision tree models by randomly selecting samples and features and recursively splitting nodes to form tree structures. The final classification result is determined through majority voting or averaging, which not only captures the diversity in the data but also significantly reduces the risk of overfitting associated with a single model. The algorithm workflow is as follows:

Given a training dataset $\mathcal{D}={\{(x_i,y_i)\}}_{i=1}^n$, where $x_i$ represents the input feature vector and $y_i$ denotes the target class label, the Random Forest algorithm begins by generating $b$ bootstrap samples $D_b$ from the original dataset $D$. Each sample $D_b$ is used to train an individual decision tree $h_b(x)$.

During the construction of each decision tree, at each node split, a subset of $m$ features ${\mathcal{M}}_b\subseteq \{1,2,\dots ,p\}$ is randomly selected from the total $p$ features. The feature $j$ and split point $s$ are chosen to minimize the Gini impurity $G(t)$ where $t$ refers to a specific node in the decision tree, for the resulting nodes:

$$(j,s)=arg\underset{(j,s)}{min}\left[\frac{N_{\text{left}}}{N}·G(\text{left})+\frac{N_{\text{right}}}{N}·G(\text{right})\right]$$ 18

Where $G(t)=1-\sum _{k=1}^K{p}_{k,t}^2$ represents the Gini impurity of node $t$, with $p_{k,t}$ being the proportion of samples in node $t$ that belong to the class $k$. $N_{\mathit{left}}$ and $N_{\mathit{right}}$ are the sample counts in the left and right child nodes, respectively.

The final classification outcome of the Random Forest is determined by majority voting across all decision trees, where the predicted class $\widehat{y}$ is:

$$\widehat{y}={\text{argmax}}_c\sum _{b=1}^B\text{I}(h_b(x)=c)$$ 19

Here, $I$ is an indicator function that equals 1 if $h_b(x)$ predicts class $c$, and 0 otherwise. This approach effectively reduces the risk of overfitting and enhances the model's generalization performance.

To further characterize neuronal firing patterns across different behavioral states and experimental groups, we constructed high-dimensional feature vectors from the spike discharge statistics obtained in the CEA and PVN regions. Principal component analysis (PCA) was then applied to project these vectors into a low-dimensional space. The first principal component (PC1), which captured the greatest variance across behaviors, was retained for subsequent clustering and visualization analysis.

One-Way Analysis of Variance and Significance Testing

We used one-way analysis of variance (ANOVA) to compare the means of multiple groups and determine whether significant differences existed [33], thereby assessing the impact of experimental conditions on the results. Hypotheses were formulated, with the null hypothesis (H₀) stating that all group means are equal (μ₁ = μ₂ = μ₃ = …), while the alternative hypothesis (Hₐ) posited that at least one group mean is different. The total variance was partitioned into between-group variance and within-group variance, quantified as the sum of squares between groups (SSB), the sum of squares within groups (SSW), and the total sum of squares (SST). The mean square between groups (MSB) and the mean square within groups (MSW) are calculated as follows:

$$MSB=SSB/\left(k-1\right)$$ 20

$$MSW=SSW/\left(N-k\right)$$ 21

Where k represents the number of groups, and N represents the total number of observations. The F-statistic is computed as the ratio of MSB to MSW.

The F-value, along with its associated degrees of freedom (k-1 for between groups and N-k for within groups), was used to determine the significance level (P-value) through statistical software. A P-value <0.05 was considered statistically significant, leading to the rejection of the null hypothesis and indicating that at least one group mean was significantly different.

Following the identification of significant group differences through ANOVA, Dunnett’s post hoc test was applied to compare each experimental group against the control group while controlling for multiple comparisons. The test computes the mean difference (Δμ) between each experimental group and the control group and evaluates its significance using the standard error (SE), which is calculated as:

$$SE=sqrt\left(MSW/n\right)$$ 22

Where n represents the sample size of each group, and MSW is the mean square within groups from the ANOVA. The test statistic $t$ is subsequently calculated as:

$$t=\mathrm{\Delta }\mu /SE$$ 23

The t value is compared against the critical values derived from the Dunnett distribution to determine statistical significance. In the figures, varying levels of significance are indicated by asterisks: ns for P > 0.05, * for P ≤ 0.05, ** for P ≤ 0.01, and *** for P ≤ 0.001.

Results

Validation of Chronic Stress Models

The open field was divided into three regions: center, edge, and corner, to accurately evaluate the mice's positional distribution (Fig. 3A). The heatmap illustrates the activity patterns of the mice. In the control group, the activity trajectories were broadly distributed across the entire experimental field, reflecting uniform exploratory behavior and normal activity levels (Fig. 3B). In contrast, the activity trajectories of the negative emotion model mice were significantly reduced and primarily concentrated in the edge areas of the field, indicating pronounced edge dependency (Fig. 3C). This behavior pattern reflects a marked suppression of exploratory behavior in the negative emotion model mice, as well as increased anxiety and avoidance behaviors. These behavioral changes are consistent with the negative emotional state induced by chronic stress modeling.

Fig. 3.

Behavioral comparisons between control and negative emotion model mice. A Diagram of Open Field Region Division. B, C Heatmaps of activity trajectories for control and negative emotion model mice in the open field test. D Comparison of shuttle crossing times between the restraint stress, LPS injection, and control groups. E Comparison of open arm residence between groups. F Comparison of sucrose preference index between groups.

In terms of exploratory behavior, we compared the number of cross-regional movements between the control group and the negative emotion model mice in the OFT (Fig. 3D). The bar graph shows that the control group mice had an average of 10 shuttle crosses, higher than the 7 in the negative emotion model mice. This result indicated that the exploratory behavior of the negative emotion model mice was severely restricted, with a significantly reduced range of activity and decreased environmental adaptability, consistent with the behavioral characteristics of negative emotional states.

The behavioral parameters (Table 2) from the OFT show that the total displacement for the restraint stress group was 8,215 pixels and for the LPS injection group was 8,932 pixels, both significantly lower than the 36,931 pixels found in the control group, indicating a marked decrease in overall activity levels in the negative emotion model mice. Similarly, the movement time for the restraint stress and LPS injection groups was 73.2 s and 61.8 s, respectively, compared to 209.8 s in the control group, further reflecting a significant reduction in exploratory behavior. This reduction in movement time and overall activity suggests that the negative emotion model mice exhibited heightened anxiety and avoidance behaviors, particularly in open spaces. In addition, the reduced time spent in the central area and the increased time spent in the corner and edge areas further support the finding of pronounced edge dependency and anxiety-like behavior. These significant changes in behavioral parameters provide strong evidence for the successful modeling of chronic stress-induced negative emotions.

Table 2.

Comparison of behavioral parameters in the open field test between negative emotion model mice and control mice.

Parameters	Control (Mean ± SEM)	Restraint stress (Mean ± SEM)	LPS injection (Mean ± SEM)	P-value (ANOVA)
Displacement (pixels)	36931±4096	8215±654	8932±930	∗∗∗
Movement time (s)	209.8±17.2	73.2±10.8	61.8±9.1	∗∗∗
Velocity (pixels/s)	123.1±12.7	31.4±3.7	25.3±1.9	∗∗
Center residence time (s)	80.1±9.0	4.7±0.7	3.2±1.2	∗∗∗
Corner residence time (s)	109.7±12.2	156.6±17.8	163.8±22.7	∗∗
Edge residence time (s)	110.2±10.5	138.7±10.0	133.0±8.2	∗

To further validate the success of the model, we conducted EPM and SPT. Data from the EPM revealed that the average time spent in the open arms was 17 s for the control mice, which was significantly longer than the 4 s in the negative emotion model groups, indicating anxiety-like behavior in the stressed mice (Fig. 3E , P = 0.001). In addition, the sucrose preference index in the restraint stress and LPS injection groups was 61%, significantly lower than the 72% recorded in the control group, suggesting a deficit in hedonic behavior in the negative emotion model mice (Fig. 3F, P = 0.007).

Algorithm Performance Comparison

The results of our model comparison demonstrated that the OF-Tracking Net outperformed other commonly-used methods on the MousePosekey dataset, including Detrs [34], D-FINE [35], Hpnet [36], RealNet [37], and AICellCounter [38], in both model loss and accuracy (Fig. 4A, B). OF-Tracking Net achieved the lowest model loss, indicating superior performance in minimizing prediction errors over successive training epochs. In addition, it exhibited the highest accuracy, maintaining consistent improvement throughout the training process. This demonstrated the effectiveness of the linear attention mechanism and dynamic snake convolution, which were crucial for handling occlusion and complex mouse behavior recognition, making OF-Tracking Net the optimal choice for precise and reliable behavior analysis in experimental settings.

Fig. 4.

Performance comparison of different algorithms in mouse behavior recognition on the MousePoseKey dataset. A Loss across training epochs, with the OF-Tracking Net achieving the lowest final loss. B Accuracy comparison, where OF-Tracking Net consistently outperforms other models after 100 epochs.

To quantitatively evaluate the endpoint detection performance, we used the precision and recall metrics [39]. A candidate point was considered correct if it fell within a circular region centered on the ground truth with a radius of $r$, and each ground truth point corresponded to only one candidate point. For a test set containing $N$ ground truth points and $M$ candidate points, with a defined threshold radius $r$, the precision and recall were calculated as follows:

1. Accuracy: Defined as the classification accuracy during model training and evaluation, reflecting the proportion of correctly predicted behavioral classes across all predictions.

2. Precision: The ratio of correctly identified candidate points to the total number of candidate points $M$, denoted as $P(r)$.

3. Recall: The ratio of correctly identified ground truth points to the total number of ground truth points $N$, denoted as $R(r)$.

   $$P(r)=\frac{1}{M}\sum _{t=1}^M\mathbb{F}\left(∥y_{}^{t\ast }-y^t∥_2^{}\le r\right)\times 100\%$$ 24

   $$R(r)=\frac{1}{N}\sum _{t=1}^M\mathbb{F}\left(∥y_{}^{t\ast }-y^t∥_2^{}\le r\right)\times 100\%$$ 25

Here, $y_{}^{t_{}^{\ast }}$ denotes the prediction and $y^t$ represents the ground truth for the test example $t$. The indicator function $\mathbb{F}$ equaled 1 when the condition was true and 0 otherwise.

We conducted ablation experiments to validate the contributions of each module to different tasks within the model (Table 3), with results obtained at r = 8 pixels. The baseline model achieved a precision of 0.91 and a recall of 0.85. Adding the LA module increased the recall to 0.88, demonstrating its role in enhancing key point prediction in multiple key-point recognition tasks. Incorporating the DSC module further improved precision to 0.95, confirming its role in addressing key-point occlusion. Finally, combining the LA and DSC modules achieved the best performance, with accuracy increasing to 0.94, precision to 0.97, and recall to 0.93, validating the synergistic effect of both modules in optimizing the overall model performance.

Table 3.

Performance comparison of different Components (r = 8 pixels).

OF-Tracking			Evaluation metrics
baseline	LA	DSC	Accuracy	Precision	Recall
√			0.87	0.91	0.85
√	√		0.90	0.93	0.88
√		√	0.89	0.95	0.91
√	√	√	0.94	0.97	0.93

A detailed performance comparison of various models based on precision and recall under different conditions was provided (Table 4). OF-Tracking Net demonstrated superior performance overall, particularly in occlusion scenarios, but some competing methods still achieved comparable results in specific evaluation metrics. Specifically, for precision, OF-Tracking Net attained values of 0.81, 0.92, and 0.97 at radii of $r$= 4, $r$= 6, and $r$= 8, respectively, outperforming models such as D-FINE and Hpnet. Similarly, for recall, OF-Tracking Net achieved values of 0.78, 0.86, and 0.93, further demonstrating its robustness and accuracy in detecting and recalling key behavior patterns. These results underscored OF-Tracking Net as the most efficient and reliable algorithm for mouse behavior recognition tasks, particularly under varying experimental conditions.

Table 4.

Performance comparison of different models on the MousePoseKey dataset in terms of precision and recall across different threshold radii (r = 4, 6, 8 pixels).

Model	Precision			Recall
	r = 4	r = 6	r = 8	r = 4	r = 6	r = 8
Detrs	0.68	0.89	0.96	0.68	0.79	0.87
D-FINE	0.71	0.86	0.91	0.66	0.78	0.88
Hpnet	0.77	0.90	0.93	0.61	0.77	0.70
RealNet	0.78	0.94	0.95	0.76	0.85	0.89
AICellCounter	0.80	0.90	0.92	0.82	0.84	0.91
OF-Tracking	0.81	0.92	0.97	0.78	0.86	0.93

Comparison of Mouse Behavioral Postures

We analyzed mouse behavioral postures by processing the open-field behavioral videos using the OF-Tracking Net model. This model accurately detected multiple key points on the mouse's body, including the snout, ears, and tail (Fig. 5A). In addition, the model accurately estimated the positions of occluded key points, ensuring reliable detection even when parts of the mouse's body were not visible, providing essential raw data for further behavioral analysis. Based on the coordinates of these key points, we classified the mouse postures using various features and applied hierarchical clustering to group similar action patterns (Fig. 5B). Through this process, we identified the typical behaviors exhibited by the mice during the experiment. The clustering results were grouped into five distinct behavioral categories: stationary, sniffing, climbing, walking, and stepping (Fig. 5C).

Fig. 5.

Mouse posture analysis and time distribution of behaviors across groups. A Key-point detection of body parts using the OF-Tracking Net model. B Hierarchical clustering of postures, categorized into five behaviors: stationary, sniffing, climbing, walking, and stepping. C Visual representation of the five behavioral types. D–F Pie charts showing the time distribution of behaviors in the restraint stress, LPS injection, and control groups.

To further investigate behavioral differences, we analyzed the time distribution of the five identified behaviors across the restraint stress group, LPS injection group, and control group (Fig. 5D–F). In the restraint stress group, stationary and sniffing behaviors accounted for 38.83% and 27.61% of the total observed time, respectively, which were significantly higher than the corresponding proportions of 22.08% and 17.38% in the control group. These results indicate a marked reduction in overall activity levels and an increased state of vigilance in response to the novel environment. Walking, stepping, and climbing behaviors in the restraint stress group constituted 21.39%, 5.56%, and 6.61%, respectively, all of which were lower than the control group, reflecting a diminished exploratory drive. Similarly, in the LPS injection group, stationary and sniffing behaviors accounted for 32.17% and 33.95%, respectively, which were also significantly higher than those in the control group. Although stepping and climbing behaviors showed a slight increase compared to the restraint stress group, their proportions remained lower than the control group, 15.93% and 7.11%. These findings collectively suggest a consistent suppression of exploratory activity in both negative emotion models.

The significantly longer stationary time in the negative emotion model mice compared to the control group indicates a marked reduction in overall activity levels. The increase in sniffing behavior and the decrease in climbing behavior in the negative emotion model groups further suggest heightened vigilance and a diminished exploratory drive under conditions of a negative emotional state. In contrast, the control group displayed more dynamic behaviors, such as walking and stepping, reflecting normal levels of activity.

These findings showed that negative emotions induced by restraint stress or LPS injection significantly modified the behavioral patterns of mice, resulting in reduced activity and diminished exploratory behavior, as demonstrated by the time distribution of behaviors.

Comparison of Neuronal Spike Discharges in Mice

Neuronal spike discharges in the CEA and PVN brain regions were examined across the restraint stress, LPS injection, and control groups. The temporal electrical signals recorded from these regions revealed distinct patterns of neuronal activity. Control group mice displayed relatively sparse and evenly spaced spike discharges in both the CEA and PVN regions (Fig. 6A), whereas negative emotion model mice exhibited significantly higher spike discharge frequencies with more densely clustered firing patterns, suggesting heightened neuronal activity under negative emotional conditions (Fig. 6B).

Fig. 6.

Comparison of multi-brain region neuronal spike discharges. A, B Time-series signals from the CEA and PVN regions show distinct discharge patterns across the groups. C Spike waveform processed with SNR = 6.8. D, E Spike occurrence and corresponding histograms highlight increased firing frequency in the CEA and PVN regions of the negative emotion groups. F, G Bar graphs showing significantly higher spike counts in the CEA and PVN of the restraint stress and LPS injection groups compared to controls.

To better distinguish the target neuronal signals from background noise generated by distant neuronal activity, we identified spike discharge segments from the temporal electrical signals with a signal-to-noise ratio greater than 6.8 [40]. (Fig. 6C). To further assess neuronal activity, the temporal occurrence of spike discharges was mapped on the same time scale, and a histogram was used to represent firing frequencies (Fig. 6D, E). The results indicated that the spike discharge frequency in the restraint stress and LPS injection groups steadily increased throughout the observation period, with discharges becoming more concentrated and intense. This pointed to more frequent and sustained high-frequency firing activity in neurons located in the CEA and PVN regions under negative emotional conditions. Conversely, the control group exhibited lower spike discharge frequencies, with relatively sparse and evenly distributed discharges, suggesting a more stable and lower-activity neural state. These findings further underscore the persistent and significant increase in neuronal activity linked to negative emotions.

To clearly illustrate the effect of negative emotions on neuronal activity, we compared the total spike counts in the CEA and PVN regions across the experimental groups, as shown in Figure 5F, G. In the CEA region, both the restraint stress and LPS injection groups exhibited significantly higher spike counts than the control group, exceeding 3,200 on average, while the control group had fewer than 1,800 spikes (Fig. 6F, P = 0.012). Similarly, in the PVN region, the restraint stress and LPS injection groups showed increased spike counts, approaching 3,000, whereas the control group had nearly 1,200 spikes (Fig. 6G, P = 0.003).

These findings indicate that negative emotions induced by restraint stress and LPS injection led to a marked increase in neuronal firing activity within both brain regions.

Synchronous Analysis of Mouse Behavioral Postures and Multi-Brain Region Neuronal Discharge

We constructed a behavioral spectrogram and precisely aligned it with spike discharge data over time (Fig. 7A), revealing the dynamic relationship between multi-brain region neural activity and behavioral patterns in mice under stress conditions. As shown in the figure, the behavioral spectrogram provides high temporal resolution and illustrates the neural discharge dynamics in various behavioral states (stationary, sniffing, stepping, walking, and climbing), clearly demonstrating the real-time correspondence between specific behaviors and spike discharges in multiple brain regions.

Fig. 7.

Synchronized analysis of spike discharges and behavior in negative emotion model mice and control mice. A Time-aligned spectrogram of animal behavior and spike discharge data. B–F Comparison of multi-brain region spike discharges between negative emotion model mice and control mice during stationary, sniffing, climbing, walking, and stepping behaviors. G Cluster analysis of neuronal activity in the CEA and PVN regions across multiple behaviors in mice following PCA dimensionality reduction.

In the stationary state, spike discharge frequencies in the CEA and PVN were notably higher in the restraint stress and LPS injection groups compared to the control group (CEA: 8.5–10.5 spikes/s in stress models vs 4.5–6.0 spikes/s in controls; PVN: 7.5–9.5 spikes/s in stress models vs 3.0–4.5 spikes/s in controls). This sustained neural activity, even in the absence of overt movement, suggests a heightened state of stress-induced arousal. During sniffing, the spike discharge frequency of the CEA increased rapidly (12.0–15.5 spikes/s), closely corresponding with the onset of the behavior, whereas the PVN maintained a relatively high and stable firing rate (9.5–12.0 spikes/s), indicating its role in stress regulation. Climbing behavior was associated with dense and highly synchronized spike discharges in both the CEA (13.5–18.0 spikes/s) and PVN (10.5–14.0 spikes/s), suggesting cross-regional coordination during complex motor execution.

The neural dynamics during behavior transitions were also evident. When shifting from stationary to active behavioral states such as walking or climbing, the CEA and PVN exhibited a rapid and synchronous increase in spike discharge frequency. Walking was characterized by sustained rhythmic activity (CEA: 10.5–14.5 spikes/s; PVN: 9.0–13.5 spikes/s), whereas stepping exhibited a slightly lower discharge rate (CEA: 9.5–12.5 spikes/s; PVN: 8.5–11.5 spikes/s). These results indicate that stress conditions not only alter baseline neural activity but also enhance interregional coordination during movement execution.

By analyzing the temporal dynamics of the behavioral spectrogram, we revealed the precise correlation between spike discharges and behavior occurrence, offering compelling evidence for the neural mechanisms underpinning emotional regulation. This approach enhances real-time analyses of behavior and neural activity, complementing traditional spike count methods and providing robust evidence for the neural mechanisms underlying emotional regulation. Subsequently, we analyzed the behaviors and spike firing frequencies of both negative emotion model mice and control mice within 5 min.

In the stationary state, spike discharges in the CEA and PVN regions of the restraint stress and LPS injection groups were significantly higher than those in the control group (Fig. 7B). In the CEA region, the restraint stress group exhibited an average of 9 spikes/s, while the LPS injection group showed nearly 10 spikes/s, compared to ~5 spikes/s in the control group. Similarly, in the PVN region, the restraint stress and LPS injection groups displayed an average of 8 spikes/s, whereas the control group had only 3 spikes/s. These findings suggested that even in the absence of external stimuli, the negative emotion model mice exhibited markedly elevated neural activity, reflecting heightened internal anxiety and negative emotions. This elevated neural activity may have corresponded to the emotional instability and anxiety symptoms commonly reported in patients with depression during periods of rest.

Exploratory behaviors such as sniffing and climbing revealed that spike discharges in the CEA and PVN regions in the restraint stress and LPS injection groups were recorded at 14 spikes/s and 10 spikes/s, respectively, while the control group had fewer than 7 spikes/s in both regions (Fig. 7C). A similar trend was found during climbing (Fig. 7D), where spike discharges in both the CEA and PVN regions significantly increased in the negative emotion groups. This elevated neural activity reflected the tension and discomfort experienced by negative emotion model mice when exposed to novel environments, indicating abnormal exploratory behavior.

Furthermore, dynamic behaviors such as walking and stepping also showed significantly higher spike discharges in the CEA and PVN regions of the depression model mice compared to the control group, suggesting abnormal psychological activation during movement.

In walking behavior, the control group generated only about 6 spikes/s in the CEA region and 4 spikes/s in the PVN region (Fig. 7E). In contrast, the CEA region exhibited 11 spikes/s in the restraint stress group and LPS injection group, and the PVN region showed 10 spikes/s. Similarly, during stepping, even more, pronounced differences in spike discharges were recorded (Fig. 7F). These findings demonstrated heightened neural activity in the negative emotion model mice during physical movements.

Following the PCA-based dimensionality reduction, the spike discharge features from five behavioral states were projected into a lower-dimensional space, and PC1 was extracted for clustering with the total spike discharge count (Fig. 7G). It was found that the control group exhibited higher principal component values in both brain regions, forming a sharp contrast with the restraint stress group and the LPS injection group. This suggested that the electrical activity in the CEA and PVN regions of mice affected by negative emotions significantly differed from that of normal mice across various behaviors. Notably, the PVN region in the same group of mice exhibited higher PCA results compared to the CEA region, indicating that although the mice within the same group shared similar neuronal activity characteristics across different brain regions, there were still some differences in their electrical activity.

A comprehensive analysis of spike discharge data across the five behaviors revealed that neuronal activity in the CEA and PVN regions of the depression model mice was significantly elevated compared to the control group. The high-frequency discharges recorded during stationary states reflected heightened internal anxiety and negative emotions. Enhanced neural activity in the CEA and PVN regions during exploratory behaviors suggested an overreaction to novel stimuli and environmental changes. In dynamic behaviors, the significant increase in neural activity within these regions indicated abnormal psychological activation during movement. The consistently elevated activity in the PVN region, as part of the HPA axis, underscored its pivotal role in depression, while the role of the CEA in processing fear, negative emotions, and regulating stress responses was further validated by these findings.

Discussion

In this study, we presented an innovative approach that integrated behavioral and electrophysiological techniques to examine the effects of negative emotions, induced by restraint stress and LPS injection, on neuronal spike discharges in key brain regions, specifically the CEA and PVN. By synchronously recording behavioral video data and multichannel electrode signals, we identified significant alterations in neural activity across different behavioral states, providing a more precise perspective on the relationship between emotional states and neuronal activity.

In terms of algorithms, our proposed enhancement to the mouse open-field behavior recognition model facilitated targeted analysis of behavioral patterns in open-field experiments, addressing the limitations of traditional behavior recognition methods. Specifically, incorporating a linear attention mechanism into the YOLO v8 model improved key-point detection by focusing on critical feature areas while filtering out irrelevant background noise. In addition, the dynamic snake convolution module allowed adaptive adjustment of the convolution kernel shape, enhancing the model‘s ability to capture detailed edge information, particularly in cases where head-mounted electrodes occluded parts of the mouse’s body. These improvements ensured precise behavioral recognition even under complex experimental conditions, reinforcing the reliability of our behavioral-neural correlation analyses.

Our findings revealed a critical neurobiological mechanism underlying the relationship between locomotor suppression and neural hyperactivity in the PVN and CEA under stress conditions. Correlation analyses demonstrated an inverse relationship between locomotor activity (stepping, climbing, and walking frequency) and neuronal spiking rates, with reduced movement correlating with heightened neural firing, particularly in stressed mice. This suggests that increased neuronal activity in these regions serves a regulatory function in stress-induced behavioral adaptations.

In the PVN, enhanced neuronal spiking likely reflects its modulatory role in autonomic and endocrine responses to stress [41]. As a key integrator of the HPA axis, the PVN regulates stress responses via the release of corticotropin-releasing hormone, which orchestrates downstream neuroendocrine signaling. Sustained hyperactivity in the PVN may lead to stress-induced behavioral suppression, wherein heightened neural activity actively inhibits locomotor engagement as part of a physiological adaptation to chronic stress exposure.

Conversely, the CEA is primarily involved in defensive behaviors and threat processing, playing a crucial role in shaping risk assessment strategies. Stress-induced hyperactivity in the CEA enhances vigilance and inhibits voluntary movement, likely as a precautionary response to perceived threats. The CEA’s direct projections to brainstem motor control centers suggest a potential mechanism whereby increased amygdala output suppresses locomotion, prioritizing defensive posturing over exploratory behavior.

These results suggest that stress-induced behavioral suppression is not merely a passive consequence of emotional distress but rather an active regulatory process governed by key limbic-hypothalamic circuits. The PVN and CEA appear to play complementary roles in this process: while PVN-mediated HPA axis activation modulates physiological responses that limit movement, CEA-driven defensive behaviors reduce voluntary exploration to enhance survival in adverse conditions.

This study also revealed the dynamic interaction between neural activity in multiple brain regions (CEA and PVN) and behavioral patterns under negative emotions. However, the role of the PVN in regulating feeding behavior was not further investigated. Due to the potential confounding effects of feeding behavior on PVN firing activity, the impact of stress and LPS injections on neuronal activity through changes in feeding behavior remains unresolved, which represents a limitation of the current experimental design. Future studies should incorporate real-time feeding behavior analysis to further clarify how PVN activity is modulated by both metabolic and stress-related inputs, providing a more comprehensive understanding of its role in negative emotion regulation and homeostasis.

In summary, our study enhances the understanding of how stress-induced negative emotions affect brain function by integrating behavioral and neural data. By delineating the distinct yet complementary roles of the PVN and CEA in regulating locomotor behavior under stress, we provide novel insights into the neurobiological mechanisms underlying stress-related disorders such as depression and anxiety [42]. These findings have significant clinical potential, particularly for brain-computer interface therapies targeting emotion-related neural circuits [43], offering new possibilities for personalized interventions in emotional disorders.

Data Availability

The data and code generated and analyzed during the current study are available from the corresponding author upon reasonable request.

Conflict of interests

The authors declare that they have no conflicts of interest.

Ethical Approval

All experiments involving mice were approved by the Animal Care and Use Committee of the University of Science and Technology of China.