Cross-crowd wearable ECG emotion recognition method based on hybrid convolution-Mama network

Abstract

The invention discloses a cross-crowd wearable ECG emotion recognition method based on a hybrid convolution-Mama network. The method comprises the following steps: S1, constructing a multimedia emotion induction experiment normal form for old people and cognitive impairment groups, and collecting wearable ECG data; s2, the collected original electrocardiosignals are subjected to standardization preprocessing; s3, constructing a hierarchical scale perception convolution module to perform multi-scale morphological feature extraction on the preprocessed electrocardiogram data; s4, performing physiological baseline remodeling on the feature map by using a non-local channel convolution attention mechanism; s5, constructing a bidirectional state space model based on a Mamba2 architecture, and carrying out long-time-history time sequence dependence modeling; and S6, fusing the spatio-temporal features to carry out emotion category probability prediction and implement personalized model fine tuning. According to the method, core challenges in ECG emotion recognition can be successfully solved, and a complete processing link from multi-dimensional sensing of signals to individualized noise filtering to efficient context understanding is constructed.

Description

Technical Field

[0001] This invention belongs to the field of affective computing and biomedical signal processing, specifically involving a cross-group wearable ECG emotion recognition method based on a hybrid convolutional-Mamba network. Background Technology

[0002] Affective computing is a core branch of human-computer interaction, aiming to endow machines with the ability to understand and respond to human emotions. Emotion recognition based on physiological signals (such as ECG and EEG) has attracted much attention due to its objectivity and reliability. ECG signals can reflect the regulation of emotional changes by the autonomic nervous system (ANS), making them an ideal biomarker for emotion recognition.

[0003] However, ECG-based emotion recognition faces three major challenges:

[0004] (1) Signal complexity: ECG signals contain information on multiple time scales, from millisecond-level QRS group morphology to second-level rhythm changes, which existing models cannot fully capture.

[0005] (2) Individual differences: The physiological baseline (such as resting heart rate) varies greatly among different individuals, making it difficult for traditional models to be generalized among different users, especially among the elderly and cognitively impaired populations with more complex physiological response patterns.

[0006] (3) Deployment efficiency requirements: For wearable Internet of Things (IoT) devices, the model must have both high accuracy and low computational complexity to achieve real-time, low-power emotion monitoring.

[0007] Existing technologies, such as traditional CNN, RNN, or Transformer models, have limitations in addressing the aforementioned problems: CNNs have limited receptive fields, making it difficult to capture long-range dependencies; RNNs suffer from gradient propagation issues; and Transformers have high computational complexity (O(L²)), making them unsuitable for long sequences and resource-constrained devices. Therefore, a novel technological solution that can systematically address these three major challenges is urgently needed to promote the application of ECG emotion recognition technology in the real world, especially among specific populations.

[0008] Existing ECG emotion recognition technologies mainly fall into two categories. The first category is based on traditional feature engineering and machine learning methods. These approaches typically require domain experts to manually extract predefined time-domain, frequency-domain, and nonlinear features (such as SDNN, LF / HF ratio, etc.) from ECG signals, and then input these features into traditional classifiers such as Support Vector Machines (SVM) for training and prediction. However, the core drawback of this method is its heavy reliance on expert knowledge; the feature extraction process is rigid and cannot adaptively learn the optimal feature representation for a specific task. Furthermore, these manually designed features often fail to capture the deep dynamic information and complex patterns contained in ECG signals, limiting the model's generalization ability and recognition accuracy.

[0009] The second category is end-to-end deep learning-based methods. These approaches directly input raw or lightly preprocessed ECG segments into deep neural networks (such as CNNs, RNNs, and Transformers) to achieve automatic feature learning and classification. While this method avoids tedious manual feature engineering, mainstream model architectures still have significant drawbacks: CNN models, due to the limitations of their local receptive fields, struggle to effectively capture long-range temporal dependencies of signals; RNN models suffer from the vanishing or exploding gradient problem, resulting in performance degradation when processing longer sequences; and while Transformer models can capture global dependencies, the quadratic computational complexity (O(L²)) of their self-attention mechanism makes them inefficient and resource-intensive when processing medium- to long sequences like ECGs, making them unsuitable for deployment on wearable devices. Furthermore, even with the addition of general attention mechanisms (such as SE-Block), the method of adjusting channel weights through global information compression is too coarse, failing to finely distinguish and suppress individual-specific physiological baseline noise, ultimately leading to poor generalization performance across different individuals. Summary of the Invention

[0010] The purpose of this invention is to provide a cross-group wearable ECG emotion recognition method based on a hybrid convolutional-Mamba network. Through a novel hybrid deep learning network model, it can accurately identify the emotional state of different groups (especially the elderly and patients with mild cognitive impairment).

[0011] To achieve the above objectives, the technical solution adopted by this invention is: a cross-group wearable ECG emotion recognition method based on a hybrid convolutional-Mamba network, comprising the following steps:

[0012] S1. Construct a multimedia emotion-inducing experimental paradigm for the elderly and cognitively impaired populations and collect wearable electrocardiogram data;

[0013] S2. Standardize and preprocess the acquired raw electrocardiogram signals;

[0014] S3. Construct a hierarchical scale-aware convolutional module to extract multi-scale morphological features from the preprocessed ECG data.

[0015] S4. Use the nonlocal channel convolutional attention mechanism to reshape the physiological baseline of the feature map;

[0016] S5. Construct a bidirectional state-space model based on the Mamba2 architecture for long-term time-series dependency modeling;

[0017] S6. Integrate spatiotemporal features to predict the probability of sentiment categories and implement personalized model fine-tuning.

[0018] Furthermore, step S1 includes:

[0019] S11. Design a video evoked paradigm that includes six basic emotional dimensions, namely neutrality, sadness, anger, happiness, boredom and tension. Select standardized video clips as stimulus materials to ensure the sufficiency of the evoked stimuli.

[0020] S12. Establish a non-invasive data acquisition environment based on wearable devices;

[0021] S13. Establish an emotion labeling mechanism based on self-assessment. After each video is played, subjects are immediately asked to select discrete emotion categories and score the valence and arousal of the SAM scale. Subjects' subjective scores are used as the true emotion labels of the corresponding electrocardiogram data.

[0022] Furthermore, in step S11, 18 standardized video clips are selected as stimulus materials, with each video clip set to a duration of 165 seconds, and each emotional dimension corresponds to 3 videos.

[0023] Furthermore, in step S12, before the experiment begins, the subjects are shown instructions and given a practice test. In the formal experiment, the subjects wear a single-lead ECG patch with a sampling rate of not less than 250Hz, with the electrode positioned in the subclavian region of the left upper chest. The single-channel ECG signal is continuously recorded while the subjects are watching the video at rest.

[0024] Furthermore, step S2 includes:

[0025] S21. Perform multi-source data frequency unification processing. For ECG signals acquired by different acquisition devices, use a resampling algorithm to uniformly adjust their sampling rate to a reference frequency of 256Hz in order to eliminate the impact of hardware specification differences on subsequent feature extraction.

[0026] S22. Use a 101st order finite impulse response (FIR) bandpass filter to filter and denoise the signal. Set the passband frequency range to 3Hz to 45Hz to filter out power line interference, high-frequency electromyographic noise and low-frequency baseline drift, and retain the electrocardiogram waveform components containing emotional characteristics.

[0027] S23. The filtered effective signal is windowed and truncated. The last 60 seconds of data before the end of each stimulus material is selected as the effective analysis segment. The signal is cut into time slices with a duration of 1 second using a non-overlapping sliding window. Each time slice contains 256 sampling points.

[0028] S24. A robust scaling algorithm is used to normalize each time slice, calculate the median and interquartile range of the current 1-second signal segment, and use the median and interquartile range to center and scale the signal amplitude to suppress outlier interference caused by motion artifacts.

[0029] Furthermore, step S3 includes:

[0030] S31. Design a three-stage hierarchical convolutional network architecture for simulating the spectral analysis of heart rate variability. The preprocessed 1×256 ECG vector is used as input, and morphological features at different time resolutions are extracted by progressive downsampling.

[0031] S32. In the first stage, a one-dimensional convolutional layer with a kernel size of 32 is used to extract long-term low-frequency trend features. In the second stage, a convolutional layer with a kernel size of 16 is used to extract mid-frequency rhythm features. In the third stage, a convolutional layer with a kernel size of 8 is used to extract high-frequency fine waveform features. Each stage includes batch normalization and LeakyReLU activation operations, and finally outputs a multi-channel feature map.

[0032] Furthermore, step S4 includes:

[0033] S41. Construct channel attention branches to suppress individual difference noise, perform global average pooling on the input feature map to obtain channel descriptors, capture the interaction between adjacent channels through local convolutional paths, and calculate the global context scalar through dot product operations. After fusing local and global information, channel weights are generated through the Sigmoid function. Feature channels are reweighted to suppress subject-specific baseline signals.

[0034] S42. Construct multi-scale spatial attention branches to enhance key waveform features. Use a serial-to-parallel convolutional structure to extract spatial features in parallel using convolutional kernels of sizes 5, 7, 11, and 21 respectively. After fusion and projection, generate a spatial attention map.

[0035] S43. By adding the features that have undergone dual modulation of channel and space to the original input features through residual connection, the output feature tensor that completes baseline reshaping and feature enhancement is obtained, thereby highlighting the common emotional patterns.

[0036] Furthermore, step S5 includes:

[0037] S51. Map the reshaped features to a high-dimensional latent space, construct a forward Mamba2 module to process the feature sequence in chronological order, and use a selective state space mechanism to capture the forward dynamic dependence of ECG signals as they evolve over time.

[0038] S52. Construct the backward Mamba2 module to process the time-reversed feature sequence, capture the reverse context information of the ECG signal, restore the time order of the backward output and fuse it with the forward output by element-level addition, so as to achieve modeling of the long-distance dependency of the whole sequence with linear computational complexity.

[0039] S53. By using residual connections, the bidirectional temporal features are fused with the spatial features output in step S4 to obtain a high-level sentiment feature representation containing rich spatiotemporal information.

[0040] Furthermore, step S6 includes:

[0041] S61. Use a global average pooling layer to aggregate high-level sentiment features into a fixed-length feature vector, input it into a fully connected classification layer and a Softmax function to calculate the probability distribution of each sentiment category, and output the category with the highest probability as the sentiment recognition result.

[0042] S62. For newly added subjects, a small-sample fine-tuning strategy is adopted, freezing the core parameters of the model and updating the parameters of the classification layer using only the first 10% of the calibration data of the new subjects, so as to achieve personalized adaptation of the model while preserving the general emotional feature representation.

[0043] The above-mentioned cross-group wearable ECG emotion recognition method based on hybrid convolutional-Mamba networks is applied in scenarios requiring real-time, objective emotion monitoring.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] High accuracy and reliability: On a self-built dataset targeting the elderly and MCI patients, the accuracy rate reached 90.6% and the F1 score reached 90.8%, significantly surpassing existing technologies and demonstrating its robustness under complex physiological signals.

[0046] Exceptional generalization and personalization capabilities: This demonstrates the value of the model as a "transferable base model." With only 10% new user data for fine-tuning, accuracy can be significantly improved from 16.9% (zero samples) to 73.4%, providing a realistic and feasible path for technology implementation.

[0047] Computational efficiency: The model has only 1.25M parameters and a single inference computation of 0.182 GFLOPS, meeting the requirements of wearable devices for real-time, low-power operation.

[0048] Hierarchical Scale-Aware Convolution (SAC): Mimicking HRV analysis, it uses convolution kernels of different sizes to extract multi-scale features of ECG signals in layers.

[0049] Non-local channel convolutional attention (NLCCA): Innovatively combines local and non-local channel information, as well as multi-scale spatial information, to "reshape" the feature map, actively suppress individual physiological baseline noise, and amplify general emotional patterns.

[0050] Bidirectional State-Space Model (BiSSM): Employing the Mamba2 module with linear computational complexity, it efficiently models the long-range bidirectional time dependencies of signals, balancing performance and efficiency. Attached Figure Description

[0051] Figure 1 This is an overall flowchart of the present invention;

[0052] Figure 2 This is a flowchart of the electrocardiogram emotion recognition process of the present invention;

[0053] Figure 3 This is a bar chart showing the ablation study results of this invention;

[0054] Figure 4 These are ECG feature maps on our self-collected dataset, DREAMER, and WESAD dataset before and after adding NLCCA.

[0055] Figure 5 This diagram illustrates the changes in the feature space caused by comparing the distribution of activation values ​​with and without the NLCCA module on three datasets. Detailed Implementation

[0056] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0057] Example 1

[0058] like Figure 1 A wearable ECG cross-group emotion recognition method based on hybrid convolutional-Mamba networks, specifically including the following steps:

[0059] S1. Construct a multimedia emotion-inducing experimental paradigm for the elderly and cognitively impaired populations and collect wearable electrocardiogram data.

[0060] S11. Design a video evoked paradigm that includes six basic emotional dimensions, namely neutrality, sadness, anger, happiness, boredom and tension. Select 18 standardized video clips as stimulus materials, with each video length set to 165 seconds. Each emotional dimension corresponds to 3 videos to ensure sufficient evoked stimuli.

[0061] S12. Set up a non-invasive data acquisition environment based on wearable devices. Before the experiment, show the instructions to the subjects and conduct practice tests. In the formal experiment, the subjects wear a single-lead ECG patch with a sampling rate of not less than 250Hz. The electrode is placed in the left upper chest below the clavicle. The single-channel ECG signal is continuously recorded while the subjects are watching the video at rest.

[0062] S13. Establish an emotion labeling mechanism based on self-assessment. After each video is played, subjects are immediately asked to select discrete emotion categories and score the valence and arousal of the SAM scale. Subjects' subjective scores are used as the true emotion labels of the corresponding electrocardiogram data.

[0063] S2. Standardize and preprocess the acquired raw electrocardiogram signals.

[0064] S21. Perform multi-source data frequency unification processing. For ECG signals acquired by different acquisition devices, use a resampling algorithm to uniformly adjust their sampling rate to a reference frequency of 256Hz in order to eliminate the impact of hardware specification differences on subsequent feature extraction.

[0065] S22. Use a 101st order finite impulse response (FIR) bandpass filter to filter and denoise the signal. Set the passband frequency range to 3Hz to 45Hz to filter out power line interference, high-frequency electromyographic noise and low-frequency baseline drift, and retain the electrocardiogram waveform components containing emotional characteristics.

[0066] S23. The filtered effective signal is windowed and truncated. The last 60 seconds of data before the end of each stimulus material is selected as the effective analysis segment. The non-overlapping sliding window technique is used to cut it into time slices with a duration of 1 second. Each time slice contains 256 sampling points.

[0067] S24. A robust scaling algorithm is used to normalize each time slice, calculate the median and interquartile range of the current 1-second signal segment, and use the median and interquartile range to center and scale the signal amplitude to suppress outlier interference caused by motion artifacts.

[0068] S3. Construct a hierarchical scale-aware convolutional module to extract multi-scale morphological features from the preprocessed ECG data.

[0069] S31. Design a three-stage hierarchical convolutional network architecture to simulate the spectral analysis of heart rate variability. The preprocessed 1×256 ECG vector is used as input, and morphological features at different time resolutions are extracted by progressive downsampling.

[0070] S32. In the first stage, a one-dimensional convolutional layer with a kernel size of 32 is used to extract long-term low-frequency trend features. In the second stage, a convolutional layer with a kernel size of 16 is used to extract mid-frequency rhythm features. In the third stage, a convolutional layer with a kernel size of 8 is used to extract high-frequency fine waveform features. Each stage includes batch normalization and LeakyReLU activation operations, and finally outputs a multi-channel feature map.

[0071] S4. Use the nonlocal channel convolutional attention mechanism to reshape the physiological baseline of the feature map.

[0072] S41. Construct channel attention branches to suppress individual difference noise, perform global average pooling on the input feature map to obtain channel descriptors, capture the interaction between adjacent channels through local convolutional paths, and calculate the global context scalar through dot product operations. After fusing local and global information, channel weights are generated through the Sigmoid function. Feature channels are reweighted to suppress subject-specific baseline signals.

[0073] S42. Construct multi-scale spatial attention branches to enhance key waveform features. Use a serial-to-parallel convolutional structure to extract spatial features in parallel using convolutional kernels of sizes 5, 7, 11, and 21 respectively. After fusion and projection, generate a spatial attention map.

[0074] S43. By adding the features that have undergone dual modulation of channel and space to the original input features through residual connection, the output feature tensor that completes baseline reshaping and feature enhancement is obtained, thereby highlighting the common emotional patterns.

[0075] S5. Construct a bidirectional state-space model based on the Mamba2 architecture for long-term time-series dependency modeling.

[0076] S51. Map the reshaped features to a high-dimensional latent space, construct a forward Mamba2 module to process the feature sequence in chronological order, and use a selective state space mechanism to capture the forward dynamic dependence of ECG signals as they evolve over time.

[0077] S52. Construct the backward Mamba2 module to process the time-reversed feature sequence, capture the reverse context information of the ECG signal, restore the time order of the backward output and fuse it with the forward output by element-level addition, so as to achieve modeling of the long-distance dependency of the whole sequence with linear computational complexity.

[0078] S53. By using residual connections, the bidirectional temporal features are fused with the spatial features output from step S4 to obtain a high-level sentiment feature representation containing rich spatiotemporal information.

[0079] S6. Integrate spatiotemporal features to predict the probability of sentiment categories and implement personalized model fine-tuning.

[0080] S61. Use a global average pooling layer to aggregate high-level sentiment features into a fixed-length feature vector, input it into a fully connected classification layer and a Softmax function to calculate the probability distribution of each sentiment category, and output the category with the highest probability as the sentiment recognition result.

[0081] S62. For newly added subjects, a small-sample fine-tuning strategy is adopted, freezing the core parameters of the model and updating the parameters of the classification layer using only the first 10% of the calibration data of the new subjects, so as to achieve personalized adaptation of the model while preserving the general emotional feature representation.

[0082] The following experiments will further illustrate this point:

[0083] 1. Quantitative analysis: Ablation experiment

[0084] On a basic SAC model, NLCCA and BiSSM modules were added progressively, and 10-fold cross-validation was performed on a self-built dataset to compare the changes in various performance metrics. The results show that adding each module brought significant performance improvements.

[0085] Table 1 Ablation Experiment Results

[0086]

[0087] 2. Qualitative Analysis: Feature Visualization

[0088] Feature maps were extracted from the intermediate layers of the model (before and after the NLCCA module) and visualized using heatmaps to visually observe the actual impact of the NLCCA module on features. For fair comparison, all heatmaps used a uniform color mapping range. Before adding NLCCA (top row), the feature maps showed large areas of warm, indistinct colors (yellow / orange), indicating that most feature channels were in a high activation state, with high information redundancy and difficulty in distinguishing key patterns. After adding NLCCA (bottom row), the feature maps changed dramatically. Large areas turned to cool colors (blue / green), indicating that the activation values ​​of many feature channels were strongly suppressed. Meanwhile, a few horizontal lines remained or showed brighter warm colors, indicating that key feature channels were preserved or enhanced. This visualization strongly and intuitively confirms our "feature reshaping" concept. NLCCA does indeed perform a "push-pull" operation, rather than a simple weighting.

[0089] 3. Mechanism Analysis: Activation Value Distribution Histogram

[0090] The distribution of activation values ​​of all features before and after the NLCCA module was statistically analyzed and histograms were plotted to quantify the effect of "feature reshaping" from a statistical perspective. Before adding NLCCA, the activation value distribution exhibited a very narrow and concentrated peak, close to the right of 0 (e.g., a mean of 0.3070 on a self-built dataset), indicating a very limited dynamic range of features. After adding NLCCA, the distribution pattern changed fundamentally. The overall distribution shifted significantly into the negative region (mean became -2.5898), while the distribution range became very wide. The central part was "pushed" into the negative region, while a long tail extended into the positive region. This statistical result provides the strongest evidence for "feature reshaping." It reveals the intrinsic mechanism of NLCCA: by systematically mapping the activation values ​​of the vast majority of redundant or noisy features to the negative domain, it achieves active suppression, thereby providing a broader dynamic space for the expression of a few key features in the positive domain.

[0091] Example 2

[0092] The method in Example 1 above can be widely applied to scenarios requiring real-time, objective emotion monitoring, specifically including:

[0093] Monitoring of elderly health and cognitive impairment: Used in wearable ECG patches and health monitoring apps built into smart bracelets / watches, it can monitor the emotional fluctuations of the elderly and patients with MCI for a long time and continuously, providing objective data support for the early screening and intervention of emotional disorders such as depression and anxiety.

[0094] Intelligent human-computer interaction and emotional robots: enabling machines to sense users' emotional states (such as stress and pleasure) and thus provide more empathetic and personalized services. For example, automatically playing soothing music when it detects that the driver is under too much stress.

[0095] Clinical research and auxiliary diagnosis: as a research tool, to objectively evaluate the impact of specific treatments (such as drugs and psychotherapy) on patients' emotional state.

[0096] In the future, this solution can be further combined with other physiological signals (such as EEG and GSR) to form a multimodal emotion recognition system, or combined with super-resolution technology to further improve the recognition performance at extremely low sampling rates and expand its application depth in clinical diagnosis and scientific research.

[0097] The following will provide further explanation using specific experimental examples:

[0098] 1. Data preparation:

[0099] Self-built dataset (elderly and MCI patients): 39 elderly participants were recruited, including 25 healthy controls (HC) and 14 MCI patients. All participants signed informed consent forms. Validated standardized video clips (18 clips, corresponding to 6 emotions: neutral, sad, angry, happy, bored, and tense) were used as emotion evoking materials. The experiment was conducted three times, with 6 videos viewed in a pseudo-random order each time. A portable single-lead ECG patch was used to continuously record ECG signals at a sampling rate of 250 Hz. After each video playback, participants rated valence and arousal using a 9-point Self-Assessment Scale (SAM) and selected the discrete emotion label that best reflected their own feelings; this label was used as the label for a classification task.

[0100] Public datasets: WESAD and DREAMER, used to validate the model's generalization ability.

[0101] Data preprocessing: ECG signals from all datasets were uniformly resampled to 256 Hz. A 101st-order FIR bandpass filter (3-45 Hz) was applied to remove baseline drift and high-frequency noise. The effective data was divided into non-overlapping 1-second windows. Robust scaling was applied to each window, which involves centering and scaling based on quantiles to mitigate the impact of outliers.

[0102] 2. Evaluation indicators:

[0103] Accuracy (Accuracy), Macro F1-Score, Unweighted Average Recall (UAR)

[0104] 3. Experimental setup and training parameters:

[0105] Hardware platform: NVIDIA GeForce RTX 4090 GPU

[0106] Software framework: PyTorch

[0107] Optimizer: AdamW, initial learning rate lr=0.001.

[0108] Loss function: Cross-entropy loss function.

[0109] Training strategy: Batch size is 256, using the ReduceLROnPlateau scheduler. The learning rate is multiplied by 0.1 if the validation set loss does not improve within 15 epochs. Early stopping is used; training is terminated early if the validation set loss does not improve within 50 epochs, and the best-performing model is saved.

[0110] Verification scheme:

[0111] Mixed 10-fold cross-validation: Used to evaluate the overall performance of the model in mixed populations.

[0112] Intra-Subject Validation: Used to evaluate the upper limit of a model's personalized performance.

[0113] Leave-one-out cross-validation (LOSO) and fine-tuning: used to evaluate the model’s zero-shot generalization ability and rapid personalization potential.

[0114] 4. Data Validation

[0115] Reproducibility: All reported experimental results were validated using 10-fold cross-validation.

[0116] Table 2 compares the performance and complexity with the benchmark model on the self-collected dataset.

[0117]

[0118] Table 3 Comparison between the WESAD dataset and the SOTA method

[0119]

[0120] Table 4 Comparison between the DREAMER dataset and the SOTA method

[0121]

[0122] Table 5. Within-subject trials in all 39 participants

[0123]

[0124] Table 6 analyzes the performance of models on self-built datasets based on a series of personalized protocols.

[0125]

[0126] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0127] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the scope of protection of the present invention in any way, and all technical solutions obtained by equivalent substitution or other means fall within the scope of protection of the present invention.

[0128] All parts not covered in this invention are the same as or can be implemented using existing technologies.

Claims

1. A cross-group wearable ECG emotion recognition method based on a hybrid convolutional-Mamba network, characterized in that, Includes the following steps: S1. Construct a multimedia emotion-inducing experimental paradigm for the elderly and cognitively impaired populations and collect wearable electrocardiogram data; S2. Standardize and preprocess the acquired raw electrocardiogram signals; S3. Construct a hierarchical scale-aware convolutional module to extract multi-scale morphological features from the preprocessed ECG data. S4. Use the nonlocal channel convolutional attention mechanism to reshape the physiological baseline of the feature map; S5. Construct a bidirectional state-space model based on the Mamba2 architecture for long-term time-series dependency modeling; S6. Integrate spatiotemporal features to predict the probability of sentiment categories and implement personalized model fine-tuning.

2. The cross-group wearable ECG emotion recognition method based on a hybrid convolutional-Mamba network according to claim 1, characterized in that, Step S1 includes: S11. Design a video evoked paradigm that includes six basic emotional dimensions, namely neutrality, sadness, anger, happiness, boredom and tension. Select standardized video clips as stimulus materials to ensure the sufficiency of the evoked stimuli. S12. Establish a non-invasive data acquisition environment based on wearable devices; S13. Establish an emotion labeling mechanism based on self-assessment. After each video is played, subjects are immediately asked to select discrete emotion categories and score the valence and arousal of the SAM scale. Subjects' subjective scores are used as the true emotion labels of the corresponding electrocardiogram data.

3. The cross-group wearable ECG emotion recognition method based on a hybrid convolutional-Mamba network according to claim 2, characterized in that, In step S11, 18 standardized video clips are selected as stimulus materials, with each video clip set to a duration of 165 seconds, and each emotional dimension corresponds to 3 videos.

4. The cross-group wearable ECG emotion recognition method based on a hybrid convolutional-Mamba network according to claim 2, characterized in that, In step S12, before the experiment begins, the subjects are shown instructions and practiced. In the formal experiment, the subjects wear a single-lead ECG patch with a sampling rate of not less than 250Hz, with the electrode placed in the left upper chest below the clavicle. The single-channel ECG signal is continuously recorded while the subjects are watching the video at rest.

5. The cross-group wearable ECG emotion recognition method based on a hybrid convolutional-Mamba network according to claim 1, characterized in that, Step S2 includes: S21. Perform multi-source data frequency unification processing. For ECG signals acquired by different acquisition devices, use a resampling algorithm to uniformly adjust their sampling rate to a reference frequency of 256Hz in order to eliminate the impact of hardware specification differences on subsequent feature extraction. S22. Use a 101st order finite impulse response (FIR) bandpass filter to filter and denoise the signal. Set the passband frequency range to 3Hz to 45Hz to filter out power line interference, high-frequency electromyographic noise and low-frequency baseline drift, and retain the electrocardiogram waveform components containing emotional characteristics. S23. The filtered effective signal is windowed and truncated. The last 60 seconds of data before the end of each stimulus material is selected as the effective analysis segment. The signal is cut into time slices with a duration of 1 second using a non-overlapping sliding window. Each time slice contains 256 sampling points. S24. A robust scaling algorithm is used to normalize each time slice, calculate the median and interquartile range of the current 1-second signal segment, and use the median and interquartile range to center and scale the signal amplitude to suppress outlier interference caused by motion artifacts.

6. The cross-group wearable ECG emotion recognition method based on a hybrid convolutional-Mamba network according to claim 1, characterized in that, Step S3 includes: S31. Design a three-stage hierarchical convolutional network architecture for simulating the spectral analysis of heart rate variability. The preprocessed 1×256 ECG vector is used as input, and morphological features at different time resolutions are extracted by progressive downsampling. S32. In the first stage, a one-dimensional convolutional layer with a kernel size of 32 is used to extract long-term low-frequency trend features. In the second stage, a convolutional layer with a kernel size of 16 is used to extract mid-frequency rhythm features. In the third stage, a convolutional layer with a kernel size of 8 is used to extract high-frequency fine waveform features. Each stage includes batch normalization and LeakyReLU activation operations, and finally outputs a multi-channel feature map.

7. The cross-group wearable ECG emotion recognition method based on a hybrid convolutional-Mamba network according to claim 1, characterized in that, Step S4 includes: S41. Construct channel attention branches to suppress individual difference noise, perform global average pooling on the input feature map to obtain channel descriptors, capture the interaction between adjacent channels through local convolutional paths, and calculate the global context scalar through dot product operations. After fusing local and global information, channel weights are generated through the Sigmoid function. Feature channels are reweighted to suppress subject-specific baseline signals. S42. Construct multi-scale spatial attention branches to enhance key waveform features. Use a serial-to-parallel convolutional structure to extract spatial features in parallel using convolutional kernels of sizes 5, 7, 11, and 21 respectively. After fusion and projection, generate a spatial attention map. S43. By adding the features that have undergone dual modulation of channel and space to the original input features through residual connection, the output feature tensor that completes baseline reshaping and feature enhancement is obtained, thereby highlighting the common emotional patterns.

8. The cross-group wearable ECG emotion recognition method based on a hybrid convolutional-Mamba network according to claim 1, characterized in that, Step S5 includes: S51. Map the reshaped features to a high-dimensional latent space, construct a forward Mamba2 module to process the feature sequence in chronological order, and use a selective state space mechanism to capture the forward dynamic dependence of ECG signals as they evolve over time. S52. Construct the backward Mamba2 module to process the time-reversed feature sequence, capture the reverse context information of the ECG signal, restore the time order of the backward output and fuse it with the forward output by element-level addition, so as to achieve modeling of the long-distance dependency of the whole sequence with linear computational complexity. S53. By using residual connections, the bidirectional temporal features are fused with the spatial features output in step S4 to obtain a high-level sentiment feature representation containing rich spatiotemporal information.

9. The cross-group wearable ECG emotion recognition method based on a hybrid convolutional-Mamba network according to claim 1, characterized in that, Step S6 includes: S61. Use a global average pooling layer to aggregate high-level sentiment features into a fixed-length feature vector, input it into a fully connected classification layer and a Softmax function to calculate the probability distribution of each sentiment category, and output the category with the highest probability as the sentiment recognition result. S62. For newly added subjects, a small-sample fine-tuning strategy is adopted, freezing the core parameters of the model and updating the parameters of the classification layer using only the first 10% of the calibration data of the new subjects, so as to achieve personalized adaptation of the model while preserving the general emotional feature representation.

10. The application of the cross-group wearable ECG emotion recognition method based on hybrid convolutional-Mamba networks as described in any one of claims 1 to 9 in real-time, objective emotion monitoring scenarios.

Images