Method and system for stress detection based on cardiovascular physiological signals

Abstract

A method for pressure detection based on cardiovascular physiological signals and a system using the method are provided. The method includes: obtaining multiple signal segments based on a cardiovascular physiological signal, where the signal segments are stable under an augmented Dickey-Fuller (ADF) test; calculating multiple trends for the signal segments; obtaining a detrended signal corresponding to the cardiovascular physiological signal based on the trends; extracting multiple first heart rate variability (HRV) features based on the detrended signal; and obtaining a stress detection result corresponding to the cardiovascular physiological signal based on the first HRV features.

Description

[Technical Field]

[0001] This invention relates to a biological signal processing and analysis technology, and more particularly to a stress detection method and system based on cardiovascular physiological signals that combines signal processing and machine learning. [Previous Technology]

[0002] In modern healthcare systems, physiological biological signals such as electrocardiography (ECG) and electroencephalography (EEG) are widely used for automated assessment and prediction. However, these studies often face the challenge of individual identity bias. When the dataset contains multiple samples from the same subject, these signals often carry individual physiological characteristics, making it difficult for the model to distinguish between individual characteristics and task-related characteristics. For example, heart rate variability in ECG or spectral characteristics in EEG may remain consistent across different contexts, causing the model to misinterpret individual characteristics as task-related signals, thereby reducing the model's generalization ability.

[0003] Decoupling techniques are widely used to separate components of interest from various signals. Methods based on autoencoders (AEs) and variational autoencoders (VAEs) are widely used in speech recognition to extract speaker features and improve system effectiveness. Similarly, VAEs are used in music signal processing to extract features and convert music styles. However, these techniques heavily rely on large datasets, and their generalization ability drops significantly when the sample size is limited. This limitation is particularly critical in physiological and biological signals where the sample size is typically small.

[0004] Currently, there is no method to effectively separate individual characteristics from task-related characteristics, especially in the biomedical field where sample sizes are limited. Therefore, developing an effective decoupling method under limited sample conditions is crucial for improving the accuracy and generalization ability of physiological and biological signal analysis. [Summary of the Invention]

[0005] In view of this, the present invention provides a pressure detection method and system based on cardiovascular physiological signals, which can effectively decouple cardiovascular physiological signals and thereby improve the accuracy of pressure detection.

[0006] A first aspect of the present invention provides a stress detection method based on cardiovascular physiological signals, comprising: obtaining multiple signal segments based on cardiovascular physiological signals, wherein these signal segments are stable under the Augmented Dickey-Fuller (ADF) test; calculating multiple trends for these signal segments respectively; obtaining detrended signals corresponding to the cardiovascular physiological signals based on these trends; extracting multiple first heart rate variability (HRV) features based on the detrended signals; and obtaining stress detection results corresponding to the cardiovascular physiological signals based on these first HRV features.

[0007] In some embodiments of the first aspect, the above method further includes: obtaining multiple candidate modulation signals based on cardiovascular physiological signals using multiple modulation schemes; obtaining multiple second HRV features extracted from the modulated signals in these candidate modulation signals; and obtaining pressure detection results corresponding to the cardiovascular physiological signals based on these first HRV features and these second HRV features.

[0008] In some embodiments of the first aspect, obtaining a plurality of second HRV features extracted from the modulated signal in these candidate modulated signals includes: extracting a plurality of second candidate HRV features from these candidate modulated signals respectively; using a selection model to obtain a plurality of weights corresponding to these second candidate HRV features; and selecting the second candidate HRV feature corresponding to the largest weight. The selected second candidate HRV feature is extracted from the modulated signal.

[0009] In some embodiments of the first aspect, each of the above modulation schemes includes at least one of magnitude warping, scaling, noise addition, jittering, time warping, window warping, and permutation.

[0010] In some embodiments of the first aspect, the above-mentioned method of obtaining stress detection results corresponding to cardiovascular physiological signals based on these first HRV features and these second HRV features includes: obtaining embedding features using a multi-head attention model based on these first HRV features and these second HRV features; and obtaining stress detection results based on the embedding features.

[0011] In some embodiments of the first aspect, the multi-head attention model described above is trained via guided contrastive learning.

[0012] In some embodiments of the first aspect, the above-mentioned method of obtaining multiple signal segments based on cardiovascular physiological signals includes: normalizing the cardiovascular physiological signals to obtain normal signals; dividing the normal signals into multiple first candidate segments based on a first segment length; using an ADF test to determine whether multiple first p values ​​corresponding to these first candidate segments are all less than a preset threshold; when the first p values ​​are all less than the preset threshold, determining that these first candidate segments are the above-mentioned multiple signal segments; and when one of the first p values ​​is not less than the preset threshold, reselecting a second segment length to divide the normal signals into multiple second candidate segments.

[0013] In some embodiments of the first aspect, the length of the first segment is less than the length of the second segment.

[0014] In some embodiments of the first aspect, the above-described calculation of multiple trends of these signal segments includes: fitting these trends based on these signal segments using the Ordinary Least Squares (OLS) method.

[0015] A second aspect of the present invention provides a pressure detection system comprising: one or more non-transitory computer-readable memories storing one or more computer-executable instructions; and at least one processor coupled to the one or more non-transitory computer-readable memories. The at least one processor is configured to execute one or more computer-executable instructions to cause the pressure detection system to perform the pressure detection method as described in the first aspect of the present invention.

Implementation Method

[0016] The following description contains specific information relating to exemplary embodiments of the present invention. The accompanying drawings and detailed description are exemplary embodiments only. However, the present invention is not limited to these exemplary embodiments. Other variations and embodiments of the invention will occur to those skilled in the art. Unless otherwise stated, the same or corresponding elements in the drawings may be indicated by the same or corresponding reference numerals. Furthermore, the drawings and illustrations in the present invention are generally not drawn to scale and are not intended to correspond to actual relative dimensions.

[0017] For the purposes of consistency and ease of understanding, the same features are indicated by reference numerals in the exemplary drawings (although not in some examples). However, features in different embodiments may differ in other respects, and therefore should not be narrowly limited to the features shown in the drawings.

[0018] The terms "at least one embodiment," "one embodiment," "multiple embodiments," "different embodiments," "some embodiments," and "this embodiment" indicate that the embodiments of the present invention described herein may include specific features, structures, or characteristics, but not every possible embodiment of the present invention must include specific features, structures, or characteristics. Furthermore, the repeated use of the phrases "in one embodiment" and "in this embodiment" does not necessarily refer to the same embodiment, although they may be the same. Moreover, the use of phrases such as "embodiment" in connection with "the present invention" does not mean that all embodiments of the present invention must include specific features, structures, or characteristics, and should be understood as "at least some embodiments of the present invention" including the specific features, structures, or characteristics stated therein. The term "coupled" is defined as a connection, whether direct or indirect through an intermediate element, and is not necessarily limited to physical connections. When the term "comprising" is used, it means "including but not limited to," which explicitly indicates an open inclusion or relationship of combinations, groups, series, and equivalents.

[0019] In addition, for purposes of explanation and non-limitation, specific details such as functional entities, technologies, agreements, standards, etc., are described to provide an understanding of the described technologies. In other examples, detailed descriptions of well-known methods, technologies, systems, architectures, etc., are omitted to avoid obscuring the explanatory narrative with unnecessary details.

[0020] The terms "first," "second," and "third," etc., in the specification and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order, and do not necessarily correspond to "first," "second," and "third" in the claims. Furthermore, the term "comprising," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or modules is not limited to the listed steps or modules, but may optionally include steps or modules not listed, or may optionally include other steps or modules inherent to these processes, methods, products, or devices.

[0021] This invention proposes a stress detection system and method that can decoupling cardiovascular physiological signals to reduce interference between identity information and task-related information in the cardiovascular physiological signals, thereby improving the accuracy of stress detection. It must be noted that in several embodiments of this invention, heart rate (HR) signals are used as examples of cardiovascular physiological signals to illustrate the technology and effects of this invention. However, this invention is not limited to the above. Those skilled in the art can apply the method proposed in this invention to other types of cardiovascular physiological signals based on the technical concepts described in these embodiments.

[0022] Figure 1 illustrates a schematic diagram of a pressure detection system according to an embodiment of the present invention.

[0023] Referring to Figure 1, the pressure detection system 10 can, for example, obtain cardiovascular physiological signals 12 from the first device 11, and then, based on the cardiovascular physiological signals 12, execute the pressure detection method proposed in this embodiment of the invention to obtain a pressure detection result 13. Furthermore, the pressure detection system 10 can, for example, send the pressure detection result 13 to the second device 14.

[0024] In some embodiments, the first device 11 may be an internal or external device of the pressure detection system 10. For example, the first device 11 may be a device including physiological signal measuring elements such as an electrocardiograph, photoplethysmograph (PPG), blood pressure monitor, phonocardiograph (PCG), and cardiac output monitor. For example, the first device 11 may include a wearable device, a portable device, and / or a fixed device.

[0025] In some embodiments, the cardiovascular physiological signal 12 is, for example, but not limited to, a heart rate signal.

[0026] In some embodiments, the pressure detection result 13 includes, for example, a binary result indicating whether there is pressure or no pressure. In some embodiments, the pressure detection result 13 includes, for example, a pressure level and / or a pressure score.

[0027] In some embodiments, the second device 14 is, for example, an internal or external device of the pressure detection system 10, having an output element utilizing means such as images or sounds. For example, the second device 14 may be a wearable device, a portable device, and / or a fixed device.

[0028] Figure 2 illustrates a flowchart of a pressure detection method according to an embodiment of the present invention; Figure 3 illustrates a schematic diagram of signal detrending performed based on the Augmented Dickey-Fuller (ADF) test according to an embodiment of the present invention. The pressure detection method in Figure 2 is presented, for example, as flow 200. Furthermore, the method is performed, for example, by the pressure detection system 10 of the embodiment of Figure 1, and the method flow includes the detrending operation of cardiovascular physiological signals based on the ADF test as shown in Figure 3. Therefore, the embodiments related to Figure 2 will be described in conjunction with Figures 1 and 3 herein.

[0029] Referring to Figure 2, the stress detection system 10 first acquires cardiovascular physiological signals 12. For example, the stress detection system 10 can acquire heart rate signals from the wearable device 12.

[0030] In action S210, the pressure detection system 10 obtains multiple signal segments based on the cardiovascular physiological signal 12, and these signal segments are stable under the ADF test.

[0031] Specifically, the stress detection system 10 may, for example, cut the cardiovascular physiological signal 12 (e.g., heart rate signal or its normalized signal) into multiple signal segments with the goal that each signal segment is stable under the ADF test.

[0032] Referring to Figure 3, in some embodiments, the cardiovascular physiological signal 12 is, for example, a heart rate signal 31. The stress detection system 10 first normalizes the heart rate signal 31 into a normal signal 32, and then cuts out multiple signal segments 321 from the normal signal 32. During the cutting process, the stress detection system 10 aims to ensure that these signal segments 321 are stable under the ADF test. When the p-value of the signal under the ADF test is less than a preset threshold, the null hypothesis is rejected, that is, the corresponding signal is determined to be stable; otherwise, the null hypothesis is accepted, that is, the corresponding signal is determined to be unstable.

[0033] In some embodiments, the length of each signal segment 321 is, for example, the same. In other embodiments, the length of each signal segment 321 may not be the same.

[0034] Taking the example that each signal segment 321 has the same length, the pressure detection system 10 first divides the normal signal 32 into multiple first candidate segments based on multiple first length segments. Then, for each first candidate segment, an ADF test can be used to obtain a p-value. Therefore, when the p-value corresponding to each first candidate segment is less than a preset threshold, the pressure detection system 10 will determine that these first candidate segments are multiple signal segments 321; when there is at least one first candidate segment whose p-value is not less than the preset threshold, the pressure detection system 10 will reselect a second length segment, and then divide the normal signal 32 into multiple second candidate segments based on multiple second length segments. When the p-value corresponding to each second candidate segment is less than the preset threshold, the pressure detection system 10 will determine that these second candidate segments are multiple signal segments 321; when there is at least one second candidate segment whose p-value is not less than the preset threshold, the pressure detection system 10 will reselect a third length segment and repeat the above actions, and so on, until multiple signal segments 321 are found.

[0035] In some embodiments, the selection of the length segment is, for example, from small to large, that is, starting from a preset minimum length to search for a suitable signal segment 321 until a preset maximum length. In some embodiments, the selection of the length segment can also be from large to small, that is, starting from a preset maximum length to search for a suitable signal segment 321 until a preset minimum length.

[0036] In action S220, the pressure detection system 10 calculates the trend of each signal segment separately.

[0037] Referring to Figure 3, since each signal segment 321 is stable, there will be a corresponding trend 331. In some embodiments, the pressure detection system 10 uses, for example, the Ordinary Least Squares (OLS) method to fit the corresponding trend 331 for each signal segment 321, thereby obtaining the trend line 33.

[0038] In action S230, the stress detection system 10 will obtain the detrended signal corresponding to the cardiovascular physiological signal 12 based on these trends.

[0039] Refer to Figure 3. For example, the pressure detection system 10 will subtract the trend line 33 from the normal signal 32 to obtain the detrending signal 34.

[0040] In some embodiments, detrending is performed on cardiovascular physiological signals (e.g., heart rate signal 31) using ADF testing. The resulting detrended signal can be considered as a signal with identity characteristics within the cardiovascular physiological signals. In other words, by executing actions S210, S220, and S230, the stress detection system 10 can successfully extract signals containing individual or identity characteristics (e.g., detrended signal 34) from the cardiovascular physiological signals.

[0041] In some embodiments, actions S210, S220, and S230 can be accomplished, for example, by the control flow presented by the following pseudo code: Requirement: Signal X, length N Ensure: Detrended signal X detrended 1: Step 1: Initialization settings 2: Initialize minimum allowable length Lmin and maximum allowable length Lmax 3: Set stability tolerance ϵ 4: Step 2: Segment detection 5: for L=Lminto Lmaxdo 6: Divide X into multiple segments of length L 7: for each segment xido 8: Perform ADF test on xi 9: if p value < ϵ then 10: Mark xi as stable 11: end if 12: end for 13: if all segments are stable then 14: Proceed to step 3 15: end if 16: end for 17: Step 3: Detrended 18: Calculate the local trend for each segment and remove it from X 19: Combine all detrended fragments to obtain Xdetrended 20: return Xdetrended

[0042] In action S240, the stress detection system 10 extracts multiple first heart rate variability (HRV) features based on the detrended signal.

[0043] Specifically, the stress detection system 10 uses a feature extractor to extract multiple first HRV features from the detrending signal.

[0044] Advantageously, these first HRV features can be considered as features that better reflect an individual or identity. In other words, multiple first HRV features, or a set of first HRV features, can be obtained for each heart rate signal, and there is a high degree of feature similarity among multiple sets of first HRV features corresponding to multiple heart rate signals of the same individual. Therefore, in this paper, first HRV features are also referred to as identity features.

[0045] In action S250, the pressure detection system 10 will obtain the pressure detection result 13 corresponding to the cardiovascular physiological signal 12 based on these first HRV features.

[0046] Specifically, obtaining the first HRV feature that reflects an individual's or identity is equivalent to obtaining additional individual or identity-related information corresponding to cardiovascular physiological signals. With the assistance of this additional information, the system will be able to perform more accurate analysis for the target task. In other words, the stress detection system 10 will be able to obtain more accurate stress detection results 13.

[0047] In some embodiments, the stress detection system 10 may, for example, first analyze the corresponding identity using a pre-trained model based on these first HRV features, then obtain (e.g., a pre-stored) the stress detection model corresponding to that identity, and input the cardiovascular physiological signal 12 into the stress detection model corresponding to that identity for analysis to obtain the stress detection result 13. In other words, the stress detection system 10 can maintain different stress detection models for different identities to reduce the impact of individual differences on stress detection.

[0048] In some embodiments, for privacy protection, the aforementioned identity is maintained in the pressure detection system 10, for example, by a number or other means.

[0049] In some embodiments, the stress detection system 10 may further utilize this first HRV feature, which reflects an individual or identity, to extract task-related features (e.g., second HRV features, also known as content features) from the cardiovascular physiological signal 12. This will be described in detail below with examples.

[0050] Figure 4 illustrates a flowchart of a pressure detection method according to an embodiment of the present invention; Figure 5 illustrates a schematic diagram of a pressure detection method according to an embodiment of the present invention. The pressure detection method in Figure 4 is presented, for example, as flow 400, in which multiple actions are identical to those in flow 200; related details will not be repeated herein. Furthermore, the pressure detection method in Figure 4 can be executed, for example, by the pressure detection system 10 of the embodiment in Figure 1, and the model adopted by the method can be presented, for example, by the architecture shown in Figure 5. Therefore, the embodiments related to Figure 4 will be described in conjunction with Figures 1 and 5 herein.

[0051] Refer to Figures 4 and 5. In actions S210 to S240, the stress detection system 10 will first obtain the detrended signal 51 corresponding to the cardiovascular physiological signal 50, and then extract multiple first HRV features (e.g., identity features) from the detrended signal 51.

[0052] In action S410, the pressure detection system 10 will obtain multiple candidate modulation signals based on the cardiovascular physiological signal 12 using multiple modulation schemes.

[0053] Specifically, heart rate in real life can be affected by various factors such as visceral pathology, user activity, and ambient temperature, resulting in variations in the timing and amplitude of the heart rate signal. To address these variations, the stress detection system 10 uses multiple modulation schemes based on the cardiovascular physiological signal 12 to obtain multiple candidate modulation signals. Each modulation scheme belongs to one of the following types: spatial modulation, temporal modulation, and spatiotemporal modulation. For the cardiovascular physiological signal 12, each modulation scheme can correspond to a candidate modulation signal. For example, as shown in Figure 5, the stress detection system 10 uses multiple modulation schemes based on the cardiovascular physiological signal 50 to obtain multiple candidate modulation signals 53.

[0054] In some embodiments, spatial modulation of this type of modulation scheme includes at least one of magnitude warping, scaling, noise addition, and jittering.

[0055] In some embodiments, the modulation scheme of this type, time modulation, includes at least one of time warping, window warping, and permutation.

[0056] In some embodiments, the modulation scheme of the time-varying type includes at least one spatial modulation scheme and at least one time modulation scheme. For example, the modulation scheme of the time-varying type includes a composite modulation scheme that starts with a time modulation scheme and is followed by a spatial modulation scheme.

[0057] In action S420, the pressure detection system 10 obtains multiple second HRV features extracted from the modulated signal among multiple candidate modulation signals.

[0058] Specifically, the stress detection system 10 uses a feature extractor to extract features from a candidate modulation signal to extract multiple second HRV features (also known as a set of second HRV features).

[0059] In some embodiments, the stress detection system 10 uses a feature extractor to extract multiple HRV features (e.g., a second HRV feature) from a signal (e.g., a candidate modulation signal). For example, the signal is first segmented into multiple segments of a preset unit length, and multiple HRV features (e.g., 61) are extracted from each segment. Then, the feature extractor performs statistical aggregation on these HRV features to obtain 13 statistical values. Therefore, one signal can correspond to 793 statistical features. Finally, the feature extractor uses, for example, a method such as Recursive Feature Elimination (RFE) to extract multiple HRV features from the 793 statistical features as the final output.

[0060] For example, the feature extractor might divide a 24-hour candidate modulation signal into 288 segments of 5 minutes each, and extract multiple HRV features from each segment. These extracted HRV features might include, for example, 8 temporal features, 6 frequency features, and 47 multi-scale features, totaling 61 HRV features. Next, statistical aggregation is performed on the 61 HRV features for each of the 288 segments to obtain 13 statistical values. Therefore, a candidate modulation signal can yield 793 statistical features. Finally, the feature extractor might use a random forest algorithm, for example, to select the 100 most important statistical features from the 793 statistical features as multiple (or a group of) second HRV features corresponding to the candidate modulation signal.

[0061] For example, the eight time features include the mean NN interval (meanNN), the standard deviation of the NN interval (sdNN), the coefficient of variation, the mean of the first-order differences, the root mean square of the sum of squares of adjacent differences (RMSSD), the standard deviation of the absolute value of the first-order differences, the proportion of adjacent NN intervals greater than 50 milliseconds (pNN50), and the normalized mean of the absolute value of the first-order differences.

[0062] For example, the six frequency characteristics include high frequency power (HF), low frequency power (LF), high frequency to low frequency power ratio (HF / LF), normalized high frequency power, and very low frequency power (VLF).

[0063] For example, the 47 multi-scale features include the following features calculated for the difference (dRR) between RR intervals: Multi-Scale Permutation Entropy (MSPE), Multi-Scale Modified Permutation Entropy (MSmPE), Permutation Entropy (PE), First-order difference, Second-order difference, Difference of First-order difference, Difference of Second-order difference, Sum of First-order difference, Sum of Second-order difference; and an additional total asymmetry index.

[0064] In some embodiments, the stress detection system 10 may, for example, use a feature extractor for each candidate modulation signal to extract multiple sets of second candidate HRV features, and then use a selection model to select one set from the multiple sets of second candidate HRV features as the multiple second HRV features corresponding to the modulated signal. For example, as shown in FIG5, the stress detection system 10 obtains multiple sets of second candidate HRV signals from multiple candidate modulation signals 53, and then uses a selection model 54 to select multiple second HRV features 55 corresponding to the modulated signal. In other embodiments, the stress detection system 10 may, for example, first use a selection model to select one of the multiple candidate modulation signals as the modulated signal, and then use a feature extractor to extract multiple (e.g., one set) second HRV features corresponding to it from the modulated signal.

[0065] In some embodiments, the selected model 54 is, for example, a model based on a convolutional neural network (CNN), which can output a corresponding weight for each set of second HRV features after inputting multiple sets of second HRV features. Therefore, the stress detection system 10 can select the set of second HRV features corresponding to the highest weight as multiple second HRV features 55 corresponding to the modulated signal.

[0066] Advantageously, this architecture allows the selection model 54, after training, to find the most suitable modulation scheme for cardiovascular physiological signals of different individuals or scenarios. In other words, compared to the first HRV feature (e.g., identity feature), the selected second HRV feature is not easily affected by different factors such as individuals or scenarios, and can be regarded as task-related feature. Therefore, the selected multiple second HRV features 55 are also referred to as content features in this paper.

[0067] In action S430, the pressure detection system 10 obtains the pressure detection result 13 corresponding to the cardiovascular physiological signal 12 based on multiple first HRV features and multiple second HRV features.

[0068] Specifically, obtaining a first HRV feature that reflects an individual's or identity is equivalent to obtaining additional individual or identity-related information corresponding to the cardiovascular physiological signal. Obtaining a second HRV feature that is not affected by factors such as the individual or the scenario is equivalent to obtaining additional task-related information corresponding to the cardiovascular physiological signal. With the assistance of this additional information, the system will be able to perform more accurate analysis for the target task. In other words, the stress detection system 10 will be able to obtain more accurate stress detection results 13.

[0069] For example, as shown in Figure 5, the stress detection system 10 may combine multiple first HRV features 52 and multiple second HRV features 55 as input to a multi-head attention model 56 to obtain embeddings 57 (or embedding space). Subsequently, the embeddings 57 may be passed through a classifier 58 to obtain a stress detection result 13.

[0070] For example, the multi-head attention model 56 may be, but is not limited to, a neural network model including an input layer, attention layer, fusion layer, and linear layer, etc. For example, the classifier 58 may be, but is not limited to, a Temporal Convolution Neural Network (TCNN) classifier.

[0071] As shown in Figure 5, the architecture of the entire stress detection method includes at least three machine learning models that can be trained simultaneously: a selection model 54, a multi-head attention model 56, and a classifier 58. Using existing datasets (e.g., but not limited to, TILES 2018), the stress detection system 10 can train these machine learning models, for example, through guided contrastive learning.

[0072] In detail, in order for the multi-head attention model 56 to separate features with similar identities or individuals and bring features with different identities or individuals closer together, the identity contrastive loss (LID) is first defined:

[0073] Where S is the (e.g., cosine) similarity between different samples, i is the sample index, k and j are the positive and negative sample pair indices, and m, α, and β are hyperparameters. In detail, the hyperparameter m is used to extrapolate the distribution of positive and negative samples, and the hyperparameters α and β are used to adjust the loss sensitivity of positive and negative samples.

[0074] Next, in order to consider the classification effect of stress detection, the stress contrast loss LStress is defined:

[0075] Where N is the total number of samples, si and sj are the pressure labels of samples i and j, is the indicator function (for example, 1 when si=sj, and 0 otherwise), zi and zj are the feature vectors of samples i and j, and τ is the temperature coefficient.

[0076] Next, the binary cross-entropy loss (LBCE) will be used in the fine-tuning phase:

[0077] Where N is the total number of samples, yi is the ground truth label of sample i, and is the stress prediction probability of sample i (e.g., between 0 and 1).

[0078] To emphasize the purpose of training at each stage in guided contrastive learning, weights are added to the total loss Ltotal:

[0079] In actual training, in the early stages of training (e.g., the training phase of the first n iterations), (α, β, γ) can be set to (0.6, 0.3, 0.1). Advantageously, a higher α value ensures the decoupling of individual or identity features, and the retention of task-related features is adjusted by adjusting the β value. Then, in the fine-tuning stage in the later stages of training (e.g., the training phase starting from the (n+1)th iteration), (α, β, γ) can be set to (0.1, 0.0, 0.9). Advantageously, having the γ value dominate the fine-tuning stage ensures the accuracy of classifier 58. In some embodiments, in the later stages of training, for example, selection model 54 and multi-head attention model 56 are frozen, focusing on fine-tuning classifier 58.

[0080] After each model is trained, the stress detection system 10 can perform a high-accuracy stress detection method using the architecture shown in the embodiment of Figure 5.

[0081] Figure 6 shows a schematic block diagram of a computing system according to an embodiment of the present invention.

[0082] Referring to Figure 6, the pressure detection method described herein can be implemented on one or more computing systems 600 having various hardware components. For example, the computing system 600 can be implemented as a first device 11, a pressure detection system 10, and / or a second device 14. In some embodiments, the computing system 600 can be implemented as an electronic device, including, but not limited to, one or more of the following components: a central processing unit (CPU) 610, a graphics processing unit (GPU) 620, an input / output element 630, a network element 640, and a memory 650, which can communicate and transmit data via the system bus 660. However, the present invention does not limit the specific type, quantity, and configuration of the components, and those skilled in the art can adjust, select, or add or remove components according to the specific needs and operating environment when implementing the present invention.

[0083] In some embodiments, the main computing core inside the computing system 600 is one or more processors 610. This processor 610 may be responsible for the main computing process and related control logic of algorithms such as deep learning. In some embodiments, the processor 610 is used to execute processing instructions (i.e., machine-executable instructions) stored in a non-volatile computer-readable medium (e.g., storage device 670).

[0084] In some embodiments, to improve the computational efficiency of deep learning, the computing system 600 may further include one or more image processors 620 specifically designed for performing massive parallel computations. These image processors 620 can effectively enhance the system's computational power during deep learning training and inference.

[0085] In some embodiments, the computing system 600 may include a variety of input / output elements 630 for receiving user input and displaying system output. For example, these input / output elements 630 may include a keyboard, mouse, touchpad, display screen, speaker, and other types of sensing devices.

[0086] In some embodiments, the computing system 600 may also include a network element 640 for network communication. For example, this network element 640 may include a network interface card for wired or wireless network connections, or a communication module for 3G, 4G, 5G or other wireless communication technologies.

[0087] In some embodiments, the computing system 600 may include one or more memory elements 650, such as volatile memory elements like random access memory (RAM). The memory 650 may be used to store parameters of a deep learning model, as well as other data and programs used to run algorithms such as deep learning.

[0088] In addition, the computing system 600 may also include one or more of the following elements: storage device 670, power management element 680, and various other elements 690.

[0089] In some embodiments, the computing system 600 may include one or more storage devices 670, such as non-volatile memory elements like hard disk drives (HDDs) or solid-state drives (SSDs). These storage devices 670 can be used to store information such as the code, training data, and model parameters of deep learning software. Furthermore, the storage devices 670 can also be used to store intermediate results and final outputs of algorithms such as deep learning.

[0090] In some embodiments, the computing system 600 may include one or more power management elements 680 for providing power to various hardware components of the computing system 600 and managing their power consumption. This power management element 680 may include a battery, a power converter, and other power management devices.

[0091] In some embodiments, the computing system 600 may also include various other (hardware) components 690, such as cooling fans, heat sinks, and other various control and monitoring devices, which are not limited thereto.

[0092] Furthermore, the embodiments described herein can also be implemented as one or more computer program products, including computer programs having one or more instructions. Specifically, a computer program (also referred to as a program, software, script, or code) can be presented in any form of programming language, and this computer program can be deployed in any form. During operation of the computing system 600 (e.g., an electronic device), the instructions, or a portion thereof, may also reside wholly or at least partially within the processor 610 to cause the processor 610 to perform the methods described herein.

[0093] In summary, the pressure detection method and system proposed in the embodiments of the present invention can effectively decouple cardiovascular physiological signals, thereby improving the accuracy of pressure detection.

[0094] Based on the above description, it is evident that various techniques can be used to implement the concepts described in this application without departing from the scope of these concepts. Furthermore, although the concepts have been described with specific reference to certain embodiments, those skilled in the art will recognize that changes in form and detail may be made without departing from the scope of these concepts. Thus, the described embodiments are to be considered illustrative rather than restrictive in all respects. Moreover, it should be understood that this application is not limited to the specific embodiments described above, but many rearrangements, modifications, and substitutions can be made without departing from the scope of the invention. 10: Stress Detection System 11: First Device 12, 50: Cardiovascular Physiological Signals 13: Stress Detection Results 14: Second Device 200, 400: Process 31: Heart Rate Signal 32: Regular Signal 321: Signal Segment 33: Trend Line 331: Trend 34, 51: Detrended Signal 52: First HRV Feature 53: Candidate Modulation Signal 54: Selection Model 55: Second HRV Feature 56: Multi-head Attention Model 57: Embedded Feature 58: Classifier 600: Computing System 610: Processor 620: Image Processor 630: Input / Output Components 640: Network Components 650: Memory 660: Bus 670: Storage Device 680: Power Management Components 690: Other Components S210, S220, S230, S240, S250, S410, S420, S430: Actions of the Stress Detection Method [Simplified Explanation of the Diagram]

[0095] Figure 1 is a schematic diagram of a pressure detection system according to an embodiment of the present invention. Figure 2 is a flowchart of a pressure detection method according to an embodiment of the present invention. Figure 3 is a schematic diagram of signal detrending performed by an Augmented Dickey-Fuller (ADF) test according to an embodiment of the present invention. Figure 4 is a flowchart of a pressure detection method according to an embodiment of the present invention. Figure 5 is a schematic diagram of a pressure detection method according to an embodiment of the present invention. Figure 6 is a schematic block diagram of a computing system according to an embodiment of the present invention.

Claims

1. A stress detection method based on cardiovascular physiological signals, comprising: Multiple signal segments were obtained based on the cardiovascular physiological signal, and these signal segments were stable under the Augmented Dickey-Fuller (ADF) test. Multiple trends were calculated for each of these signal segments. Based on these trends, the detrended signal corresponding to the cardiovascular physiological signal was obtained. Based on the detrended signal, multiple first heart rate variability (HRV) features were extracted. Based on these first HRV features, the stress detection result corresponding to the cardiovascular physiological signal was obtained.

2. The pressure detection method as described in claim 1 further includes: Based on this cardiovascular physiological signal, multiple candidate modulation signals were obtained using multiple modulation schemes; Multiple second HRV features are extracted from the modulated signals in the candidate modulation signals; and based on the first HRV features and the second HRV features, the pressure detection result corresponding to the cardiovascular physiological signal is obtained.

3. The pressure detection method as described in claim 2, wherein obtaining the second HRV features extracted from the modulated signal among the candidate modulated signals includes: Multiple second candidate HRV features were extracted from these candidate modulation signals respectively; Using a selection model, multiple weights corresponding to the second candidate HRV features are obtained; and the second candidate HRV feature corresponding to the largest of these weights is selected, wherein the selected second candidate HRV feature is extracted from the modulated signal.

4. The pressure detection method as described in claim 2, wherein each of the modulation schemes includes at least one of magnitude warping, scaling, noise addition, jittering, time warping, window warping, and permutation.

5. The stress detection method as described in claim 2, wherein obtaining the stress detection result corresponding to the cardiovascular physiological signal based on the first HRV features and the second HRV features includes: Based on these first HRV features and these second HRV features, an embedding feature is obtained using a multi-head attention model; and based on this embedding feature, the stress detection result is obtained.

6. The stress detection method as described in claim 5, wherein the multi-head attention model is trained via guided contrastive learning.

7. The pressure detection method as claimed in claim 1, wherein obtaining the signal segments based on the cardiovascular physiological signal includes: The cardiovascular physiological signal is normalized to obtain a normalized signal; based on the first segment length, the normalized signal is segmented into multiple first candidate segments; using the ADF test, it is determined whether multiple first p values ​​corresponding to these first candidate segments are all less than a preset threshold; when all of these first p values ​​are less than the preset threshold, these first candidate segments are determined to be these signal segments; and when at least one of these first p values ​​is not less than the preset threshold, a second segment length is reselected to segment the normalized signal into multiple second candidate segments.

8. The pressure detection method as described in claim 7, wherein the length of the first segment is less than the length of the second segment.

9. The pressure detection method as described in claim 1, wherein calculating the trends of the signal segments respectively includes: Based on these signal segments, the trends were fitted using the Ordinary Least Squares (OLS) method.

10. A pressure detection system, comprising: One or more non-transitory computer-readable memories that store one or more instructions that can be executed by a computer; And at least one processor, coupled to the one or more non-transitory computer-readable memories, for executing the one or more computer-executable instructions to cause the stress detection system to perform the stress detection method as described in any one of claims 1 to 9.

Images