Physiological age prediction device based on multi-modal physiological signal fusion

Abstract

A physiological age prediction device based on multimodal physiological signal fusion includes: a signal preprocessing unit that resamples, filters, standardizes, and segments the original ECG and EEG signals; a feature extraction unit that uses a dual-branch feature extraction network to extract features from the preprocessed signals to obtain corresponding temporal feature sequences; each branch includes: a channel mixing layer, a lightweight MobileNetV2 feature extraction module, and a BiLSTM feature extraction module; a feature fusion unit that uses a multi-head bidirectional cross-attention fusion module to perform bidirectional interaction between the temporal feature sequences of the two modal signals to obtain fused features; and a multi-level temporal modeling and age prediction unit that models the complex temporal relationship of the fused features through a multi-level BiLSTM network, uses an additive attention mechanism to weight and aggregate the temporal features to obtain a global feature vector, concatenates the global feature vector with a conditional label vector to obtain the fused features, and inputs them into a two-layer fully connected network to obtain the final prediction result.

Description

Technical Field

[0001] This application relates to physiological age prediction, and more particularly to a physiological age prediction device based on multimodal physiological signal fusion. Background Technology

[0002] Age is a key indicator reflecting the physiological state of the human body and has important applications in clinical diagnosis, health management, and disease early warning. Traditional age assessment methods mainly rely on birth date records or manual observation of physiological characteristics (such as facial wrinkles and bone density tests), which have limitations such as high subjectivity, high invasiveness, and poor timeliness.

[0003] With the development of medical monitoring technology and artificial intelligence, physiological signals have been shown to contain rich age-related information. Existing deep learning-based physiological signal age prediction technologies mainly include: 1. Single-modal deep learning methods: These methods use only a single signal source. For example, Paliwal et al. [1] used CNN to predict age using sleep EEG, with a mean absolute error (MAE) of about 13.76 years; Jusseaume et al. [2] used LSTM / GRU to predict age using resting EEG, with an MAE of about 7.0 years. These methods ignore the correlation between physiological age and multiple important organs. For example, ECG reflects cardiovascular aging, while EEG reflects nervous system aging. Single-modal methods that predict physiological age based solely on the physiological signals of a single important organ cannot comprehensively assess overall physiological age.

[0004] 2. Single-modal fusion method: Most existing single-modal fusion methods use early splicing or late weighting, such as Moguilner et al. [3] who processed fMRI and EEG separately and then fused the results. This method is essentially still a single signal source. fMRI and EEG are both signals that reflect the state of the brain. Such methods can only make better use of single-modal signals, but cannot achieve dynamic correlation between multiple modalities.

[0005] 3. Insufficient application of attention mechanisms: Some studies have introduced self-attention to extract single-modal features, but lack cross-modal attention mechanisms, and have not utilized cross-modal attention mechanisms to achieve information interaction and mutual enhancement between signals from different modalities.

[0006] 4. One-sided feature extraction: Existing methods only extract local structural features through CNN or only extract temporal features through LSTM, failing to simultaneously take into account the local transient nature (such as waveform abrupt changes) and long-term dependence (such as circadian rhythm and sleep cycle) of physiological signals.

[0007] 5. Insufficient robustness: It is sensitive to signal noise, poor electrode contact, and individual differences, resulting in large prediction errors and limited clinical applicability. Summary of the Invention

[0008] In view of the above problems, this application aims to propose a physiological age prediction method and device based on multimodal physiological signal fusion, which can assess physiological age based on physiological signals of multiple dimensions.

[0009] The physiological age prediction method based on multimodal physiological signal fusion in this application includes: Signal preprocessing steps: The original ECG and EEG signals acquired from the subject are resampled, filtered, standardized, and segmented to output preprocessed ECG and EEG signals; wherein the original ECG signal is a single-channel signal, the original EEG signal is a dual-channel signal, and the original ECG and EEG signals are synchronous signals. Feature extraction steps: A dual-branch feature extraction network is used to extract features from the preprocessed ECG and EEG signals to obtain corresponding time-series feature sequences. The dual-branch feature extraction network includes an EEG branch and an ECG branch, each comprising: a channel mixing layer, a lightweight MobileNetV2 feature extraction module, and a BiLSTM feature extraction module. The channel mixing layer transforms the preprocessed ECG or EEG signal to obtain an ECG or EEG signal with uniform feature dimensions. The ECG or EEG signal with uniform feature dimensions is processed by the lightweight MobileNetV2 feature extraction module, outputting a corresponding two-dimensional feature vector. This two-dimensional feature vector is then processed by the BiLSTM feature extraction module to output the corresponding ECG signal time-series feature sequence or EEG signal time-series feature sequence. Feature fusion steps: A multi-head bidirectional cross-attention fusion module is used to perform bidirectional interaction between the ECG signal time-series feature sequence and the EEG signal time-series feature sequence to obtain fused features; Multi-level temporal modeling and age prediction steps: In the multi-level BiLSTM modeling sub-step, the complex temporal relationships of the fused features are modeled through a multi-level BiLSTM network to obtain aggregated temporal features; in the attention pooling sub-step, the aggregated temporal features are weighted using an additive attention mechanism to obtain a global feature vector; in the conditional feature fusion sub-step, the global feature vector is concatenated with the conditional label vector obtained from the input conditional labels to obtain fused features; in the age prediction sub-step, the fused features are input into a two-layer fully connected network to obtain the final prediction result.

[0010] The physiological age prediction device based on multimodal physiological signal fusion of this application includes: The signal preprocessing unit is used to resample, filter, standardize, and segment the raw ECG and EEG signals acquired from the subject, and output the preprocessed ECG and EEG signals; wherein the raw ECG signal is a single-channel signal, the raw EEG signal is a dual-channel signal, and the raw ECG and EEG signals are synchronous signals. The feature extraction unit utilizes a dual-branch feature extraction network to extract features from the preprocessed ECG and EEG signals, respectively, to obtain corresponding time-series feature sequences. The dual-branch feature extraction network includes an EEG branch and an ECG branch, each comprising: a channel mixing layer, a lightweight MobileNetV2 feature extraction module, and a BiLSTM feature extraction module. The channel mixing layer transforms the preprocessed ECG or EEG signal to obtain an ECG or EEG signal with uniform feature dimensions. The ECG or EEG signal with uniform feature dimensions is processed by the lightweight MobileNetV2 feature extraction module, outputting a corresponding two-dimensional feature vector. This two-dimensional feature vector is then processed by the BiLSTM feature extraction module, outputting the corresponding ECG signal time-series feature sequence or EEG signal time-series feature sequence. The feature fusion unit uses a multi-head bidirectional cross-attention fusion module to perform bidirectional interaction between the ECG signal time-series feature sequence and the EEG signal time-series feature sequence to obtain fused features. The multi-level temporal modeling and age prediction unit, in the multi-level BiLSTM modeling module, models the complex temporal relationships of fused features through a multi-level BiLSTM network to obtain aggregated temporal features; in the attention pooling module, an additive attention mechanism is used to weight and aggregate the temporal features to obtain a global feature vector; in the conditional feature fusion module, the global feature vector is concatenated with the conditional label vector obtained from the input conditional labels to obtain fused features; in the age prediction module, the fused features are input into a two-layer fully connected network to obtain the final prediction result.

[0011] Preferably, before bidirectional interaction, the ECG signal timing feature sequence and the EEG signal timing feature sequence are aligned in the time dimension.

[0012] Preferably, if the timing feature sequences of the ECG signal and the EEG signal are inconsistent in length, the least common length is taken for truncation, and then zero-padding is used to unify the sequence length to a fixed value.

[0013] Preferably, the multi-head bidirectional cross-attention fusion module includes: a first attention path, a second attention path, and a feature fusion layer; the first attention path corresponds to the time-series feature sequence of ECG signals aligned in the time dimension, and the second attention path corresponds to the time-series feature sequence of EEG signals aligned in the time dimension; both the first and second attention paths include: a multi-head attention layer, a residual connection and normalization layer, and a feedforward network layer; the output features of the first and second attention paths are respectively sent to the feature fusion layer; In the multi-head attention layer, the core formula for calculating attention is: Q is the query matrix, representing the information that the current signal wants to focus on; K is the key matrix, representing the features of all signals; V is the value matrix, representing the final features after attention weighting. For each dimension of attention head, The scaling factor is used as follows: For the first attention path, the attention weight is calculated using the query matrix calculated from the time-aligned ECG signal time-series feature sequences, the key matrix calculated from the time-aligned EEG signal time-series feature sequences, and the value matrix calculated from the time-aligned EEG signal time-series feature sequences. This updates the context representation of the time-aligned ECG signal time-series feature sequences, resulting in ECG context features that integrate the correlation information between the ECG signal time-series feature sequences and the EEG signal time-series feature sequences. For the second attention path, the attention weight is calculated using the query matrix calculated from the time-aligned EEG signal time-series feature sequences, the key matrix calculated from the time-aligned ECG signal time-series feature sequences, and the value matrix calculated from the time-aligned ECG signal time-series feature sequences. This updates the context representation of the time-aligned EEG signal time-series feature sequences, resulting in EEG context features that integrate the correlation information between the EEG signal time-series feature sequences and the ECG signal time-series feature sequences. In the residual connection and normalization layer, for the first attention path, the ECG context features and the time-aligned ECG signal temporal feature sequence are connected; for the second attention path, the EEG context features and the time-aligned EEG signal temporal feature sequence are connected; then, the internal covariate offset is mitigated through layer normalization to obtain the normalized attention context features. In the feedforward network layer, for the first attention path, the normalized attention context features are subjected to nonlinear transformation and feature purification to obtain the output features of the first attention path; for the second attention path, the normalized attention context features are subjected to nonlinear transformation and feature purification to obtain the output features of the second attention path. In the feature fusion layer, feature splicing is applied to fuse the output features of the first attention path and the second attention path to obtain the final features after fusing the ECG signal time sequence and the EEG signal time sequence.

[0014] Preferably, in the multi-head attention layer, association patterns of different dimensions are learned in parallel by multiple attention heads.

[0015] This application presents a physiological age prediction method and device based on multimodal physiological signal fusion. It simultaneously extracts local structural features of physiological signals (such as QRS waveform details of ECG and alpha wave morphology of EEG) and long-term temporal dependence features (such as heart rate fluctuation trends and changes in EEG sleep rhythm). It establishes a dynamic bidirectional correlation between ECG and EEG signals (ECG→EEG and EEG→ECG), realizes cross-modal information enhancement and complementarity, improves the robustness of the model to signal noise, individual differences and data gaps, reduces age prediction errors, and provides a general framework that can be extended to multimodal physiological signal fusion analysis. Attached Figure Description

[0016] Figure 1 This is a diagram illustrating the overall framework of the physiological age prediction method based on multimodal physiological signal fusion proposed in this application.

[0017] Figure 2 This is a structural diagram of the MobileNetV2 feature extraction module.

[0018] Figure 3 This is a structural diagram of the BiLSTM feature extraction module.

[0019] Figure 4 This is a structural diagram of the bidirectional cross-attention fusion module.

[0020] Figure 5 Scatter plot for predicting physiological age.

[0021] Figure 6 This is a Bland-Altman bias analysis chart.

[0022] Figure 7 This is a graph showing the error distribution in physiological age prediction.

[0023] Figure 8 A distribution of prediction errors for different age groups.

[0024] Figure 9 This is a performance evaluation chart for physiological age prediction in Example 2. Detailed Implementation

[0025] Detailed implementation method.

[0026] This application provides a physiological age prediction method based on multimodal physiological signal fusion. It achieves deep fusion of ECG and EEG signals through a bidirectional cross-attention mechanism to predict age. The overall framework of this method is as follows: Figure 1 As shown, it includes the following steps: Step 1: Preprocessing of multimodal physiological signals The original ECG and EEG synchronization signals are acquired and then resampled, filtered, normalized, and segmented. (1) Resampling To meet the requirement of a constant sampling rate for convolutional neural networks (CNNs), a finite impulse response (FIR) low-pass filter with a Kaiser window is used to resample all signals to 128Hz. This frequency reduces the computational load while preserving key physiological information.

[0027] (2) Filtering The signal is filtered using a 16th-order infinite impulse response (IIR) elliptic filter to eliminate noise artifacts and unify spectral characteristics. Specifically, the EEG uses a bandpass filter of 0.3–45 Hz, and the ECG uses a high-pass filter of 0.3 Hz.

[0028] (3) Standardization Linear scaling of signal amplitude and uniform dynamic range are used to improve neural network training efficiency: the 5th percentile value is standardized to -1 and the 95th percentile value is standardized to 1.

[0029] (4) Data segmentation By pruning in the middle or padding with zeros at both ends, all signals are standardized to a fixed length.

[0030] Step 2: Extraction of Temporal Features from Two Branches Construct a dual-branch feature extraction network for ECG and EEG, with each channel using a MobileNetV2+BiLSTM cascaded structure: (1) Convert single-channel ECG signals or dual-channel EEG signals into a unified feature dimension: use 1×1 convolutional layers to map the input channels to a 16-dimensional feature space; batch normalization (BatchNorm) and ReLU6 activation function ensure training stability.

[0031] (2) Lightweight MobileNetV2 feature extraction A lightweight MobileNetV2 network with a width multiplier of 0.5 is used, and its structure is as follows: Figure 2 As shown, computational complexity is reduced through depthwise separable convolution; it contains multiple inverse residual blocks, each containing extended convolution, depthwise convolution, and projective convolution; the output feature dimension is [batch_size, 64, time step], where the time step is obtained by downsampling the original signal by 128 times.

[0032] (3) BiLSTM timing modeling Long-term time-series dependencies are modeled using a single-layer bidirectional LSTM network, with the structure as follows: Figure 3 As shown. The hidden layer has a dimension of 64, which becomes 128 after bidirectional concatenation; the two-dimensional feature sequence output by MobileNetV2 is reshaped into a temporal feature vector; the temporal feature vector for each 5-minute epoch is output, with a dimension of 128.

[0033] (4) Dual-mode feature output Supports two output modes: Temporal mode: Returns the complete temporal feature sequence ([batch_size, seq_len, 128]) for subsequent cross-attention fusion; Aggregation mode: Returns an aggregated feature vector through attention pooling, and outputs the age prediction value and uncertainty parameter.

[0034] Step 3: Bidirectional cross-attention feature fusion A multi-head bidirectional cross-attention mechanism is used to model the bidirectional interaction between ECG and EEG modalities, with the structure as follows: Figure 4 As shown.

[0035] (1) Input feature alignment Ensure that the ECG and EEG feature sequences are aligned in the time dimension: if the feature sequence lengths of the two modalities are inconsistent, take the least common length for truncation; unify the sequence length to a fixed value (default 120 epochs) by padding with zeros.

[0036] (2) The core calculation formula for Multihead Attention is: ; Where: Q (Query) is the query matrix, representing the information we want to focus on in the current word or sentence. K (Key) is the key matrix, representing the features of all words. V (Value) is the value matrix, representing the final features after attention weighting. It is the dimension of each attention head. As a scaling factor, it prevents gradient vanishing / exploding.

[0037] This invention establishes two-directional attention interaction pathways: ECG→EEG direction: Using ECG feature sequences as queries and EEG feature sequences as keys and values, calculate attention weights and update the ECG context representation; EEG→ECG direction: Using EEG feature sequences as queries and ECG feature sequences as keys and values, calculate attention weights and update the EEG context representation.

[0038] The association patterns of different dimensions are learned in parallel through 8 attention heads: the dimension of each head is hidden_dim / num_heads (default 128 / 8=16).

[0039] (3) Residual connectivity and layer normalization Add & Norm Residual connectivity includes bidirectional attention output, ECG raw features, and EEG raw features; Layer normalization (LayerNorm) mitigates internal covariate bias; (4) Feedforward network layer FFN Feedforward network layers use linear transformation, ReLU activation, and Dropout operations to perform nonlinear transformation and feature refinement on attention features, thereby improving the model's feature representation ability.

[0040] (5) Feature fusion layer Feature fusion is achieved by using feature splicing to obtain the final features after fusing the ECG signal time-series feature sequence and the EEG signal time-series feature sequence.

[0041] Step 4: Multi-level time series modeling and age prediction Multi-level temporal modeling and attention aggregation are performed on the fused features: (1) Multi-level BiLSTM modeling Complex temporal relationships of fused features are modeled using a 3-layer bidirectional LSTM network: the dimension of each hidden layer is 32 × network size scaling factor (default 32 × 8 = 256); the output dimension after bidirectional concatenation is 512; and variable-length temporal inputs are supported.

[0042] (2) Attention pooling The additive attention mechanism is used to weight and aggregate temporal features: the attention weight of each time step is calculated to reflect its importance to age prediction; the weighted sum is used to obtain the global feature vector with a dimension of 512.

[0043] (3) Conditional feature fusion Supports optional conditional label input (such as BMI, gender, etc.): concatenates the global feature vector with the conditional label vector; fuses multi-source information through a linear layer, with an output dimension of 512.

[0044] (4) Age prediction head The final prediction is output through a two-layer fully connected network: First layer: 512-dimensional input, 256-dimensional output, ReLU activation, Dropout regularization (default 0.5). The second layer has 256-dimensional input, 1-dimensional output, and no activation function, yielding the predicted age value. Optional output: Uncertainty parameter (standard deviation), used for heteroscedasticity loss calculation.

[0045] The formula for the heteroscedastic Huber loss function is: ； in, This is the actual age. To predict age, For uncertain parameters, This is the Huber loss threshold.

[0046] Step 5: Model Training and Optimization A phased training strategy and an adaptive loss function are adopted: (1) Phased training strategy The training process is divided into five stages: Phase 1: Pre-training the ECG feature extraction branch, with a fixed input channel of 1, for 20 epochs; Phase 2: Pre-training the EEG feature extraction branch, with a fixed input channel of 2, for 20 epochs; Phase 3: Extract and save the temporal features of ECG and EEG, ensuring feature alignment; Phase 4: Train the cross-attention fusion network, fix the feature extractor parameters, and train for 200 epochs; Phase 5: Comprehensively evaluate model performance and calculate multi-dimensional evaluation indicators.

[0047] (2) Adaptive loss function Select different types of loss functions based on the configuration: Huber loss combines the advantages of L1 and L2 losses, penalizing large errors less and penalizing small errors moderately. Negative log-likelihood loss (NLL): Considers prediction uncertainty and supports the assumption of normal or gamma distribution; Age-specific loss: Optimized for age prediction tasks, balancing the weights of samples from different age groups.

[0048] (3) Optimization strategy The Adam optimizer is used, and hierarchical weight decay is supported: L2 regularization is applied to the weight parameters (default 1e-5); weight decay is not applied to the bias parameters; the learning rate is initialized to 5e-4, and an early stopping strategy is used to prevent overfitting.

[0049] Step 6: Multi-dimensional model evaluation Apply multi-dimensional evaluation metrics to comprehensively assess model performance: (1) Basic regression indicators Mean Absolute Error (MAE): The average of the absolute values ​​of the differences between the predicted age and the actual age, reflecting the average degree of deviation in the prediction.

[0050] ； Root Mean Square Error (RMSE): Calculated by taking the square root of the square mean of the prediction error, it is more sensitive to outliers.

[0051] ; Coefficient of determination (R) 2 ): Measures the proportion of variance explained by the model, reflecting the goodness of fit, where This represents the average of real ages.

[0052] ； (2) Deviation analysis indicators Average deviation: The average of the differences between the predicted and actual values, reflecting the direction of systematic deviation; Standard deviation of the deviation: the degree of dispersion of the prediction deviation, reflecting the stability of the prediction; Bland-Altman analysis: Visualizing the relationship between prediction bias and average age.

[0053] (3) Age group analysis Performance of the model was assessed by age grouping: Youth group (20-40 years old): Evaluate the predictive accuracy of the model in a young population; Middle-aged group (40-60 years old): Evaluate the predictive accuracy of the model in the middle-aged population; Older age group (60-90 years): To evaluate the predictive accuracy of the model in the older age population.

[0054] (4) Statistical correlation analysis Pearson correlation coefficient: measures the degree of linear correlation between predicted age and actual age; Significance test: assesses the statistical significance of the correlation.

[0055] The present invention also protects an apparatus for implementing the above method, comprising: (1) Signal preprocessing unit: used for signal filtering, resampling, normalization and segment alignment, supporting multiple queue data formats (SHHS, WSC, SOF, etc.). (2) Feature extraction unit: includes ECG branch and EEG branch, each branch is composed of lightweight MobileNetV2 and BiLSTM network, supporting two modes: time-series feature output and aggregated feature output; (3) Feature fusion unit: Deep interactive fusion of ECG and EEG features is achieved through a multi-head bidirectional cross-attention mechanism, including residual connections and layer normalization; (4) Multi-level temporal modeling and age prediction unit: includes multi-level BiLSTM, attention pooling and conditional feature fusion, outputting age prediction value and uncertainty parameters; (5) Phased training unit: Supports a five-stage training process, including dual-branch pre-training, feature extraction, fusion training and evaluation; (6) Multi-dimensional evaluation unit: calculates basic regression indicators, deviation analysis indicators, age group indicators and statistical correlation indicators, and supports visualization output.

[0056] Example 1 To measure the effectiveness of this invention, an age prediction and evaluation study was conducted using publicly available datasets.

[0057] 1. Data Preprocessing Experimental data were obtained from the SHHS database (https: / / jhuccs1.us / shhs / index.htm), a multicenter cohort study initiated by the National Heart, Lung, and Blood Institute (NWHI) to determine the impact of sleep apnea on cardiovascular and other health outcomes. Participants were drawn from nine existing population studies that had conducted cardiovascular risk factor surveys. From these cohorts, participants meeting the inclusion criteria (age ≥40 years; no history of sleep apnea treatment; no tracheotomy; no home oxygen therapy) were invited to participate in the SHHS baseline assessment and polysomnography (PSG) collection. This study collected PSG data from 5713 participants. The selected channels included: dual-channel electroencephalography (EEG) with C3 / A2 and C4 / A1 leads at a sampling rate of 125 Hz; and single-channel bipolar electrocardiography (ECG) with a sampling rate of 125 Hz. The dataset was divided into training, validation, and test sets in a 7:1.5:1.5 ratio.

[0058] Signal preprocessing: The original signal is processed using the preprocessing module of this invention.

[0059] Resampling: The signal is resampled to 128Hz using an FIR low-pass filter with a Kaiser window.

[0060] Filtering: A bandpass filter of 0.3-45Hz is used for the EEG signal, and a high-pass filter of 0.3Hz is used for the ECG signal. Both are 16th-order elliptic IIR filters.

[0061] Standardization: Calculate the 5th and 95th percentiles of each signal, standardize the corresponding value of the 5th percentile to -1, and standardize the corresponding value of the 95th percentile to 1.

[0062] Data segmentation: The entire night's signal is divided into 5-minute segments, which are then uniformly converted into 38,400 sampling points (5 minutes × 128Hz) by cropping in the middle or padding with zeros at both ends.

[0063] 2. Training a dual-branch feature extraction network (1) ECG branch pre-training: Network structure: The lightweight MobileNetV2 (width_mult=0.5) is used as the backbone network, followed by a single-layer bidirectional LSTM (64 hidden layers) and an attention pooling layer.

[0064] Input: ECG signal segment (1 channel, 38400 sampling points).

[0065] Training parameters: Adam optimizer, initial learning rate 1e-3, batch size 16, training for 20 epochs, early stopping strategy (patience value 5).

[0066] Loss function: Heteroscedastic Huber loss is used.

[0067] Save the model: Save the trained ECG feature extraction network.

[0068] (2) EEG branch pre-training: Network structure: Same as ECG branch, but with 2 input channels (C3 and C4).

[0069] Input: EEG signal segment (2 channels, 38,400 sampling points).

[0070] Training parameters: Same as ECG branch.

[0071] Save the model: Save the trained EEG feature extraction network.

[0072] 3. Temporal Feature Extraction and Preservation (1) Load the pre-trained model: Load the best models for the ECG and EEG branches respectively.

[0073] (2) Feature extraction: ECG and EEG feature extraction networks are used to extract temporal features for each sample in the training set, validation set and test set respectively.

[0074] For each 5-minute segment, the ECG branch outputs time-series features with dimensions [15, 128] (15 time steps, 128 features per time step), and the EEG branch also outputs time-series features with dimensions [15, 128].

[0075] (3) Feature alignment: Ensure that the ECG and EEG feature sequences of each sample are of the same length (15 time steps each). If they are inconsistent, the least common length is used for truncation.

[0076] (4) Feature saving: Save the ECG feature sequence and EEG feature sequence of each sample as an HDF5 file, and save labels such as age, gender, and BMI.

[0077] 4. Training of Cross-Attention Fusion Networks (1) Dataset preparation: Construct a dataset using the time series feature files saved in step 3. Each sample includes an ECG feature sequence, an EEG feature sequence, an age label, and conditional labels (such as BMI, gender, etc.).

[0078] (2) Network structure: A bidirectional cross-attention fusion network is constructed, including: a bidirectional cross-attention module (hidden_dim=128, num_heads=8, dropout=0.2), a 3-layer bidirectional LSTM (each hidden layer has a dimension of 256, and the bidirectional concatenation results in a dimension of 512), an additive attention pooling layer, and a prediction head (2 fully connected layers that output age and uncertainty parameters). (3) Training parameters: Adam optimizer was used, with an initial learning rate of 5e-4, weight decay of 1e-5, batch size of 64, training for 200 epochs, and early stopping patience value of 20.

[0079] (4) Loss function: Use heteroscedastic Huber loss.

[0080] (5) Training process: Input the features of the training set into the network, calculate the loss and backpropagate to optimize the network parameters. The validation set is used to monitor the training process and select the best model.

[0081] (6) Save the model: Save the trained fusion network.

[0082] 5. Model Evaluation (1) Test set prediction: The trained fusion network is used to predict the test set samples to obtain the predicted age and uncertainty parameters of each sample.

[0083] (2) Performance evaluation: The following evaluation indicators were calculated: Mean absolute error (MAE): 5.67 years Root mean square error (RMSE): 7.36 years Coefficient of determination (R) 2 ): 0.393 Pearson correlation coefficient: 0.66 (p<0.001) (3) Visualization: Generate scatter plots and Bland-Altman plots of the actual and predicted ages, and display the prediction errors by age group (youth, middle-aged, elderly). Age prediction scatter plot display (e.g.) Figure 5 As shown in the figure, the regression slope between predicted age and actual age reaches 0.56, R0. 2 The value is 0.393, and the overall trend follows the ideal line (y=x); Bland-Altman bias analysis (e.g.) Figure 6 As shown in the figure, the average prediction bias is -0.83 years, the 95% threshold is ±14.34 years, and the bias distribution is generally concentrated around 0; the prediction error distribution (as shown in the figure) Figure 7 The prediction error (as shown) exhibits an approximately normal distribution, with the peak value close to the zero error line. The prediction errors for different age groups (20-40 years old (2 subjects aged 40), 40-60 years old (160 subjects), and 60-90 years old (764 subjects) are as follows: Figure 8 All values ​​(as shown) are within a reasonable fluctuation range. These results indicate that the bidirectional cross-attention fusion model possesses age prediction capabilities, and its bias level and distribution characteristics conform to the expectations of the initial modeling.

[0084] The above embodiments demonstrate that the age prediction method based on multimodal physiological signal fusion provided by this invention can accurately predict an individual's age and has good generalization ability. This method can be used in fields such as health assessment and disease risk prediction through age prediction.

[0085] 6. Training Process Hardware environment: NVIDIA RTX 3090 GPU (24GB VRAM), Intel i9-10900K CPU, 64GB RAM; Software environment: PyTorch 1.12.0, Python 3.8.10, CUDA 11.6.

[0086] Example 2: Multimodal physiological signal fusion analysis of the effects of plateau environment on age To verify the sensitivity of this invention to age changes and its transfer learning ability under specific environmental stress (high altitude and low oxygen) conditions, this embodiment uses the bidirectional cross-attention fusion model pre-trained in Embodiment 1 to predict and compare the age of people in high-altitude areas before and after going to high altitudes through transfer learning.

[0087] 1. Data Sources and Preprocessing The experimental data came from a high-altitude medicine research cohort, which included 221 healthy subjects (aged 18-25 years, mean 21.2±1.7 years, male). EEG and ECG physiological signal data were collected at two time points: before and after going to the high-altitude area.

[0088] Data acquisition parameters: EEG: Single channel, raw sampling rate 2000 Hz ECG: Single-channel bipolar lead, raw sampling rate 2000 Hz Recording duration: 5 minutes Signal preprocessing adopts the SHHS standardized process compatible with Example 1: Resampling: After using an 8th-order Butterworth low-pass filter to resist aliasing, the signal is downsampled from 2000 Hz to 128 Hz; Standardization: Calculate the 5th and 95th percentiles of each signal, and map the 5th percentile to -1 and the 95th percentile to 1; Data segmentation: The physiological signals of each subject consist of a 5-minute segment (38,400 sampling points, 5 minutes × 128 Hz).

[0089] 2. Transfer learning and model fine-tuning Based on the model parameters pre-trained on the SHHS dataset (5713 cases) in Example 1, this example employs a transfer learning strategy to adapt to the data distribution characteristics of the plateau population: (1) Feature extraction network loading: Load the pre-trained ECG feature extraction network and EEG feature extraction network. Both branches maintain the same network structure as in Example 1. Freeze the parameters of the bottom convolutional layers to retain the pre-trained feature extraction capabilities.

[0090] (2) Fine-tuning of the fusion network: Network structure: Load the bidirectional cross-attention fusion network trained in Example 1, which includes a bidirectional cross-attention module (hidden_dim=128, num_heads=8), 3 layers of bidirectional LSTM (hidden layer dimension 256) and additive attention pooling layers.

[0091] Fine-tuning strategy: Freeze the cross-attention layer, unfreeze the LSTM layer, attention layer and prediction head, and use a small learning rate (5e-5) for fine-tuning to avoid overfitting.

[0092] Training parameters: Adam optimizer, weight decay 1e-5, batch size 16, training for 100 epochs, early stopping patience value 15.

[0093] Loss function: Mean squared error (MSE) loss.

[0094] (3) Data augmentation: Timing offset: random offset ±3 time steps.

[0095] Gaussian noise: Add Gaussian noise with a standard deviation of 0.005.

[0096] 3. Feature Extraction and Age Prediction (1) Temporal feature extraction: Using a finely tuned feature extraction network, temporal features were extracted from data at two time points before and after each subject went to the plateau: ECG branch output: Temporal features with dimensions [time steps, 128]. EEG branch output: Temporal features with dimensions [time steps, 128]. (2) Age prediction: The extracted ECG and EEG temporal features are input into a fine-tuned bidirectional cross-attention fusion network to calculate the predicted age before and after going to the plateau.

[0097] 4. Model Evaluation and Statistical Analysis (1) Predictive performance evaluation: The prediction results at two time points, before and after going to the plateau, were evaluated separately: Before going to the plateau: Mean absolute error (MAE): 1.32 years Root mean square error (RMSE): 1.64 years Coefficient of determination (R) 2 ): 0.006 Pearson correlation coefficient: 0.130 (p=0.0542) After arriving at the plateau: Mean absolute error (MAE): 1.31 years Root mean square error (RMSE): 1.62 years Coefficient of determination (R) 2 ): 0.038 Pearson correlation coefficient: 0.203 (p=0.0024) (2) Visual analysis (e.g.) Figure 9 (as shown) Age prediction scatter plot: shows the correlation between predicted age and actual age before and after going to the plateau. Both are in line with the ideal line (y=x). The fine-tuned model maintains good predictive stability in the young population (18-25 years old).

[0098] Bland-Altman consistency analysis: Before reaching the plateau, the average bias was -0.03 years, with a 95% consensus threshold of ±3.22 years; after reaching the plateau, the average bias was 0.01 years, with a 95% consensus threshold of ±3.17 years. The bias distribution was concentrated around 0, indicating good model consistency.

[0099] Compared with the prior art, this application has the following beneficial effects: 1. Multimodal deep interaction: Deep interaction between ECG and EEG signals is achieved through a bidirectional cross-attention mechanism, rather than simple splicing, which fully explores the complementary information between modes and solves the problem of low information utilization in traditional splicing fusion. 2. Complete temporal modeling: Combining the local feature extraction capability of lightweight MobileNetV2 and the long-term temporal modeling capability of BiLSTM, the temporal dynamic characteristics of physiological signals are captured in multiple levels; 3. Lightweight model design: The MobileNetV2 network with a width multiplier of 0.5 is adopted, which significantly reduces the computational resource requirements while ensuring performance; 4. Training strategy optimization: The phased training strategy avoids the optimization difficulties of joint training and improves training stability and convergence speed; 5. Modeling of prediction uncertainty: Supports heteroscedasticity loss function to model prediction uncertainty, enhancing the robustness and interpretability of the model; 6. Comprehensive evaluation system: Provides multi-dimensional evaluation indicators and visualization tools to fully reflect the model's performance in different populations and age groups.

[0100] 7. Strong generalization ability: It can be extended to prediction tasks of other multimodal physiological signals (such as respiration and electromyography), and is suitable for scenarios such as sleep monitoring and health management.

[0101] References [1] Paliwal V, Das K, Doesburg SM, et al. Classifying RoutineClinical Electroencephalograms With Multivariate Iterative Filtering andConvolutional Neural Networks. IEEE Trans Neural Syst Rehabil Eng. 2024;32:2038-2048. [2] Jusseaume K , Valova I .Brain Age Prediction / Classificationthrough Recurrent Deep Learning with Electroencephalogram Recordings ofSeizure Subjects.Sensors, 2022, 22(21),8112. [3] Moguilner S, Baez S, Hernandez H, et al. Brain clocks capturediversity and disparities in aging and dementia across geographically diversepopulations. Nat Med. 2024;30(12):3646-3657.

Claims

1. A physiological age prediction device based on multimodal physiological signal fusion, comprising: The signal preprocessing unit is used to resample, filter, standardize, and segment the raw ECG and EEG signals acquired from the subject, and output the preprocessed ECG and EEG signals; wherein the raw ECG signal is a single-channel signal, the raw EEG signal is a dual-channel signal, and the raw ECG and EEG signals are synchronous signals. The feature extraction unit utilizes a dual-branch feature extraction network to extract features from the preprocessed ECG and EEG signals, respectively, to obtain corresponding time-series feature sequences. The dual-branch feature extraction network includes an EEG branch and an ECG branch, each comprising: a channel mixing layer, a lightweight MobileNetV2 feature extraction module, and a BiLSTM feature extraction module. The channel mixing layer transforms the preprocessed ECG or EEG signal to obtain an ECG or EEG signal with uniform feature dimensions. The ECG or EEG signal with uniform feature dimensions is processed by the lightweight MobileNetV2 feature extraction module, outputting a corresponding two-dimensional feature vector. This two-dimensional feature vector is then processed by the BiLSTM feature extraction module, outputting the corresponding ECG signal time-series feature sequence or EEG signal time-series feature sequence. The feature fusion unit uses a multi-head bidirectional cross-attention fusion module to perform bidirectional interaction between the ECG signal time-series feature sequence and the EEG signal time-series feature sequence to obtain fused features. The multi-level temporal modeling and age prediction unit, in the multi-level BiLSTM modeling module, models the complex temporal relationships of fused features through a multi-level BiLSTM network to obtain aggregated temporal features; in the attention pooling module, an additive attention mechanism is used to weight and aggregate the temporal features to obtain a global feature vector; in the conditional feature fusion module, the global feature vector is concatenated with the conditional label vector obtained from the input conditional labels to obtain fused features; in the age prediction module, the fused features are input into a two-layer fully connected network to obtain the final prediction result.

2. The physiological age prediction device based on multimodal physiological signal fusion according to claim 1, characterized in that: Before bidirectional interaction, the timing feature sequences of ECG and EEG signals are aligned in the time dimension.

3. The physiological age prediction device based on multimodal physiological signal fusion according to claim 2, characterized in that: If the timing feature sequences of the ECG signal and the EEG signal are not of the same length, the least common length is taken for truncation, and then the sequence length is unified to a fixed value by padding with zeros.

4. The physiological age prediction device based on multimodal physiological signal fusion according to claim 1, characterized in that: The multi-head bidirectional cross-attention fusion module includes: a first attention path, a second attention path, and a feature fusion layer; the first attention path corresponds to the time-aligned ECG signal temporal feature sequence, and the second attention path corresponds to the time-aligned EEG signal temporal feature sequence; both the first and second attention paths include: a multi-head attention layer, a residual connection and normalization layer, and a feedforward network layer; the output features of the first and second attention paths are respectively sent to the feature fusion layer; In the multi-head attention layer, the core formula for calculating attention is: Q is the query matrix, representing the information that the current signal wants to focus on; K is the key matrix, representing the features of all signals; V is the value matrix, representing the final features after attention weighting. For each dimension of attention head, The scaling factor is used. For the first attention path, the attention weight is calculated using the query matrix calculated from the time-aligned ECG signal time-series feature sequences, the key matrix calculated from the time-aligned EEG signal time-series feature sequences, and the value matrix calculated from the time-aligned EEG signal time-series feature sequences. This updates the context representation of the time-aligned ECG signal time-series feature sequences, resulting in ECG context features that integrate the correlation information between the ECG signal time-series feature sequences and the EEG signal time-series feature sequences. For the second attention path, the attention weight is calculated using the query matrix calculated from the time-aligned EEG signal time-series feature sequences, the key matrix calculated from the time-aligned ECG signal time-series feature sequences, and the value matrix calculated from the time-aligned ECG signal time-series feature sequences. This updates the context representation of the time-aligned EEG signal time-series feature sequences, resulting in EEG context features that integrate the correlation information between the EEG signal time-series feature sequences and the ECG signal time-series feature sequences. In the residual connection and normalization layer, for the first attention path, the ECG context features and the time-aligned ECG signal temporal feature sequence are connected; for the second attention path, the EEG context features and the time-aligned EEG signal temporal feature sequence are connected; then, the internal covariate offset is mitigated through layer normalization to obtain the normalized attention context features. In the feedforward network layer, for the first attention path, the normalized attention context features are subjected to nonlinear transformation and feature purification to obtain the output features of the first attention path; for the second attention path, the normalized attention context features are subjected to nonlinear transformation and feature purification to obtain the output features of the second attention path. In the feature fusion layer, feature splicing is applied to fuse the output features of the first attention path and the second attention path to obtain the final features after fusing the ECG signal time sequence and the EEG signal time sequence.

5. The physiological age prediction device based on multimodal physiological signal fusion according to claim 4, characterized in that: In the multi-head attention layer, different dimensions of association patterns are learned in parallel by multiple attention heads.

Images