Cognitive state real-time estimation method driven by physiological data under plateau hypoxia
By using a hybrid expert model framework that combines Transformer, CNN-LSTM, and residual MLP networks, the problem of estimating multi-dimensional cognitive states in high-altitude hypoxic environments was solved. This enabled real-time and accurate estimation of the multi-dimensional cognitive states of workers, adapting to dynamic scheduling of different physiological modes and improving the prediction accuracy of the model in extreme environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA JAPAN FRIENDSHIP HOSPITAL
- Filing Date
- 2026-01-30
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies struggle to accurately estimate the multidimensional cognitive states of workers in real time under high-altitude, low-oxygen conditions, especially given the high heterogeneity of physiological-cognitive relationships in extreme environments. A single network structure is unlikely to achieve optimal performance, and a multi-task learning framework is lacking.
A hybrid expert model framework is adopted, including a gating network and several heterogeneous expert networks, such as Transformer, CNN-LSTM hybrid expert network and residual MLP expert network. Expert weight distribution is generated through self-attention mechanism, and feature extraction is performed in parallel to estimate the cognitive state of multiple tasks.
It achieves real-time and accurate estimation of the multi-dimensional cognitive state of workers in high-altitude and low-oxygen environments, can accurately predict the reaction time and accuracy of eight core cognitive dimensions, dynamically allocates expert weights to adapt to different physiological modes, and improves the applicability and accuracy of the model in extreme environments.
Smart Images
Figure CN122004764A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of physiological signal processing technology, and in particular to a method for real-time estimation of cognitive state driven by physiological data under high-altitude hypoxia. Background Technology
[0002] The low-oxygen environment of high-altitude areas poses a serious threat to human cognitive function. As altitude increases, atmospheric pressure and oxygen partial pressure decrease, leading to reduced blood oxygen saturation and triggering a series of negative physiological and cognitive changes. Empirical studies have shown that at altitudes above 2500 meters, core cognitive functions such as episodic memory, working memory, and inhibitory control are significantly impaired. This central nervous system dysfunction, directly or indirectly caused by hypoxia, manifests as inattention, slowed information processing, and decreased judgment and decision-making abilities. In safety-critical areas such as high-altitude operations, this can easily lead to human error, posing a severe challenge to operational safety and efficiency. In recent years, with the development of wearable sensing technology, the real-time dynamic assessment of cognitive status using continuously and non-invasively collected physiological signals (such as heart rate and blood oxygen) has become a highly promising technological direction.
[0003] Early studies often employed classic machine learning algorithms such as linear regression, support vector regression, and random forest. However, their limitation lies in the fact that feature engineering is highly dependent on expert knowledge and is difficult to fully capture the complex interaction patterns of high-dimensional, nonlinear, and temporal dynamics in physiological signals.
[0004] To enable automatic learning of feature representations, researchers have turned to deep learning. Recurrent neural networks and their variant, Long Short-Term Memory networks, have been widely used due to their temporal modeling capabilities. Single-architecture models have demonstrated performance that surpasses traditional methods, but their capabilities have clear limitations. When faced with the high heterogeneity of physiological-cognitive relationships in extreme environments (different tasks, different individuals, different stress stages), a universal single network structure often struggles to achieve optimal performance.
[0005] Furthermore, existing research focuses on estimating the state of a single cognitive task (such as fatigue and attention), lacking a multi-task learning framework that simultaneously predicts and models multiple cognitive dimensions (such as attention, working memory, and processing speed). This limits the model's ability to comprehensively assess the cognitive state of individuals in complex work environments. Summary of the Invention
[0006] The purpose of this invention is to provide a real-time estimation method for cognitive state driven by physiological data under high-altitude hypoxia, so as to realize the real-time and accurate estimation of the multi-dimensional cognitive state of workers in extreme environments such as high-altitude hypoxia.
[0007] To achieve the above objectives, this invention provides a real-time cognitive state estimation method driven by physiological data under high-altitude hypoxia conditions, comprising the following steps: S1. Obtain physiological signals of workers in a high-altitude, low-oxygen environment, and preprocess the physiological signals to obtain time series data of physiological signals; S2. Input physiological signal time series data into a task-aware hybrid expert model framework, which includes a gating network and several heterogeneous expert networks. S3. Using the gating network in the hybrid expert model framework, the corresponding expert weight distribution is generated based on the input physiological signal time series data through a self-attention mechanism. S4. Parallel feature extraction of the input physiological signal time series data is performed using several heterogeneous expert networks in the hybrid expert model framework. S5. Based on the expert weight distribution, perform a weighted summation of the outputs of several heterogeneous expert networks to output the estimated results of reaction time and accuracy for multiple cognitive dimensions.
[0008] Preferably, in step S1, the preprocessing specifically includes: A linear interpolation method was used to align the physiological signal data acquired at different time points with the cognitive test time points to obtain physiological signal time series data; Abnormal samples with a heart rate less than 30 bpm or blood oxygen saturation less than 60% were excluded.
[0009] Preferably, the heterogeneous expert networks include Transformer expert networks, CNN-LSTM hybrid expert networks, and residual MLP expert networks, each with an independent output layer.
[0010] Preferably, in step S3, the process of the gating network generating the expert weight distribution includes: A self-attention mechanism is applied to the input physiological signal time series data to obtain a globally context-weighted feature representation; The feature representation is compressed into an E-dimensional logits vector by using a multilayer perceptron, where E represents the number of expert networks; The softmax function is used to process the logits vector to generate the weight coefficients of each expert network.
[0011] Preferred methods for constructing Transformer expert networks include: For the input physiological signal time series The query, key, and value matrix is obtained through linear projection: ; ; ; in, Represents the query matrix. Represents the trainable query projection weight matrix. Represents the bond matrix. Represents the trainable key-projection weight matrix. Representative value matrix, The projected weight matrix represents the trainable values.
[0012] Preferred methods for constructing CNN-LSTM hybrid expert networks include: First, a one-dimensional convolutional layer is used to slide along the time dimension to detect local features; Then, BiLSTM is used to concatenate the final hidden states of the forward LSTM and the backward LSTM to extract features based on historical and future context.
[0013] Preferably, the mapping function implemented by each expert network is: ; in, Represents the time series of the input physiological signal. Parameters representing the expert network, Represents the length of the time series; The mapping function implemented by the gating network is: ; in, The set of trainable parameters representing the gating network. The number of representative expert networks.
[0014] Preferably, in step S5, the formula for weighted summation of the outputs of several heterogeneous expert networks based on the expert weight distribution is as follows: ; in, Represents the final predicted output. The corresponding output of the gating network The weights of an expert network, Representing the The output of an expert network.
[0015] Preferably, the total loss function of the task-aware hybrid expert model framework is: ; in, Represents the total number of training samples. Represents the sample index. Representing the When the actual response of a sample is... Representing the Predicted response time for each sample Representing the The true accuracy of a sample Representing the Prediction accuracy per sample Represents the L2 regularization coefficient. This represents the set of trainable parameters for all expert networks and gated networks.
[0016] Therefore, the present invention employs the above-mentioned real-time cognitive state estimation method driven by physiological data under high-altitude hypoxia, which has the following advantages: (1) In this invention, by integrating three complementary heterogeneous expert networks, namely Transformer, CNN-LSTM and residual MLP, the modeling advantages of different networks in long-range dependence, local temporal features and nonlinear mapping are fully utilized, and the reaction time and accuracy of eight core cognitive dimensions (such as attention, working memory, processing speed, etc.) can be accurately predicted.
[0017] (2) In this invention, intelligent scheduling is performed using a gating network based on a self-attention mechanism, which can dynamically allocate expert weights according to the real-time pattern of the input physiological signal (such as long-term trends related to hypoxia accumulation or sudden fluctuations in heart rate).
[0018] (3) In this invention, a design paradigm of multi-task learning and conditional computation is adopted. Each expert network is equipped with an independent output layer, which enables parallel processing of the prediction needs of heterogeneous cognitive functions such as visuospatial attention, working memory, and information processing speed.
[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0020] Figure 1 This is a flowchart illustrating the real-time estimation method for cognitive state driven by physiological data under high-altitude hypoxia according to the present invention. Figure 2 Scatter plots and Bland-Altman consistency analysis plots corresponding to the accuracy and reaction time of the matching memory test provided in the embodiments of the present invention; Figure 3 Scatter plots and Bland-Altman consistency analysis plots of the accuracy and reaction time of the visuospatial attention test provided in the embodiments of the present invention; Figure 4 Scatter plots and Bland-Altman consistency analysis plots corresponding to the accuracy and reaction time of the sequential memory test provided in the embodiments of the present invention; Figure 5Scatter plots and Bland-Altman consistency analysis plots corresponding to the accuracy and reaction time of the digital cancellation test provided in the embodiments of the present invention; Figure 6 Scatter plots and Bland-Altman consistency analysis plots corresponding to the accuracy and reaction time of symbol decoding tests provided in embodiments of the present invention; Figure 7 Scatter plots and Bland-Altman consistency analysis plots of the accuracy and reaction time of the computational skills test provided in the embodiments of the present invention; Figure 8 Scatter plots and Bland-Altman consistency analysis plots corresponding to the accuracy and reaction time of sensitivity tests provided in embodiments of the present invention; Figure 9 The scatter plot and Bland-Altman consistency analysis plot are provided for the accuracy and reaction time of the attention detachment test in the embodiments of the present invention. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Specific model specifications need to be selected and determined according to the actual specifications of the device, etc. The specific selection calculation method adopts the prior art, and therefore will not be described in detail.
[0022] Example like Figure 1 As shown, this invention provides a real-time cognitive state estimation method driven by physiological data under high-altitude hypoxia conditions, comprising the following steps: S1. Acquire physiological signals of workers in a high-altitude, low-oxygen environment, and preprocess the physiological signals to obtain time-series data. The preprocessing specifically includes: A linear interpolation method was used to align the physiological signal data acquired at different time points with the cognitive test time points to obtain physiological signal time series data; Abnormal samples with a heart rate less than 30 bpm or blood oxygen saturation less than 60% were excluded.
[0023] S2. Input physiological signal time series data into a task-aware hybrid expert model framework, which includes a gating network and several heterogeneous expert networks. S3. Using the gating network in the hybrid expert model framework, the corresponding expert weight distribution is generated based on the input physiological signal time series data through a self-attention mechanism. S4. Parallel feature extraction of the input physiological signal time series data is performed using several heterogeneous expert networks in the hybrid expert model framework. S5. Based on the expert weight distribution, perform a weighted summation of the outputs of several heterogeneous expert networks to output the estimated results of reaction time and accuracy for multiple cognitive dimensions.
[0024] Heterogeneous expert networks include Transformer expert networks, CNN-LSTM (convolutional neural network-long short-term memory) hybrid expert networks, and residual MLP expert networks, each with an independent output layer.
[0025] Transformer expert networks specialize in long-range dependency modeling. Their core is the multi-head self-attention mechanism. Transformer expert networks can be constructed in the following ways: For the input physiological signal time series The query, key, and value matrix is obtained through linear projection: ; ; ; in, Represents the query matrix. Represents the trainable query projection weight matrix. Represents the bond matrix. Represents the trainable key-projection weight matrix. Representative value matrix, The weight matrix represents the projected values of the trainable values. By stacking multiple layers of such encoders, the Transformer expert network is able to capture complex physiological interactions and delay effects across the entire time window.
[0026] CNN-LSTM hybrid expert networks specialize in multi-scale local temporal pattern extraction, and their construction methods include: First, a one-dimensional convolutional layer is used to slide along the time dimension to detect local features; Then, a BiLSTM (Bidirectional Long Short-Term Memory) network is used to concatenate the final hidden states of the forward LSTM and the backward LSTM to extract features based on historical and future context.
[0027] Residual MLP (Multilayer Perceptron) expert networks focus on deep nonlinear mappings. This structure effectively alleviates the gradient vanishing problem in deep networks and allows the network to learn complex high-order interactions between input features.
[0028] The process of generating expert weight distributions using a gating network includes: applying a self-attention mechanism to the input physiological signal time series data to obtain a globally context-weighted feature representation; compressing the feature representation into an E-dimensional logits vector using a multilayer perceptron, where E represents the number of expert networks; and processing the logits vector using the softmax function to generate the weight coefficients of each expert network.
[0029] This design enables the gating network to dynamically allocate weights based on the real-time characteristics of the input physiological patterns. The gating network plays a core scheduling role in this process. Its weights, generated based on the self-attention mechanism, essentially reflect the optimal expert matching relationship between the current input physiological pattern and each cognitive task. For example, when the physiological signal shows a long-term slow trend related to the accumulation of hypoxia, the gating network will assign higher weights to the Transformer expert, which is good at global dependency modeling, so as to better predict memory tasks affected by continuous hypoxia (such as sequential memory and matching memory). When there are sudden fluctuations in heart rate in the signal, the CNN-LSTM hybrid expert, which is good at capturing local temporal patterns, will be activated first to achieve accurate prediction of reaction time indicators that are sensitive to transient physiological changes (such as attentional blink and sensitivity test).
[0030] The mapping function implemented by each expert network is: ; in, Represents the time series of the input physiological signal. Parameters representing the expert network, Represents the length of the time series; The mapping function implemented by the gating network is: ; in, The set of trainable parameters representing the gating network. The number of representative expert networks.
[0031] In step S5, the formula for weighted summation of the outputs of several heterogeneous expert networks based on the expert weight distribution is as follows: ; in, Represents the final predicted output. The corresponding output of the gating network The weights of an expert network, Representing the The output of an expert network.
[0032] The training objective of the entire system is to minimize the difference between the predicted cognitive state and the true value. The total loss function of the task-aware hybrid expert model framework is: ; in, Represents the total number of training samples. Represents the sample index. Representing the When the actual response of a sample is... Representing the Predicted response time for each sample Representing the The true accuracy of a sample Representing the Prediction accuracy per sample Represents the L2 regularization coefficient. This represents the set of trainable parameters for all expert networks and gated networks.
[0033] The training process employs the AdamW optimizer, coupled with cosine annealing learning rate scheduling, to achieve stable convergence and good generalization on limited and noisy extreme environment data.
[0034] To verify the effectiveness of the method, the test results of 113 operators in eight cognitive state estimation tests (visuospatial attention test, matching memory, symbol decoding, attentional blink, sensitivity test, digit cancellation, calculation proficiency test, and sequential memory), along with their heart rate and blood oxygen data during these tests, were used as the physiological data selected for this embodiment. These data and the test results (reaction time and accuracy) constituted the preliminary dataset. Since the wearable device recorded physiological data every 10 minutes, a linear interpolation method was used to align the physiological signals with the cognitive test time points, ensuring that each test sample had corresponding real-time physiological data, and abnormal samples with a heart rate less than 30 bpm or blood oxygen saturation less than 60% were removed.
[0035] In this embodiment, physiological data is randomly divided into a training set (60%), a validation set (20%), and a test set (20%). Since hybrid expert models demonstrate good performance in complex pattern learning, a high-level hybrid architecture comprising six experts was constructed. Specifically, two Transformer expert networks were designed to capture global physiological pattern dependencies, two CNN-LSTM hybrid expert networks were designed to extract local temporal features, two residual MLP expert networks were designed to learn deep nonlinear mappings, and a gating network dynamically allocates expert weights based on a self-attention mechanism. All expert networks had a uniform hidden layer dimension of 128, were trained using the AdamW optimizer with a learning rate of 0.001, and were coupled with cosine annealing scheduling. The validation set... The indicator implements an early termination strategy, selecting the model with the best generalization ability.
[0036] Accuracy of eight tests and reaction time The results are shown in Table 1: Table 1. Accuracy of the Eight Tests and reaction time
[0037] The Pearson correlation coefficients for the eight tests are shown in Table 2: Table 2. Pearson correlation coefficients for the eight tests.
[0038] The percentage of data points falling within the consistency limits is shown in Table 3: Table 3. Percentage of data points falling within the consistency bounds
[0039] A more intuitive graph showing the accuracy and reaction time prediction results for the eight tests is shown below. Figures 2-9 As shown in Tables 1 and 2, both tables reveal the association between predictive efficacy and the neurophysiological basis of cognitive tasks, providing direct evidence for evaluating the strength of the association between the predicted values of the model proposed in this invention and the actual cognitive behavioral measurements. It exhibits extremely high predictive accuracy on most cognitive tasks. Table 3 clearly shows that for all tasks, the proportion of predictions in reaction time and accuracy exceeds 93%, with most exceeding 94%. This indicates that even... On relatively low-level tasks, the model's prediction error is also limited and uniformly distributed, without producing extreme miscalculations. This high consistency is crucial for security monitoring applications, as it means that the model output is stable and reliable.
[0040] Therefore, the present invention adopts the above-mentioned physiological data-driven real-time estimation method for cognitive state under high altitude and low oxygen conditions, and realizes real-time and accurate estimation of the multi-dimensional cognitive state of workers in extreme environments such as high altitude and low oxygen conditions.
[0041] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A real-time cognitive state estimation method driven by physiological data under high-altitude hypoxia conditions, characterized in that: Includes the following steps: S1. Obtain physiological signals of workers in a high-altitude, low-oxygen environment, and preprocess the physiological signals to obtain time series data of physiological signals; S2. Input physiological signal time series data into a task-aware hybrid expert model framework, which includes a gating network and several heterogeneous expert networks. S3. Using the gating network in the hybrid expert model framework, the corresponding expert weight distribution is generated based on the input physiological signal time series data through a self-attention mechanism. S4. Parallel feature extraction of the input physiological signal time series data is performed using several heterogeneous expert networks in the hybrid expert model framework. S5. Based on the expert weight distribution, perform a weighted summation of the outputs of several heterogeneous expert networks to output the estimated results of reaction time and accuracy for multiple cognitive dimensions.
2. The method for real-time estimation of cognitive state driven by physiological data under high-altitude hypoxia as described in claim 1, characterized in that: In step S1, the preprocessing specifically includes: A linear interpolation method was used to align the physiological signal data acquired at different time points with the cognitive test time points to obtain physiological signal time series data; Abnormal samples with a heart rate less than 30 bpm or blood oxygen saturation less than 60% were excluded.
3. The method for real-time estimation of cognitive state driven by physiological data under high-altitude hypoxia as described in claim 2, characterized in that: Heterogeneous expert networks include Transformer expert networks, CNN-LSTM hybrid expert networks, and residual MLP expert networks, each with an independent output layer.
4. The method for real-time estimation of cognitive state driven by physiological data under high-altitude hypoxia as described in claim 3, characterized in that: In step S3, the process of the gating network generating the expert weight distribution includes: A self-attention mechanism is applied to the input physiological signal time series data to obtain a globally context-weighted feature representation; The feature representation is compressed into an E-dimensional logits vector by using a multilayer perceptron, where E represents the number of expert networks; The softmax function is used to process the logits vector to generate the weight coefficients of each expert network.
5. The method for real-time estimation of cognitive state driven by physiological data under high-altitude hypoxia as described in claim 4, characterized in that: The construction methods for Transformer expert networks include: For the input physiological signal time series The query, key, and value matrix is obtained through linear projection: ； ； ； in, Represents the query matrix. Represents the trainable query projection weight matrix. Represents the bond matrix. Represents the trainable key-projection weight matrix. Representative value matrix, The projected weight matrix represents the trainable values.
6. The method for real-time estimation of cognitive state driven by physiological data under high-altitude hypoxia as described in claim 5, characterized in that: The construction methods of CNN-LSTM hybrid expert networks include: First, a one-dimensional convolutional layer is used to slide along the time dimension to detect local features; Then, BiLSTM is used to concatenate the final hidden states of the forward LSTM and the backward LSTM to extract features based on historical and future context.
7. The method for real-time estimation of cognitive state driven by physiological data under high-altitude hypoxia as described in claim 6, characterized in that: The mapping function implemented by each expert network is: ； in, Represents the time series of the input physiological signal. Parameters representing the expert network, Represents the length of the time series; The mapping function implemented by the gating network is: ； in, The set of trainable parameters representing the gating network. The number of representative expert networks.
8. The method for real-time estimation of cognitive state driven by physiological data under high-altitude hypoxia as described in claim 7, characterized in that: In step S5, the formula for weighted summation of the outputs of several heterogeneous expert networks based on the expert weight distribution is as follows: ； in, Represents the final predicted output. The corresponding output of the gating network The weights of an expert network, Representing the The output of an expert network.
9. The method for real-time estimation of cognitive state driven by physiological data under high-altitude hypoxia as described in claim 8, characterized in that: The total loss function of the task-aware hybrid expert model framework is: ； in, Represents the total number of training samples. Represents the sample index. Representing the When the actual response of a sample is... Representing the Predicted response time for each sample Representing the The true accuracy of a sample Representing the Prediction accuracy per sample Represents the L2 regularization coefficient. This represents the set of trainable parameters for all expert networks and gated networks.