Fair-oriented physiological measurement model training method and system

Abstract

The invention relates to the technical field of artificial intelligence, discloses a fairness-oriented physiological measurement model training method and system, and aims to solve the problem of insufficient performance of camera-based human physiological sign detection in deep skin color crowds in the prior art. The method comprises the steps that face video data and synchronous physiological signals of a multi-skin-color individual are collected, and skin color category labels are distributed according to a Fitzpatrick scale; constructing a balanced training set, and adjusting sample distribution of each skin color category through an oversampling and synthetic sample generation algorithm; a hybrid model of a deep convolutional neural network and a recurrent neural network is constructed, and physiological signal time sequence features are extracted; a skin color bias corrector based on the maximum mean value difference is introduced, and the dependence of the model on the specific skin color is inhibited; designing a multi-task learning framework, and optimizing physiological parameter prediction and skin color fairness constraint loss; and performing end-to-end training by using the balanced training set, and evaluating the model performance on the independent test set. The system comprises a skin color perception data acquisition module, a skin color balance training set construction module, a physiological signal feature extraction network, a skin color bias correction module, a multi-task learning and fusion module and a model verification and deployment module. By adopting the technical scheme, the measurement performance of people with deep skin color can be remarkably improved, and the physiological detection technology is promoted to develop in a more fair and universal direction.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a method and system for training physiological measurement models with a focus on fairness. Background Technology

[0002] Non-contact physiological measurement technology based on cameras has shown broad application prospects in health monitoring, telemedicine, and human-computer interaction in recent years. This technology achieves non-invasive estimation of key physiological parameters such as heart rate and blood oxygen saturation by analyzing subtle color changes in facial or skin areas in videos. However, existing technologies have revealed significant group equity deficiencies in practical deployments, especially with extremely prominent performance differences among people of different skin colors, severely limiting their universality and reliability in the real world.

[0003] Among these, training physiological measurement models for fairness has become a key direction for improving technological inclusivity. Its core lies in building a data foundation that covers diverse skin color phenotypes and actively suppressing model preferences or biases towards specific skin color groups through algorithmic mechanisms. Current mainstream methods typically rely on naturally collected video datasets for model training. However, such datasets generally suffer from severe imbalances in skin color distribution, with light-skinned samples dominating while dark-skinned samples are scarce and of inconsistent quality. This makes it difficult for models to learn robust physiological signal representations for dark-skinned individuals during training.

[0004] Current technologies lack systematic skin color balancing strategies at the data level and have not effectively introduced targeted bias correction mechanisms at the algorithm level, causing a sharp decline in model performance when generalizing across skin colors. Simultaneously, due to the difficulty in obtaining real dark skin data, existing augmentation methods struggle to generate physiologically accurate synthetic samples, further exacerbating measurement bias in minority groups. These problems are particularly severe in high-risk applications such as healthcare, potentially leading to misjudgments of the physiological state of individuals with dark skin and even systemic discrimination. Therefore, there is an urgent need for a physiological measurement model training method that integrates balanced data construction, synthetic data augmentation, and skin color bias correction to achieve truly fair, reliable, and universally applicable non-contact physiological monitoring. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a physiological measurement model training method and system oriented towards fairness, which can effectively solve the problems in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: In one aspect, a physiological measurement model training system oriented towards fairness, the system comprising the following components:

[0007] The skin color perception data acquisition module acquires raw video data and corresponding real physiological signal labels for individuals with various skin colors, and automatically labels each individual sample with its skin color category. Skin color categories are divided into 6 levels according to the Fitzpatrick skin typing scale. The skin color balance training set construction module, connected to the skin color perception data acquisition module, receives the raw data and adjusts the sample distribution of each skin color category in the dataset to meet a preset balance ratio through resampling and synthetic data generation techniques. The physiological signal feature extraction network, connected to the skin color balance training set construction module, extracts physiological signal features from the input balance training set video frame sequence. The system includes: a temporal feature representation related to the skin color bias; a skin color bias correction module, integrated into the training process of the physiological signal feature extraction network, used to calculate and minimize the distributional differences in the model's response to features of different skin color categories; a multi-task learning and fusion module, connected to the physiological signal feature extraction network and the skin color bias correction module, used to jointly optimize the physiological parameter regression loss and the skin color fairness constraint loss, and to fuse the gradient information of the two to update the network parameters; and a model validation and deployment module, used to evaluate the performance of the trained model on a cross-skin color test set independent of the training set, ensuring that it achieves the preset accuracy and fairness indicators across all skin color groups, and deploying the qualified model to the actual application environment.

[0008] On the other hand, a fairness-oriented physiological measurement model training method is proposed, comprising the following steps: Step S110, collecting facial video data and synchronous physiological signals from individuals with multiple skin colors, and assigning skin color category labels to each sample according to the Fitzpatrick scale; Step S120, performing skin color category statistical analysis on the collected raw dataset, and applying oversampling and synthetic sample generation algorithms to construct a training dataset with a balanced number of samples for each skin color category; Step S130, constructing a hybrid model of deep convolutional neural network and recurrent neural network as a physiological signal feature extractor, which takes video frame sequences as input and outputs corresponding physiological signal temporal features; Step S140, introducing a skin color bias correction based on the maximum mean difference on the basis of the feature extraction network. The corrector suppresses the model's dependence on a specific skin color by measuring and reducing the distribution distance between different skin color groups in the feature space; step S150: design a multi-task learning framework that simultaneously optimizes the main task loss for physiological parameter prediction and the auxiliary task loss for skin color fairness constraints, and achieves collaborative training of the two through a gradient reversal layer or weight sharing mechanism; step S160: use the constructed balanced training set to perform end-to-end training on the above multi-task model until the model's overall performance on the validation set converges; step S170: comprehensively evaluate the heart rate estimation error, blood oxygen saturation measurement accuracy, and cross-skin color group performance standard deviation of the trained model on an independent test set containing unseen dark and light skin color samples, and after ensuring that the model meets the requirements of fairness and reliability, it is deployed in practice.

[0009] Preferably, the skin tone equalization training set construction module specifically performs the following operations: First, it counts the number of samples for each Fitzpatrick level (I to VI) in the initial dataset, identifying a few skin tone categories with insufficient sample size; second, for skin tone categories with limited sample size (usually levels IV, V, and VI), it uses SMOTE oversampling technology to generate synthetic samples; the generation process of the synthetic samples is based on the nearest neighbor interpolation principle in the feature space, which is composed of high-level semantic features (dimension 512) of facial images extracted by the pre-trained VGG-16 network, and its mathematical expression is x new =x i +λ(x j -x i ), where x i It is a sample from the minority class, x j It is one of its randomly selected k nearest neighbors (k=5), and λ is a random number in the range [0,1]. Finally, the synthetic sample generated by oversampling is merged with the original sample to form a balanced training set with approximately equal number of samples for each skin color category.

[0010] Furthermore, the physiological signal feature extraction network employs a dual-branch architecture. The first branch is the spatial feature extraction branch, which uses a pre-trained ResNet-50 network as its backbone and is responsible for extracting high-level spatial features from each frame of facial images, focusing on the color and texture patterns of the skin region. The second branch is the temporal dynamic modeling branch, which uses a two-layer bidirectional long short-term memory network (Bi-LSTM). The input of this network is the feature sequence extracted by the spatial branch, and its hidden layer unit number is set to 128. Bi-LSTM captures the temporal dependence of subtle periodic changes in skin color caused by physiological signals (such as heart rate pulses) through its gating mechanism (input gate, forget gate, output gate).

[0011] Furthermore, the core of the skin color bias correction module is a metric learning component based on Maximum Mean Discrepancy (MMD). This component is activated after the bottleneck layer of the physiological signal feature extraction network. Its calculation process is as follows: First, the features output by the feature extraction network are grouped according to the skin color label of the samples; then, the MMD distance between the feature distributions of any two different skin color groups (e.g., group P and group Q) is calculated using the following formula:

[0012]

[0013] Where φ(·) passes through the Gaussian kernel function The implementation involves adaptively setting the bandwidth parameter σ to 0.5 based on the feature dimension and optimizing it on the validation set via grid search. Finally, the MMD distance is added as a regularization term to the model's total loss function, enabling the model to learn feature representations that are insensitive to skin color through backpropagation.

[0014] Preferably, the total loss function L defined by the multi-task learning and fusion module is... total It consists of three parts: L total =L regression +αL fairness +βL regularization Among them, L regularization This is the physiological parameter regression loss, calculated using the smoothed L1 loss function to determine the error between the model-predicted physiological signals (such as heart rate) and the true labels. fairness This is the fairness constraint loss, which is the sum of the MMD distances calculated by the aforementioned skin color bias correction module. regularization This is the weight decay regularization term, used to prevent the model from overfitting. Hyperparameters α and β are used to balance the importance of different loss terms, with typical values ​​ranging from [0.1, 1.0] to [1e-5, 1e-3], respectively. During training, gradient descent is used to optimize these three loss terms simultaneously.

[0015] Furthermore, in addition to SMOTE, the synthetic data generation technology also includes a skin-specific data augmentation unit based on a Generative Adversarial Network (GAN). This unit is specifically trained for a very small number of skin color categories (such as Fitzpatrick VI level). Its generator fuses the skin color category encoding with a noise vector through an embedding layer and uses a temporal convolutional layer to generate simulated physiological cycle signals (such as a sine wave corresponding to heart rate, with a frequency range of 60-100 bpm). This signal is weighted and superimposed with the skin color appearance texture in the RGB channels, outputting a facial video clip with the corresponding skin color appearance and containing simulated physiological cycle signals. The discriminator is responsible for distinguishing the generated video clip from the real video clip and ensuring that the generated physiological signals are consistent with the real signals in terms of frequency domain characteristics (such as the dominant frequency peak corresponding to heart rate). This GAN uses Wasserstein distance as the optimization objective and applies gradient penalties to improve training stability.

[0016] Furthermore, the fairness evaluation metrics used in the model validation and deployment module include: the mean absolute error (MAE) and root mean square error (RMSE) for each skin color subgroup, and the performance difference coefficient among skin colors across all samples. The difference coefficient is defined as the ratio of the standard deviation of the MAE for each skin color subgroup to the average MAE of all samples. During deployment, this module encapsulates the trained model parameters, network structure, and necessary preprocessing code into a standalone software library or microservice, providing a standardized application programming interface (API) for upper-layer health monitoring applications to call.

[0017] Compared with existing technologies, this invention has the following beneficial effects: 1. By systematically constructing a skin color balance training set and synthesizing data augmentation, the core problem of uneven skin color distribution in the training data of physiological measurement models is fundamentally solved, providing a sufficient data foundation for the model to learn cross-skin color robustness features. 2. By introducing a skin color bias correction mechanism and a multi-task learning framework based on MMD, the model feature representation is actively constrained at the algorithm level, effectively suppressing the implicit bias of the model towards specific skin color groups and significantly improving the measurement accuracy and reliability of the model on dark-skinned populations. 3. The dual-branch structure of the physiological signal feature extraction network can effectively capture the spatial skin color texture features and temporal dynamic change features related to physiological signals, thereby achieving more accurate modeling of weak physiological signals. 4. The entire method and system form a complete technical link from data preparation, model training to validation and deployment, ensuring the universality, fairness, and technical inclusivity of the finally deployed physiological measurement model in diverse real-world populations, meeting the urgent needs of the medical and health fields for highly reliable non-contact physiological monitoring. Attached Figure Description

[0018] Figure 1This is a schematic diagram of the overall technical architecture of the physiological measurement model training method and system for fairness proposed in this invention;

[0019] Figure 2 This is a schematic diagram of the core principle framework of skin color bias correction and multi-task learning in this invention;

[0020] Figure 3 This is a flowchart illustrating the logical flow of constructing the skin tone balance training set in this invention.

[0021] Figure 4 This is a schematic diagram of the dual-branch architecture of the physiological signal feature extraction network in this invention;

[0022] Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow between model verification and deployment in this invention; Detailed Implementation

[0023] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the specific embodiments according to the present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments.

[0024] Example 1

[0025] In remote medical health monitoring applications, particularly in non-contact physiological parameter measurement services for diverse global populations, the skin color perception data acquisition module is the first to undertake large-scale data collection. This module uses high-definition cameras deployed on smart devices to collect facial video data from individuals with various skin tones. Simultaneously, it utilizes medical-grade contact sensors to record the corresponding individuals' real physiological signal labels, including heart rate waveforms and blood oxygen saturation values. Each collected individual sample is automatically labeled with its skin color category according to the Fitzpatrick skin typing scale, which precisely classifies human skin color into six levels, from Level I (lightest) to Level VI (darkest), ensuring that each sample has a clear skin color attribute identifier. During the acquisition process, the module strictly controls environmental factors such as lighting conditions, shooting angle, and distance to guarantee the quality and consistency of the raw video data.

[0026] See Figure 1 After receiving the raw data from the data acquisition module, the skin tone equalization training set construction module begins data equalization processing. This module first performs statistical analysis on the number of samples for each Fitzpatrick level in the initial dataset, identifying a few skin tone categories with severely insufficient sample sizes, typically dark skin tones such as levels IV, V, and VI. For these skin tone categories with limited sample sizes, the module uses the SMOTE oversampling technique to generate synthetic samples. The generation process of synthetic samples is based on the nearest neighbor interpolation principle in feature space, mathematically expressed as x... new =xi +λ(x j -x i ), where x i It is a sample from the minority class, x j The sample is one of its randomly selected k-nearest neighbors, where k is fixed at 5 and λ is a random number between 0 and 1. This interpolation operation is performed in the feature space, ensuring that the generated synthetic samples retain the statistical properties of the original data while effectively increasing the sample size for a few skin color categories. Finally, the module merges the oversampled synthetic samples with the original samples to form a balanced training set with approximately equal sample sizes for each skin color category, providing a fair data foundation for subsequent model training.

[0027] like Figure 4 As shown, the physiological signal feature extraction network employs a carefully designed dual-branch architecture to process input facial video sequences. The first branch is the spatial feature extraction branch, using a ResNet-50 network pre-trained on a large image dataset as its backbone. This branch is specifically responsible for extracting high-level spatial features from each frame of the facial image. During feature extraction, the network focuses on subtle color changes and texture patterns in the skin region through its deep convolutional layers. These features are closely related to changes in blood flow in subcutaneous vessels. The second branch is the temporal dynamic modeling branch, constructed using a two-layer bidirectional long short-term memory network with a strictly set number of hidden layer units of 128. The temporal branch receives the feature sequence extracted by the spatial branch as input and, through its sophisticated gating mechanism including input gate, forget gate, and output gate, specifically captures the long-term dependence of subtle periodic changes in skin color caused by physiological signals in the temporal dimension. The features from the two branches are fused in the deep layers of the network to form a comprehensive physiological signal feature representation that includes both spatial details and temporal dynamics.

[0028] See Figure 2 The skin color bias correction module is integrated into the training process of the physiological signal feature extraction network. Its core is a metric learning component based on the maximum mean difference. This component is activated after the bottleneck layer of the feature extraction network. First, it precisely groups the network's output feature representations according to the skin color labels of the samples, forming feature distribution sets for different skin color groups. Next, the module calculates the maximum mean difference distance between any two different skin color group feature distributions, using the following formula: Where φ(·) is a nonlinear mapping function that maps features to the regenerating kernel Hilbert space. In practice, this mapping process is approximated by a Gaussian kernel function. This distance metric directly reflects the degree of distributional difference between different skin color groups in the model's feature space. The module adds this distance as a regularization term to the model's total loss function. Through optimization of the backpropagation process, it enables the feature extraction network to learn robust feature representations that are insensitive to skin color factors.

[0029] The multi-task learning and fusion module is responsible for coordinating the joint optimization of the physiological parameter regression task and the fairness constraint task. The total loss function defined by this module consists of three key parts, mathematically expressed as L... total =L regression +αL fairness +βL regularization The regression loss term uses a smoothed L1 loss function to calculate the error between the model's predicted physiological parameter values ​​and the true labels, ensuring the model's basic accuracy in physiological signal measurement. The fairness constraint loss term is directly derived from the sum of the maximum mean difference distances calculated by the skin color bias correction module, used to force the model to maintain consistent performance across different skin color groups. The regularization term uses a standard weight decay form to prevent overfitting during training. Hyperparameters α and β are used to balance the relative importance of fairness and regularization constraints, respectively. In actual training, the typical value range for α is 0.1 to 1.0, and the value range for β is 1e-5 to 1e-3. During training, the module uses a momentum-based stochastic gradient descent algorithm to simultaneously optimize these three losses, ensuring that the model meets fairness requirements while maintaining high accuracy.

[0030] For a very small number of skin color categories, such as Fitzpatrick VI, the system also includes a skin color-specific data augmentation unit based on a generative adversarial network (GAN). The generator network of this unit receives a Gaussian-distributed random noise vector and a specified skin color category code, and outputs a facial video clip with the corresponding skin color appearance and simulated physiological cycle signals through multiple deconvolution operations. The discriminator network is responsible for distinguishing the generated video clips from real video clips, and simultaneously ensuring, through frequency domain analysis, that the generated physiological signals are consistent with the real signals in terms of frequency domain characteristics, especially that the peak position and amplitude of the dominant frequency corresponding to heart rate must conform to physiological laws. This GAN uses Wasserstein distance as the optimization objective and applies a gradient penalty term to improve the stability of the training process, ensuring that the generated data is both visually realistic and contains correct physiological signal features.

[0031] like Figure 5As shown, the model validation and deployment module comprehensively evaluates the trained model on a cross-skin color test set independent of the training set. The fairness evaluation system used in this module includes the mean absolute error (MAO), root mean square error (RMSE), and the performance difference coefficient between skin colors across all samples. The difference coefficient is defined as the ratio of the standard deviation of the MAO for each skin color subgroup to the MAO for the entire sample; this metric quantifies the performance fluctuation of the model across different skin color groups. During the deployment phase, the module encapsulates the trained model parameters, network structure definitions, and necessary data preprocessing code into an independent software library, providing standardized application programming interfaces (APIs). These interfaces support both real-time video stream input and batch data processing modes, with output formats including heart rate values, blood oxygen saturation percentage, and corresponding confidence scores, which can be directly called by upper-level health monitoring applications.

[0032] In the specific training process, the system executes a fairness-oriented physiological measurement model training method, which includes seven rigorously ordered steps. Step S110 first collects facial video data and synchronous physiological signals from individuals with multiple skin colors, and assigns a precise skin color category label to each sample according to the Fitzpatrick scale. Step S120 performs skin color category statistical analysis on the collected raw dataset, and applies oversampling and synthetic sample generation algorithms to construct a training dataset with a balanced number of samples for each skin color category. Step S130 constructs a hybrid model of deep convolutional neural network and recurrent neural network as a physiological signal feature extractor. This network takes video frame sequences as input and outputs the corresponding temporal features of physiological signals. Step S140, based on the feature extraction network, introduces a skin color bias corrector based on the maximum mean difference. By measuring and reducing the distribution distance of different skin color groups in the feature space, it suppresses the model's dependence on a specific skin color. Step S150 designs a multi-task learning framework, simultaneously optimizing the main task loss for physiological parameter prediction and the auxiliary task loss for skin color fairness constraints, and achieving collaborative training of the two through a gradient inversion mechanism. Step S160 uses the constructed balanced training set to train the multi-task model end-to-end until the model's overall performance on the validation set converges. Step S170 comprehensively evaluates the heart rate estimation error, blood oxygen saturation measurement accuracy, and cross-skin color group performance standard deviation of the trained model on an independent test set containing unseen dark and light skin samples. After ensuring that the model meets the requirements of fairness and reliability, it is deployed in practice.

[0033] Throughout the training process, the system rigorously monitors and records the execution details of each step. In the data preprocessing stage, all facial videos undergo standardization, including face detection alignment, illumination normalization, and resolution unification. In the feature extraction stage, network parameters are initialized using the He initialization method, and the activation function uses ReLU and its variants. Training optimization employs a phased learning rate scheduling strategy, with an initial learning rate set to 0.001, decaying by a factor of 0.5 every 50 training epochs. The batch size is dynamically adjusted based on available GPU memory, typically set to 32 or 64. An early training termination mechanism is implemented based on comprehensive performance metrics on the validation set; training automatically terminates when there is no significant performance improvement after 20 consecutive training epochs. Model compression and acceleration techniques include weight pruning and quantization, ensuring that the final deployed model meets both accuracy requirements and real-time inference capabilities.

[0034] Example 2

[0035] In the quality control laboratory environment of a mobile health monitoring device manufacturer, a fairness-oriented physiological measurement model training system is used to develop and optimize contactless health monitoring functions embedded in smartphones and smartwatches. The skin color perception data acquisition module specifically collects facial video data captured in various real-world usage environments, including samples under different lighting conditions, shooting angles, and backgrounds, ensuring the diversity and representativeness of the training data. The module also records metadata for each sample, such as device model, sensor specifications, and environmental parameters, supporting subsequent analysis of the model's generalization ability.

[0036] See Figure 3 In this embodiment, the skin tone equalization training set construction module employs a more refined data equalization strategy. In addition to the basic SMOTE oversampling technique, the module introduces a clustering-based sample generation method. First, it performs cluster analysis on samples within each skin tone category to identify sparse regions in the feature space. Then, it selectively generates synthetic samples in these regions, ensuring that the synthetic samples effectively cover blind spots in the feature space. The module also establishes a data quality assessment mechanism, performing multiple verifications on the generated synthetic samples, including visual realism checks, physiological signal rationality verification, and feature distribution consistency detection. Only synthetic samples that pass all verifications are added to the final training dataset.

[0037] In this embodiment, the physiological signal feature extraction network is optimized for the computational resource constraints of mobile devices. The spatial feature extraction branch uses the lightweight MobileNetV3 architecture instead of the standard ResNet-50, significantly reducing computation and the number of parameters while maintaining feature extraction capabilities. The temporal dynamic modeling branch uses gated recurrent units instead of long short-term memory networks, achieving higher computational efficiency with similar performance. The network also introduces an attention mechanism, allowing the model to adaptively focus on the most physiologically relevant parts of the facial region, such as the cheekbone and forehead regions, ignoring background and irrelevant features.

[0038] In this embodiment, the skin color bias correction module expands the scope of fairness constraints. In addition to the distribution alignment constraint based on the maximum mean difference, the module introduces an independence measure based on mutual information, requiring the minimization of mutual information between the feature representation learned by the model and the skin color category label. This further ensures, from an information theory perspective, that the feature representation is insensitive to skin color factors. The module also implements a dynamic weight adjustment mechanism, adaptively adjusting the strength of the fairness constraint terms based on changes in performance differences among different skin color groups during training, applying varying degrees of fairness guidance at different stages of model training.

[0039] In this embodiment, the multi-task learning and fusion module explores more advanced optimization strategies. The module employs a gradient surgery algorithm to handle conflicts between loss gradients from different tasks, eliminating gradient components that hinder common progress by projecting the gradient directions. The module also implements an uncertainty weighting mechanism, automatically learning the optimal weights for each loss term, avoiding the subjectivity and time-consuming nature of manually adjusting hyperparameters. During the gradient fusion stage, the module uses adaptive gradient pruning to prevent gradient explosion during training, ensuring numerical stability during the training process.

[0040] In this embodiment, the model validation and deployment module establishes a comprehensive model lifecycle management system. The module implements a continuous performance monitoring mechanism, periodically evaluating the performance of deployed models on newly added cross-skin color test data and detecting potential performance degradation or fairness deterioration. The module also establishes a model update pipeline, automatically triggering a model retraining process when performance degradation is detected, using new data for incremental learning to ensure the model remains in optimal condition. The deployment solution supports multiple hardware platforms, including edge inference on mobile devices and centralized processing on cloud servers, providing a flexible solution to adapt to different application needs.

[0041] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A physiological measurement model training system oriented towards fairness, characterized in that, The system includes the following components: The skin color perception data acquisition module is used to acquire raw video data and corresponding real physiological signal labels of individuals with multiple skin colors, and automatically label each individual sample with its skin color category label. The skin color category is divided into 6 levels according to the Fitzpatrick skin typing scale. The skin color balance training set construction module is connected to the skin color perception data acquisition module. It is used to receive raw data and adjust the sample distribution of each skin color category in the dataset through resampling and synthetic data generation technology to meet the preset balance ratio. The physiological signal feature extraction network is connected to the skin color balance training set construction module. It is used to extract the temporal feature representation related to physiological signals from the input balance training set video frame sequence. The skin color bias correction module, integrated into the training process of the physiological signal feature extraction network, is used to calculate and minimize the distributional differences in the model's response to features of different skin color categories. The multi-task learning and fusion module, connected to the physiological signal feature extraction network and the skin color bias correction module, is used to jointly optimize the physiological parameter regression loss and the skin color fairness constraint loss, and fuse their gradient information to update the network parameters. The model validation and deployment module is used to evaluate the performance of the trained model on a cross-skin color test set independent of the training set, ensuring that it achieves the preset accuracy and fairness indicators across all skin color groups, and deploying the qualified model to the real-world application environment.

2. The fairness-oriented physiological measurement model training system according to claim 1, characterized in that, The skin tone equalization training set construction module performs the following operations: First, it counts the number of samples for each Fitzpatrick level in the initial dataset and identifies a few skin tone categories with insufficient sample size; second, for skin tone categories with limited sample size, it uses the SMOTE oversampling technique to generate synthetic samples; the generation process of the synthetic samples is based on the nearest neighbor interpolation principle in the feature space, which is composed of high-level semantic features (dimension 512) of facial images extracted by the pre-trained VGG-16 network, and its mathematical expression is x. new =x i +λ(x j -x i ), where x i It is a sample from the minority class, x j It is one of its randomly selected k nearest neighbors, where k is 5 and λ is a random number in the range [0,1]. Finally, the synthetic sample generated by oversampling is merged with the original sample to form a balanced training set in which the number of samples of each skin color category is approximately equal.

3. The fairness-oriented physiological measurement model training system according to claim 1, characterized in that, The physiological signal feature extraction network adopts a dual-branch architecture. The first branch is the spatial feature extraction branch, which uses a pre-trained ResNet-50 network as the backbone and is responsible for extracting high-level spatial features from each frame of facial image, focusing on the color and texture patterns of the skin region. The second branch is the temporal dynamic modeling branch, which uses a two-layer bidirectional long short-term memory network. The input of this network is the feature sequence extracted by the spatial branch, and its hidden layer unit number is set to 128. Through its gating mechanism, it captures the temporal dependence of the small periodic changes in skin color caused by physiological signals.

4. The fairness-oriented physiological measurement model training system according to claim 1, characterized in that, The core of the skin color bias correction module is a metric learning component based on the maximum mean difference. This component is activated after the bottleneck layer of the physiological signal feature extraction network. Its calculation process is as follows: First, the features output by the feature extraction network are grouped according to the skin color label of the samples. Then, the maximum mean difference distance between the feature distributions of any two different skin color groups is calculated using the following formula: Where φ(·) is a function that maps features to the regenerating kernel Hilbert space; finally, the maximum mean difference distance is added as a regularization term to the model's total loss function.

5. The fairness-oriented physiological measurement model training system according to claim 1, characterized in that, The total loss function L defined by the multi-task learning and fusion module total It consists of three parts: L total =L regression +αL fairness +βL regularization Among them, L regularization It is a physiological parameter regression loss, which uses a smoothed L1 loss function to calculate the error between the physiological signal predicted by the model and the true label; L fairness It is the fairness constraint loss, which is the sum of the distances between the maximum mean differences calculated by the skin color bias correction module; L regularization It is the weight decay regularization term; the hyperparameters α and β are used to balance the importance of different loss terms, and their typical values ​​are [0.1, 1.0] and [1e-5, 1e-3], respectively.

6. The fairness-oriented physiological measurement model training system according to claim 2, characterized in that, The synthetic data generation technique also includes a skin-specific data augmentation unit based on a generative adversarial network; this unit is specifically trained for a very small number of skin color categories; Its generator fuses skin color category encoding with noise vector through an embedding layer, and uses a temporal convolutional layer to generate simulated physiological cycle signals (such as a sine wave corresponding to heart rate, with a frequency range of 60-100 bpm). This signal is weighted and superimposed with the skin color appearance texture in the RGB channels, and outputs a facial video clip with the corresponding skin color appearance and containing simulated physiological cycle signals. The discriminator is responsible for distinguishing the generated video clip from the real video clip and ensuring that the generated physiological signal is consistent with the real signal in the frequency domain.

7. The fairness-oriented physiological measurement model training system according to claim 1, characterized in that, The fairness evaluation metrics used in the model validation and deployment module include: the mean absolute error, root mean square error, and performance difference coefficient among skin colors for each skin color subgroup; the difference coefficient is defined as the ratio of the standard deviation of the mean absolute error of each skin color subgroup to the mean absolute error of the entire dataset; during deployment, the module encapsulates the trained model parameters, network structure, and necessary preprocessing code into an independent software library or microservice, providing a standardized application programming interface.

8. A method for training a physiological measurement model oriented towards fairness, characterized in that, The method includes the following steps: S110: Collect facial video data and synchronized physiological signals from individuals with multiple skin colors, and assign skin color category labels to each sample according to the Fitzpatrick scale; S120: Perform skin color category statistical analysis on the collected raw dataset, and apply oversampling and synthetic sample generation algorithms to construct a training dataset with a balanced number of samples for each skin color category; S130: Construct a hybrid model of deep convolutional neural network and recurrent neural network as a physiological signal feature extractor. This network takes video frame sequences as input and outputs the corresponding temporal features of physiological signals; S140: Based on the feature extraction network, introduce a skin color bias corrector based on maximum mean difference. This corrector measures and reduces the bias of different skin color groups. The distribution distance of the bodies in the feature space is used to suppress the model's dependence on specific skin colors; S150, a multi-task learning framework is designed, which simultaneously optimizes the main task loss for physiological parameter prediction and the auxiliary task loss for skin color fairness constraints, and achieves co-training of the two through gradient reversal layers or weight sharing mechanisms; S160, the above multi-task model is trained end-to-end using the constructed balanced training set until the model's overall performance on the validation set converges; S170, on an independent test set containing unseen dark and light skin color samples, the heart rate estimation error, blood oxygen saturation measurement accuracy, and performance standard deviation across skin color groups of the trained model are comprehensively evaluated to ensure that the model meets the requirements of fairness and reliability before actual deployment.

9. The fairness-oriented physiological measurement model training method according to claim 8, characterized in that, In step S120, the application of the oversampling and synthetic sample generation algorithm specifically includes: counting the number of samples for each Fitzpatrick level in the initial dataset, and identifying a few skin color categories with insufficient sample size; for skin color categories with a small sample size, using the SMOTE oversampling technique to generate synthetic samples; the generation process of the synthetic samples is based on the nearest neighbor interpolation principle in the feature space, which is composed of high-level semantic features (dimension 512) of facial images extracted by the pre-trained VGG-16 network, and its mathematical expression is x new =x i +λ(x j -x i ), where x i It is a sample from the minority class, x j It is one of its randomly selected k nearest neighbors, where k is 5 and λ is a random number in the range [0,1]. The synthetic sample generated by oversampling is merged with the original sample to form a balanced training set.

10. The fairness-oriented physiological measurement model training method according to claim 8, characterized in that, In step S150, the total loss function L defined by the multi-task learning framework total It consists of three parts: L total =L regression +αL fairness +βL regularization The loss is a regression loss based on physiological parameters, using a smoothed L1 loss function; L fairness It is the fairness constraint loss, which is the sum of the skin color fairness constraint losses calculated based on the distance of the maximum mean difference; L regularization It is the weight decay regularization term; hyperparameters α and β are used to balance the importance of different loss terms.

Images