Pulse blood oxygen saturation degree detection method based on image processing

Abstract

The invention relates to the technical field of blood oxygen detection, in particular to a pulse blood oxygen saturation detection method based on image processing, a transfer learning strategy and a domain adaptation network are introduced, adjustment can be carried out according to personalized physiological parameters of a user, and therefore the applicability to different crowds is improved. Especially for people with deep skin color, the measuring error is obviously reduced, and the error rate is reduced by more than 30%. A set of mechanism for calculating the current signal quality index Q in real time is designed, dynamic calibration is carried out in combination with a Bayesian filtering model, and it is ensured that a reliable blood oxygen saturation degree estimation value can be obtained even under the condition of the low signal-to-noise ratio. And when the signal quality is reduced, the current detection value is compensated by using historical data through the LSTM network prediction module, so that the accuracy and continuity of the result are further ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of blood oxygen detection, and in particular to a pulse blood oxygen saturation detection method based on image processing. Background Art

[0002] Currently, traditional blood oxygen saturation measurement relies primarily on photoplethysmography (PPG), which uses a finger-clip probe to measure changes in red and infrared light absorption. Typical examples include Masimo's Signal Extraction Technology (patent US20150351697A1) and the Apple Watch's optical sensor solution (patent US20200323435A1).

[0003] However, there are three major defects in the above-mentioned prior art:

[0004] 1. Limitations of contact measurement: Probe pressure leads to poor comfort, and long-term monitoring can easily cause skin damage.

[0005] 2. Insufficient anti-interference ability: Motion artifacts and ambient light noise seriously affect measurement accuracy (see IEEE TBME 2021 research data, the error can reach ±5%).

[0006] 3. Lack of personalized adaptation: Individual differences such as skin color and thickness of the subcutaneous fat layer are not taken into account (Nature Biomedical Engineering 2022 pointed out that the error increases by 30% in people with dark skin).

[0007] Therefore, a pulse oxygen saturation detection method based on image processing that can solve the above problems is needed. Summary of the Invention

[0008] This invention provides a pulse oximetry detection method based on image processing. By incorporating a transfer learning strategy and a domain adaptation network, the method can be adjusted based on the user's individual physiological parameters (such as skin color, age, and body fat percentage), thereby improving its applicability to different populations. In particular, for dark-skinned individuals, the invention significantly reduces measurement errors, reducing the error rate by over 30%.

[0009] The present invention solves the above-mentioned technical problems by adopting a technical solution: a pulse oxygen saturation detection method based on image processing, comprising the following steps:

[0010] S1. Capture the fingertip video stream of the subject through a multispectral imaging device and simultaneously obtain the initial blood oxygen saturation baseline value;

[0011] S2. performing spatiotemporal preprocessing on the fingertip video stream, including: using adaptive weighted guided filtering to eliminate ambient light interference, and constructing a motion artifact compensation model to eliminate micro-motion interference;

[0012] S3. Extracting dual-channel feature parameters from the preprocessed video, including: extracting the time-frequency features of the PPG signal through wavelet transform, and extracting the spatial features of the skin microvascular distribution through principal component analysis;

[0013] S4. Build a two-stream convolutional neural network model, where the first branch uses ResNet-50 to extract time-frequency features, and the second branch uses U-Net to extract spatial features, and the two-branch features are fused through the attention mechanism;

[0014] S5. Calculate the current signal quality index Q in real time. The Q value is generated by fusing three parameters: signal-to-noise ratio, waveform periodicity, and skin transmittance, using the entropy weight method.

[0015] S6. Establish a dynamic calibration module to input the initial blood oxygen saturation baseline value, real-time Q value and neural network output value into the Bayesian filter model for iterative optimization;

[0016] S7. When the signal quality index Q is lower than the threshold, the time series prediction module based on the LSTM network is triggered to compensate the current detection value using historical data;

[0017] S8. Generate a contribution distribution diagram of each feature parameter through the interpretability analysis module and dynamically adjust the weight coefficient of each input parameter;

[0018] S9, adopting a transfer learning strategy, using personalized physiological parameters to build a domain adaptation network and achieve online fine-tuning of model parameters;

[0019] S10. Output the final blood oxygen saturation value and simultaneously generate graded prompt information including a confidence interval and abnormal fluctuation warning.

[0020] Furthermore, the motion artifact compensation model in step S2 specifically includes:

[0021] S21. Construct a hand motion trajectory prediction sub-model based on the inertial measurement unit;

[0022] S22, establishing a spatiotemporal mapping relationship between the video frame sequence and the motion trajectory;

[0023] S23, using the optical flow method to calculate the motion vector field, and combining the trajectory prediction results to perform reverse compensation;

[0024] S24. Verify the quality of the compensated video sequence through convolutional recurrent network.

[0025] Furthermore, the method for establishing the attention mechanism fusion in step S4 specifically includes:

[0026] S41. Perform channel attention weighting on the time-frequency feature map to generate a feature heat map;

[0027] S42, performing spatial attention weighting on the spatial feature map to generate a regional attention distribution map;

[0028] S43, dynamically fuse the dual-channel features through the gating mechanism, and the fusion weight is dynamically adjusted by the real-time Q value. Further, the specific method for establishing the Bayesian filter model in step S6 is:

[0029] S61. Establish a state space model including equipment error terms and physiological fluctuation terms;

[0030] S62, using the neural network output as the observation value and the initial reference value as the prior distribution;

[0031] S63, using the Q value to adjust the process noise covariance matrix;

[0032] S64. Use particle filtering algorithm to estimate the posterior probability.

[0033] Furthermore, the method for constructing the time series prediction module in step S7 specifically includes:

[0034] S71. Build a bidirectional LSTM network architecture with an input window containing the time series features of the previous 60 seconds.

[0035] S72. Introduce temporal convolutional networks to extract multi-scale temporal patterns;

[0036] S73. Enhance the robustness of prediction results through adversarial training strategies;

[0037] S74. Design a residual connection structure to compensate for prediction errors.

[0038] Furthermore, the method for establishing the interpretability analysis module in step S8 includes the following steps:

[0039] S81. Calculate the contribution of the input feature using the integral gradient method;

[0040] S82, locating feature regions through class activation mapping;

[0041] S83, establishing dynamic association rules between contribution and Q value;

[0042] S84. Automatically generate model optimization suggestions based on contribution distribution.

[0043] Furthermore, the method for constructing the domain adaptation network in step S9 specifically includes:

[0044] S91. Build a domain conversion module based on generative adversarial network;

[0045] S92. Design a personalized feature encoder including skin color, age, and body fat percentage;

[0046] S93, using meta-learning strategy to achieve rapid sample adaptation;

[0047] S94. Enhancing cross-domain feature representation capabilities through contrastive learning.

[0048] Furthermore, the generation of the graded prompt information in step S10 includes:

[0049] S101. Establish a multi-level warning threshold system based on clinical guidelines;

[0050] S102. Identify abnormal conditions by combining trend analysis and fluctuation patterns;

[0051] S103, generating a dynamic confidence interval including time evolution characteristics;

[0052] S104. A multimodal warning method integrating voice, vision and touch.

[0053] The advantages of the present invention are:

[0054] 1. This method uses a non-contact multispectral imaging device to capture fingertip video streams without direct skin contact, significantly improving user comfort and making it suitable for long-term continuous monitoring. Compared to traditional finger-clip probes, this method avoids the risk of skin damage caused by compression.

[0055] 2. The present invention eliminates ambient light interference through adaptive weighted guided filtering and constructs a motion artifact compensation model to eliminate micro-motion interference, significantly improving the detection stability and accuracy in complex environments.

[0056] 3. This invention incorporates a transfer learning strategy and domain adaptation network, which can be adjusted based on the user's individual physiological parameters (such as skin color, age, body fat percentage, etc.), thereby improving its applicability to different populations. In particular, for dark-skinned individuals, this invention significantly reduces measurement errors, reducing the error rate by over 30%.

[0057] 4. This invention incorporates a mechanism for real-time calculation of the current signal quality index (Q), combined with a Bayesian filter model for dynamic calibration, ensuring reliable blood oxygen saturation estimates even under low signal-to-noise ratio conditions. When signal quality degrades, the LSTM network prediction module leverages historical data to compensate for the current measurement, further ensuring the accuracy and consistency of the results.

[0058] 5. To facilitate understanding and improve algorithm performance, the present invention uses an interpretable analysis module to generate a contribution distribution diagram for each feature parameter and automatically adjust the input parameter weights accordingly. This transparent processing approach helps quickly identify problems and optimize the model, thereby continuously improving the overall performance of the system.

[0059] 6. This invention not only provides accurate blood oxygen saturation numerical output, but also simultaneously generates graded prompt information including confidence intervals and abnormal fluctuation warnings. Users can not only understand their health status, but also receive timely warnings of potential risks. DETAILED DESCRIPTION

[0060] The following is a clear and complete description of the technical solution of the present invention. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0061] Example 1:

[0062] The present invention provides a pulse oxygen saturation detection method based on image processing.

[0063] A pulse oximetry detection method based on image processing comprises the following steps:

[0064] S1. A multispectral imaging device is used to capture the subject's fingertip video stream and simultaneously obtain the initial blood oxygen saturation baseline value. Specifically, the multispectral imaging device uses a customized multispectral camera (see Table 1 for a list of components), which includes: 8 sets of narrow-band LED arrays (center wavelength: 530nm / 660nm / 810nm / 940nm); a global shutter CMOS sensor (frame rate 60fps, resolution 1280×960); and a six-axis IMU (MPU-6050, sampling rate 200Hz).

[0065] Table 1 Parameters of multispectral imaging device (multispectral camera)

[0066]

[0067] S2. Preprocessing the fingertip video stream in the spatiotemporal domain includes: using adaptive weighted guided filtering to eliminate ambient light interference and building a motion artifact compensation model to eliminate micro-motion interference. Specifically, the adaptive weighted guided filtering includes:

[0068] The light intensity distribution model is established according to the following formula: (x, y, t) = α(t)·R(x, y, t) + β(t·L env(x, y, t), where R is the reflection component, which represents the intensity of light reflected from the object surface at position x, y and time t, and L env The ambient light component represents the intensity of background or scattered light in the environment at position x, y and time t. The weight coefficients α and β are solved through iterative optimization. x is the horizontal coordinate, and y is the vertical coordinate. These two variables describe the position of a point in an image or scene. The variable t represents time. Specifically, t represents a specific moment in time. In a video stream, lighting conditions may change over time, so the time variable t is introduced to describe the light intensity distribution at different points in time. This method can determine the most appropriate weight coefficients α and β at different times and locations, thereby more accurately modeling the light intensity distribution.

[0069] Furthermore, the motion artifact compensation model in step S2 specifically includes:

[0070] S21. Construct a hand motion trajectory prediction sub-model based on an inertial measurement unit, specifically including:

[0071] S211. Data Collection: An inertial measurement unit (IMU) device (such as an accelerometer or gyroscope) is attached to the subject's hand to record hand motion data in real time. This data includes parameters such as acceleration and angular velocity, which are used to predict the hand's motion trajectory.

[0072] S212, Feature Extraction: Extract key features from the IMU data, such as the rate of change of acceleration and the trend of angular velocity. These features will be used as input to train the hand motion trajectory prediction model.

[0073] S213, Model training: Use machine learning algorithms (such as long short-term memory network LSTM or convolutional neural network CNN) to train the extracted features and build a sub-model that can predict hand movement trajectory.

[0074] S214, prediction output: In actual applications, the trained model is used to predict the hand movement trajectory in the future based on the current IMU data.

[0075] S22: Establishing a spatiotemporal mapping relationship between the video frame sequence and the motion trajectory, specifically including:

[0076] S223, Time Synchronization: Ensure that the timestamps between the IMU data and the video frames are fully synchronized so that each frame of video accurately corresponds to the corresponding hand motion state.

[0077] S224, Spatial Mapping: Based on the hand motion trajectory prediction results, calculate the fingertip position change corresponding to each frame of video. Spatiotemporal Mapping Matrix: Construct a spatiotemporal mapping matrix that records the offset of the fingertip position relative to the reference position in each frame of video. This matrix is used to adjust the video frame to make it more consistent with the actual fingertip position.

[0078] S23. Calculate the motion vector field using the optical flow method and perform reverse compensation based on the trajectory prediction result. Specifically, the following steps are performed:

[0079] S231. Optical flow calculation: Use the optical flow method to calculate the motion vector field between adjacent frames and determine the moving direction and speed of the pixels in each frame. The optical flow method can be implemented using the Lucas-Kanade algorithm or the Farneback algorithm.

[0080] S232, motion vector correction: Combine the motion vector field calculated by the optical flow method with the hand motion trajectory prediction result. For each pixel, calculate the reverse displacement that should be compensated based on its motion vector and the predicted motion trajectory.

[0081] S233, reverse compensation: Apply a reverse compensation algorithm to adjust the pixels in each frame of video according to the calculated reverse displacement, thereby eliminating artifacts caused by hand movement.

[0082] S24. Verify the quality of the compensated video sequence through a convolutional recurrent network, specifically including:

[0083] S241. Construct a Convolutional Recurrent Network (ConvRNN): Design a convolutional recurrent network (e.g., a combination of convolutional and LSTM layers) to evaluate the quality of compensated video sequences. The input is the compensated video frame sequence, and the output is a quality score.

[0084] S242. Training the network: Use high-quality (no motion artifacts, high definition, and uniform lighting) video frame sequences as positive samples and low-quality (artifacts, low definition, and uneven lighting) video frame sequences as negative samples to train a convolutional recurrent network. The goal is to enable the network to distinguish between high-quality and low-quality video frame sequences and assign corresponding scores.

[0085] S243. Quality Assessment: In practical applications, the compensated video frame sequence is fed into a trained convolutional recurrent network to obtain a quality score. If the score is below a pre-set threshold, the compensation is considered ineffective and further optimization of the compensation algorithm is required. If the score is above the pre-set threshold, the compensation is considered effective and further processing can proceed.

[0086] S3. Extracting dual-channel feature parameters from the preprocessed video, including: extracting the time-frequency features of the PPG signal through wavelet transform, and extracting the spatial features of the skin microvascular distribution through principal component analysis. Specifically, extracting the time-frequency features of the PPG signal through wavelet transform includes:

[0087] S31. Data preparation: extracting an image sequence of the fingertip area from the preprocessed video frame sequence.

[0088] S32. Wavelet Transform: Use Morlet wavelets or other suitable wavelet basis functions to perform time-frequency analysis on the image sequence and extract the time-frequency characteristics of the PPG signal. These characteristics include frequency components and amplitude variations. Set the frequency range to 0.5Hz to 3Hz, use the continuous wavelet transform (CWT) to extract these frequency components, and generate a time-frequency plot.

[0089] The spatial characteristics of skin microvascular distribution extracted by principal component analysis include:

[0090] S33. Data preparation: similarly extract an image sequence of the fingertip area from the preprocessed video frame sequence.

[0091] S34. PCA analysis: The principal component analysis (PCA) method is applied to extract the main spatial features in the image sequence, which can reflect the distribution of skin microvessels.

[0092] S4. Construct a two-stream convolutional neural network model, in which the first branch uses ResNet-50 to extract time-frequency features, uses ResNet-50 as the basic network structure, and inputs the time-frequency map generated after wavelet transform. Then, the ResNet-50 model is pre-trained on a large-scale dataset and fine-tuned for specific tasks. Generate time-frequency feature maps. The second branch uses U-Net to extract spatial features, uses U-Net as the basic network structure, and inputs the spatial feature map extracted by principal component analysis (PCA). Pre-train the U-Net model on a large-scale dataset, fine-tune it for specific tasks, and generate spatial feature maps. Then, the dual-branch features are fused through the attention mechanism, where the method for establishing the attention mechanism fusion specifically includes:

[0093] S41. Perform channel attention weighting on the time-frequency feature map to generate a feature heat map;

[0094] S42, performing spatial attention weighting on the spatial feature map to generate a regional attention distribution map;

[0095] S43. Dynamically fuse dual-channel features through a gating mechanism, and the fusion weight is dynamically adjusted by the real-time Q value.

[0096] S5. Calculate the current signal quality index Q in real time. The Q value is generated by fusing three parameters, signal-to-noise ratio, waveform periodicity, and skin transmittance, using the entropy weight method. Specifically, it includes:

[0097] S51. Calculate the signal-to-noise ratio (SNR) using the following formula: Among them, P signal is the signal power, P noise is the noise power. By performing frequency domain analysis on the signal, the signal part and the noise part are separated and their power ratio is calculated.

[0098] S52. Calculate waveform periodicity: determine the main period of the waveform through the existing autocorrelation function or Fourier transform.

[0099] S53. Calculate skin transmittance: Calculate the transmittance at different wavelengths based on the data collected by the multispectral imaging device; use different wavelengths of the narrowband LED array (such as 530nm, 660nm, 810nm, 940nm) to calculate the transmittance at each wavelength.

[0100] S54, Entropy Weight Fusion: Use the entropy weighting method to perform a weighted fusion of the three parameters of signal-to-noise ratio, waveform periodicity, and skin transmittance to obtain the final Q value. The entropy value of each parameter is calculated, and then its weight is determined. Finally, the weighted sum is used to obtain the Q value.

[0101] S6. Establish a dynamic calibration module, input the initial blood oxygen saturation baseline value, real-time Q value and neural network output value into the Bayesian filter model for iterative optimization, wherein the specific method for establishing the Bayesian filter model is as follows:

[0102] S61. Establish a state space model including equipment error terms and physiological fluctuation terms. Specifically, the formula is: t =A xt-1 +B ut +w t , where x t is the state vector, A is the state transfer matrix, B is the control input matrix, ut is the control input vector, and wt is the process noise.

[0103] S62. Use the neural network output as the observation value and the initial reference value as the prior distribution.

[0104] S63. Utilize the Q value to adjust the process noise covariance matrix, and dynamically adjust the noise matrix according to the change of the Q value to reflect the change of the signal quality.

[0105] S64. Use the particle filter algorithm to estimate the posterior probability, and obtain the optimal state estimation value by resampling and updating particles.

[0106] S7. When the signal quality index Q is lower than the threshold, trigger the time series prediction module based on the LSTM network and use historical data to compensate the current detection value. The method for constructing the time series prediction module in step S7 specifically includes:

[0107] S71. Build a bidirectional LSTM network architecture with an input window containing the time series features of the previous 60 seconds.

[0108] S72. Introduce temporal convolutional networks to extract multi-scale temporal patterns;

[0109] S73. Enhance the robustness of prediction results through adversarial training strategies;

[0110] S74. Design a residual connection structure to compensate for prediction errors. Specifically, it includes setting up a basic bidirectional LSTM (Bi-LSTM) unit to process the input time series data. Set the input time series length to T and the feature dimension to d. Implement it through Python:

[0111]

[0112]

[0113] Combining the above components, we can build a complete time series prediction model, which includes a bidirectional LSTM layer, residual connections, and a final prediction layer. In deep networks, the gradient can become very small during backpropagation, making it difficult to effectively update the parameters of earlier layers. By introducing residual connections, the gradient can be directly propagated to earlier layers, alleviating the vanishing gradient problem.

[0114] S8. Generate a contribution distribution diagram for each feature parameter through the interpretability analysis module. The feature parameters include: physiological parameters: heart rate, blood pressure, respiratory rate, body temperature; environmental factors: temperature, humidity; signal quality related indicators: signal strength, noise level, signal-to-noise ratio; time series data: data collection timestamp, periodic changes in the data (such as daily or weekly change patterns), and then dynamically adjust the weight coefficient of each input parameter. Specifically, the following steps are included:

[0115] S81. Use the integrated gradient method to calculate the contribution of each input feature to the model output. By differentiating the input feature, calculate its impact on the output.

[0116] S82. Locate Feature Regions Using Class Activation Mapping: Use Class Activation Mapping (CAM) to locate the feature regions that have the greatest impact on the classification results. Generate a class activation map by taking a weighted average of the convolutional layer outputs.

[0117] S83. Establish dynamic association rules between contribution and Q value: Based on the contribution distribution, establish dynamic association rules with Q value. Determine the relationship between contribution and Q value through statistical analysis. Specifically, it includes:

[0118] S831. Calculate the contribution of each feature to the model prediction results using the integral gradient method

[0119] S832. Select a baseline input (such as an all-zero vector), interpolate between the baseline input and the actual input, and generate a series of intermediate inputs.

[0120] S833. Calculate the gradient of each intermediate input and integrate all gradients. The final contribution is the result of the integrated gradient.

[0121] S834. Calculate the Pearson correlation coefficient between each feature contribution and the Q value to measure the linear relationship between them.

[0122] S835. Using the existing linear regression model to fit the relationship between contribution and Q value can help understand which features have a significant impact on Q value.

[0123] S836. Use clustering algorithms (such as K-means and hierarchical clustering) to divide the samples into different groups, and then analyze the relationship between the contribution and Q value within each group.

[0124] Through the above steps, a dynamic association rule between contribution and Q value can be established, so as to better understand and optimize the model in the blood oxygen saturation detection system.

[0125] S84. Automatically generate model optimization suggestions based on contribution distribution: Automatically adjust model parameters based on contribution distribution and provide optimization suggestions. Provide interpretable analysis results to help understand the working principles of the model and optimize it.

[0126] S9. Adopt a transfer learning strategy and build a domain adaptation network using personalized physiological parameters to achieve online fine-tuning of model parameters. Specifically, the following steps are involved:

[0127] S91. Build a domain conversion module based on generative adversarial network;

[0128] S92. Design a personalized feature encoder including skin color, age, and body fat percentage;

[0129] S93, using meta-learning strategy to achieve rapid sample adaptation;

[0130] S94. Enhancing cross-domain feature representation capabilities through contrastive learning.

[0131] S10: Output the final blood oxygen saturation value and simultaneously generate graded prompt information including a confidence interval and abnormal fluctuation warning. The generation of the graded prompt information includes:

[0132] S101. Establish a multi-level warning threshold system based on clinical guidelines. Specifically, this includes setting different levels of warning thresholds based on clinical guidelines. For example, a level 1 warning is issued when SpO2 falls below 90%, and a level 2 warning is issued when it falls below 85%.

[0133] S102. Identify abnormal fluctuation patterns through time series analysis. Use an ARIMA model or other existing time series analysis methods to detect abnormal fluctuations.

[0134] S103. Calculate dynamic confidence intervals to reflect temporal evolution characteristics. Generate confidence intervals using Bayesian inference methods.

[0135] S104. Remind the user of potential health risks through voice prompts, visual alarms (such as flashing lights), tactile feedback (such as vibration), etc.

[0136] The adaptive weighted guided filtering of the present invention improves the ambient light noise suppression rate to 92.3% (compared to 68.5% of traditional morphological filtering); the motion artifact compensation model improves the signal-to-noise ratio (SNR) under micro-motion interference from 4.2dB to 12.7dB. Breakthrough in measurement accuracy: Dual-stream neural network feature fusion reduces SpO2 estimation error by 42%; the LSTM time series prediction module maintains 85% prediction accuracy when the signal is interrupted. Personalized adaptation enhancement: The domain adaptation network reduces the measurement deviation across skin color groups from ±3.1% to ±

[0137] 1.5%; the interpretability analysis module automatically generates model optimization recommendations with 98.2% effectiveness.

[0138] The detection method of the present invention is deployed on the Huawei Watch 5 Pro hardware platform: processor: HiSiliconHi3861V100, dedicated NPU to accelerate dual-stream network inference; power consumption optimization: dynamic Q value control reduces average power consumption from 12.3mW to 7.8mW; user interface: real-time display of SpO2 trend graph (sampling interval 1 second), abnormal fluctuations trigger vibration alerts. Deployment and testing in a hospital ICU: Data collection: 50 postoperative patients, continuously monitored for 72 hours;

[0139] The performance indicators are:

[0140]

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A pulse oximetry detection method based on image processing, characterized in that: The following steps are involved: S1. Capture the fingertip video stream of the subject through a multispectral imaging device and simultaneously obtain the initial blood oxygen saturation baseline value; S2. performing spatiotemporal preprocessing on the fingertip video stream, including: using adaptive weighted guided filtering to eliminate ambient light interference, and constructing a motion artifact compensation model to eliminate micro-motion interference; S3. Extracting dual-channel feature parameters from the preprocessed video, including: extracting the time-frequency features of the PPG signal through wavelet transform, and extracting the spatial features of the skin microvascular distribution through principal component analysis; S4. Build a two-stream convolutional neural network model, where the first branch uses ResNet-50 to extract time-frequency features, and the second branch uses U-Net to extract spatial features, and the two-branch features are fused through the attention mechanism; S5. Calculate the current signal quality index Q in real time. The Q value is generated by fusing three parameters: signal-to-noise ratio, waveform periodicity, and skin transmittance, using the entropy weight method. S6. Establish a dynamic calibration module to input the initial blood oxygen saturation baseline value, real-time Q value and neural network output value into the Bayesian filter model for iterative optimization; S7. When the signal quality index Q is lower than the threshold, the time series prediction module based on the LSTM network is triggered to compensate the current detection value using historical data; S8. Generate a contribution distribution diagram of each feature parameter through the interpretability analysis module and dynamically adjust the weight coefficient of each input parameter; S9, adopting a transfer learning strategy, using personalized physiological parameters to build a domain adaptation network and achieve online fine-tuning of model parameters; S10. Output the final blood oxygen saturation value and simultaneously generate graded prompt information including a confidence interval and abnormal fluctuation warning.

2. The pulse oximetry detection method based on image processing according to claim 1, characterized in that: The motion artifact compensation model in step S2 specifically includes: S21. Construct a hand motion trajectory prediction sub-model based on the inertial measurement unit; S22, establishing a spatiotemporal mapping relationship between the video frame sequence and the motion trajectory; S23, using the optical flow method to calculate the motion vector field, and combining the trajectory prediction results to perform reverse compensation; S24. Verify the quality of the compensated video sequence through convolutional recurrent network.

3. The pulse oximetry detection method based on image processing according to claim 1, characterized in that: The method for establishing the attention mechanism fusion in step S4 specifically includes: S41. Perform channel attention weighting on the time-frequency feature map to generate a feature heat map; S42, performing spatial attention weighting on the spatial feature map to generate a regional attention distribution map; S43. Dynamically fuse dual-channel features through a gating mechanism, and the fusion weight is dynamically adjusted by the real-time Q value.

4. The pulse oximetry detection method based on image processing according to claim 1, characterized in that: The specific method for establishing the Bayesian filtering model in step S6 is: S61. Establish a state space model including equipment error terms and physiological fluctuation terms; S62, using the neural network output as the observation value and the initial reference value as the prior distribution; S63, using the Q value to adjust the process noise covariance matrix; S64. Use particle filtering algorithm to estimate the posterior probability.

5. The pulse oximetry detection method based on image processing according to claim 1, characterized in that: The method for constructing the time series prediction module in step S7 specifically includes: S71. Build a bidirectional LSTM network architecture with an input window containing the time series features of the previous 60 seconds. S72. Introduce temporal convolutional networks to extract multi-scale temporal patterns; S73. Enhance the robustness of prediction results through adversarial training strategies; S74. Design a residual connection structure to compensate for prediction errors.

6. The pulse oximetry detection method based on image processing according to claim 1, characterized in that: The method for establishing the interpretability analysis module in step S8 includes the following steps: S81. Calculate the contribution of the input feature using the integral gradient method; S82, locating feature regions through class activation mapping; S83, establishing dynamic association rules between contribution and Q value; S84. Automatically generate model optimization suggestions based on contribution distribution.

7. The pulse oximetry detection method based on image processing according to claim 1, characterized in that: The method for constructing the domain adaptation network in step S9 specifically includes: S91. Build a domain conversion module based on generative adversarial network; S92. Design a personalized feature encoder including skin color, age, and body fat percentage; S93, using meta-learning strategy to achieve rapid sample adaptation; S94. Enhancing cross-domain feature representation capabilities through contrastive learning.

8. The pulse oximetry detection method based on image processing according to claim 1, characterized in that: The generation of the graded prompt information in step S10 includes: S101. Establish a multi-level warning threshold system based on clinical guidelines; S102. Identify abnormal conditions by combining trend analysis and fluctuation patterns; S103, generating a dynamic confidence interval including time evolution characteristics; S104. A multimodal warning method integrating voice, vision and touch.

Images