Deep tissue blood oxygen saturation degree estimation method and flexible wearable carotid artery blood oxygen detection system

Abstract

The invention discloses a deep tissue blood oxygen saturation degree estimation method which comprises the following steps: acquiring a multi-channel photoelectric volume pulse wave signal, labeling the multi-channel photoelectric volume pulse wave signal according to the blood oxygen saturation degree and a tissue structure category, and forming a data set by a label and the multi-channel photoelectric volume pulse wave signal; constructing an initial model which comprises a feature extraction module, an organization classification module and a prediction module; and training the initial model by using the data set to obtain a prediction model for estimating the oxyhemoglobin saturation of the deep tissue. The invention further provides a flexible wearable carotid artery blood oxygen detection system. According to the method provided by the invention, structure-guided feature decoupling is realized, and system deviation caused by individual difference is effectively reduced.

Description

Technical Field

[0001] This invention belongs to the field of medical electronic equipment, and particularly relates to a method for estimating deep tissue blood oxygen saturation and a flexible wearable carotid artery blood oxygen detection system. Background Technology

[0002] Currently, domestic patents focus on fingertip or wristband pulse oximeters, which can only detect blood oxygen saturation in the micro-arteries of the fingertips or wrists. There are no pulse oximeters for detecting deep arteries such as the carotid artery, and most are rigid devices. Current commercially available fingertip pulse oximeters are based on the transmission PPG principle, offering high accuracy, but their poor wearing comfort prevents them from entering the wearable field. Smartwatches, based on the reflection PPG principle, achieve wearable blood oxygen detection, but they are highly sensitive to wearing conditions; wearing a watch too loosely or too tightly significantly affects the results, often leading to falsely low readings or undetectable readings for abnormally low blood oxygen levels. Therefore, they are more suitable for monitoring blood oxygen trends during sleep. In the sensor design of the two blood oxygen detection devices mentioned above, the detection depth is fixed because the distance between the light source and the fingertip of the detector is fixed. This means that the blood oxygen saturation of the superficial microarteries at the end of the upper limb can only be detected, and it cannot be used for blood oxygen detection in other parts of the body. Therefore, it is not suitable for situations that require blood oxygen monitoring of deep arteries near the heart, such as limb loss due to accidents, insufficient blood flow in the extremities of the elderly, and carotid artery blood oxygen feedback during chest compressions after cardiac arrest.

[0003] Patent document CN120381269A discloses a wearable detection device design for deep tissue cerebral blood oxygenation. It uses a U-shaped bracket to integrate multiple light source emitting probes and multiple photodetectors into the inner ring surface, adjusts the distance between the light source and the detector to not less than 3cm, and effectively improves the detection density through a special layout, realizing high-density sampling in the lateral direction of the head. However, since the device is rigid and has a fixed structure, it is only suitable for head wear. Furthermore, this design does not mention the implementation methods of the device's data acquisition, transmission, processing, and result display processes.

[0004] Patent document CN121003438A discloses a design for a brain blood oxygen detection sensing system, which uses sensor nodes arranged in a matrix. The corresponding optical detection unit is synchronously driven by a wireless MCU and a corresponding PPG chip, realizing low-noise multi-node synchronous signal acquisition and processing. It also includes a wireless interface for transmitting information with a host computer. However, it does not mention the program method for the host computer to process blood oxygen saturation information and display the results.

[0005] Patent document CN110974249A discloses a design for a skin-attached blood oxygen saturation detection system. The front-end blood oxygen saturation detection circuit uses a flexible encapsulation material as the encapsulation layer, giving the front-end detection circuit the ability to fit tightly to the human skin and conform to the skin. The back-end signal processing module can send the signal to the mobile terminal through the Bluetooth antenna transmission module, realizing the real-time transmission and display of blood oxygen saturation information. However, it only sets up a set of light sources and detectors with a fixed spacing, which cannot realize the signal acquisition of blood oxygen saturation information of deep tissues. It can only be attached to relatively flat parts of the human skin surface such as fingers, arms and face to detect the blood oxygen saturation of superficial tissue microarteries. Summary of the Invention

[0006] The purpose of this invention is to provide a method for estimating deep tissue blood oxygen saturation and a flexible wearable carotid artery blood oxygen detection system. This method achieves structure-guided feature decoupling and effectively reduces system bias caused by individual differences.

[0007] To achieve the first objective of this invention, the following technical solution is provided: a method for estimating blood oxygen saturation in deep tissues, comprising the following steps: Acquire multi-channel photoplethysmography (PPG) signals, and label the PPG signals with blood oxygen saturation and tissue structure category. Combine the labels and the PPG signals to form a dataset. Build an initial model, including a feature extraction module, an organization and classification module, and a prediction module; The feature extraction module is used to extract data features from the multi-channel photoplethysmography (PPG) signal to obtain the corresponding multimodal feature vector. The tissue classification module generates corresponding tissue structure categories based on the input multimodal feature vector; The prediction module makes predictions based on organizational structure category and multimodal feature vectors to obtain corresponding prediction results; The initial model was trained using the dataset to obtain a predictive model for estimating blood oxygen saturation in deep tissues.

[0008] This invention enables the model to capture complex nonlinear relationships between signals through multi-dimensional features, thereby better extracting features from layers such as skin, muscle, and blood vessels, which helps improve the model's generalization ability.

[0009] Specifically, the multi-channel photoplethysmography (PPG) signal used for training is generated using the Monte Carlo optical simulation method.

[0010] Specifically, the steps of the Monte Carlo optical simulation method are as follows: A multi-layered tissue anatomy model constructed based on a three-dimensional rectangular coordinate system, including the skin and subcutaneous fat layer, muscle layer and composite layer with embedded cylindrical blood vessels; A specific signal statistical region is defined by the overlap between the annular detection zone and the orthographic projection region of the blood vessel. The scattering deflection angle of the specific signal statistical region is determined by the HG phase function, and energy tracking is performed using a single continuous weight attenuation strategy. The energy of photons emitted from the specific signal statistical region is statistically analyzed to construct a high signal-to-noise ratio multichannel photoplethysmography (PPG) signal.

[0011] Specifically, the multimodal feature vector includes the ratio, logarithmic ratio, and difference in the multichannel photoplethysmography pulse wave signal, as well as the absolute reflectance feature space.

[0012] Specifically, the organization classification module uses a random forest classifier to classify the organizational structure of the input features, outputs the organizational structure category and its posterior probability distribution, and finally selects the top-3 candidate categories with the highest probabilities as the organizational structure category for output.

[0013] To achieve the second objective of this invention, the following technical solution is provided: a flexible wearable carotid artery blood oxygen detection system, used to perform the steps of the above-mentioned method for estimating deep tissue blood oxygen saturation, including: A blood oxygen probe is used to emit multiple dual-wavelength optical signals; The analog front-end unit analyzes the electrical signals to generate corresponding detection results; The Bluetooth MCU master controller is used to receive remote commands and send detection results. Power supply, used for supplying power; A photodetector is used to convert an input optical signal into an electrical signal.

[0014] Specifically, the pulse oximeter includes three sets of LEDs, which are arranged in a ring around the photodetector as the origin, and the distance between the three sets of LEDs and the photodetector is arranged in a gradient, with the value ranging from 2.0cm to 3.2cm.

[0015] Specifically, the analog front-end unit needs to perform preprocessing before inputting electrical signals. The preprocessing includes removing DC components, signal inversion, and moving average filtering.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: By employing a dual-wavelength light source and different detection spacing, multidimensional extraction of deep tissue oxygenation information is achieved. A fine-grained, multi-category adaptive oxygenation regression model is introduced for oxygenation calculation, improving the accuracy of cross-individual oxygenation detection. A general modeling approach for deep tissue oxygenation detection is provided, exhibiting good method transferability and scalability potential. The use of a flexible FPC circuit board and flexible material packaging effectively enhances wearing comfort and signal quality. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a method for estimating deep tissue oxygen saturation provided in this embodiment; Figure 2 A schematic diagram illustrating the parametric modeling of the multi-layered cervical tissue structure provided in this embodiment; Figure 3 This is a demonstration diagram of the Monte Carlo optical simulation method in the model training phase provided in this embodiment; Figure 4 This is a schematic diagram of the pulse oximeter probe and a schematic diagram illustrating the principle of how the detection spacing affects the penetration depth, as provided in this embodiment. Figure 5 This is the design drawing of the FPC for the blood oxygen detection patch provided in this embodiment. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, this embodiment provides a method for estimating deep tissue oxygen saturation, comprising the following steps: Acquire multi-channel photoplethysmography (PPG) signals, and label the PPG signals with blood oxygen saturation and tissue structure category. Combine the labels and the PPG signals to form a dataset. Build an initial model, including a feature extraction module, an organization and classification module, and a prediction module; The feature extraction module is used to extract data features from the multi-channel photoplethysmography (PPG) signal to obtain the corresponding multimodal feature vector. The tissue classification module generates corresponding tissue structure categories based on the input multimodal feature vector; The prediction module makes predictions based on organizational structure category and multimodal feature vectors to obtain corresponding prediction results; The initial model was trained using the dataset to obtain a predictive model for estimating blood oxygen saturation in deep tissues.

[0020] More specifically, reflective PPG signals corresponding to the distances between three light sources and detectors under alternating two-wavelength emission are collected, and an input feature vector X is constructed.

[0021] In the classification stage, X is input into the organizational structure classifier, which outputs the organizational structure category ĝ and its posterior probability P(g|X); In the regression phase, X and the aforementioned category information are used as input conditions into the blood oxygen regression model, and the predicted blood oxygen value ŝ is output. Further, in the classification phase, the top-3 candidate categories with the highest probabilities are output, and combined with the corresponding blood oxygen regression results for joint scoring and selection, yielding the final tissue structure result g* and the final predicted blood oxygen value s*.

[0022] like Figure 2 The diagram shown illustrates the parametric modeling of the multi-layered cervical tissue structure provided in this embodiment. The cervical tissue, from superficial to deep, includes the skin and subcutaneous fat layer (L1), the muscle layer (sternocleidomastoid muscle) (L2), and the carotid artery layer (L3). Five thickness gradient values ​​are set for each of L1, L2, and L3, and these are combined to form 125 types of tissue structure models. On the tissue structure model, three sets of photon incident and exit point distances (D={D1,D2,D3}) are set, and the emitted photons include two wavelengths (λ={660nm,885nm}). The discrete state of blood oxygen saturation in the carotid artery layer is represented by S={S1,S2,…,S… 30 The model obtains the set of reflected signal intensities {R^λ_{D1}, R^λ_{D2}, R^λ_{D3}} at different wavelengths and spacings through Monte Carlo optical simulation, enabling the model to fully learn the mapping relationship between "deep tissue structure - optical signal response - blood oxygenation changes".

[0023] The model provided in this embodiment adopts a two-stage learning framework of "tissue structure classification - blood oxygenation regression". In the first stage, a random forest classifier is used to classify the input features into tissue structures, outputting the tissue structure categories and their posterior probability distributions. In the second stage, the tissue category information from the first stage is introduced as a conditional variable during blood oxygenation regression, explicitly incorporating structural differences into the blood oxygenation estimation process, thereby achieving adaptive modeling of individual differences in deep carotid artery structures. Furthermore, to improve the robustness of the inference stage, the top-3 candidate categories with the highest probabilities are output during the classification stage, and these categories are jointly scored and filtered with the blood oxygenation regression results. Taking into account both category confidence and the reasonableness constraints of blood oxygenation prediction, the final tissue classification result and blood oxygenation prediction value are output. Compared to the traditional single-fit formula relying on PPG signal calibration, by introducing fine-grained multi-category tissue structure modeling and category-conditional adaptive regression, decoupling and compensation for the complex structures of differentiated deep tissues in the human body are achieved, significantly improving the accuracy and cross-individual generalization ability of deep tissue blood oxygenation detection.

[0024] In this embodiment, a 15-dimensional input feature vector containing ratio, logarithmic ratio, and difference is constructed from the detected PPG signal; In cardiac arrest scenarios, the system enters a quasi-static monitoring mode: since there is no natural pulse at this time, the system extracts the absolute reflectivity feature space from the multi-channel signal. Combining the transient displacement of blood caused by chest compressions, the model uses a pre-trained deep learning network to infer the oxygenation status of the carotid artery even without an AC component, providing critical vital sign feedback for emergency treatment.

[0025] like Figure 3 The image shown is a demonstration diagram of the Monte Carlo optical simulation method during training. Figure 3 The left side of the middle section is a multi-layer tissue anatomy model constructed based on a three-dimensional rectangular coordinate system, including the skin and subcutaneous fat layer (L1), muscle layer (L2) and composite layer (L3) with embedded cylindrical blood vessels, and the specific signal statistical region defined by the overlap of the annular probe zone and the orthographic projection area of ​​the blood vessels is clearly defined. Figure 3 The right side of the middle section illustrates the photon transport mechanism, where the scattering deflection angle is determined using the HG phase function. And adopt a single continuous weight decay strategy. Energy tracking is performed, and finally, high signal-to-noise ratio deep tissue simulation data is generated by statistically analyzing the photon energy emitted from the overlapping region.

[0026] This method first constructs a parameterized three-dimensional multilayer tissue optical geometry model, in which deep aortas are specifically modeled as cylindrical structures embedded in connective tissue. Independent optical parameters are assigned to the blood inside the cylinder and the connective tissue outside, respectively, to simulate the geometry and optical heterogeneity of deep tissues. During the three-dimensional photon transport simulation, the Henyey-Greenstein phase function is used to determine the deflection angle after photon scattering, characterizing the forward scattering characteristics of biological tissues. Simultaneously, during photon energy interaction, a continuous weight attenuation strategy based on single-scattering albedo is employed. Specifically, when a photon collides with tissue, the photon weight is iteratively reduced according to the ratio of the local scattering coefficient to the total attenuation coefficient until the weight falls below a preset threshold. Finally, in the signal extraction stage, the specific signal statistical region is defined as the spatial overlap between the orthographic projection of the blood vessel cylinder onto the tissue surface and the annular detection band centered on the light source. Only the photon energy emitted from this overlapping region is counted as the effective reflected signal, thereby generating high-fidelity, high signal-to-noise ratio deep tissue photoplethysmography (PPG) simulation data.

[0027] The specific steps are as follows: A three-dimensional non-uniform parametric anatomical model was constructed: a three-dimensional Cartesian coordinate system was established, and a three-layer tissue model was constructed, comprising the skin and subcutaneous fat layer (L1), the muscle layer (L2), and the vascular complex layer (L3). Layers L1 and L2 were modeled as flat tissues with uniform optical parameters; layer L3 was constructed as a non-uniform medium layer, with the target blood vessel modeled as a cylinder located at a specific depth z, extending along the y-axis, with a radius of R_v, and the cylinder filled with connective tissue. Discrete gradient partitioning and full factorial combination were performed on the thickness of each layer and the diameter of the blood vessels to generate a fine-grained anatomical structure model library covering individual differences.

[0028] Perform weighted decay-based three-dimensional photon transport: Initial weights are perpendicularly incident at the origin. The photon packets are used for cyclic tracking: Step size generation: Based on the Lambert-Beer law, using the total decay coefficient Generate random step size .

[0029] Scattering deflection: After a photon collides with another photon, its scattering direction changes from the azimuth angle. and deflection angle Decision. Among them, cos The sampling strictly follows the Henyey-Greenstein (HG) phase function to accurately describe the forward scattering behavior of photons in biological tissues.

[0030] Weight update: A continuous weight decay mechanism is adopted. When a photon moves one step and collides at an interaction point, the photon does not terminate immediately, but rather its weight is updated according to the formula... = [ / ( + Update the weights (i.e., multiply by the single-scatter albedo). The Russian Roulette game will only terminate if the weights fall below a preset threshold (e.g., 10^{-8}).

[0031] Signal statistics based on projection overlap; To simulate a specific detection scenario where the patch is placed parallel to a blood vessel, a localized detection strategy is employed: Region definition: The effective statistical region is defined as the spatial overlap between the "elongated orthographic projection of the vascular cylinder on the tissue surface (xy plane)" and the "annular detection zone with a source-detection distance of D".

[0032] Energy integral: Only the total weight of photons escaping from the tissue surface in the overlapping region is counted as the effective reflection intensity R(D) at the detection distance, thereby filtering out background noise from lateral non-vascular tissues at the physical level.

[0033] Adaptive point selection: Within a preset strong signal radial range, analyze the signal sensitivity under different structural models, and automatically select the detection spacing combination that is most sensitive to structural differences and changes in blood oxygenation.

[0034] This embodiment also provides a flexible wearable carotid artery blood oxygen detection system, for performing the steps of the above-described method for estimating deep tissue blood oxygen saturation, including: A blood oxygen probe is used to emit multiple dual-wavelength optical signals; The analog front-end unit analyzes the electrical signals to generate corresponding detection results; The Bluetooth MCU master controller is used to receive remote commands and send detection results. Power supply, used for supplying power; A photodetector is used to convert an input optical signal into an electrical signal.

[0035] More specifically, such as Figure 4The diagram shows a pulse oximeter and the principle of how the detector spacing affects the penetration depth. The distances between the three LEDs and the detector are 2.2cm, 2.5cm, and 2.8cm, respectively. Extensive experiments have demonstrated that the detection range with this combination of spacing can cover the depth of the carotid artery in most people. Studies have shown that the distance between the light source and the detector can significantly affect the detection depth. Because of differences in the thickness of subcutaneous fat and muscle at the carotid artery in different individuals, the trajectory of photons propagating in the tissue varies. Therefore, the maximum depth to which photons can reach and ultimately exit the skin surface varies considerably. Furthermore, studies have shown that the photon propagation trajectory is statistically approximately a parabola, and the average diameter of the carotid artery is approximately 6mm. The more the photon propagation trajectory overlaps with the carotid artery, the more carotid blood oxygen information the photon carries. Experiments show that if only one detection spacing that can achieve the maximum detection depth is retained, even if the carotid artery depth of different individuals is covered, the photon carries less carotid blood oxygen information because the overlap between the photon trajectory and the shallower carotid artery is less. The detected PPG signal has a lot of interference from deep tissues, and may even submerge the PPG signal of the blood vessel, making it impossible to extract effective information and accurately calculate blood oxygen. Therefore, each carotid artery depth position has an optimal detection spacing range that matches it, and multiple detection spacings are needed to cover the differentiated groups of different ages, heights and weights.

[0036] like Figure 5 The diagram shown is an FPC design for a pulse oximetry patch. The device is modularly laid out, comprising a power supply, a Bluetooth MCU unit, an analog front-end unit, and a pulse oximetry probe. The power supply module provides power to the entire patch and includes a switch for user control of device startup and shutdown. The Bluetooth MCU unit preprocesses the detected PPG signal and transmits the data via Bluetooth to a mobile app for further processing and calculation. The analog front-end unit sets the LED drive current and performs bit-by-bit control on the driving timing of the dual-wavelength LEDs and the detection timing of the PD, using a high-precision algorithm to eliminate the influence of background light. The pulse oximeter probe is the core component of the detection patch. It carries three dual-wavelength LED light sources and a PD detector, forming three sets of detection intervals of 2.2cm, 2.5cm, and 2.8cm. When in use, the side with the sensor device (the back of the patch) is attached to the skin near the carotid artery using medical double-sided tape. The pulse oximeter probe is attached along the narrow gap between the sternocleidomastoid muscle and the thyroid cartilage as shown in the diagram. The other parts of the patch are attached to the back of the neck along the direction around the neck. Due to the numerous components on the circuit board, FPC soldering requires reinforcement behind the component pads, which to some extent weakens the flexibility of the patch. Therefore, the blood oxygen detection patch adopts a partitioned FPC design. The FPC parts that do not support components between each unit form a buffer gap, which disperses the dense stress laterally. It is also encapsulated with flexible materials. Compared with the design of integrating all components into a single FPC, the blood oxygen detection patch has greater flexibility and deformability, and is easier to fit the uneven structure of the human neck. This solves the problem of poor PPG signal quality or signal loss caused by poor adhesion between reflective PPG devices and the skin.

[0037] More specifically, the pulse oximeter probe uses a dual-wavelength LED, alternately emitting 660 nm red light and 885 nm infrared light. The 660 nm wavelength is generally more suitable for measuring superficial tissues (such as skin and subcutaneous fat), while the 885 nm wavelength can penetrate deeper tissue layers (such as muscle and blood vessels). Alternating these two wavelengths helps obtain blood oxygenation information from multiple tissue layers. Furthermore, compared to a single wavelength, the six sets of PPG signals obtained from three detection intervals under dual-wavelength conditions significantly expand the dimensionality of the aforementioned model's input feature vector. This multi-dimensional feature allows the model to capture complex nonlinear relationships between signals, thereby better extracting features from layers such as skin, muscle, and blood vessels, and contributing to improved model generalization ability.

[0038] The analog front-end unit uses the AFE4490 chip for sensor driving and signal acquisition, featuring functions such as signal filtering, gain adjustment, and automatic gain control. In contrast, the integrated MAX30102 pulse oximeter chip used in the aforementioned patent has a fixed selection of light source and detector, and the processor can only perform simple preprocessing and filtering of the signal. The AFE4490, however, can achieve high-resolution and low-noise signal acquisition and offers greater flexibility, allowing for configuration of LED light intensity, dual-wavelength drive timing, and setting of the distance between the light source and detector, etc., according to requirements. Furthermore, compared to traditional discrete devices, using an analog front-end can reduce debugging time, improve product development efficiency, and shorten the market launch cycle.

[0039] The Bluetooth MCU master controller uses the GR5515 Bluetooth chip, rather than a combination of a master MCU and a separate Bluetooth module. This design aims to achieve low-power Bluetooth transmission and reduce device size through a single-chip design. The chip integrates a Bluetooth Low Energy protocol stack, reducing the number of peripheral components, significantly lowering power consumption, and improving transmission efficiency through optimized communication performance. It is suitable for wearable blood oxygen monitoring devices that require long-term operation.

[0040] In this embodiment, the system first preprocesses the PPG signal, including removing the DC component, signal inversion, and moving average filtering, to eliminate background components and smooth the pulse wave curve, facilitating subsequent feature point detection. Then, a threshold is applied to the preprocessed signal to limit the signal amplitude within the normal physiological range, eliminating abnormal data caused by poor sensor contact or external interference, thus improving the stability and reliability of the algorithm. After signal preprocessing and thresholding, the algorithm performs peak detection on the processed PPG signal. By sequentially judging the left edge, right edge, and width of the flat-top peak, the effective pulse wave peak positions are determined, and adjacent peaks with too small an interval are removed to avoid false detections or duplicate counting. The remaining peak positions are then sorted and rearranged to ensure the temporal rationality of the peak sequence. Finally, the algorithm calculates the heart rate value based on the average interval between adjacent effective peaks and the signal sampling frequency; when the number of detected peaks is insufficient to support heart rate calculation, the algorithm marks the heart rate result as invalid.

[0041] Furthermore, the terms "upper," "lower," "inner," "outer," "front," and "rear" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0042] Of course, the above description is only a specific embodiment of the present invention and is not intended to limit the scope of the present invention. All equivalent changes or modifications made to the structure, features and principles described in the claims of the present invention should be included in the scope of the claims of the present invention.

[0043] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for estimating blood oxygen saturation in deep tissues, characterized in that, Includes the following steps: Acquire multi-channel photoplethysmography (PPG) signals, and label the PPG signals with blood oxygen saturation and tissue structure category. Combine the labels and the PPG signals to form a dataset. Build an initial model, including a feature extraction module, an organization and classification module, and a prediction module; The feature extraction module is used to extract data features from the multi-channel photoplethysmography (PPG) signal to obtain the corresponding multimodal feature vector. The tissue classification module generates corresponding tissue structure categories based on the input multimodal feature vector; The prediction module makes predictions based on organizational structure category and multimodal feature vectors to obtain corresponding prediction results; The initial model was trained using the dataset to obtain a predictive model for estimating blood oxygen saturation in deep tissues.

2. The method for estimating deep tissue oxygen saturation according to claim 1, characterized in that, The multi-channel photoplethysmography (PPG) signal used for training was generated using a Monte Carlo optical simulation method.

3. The method for estimating deep tissue oxygen saturation according to claim 2, characterized in that, The specific steps of the Monte Carlo optical simulation method are as follows: A multi-layered tissue anatomy model constructed based on a three-dimensional rectangular coordinate system, including the skin and subcutaneous fat layer, muscle layer and composite layer with embedded cylindrical blood vessels; A specific signal statistical region is defined by the overlap between the annular detection zone and the orthographic projection region of the blood vessel. The scattering deflection angle of the specific signal statistical region is determined by the HG phase function, and energy tracking is performed using a single continuous weight attenuation strategy. The energy of photons emitted from the specific signal statistical region is statistically analyzed to construct a high signal-to-noise ratio multichannel photoplethysmography (PPG) signal.

4. The method for estimating deep tissue oxygen saturation according to claim 1, characterized in that, The multimodal feature vector includes the ratio, logarithmic ratio, and difference in the multichannel photoplethysmography pulse wave signal, as well as the absolute reflectivity feature space.

5. The method for estimating deep tissue oxygen saturation according to claim 1, characterized in that, The organization classification module uses a random forest classifier to classify the organizational structure of the input features, outputs the organizational structure category and its posterior probability distribution, and finally selects the top-3 candidate categories with the highest probabilities as the organizational structure category for output.

6. A flexible wearable carotid artery blood oxygen detection system, characterized in that, The steps for performing the method for estimating deep tissue oxygen saturation as described in any one of claims 1 to 5 include: A blood oxygen probe is used to emit multiple dual-wavelength optical signals; The analog front-end unit analyzes the electrical signals to generate corresponding detection results; The Bluetooth MCU master controller is used to receive remote commands and send detection results. Power supply, used for supplying power; A photodetector is used to convert an input optical signal into an electrical signal.

7. The flexible wearable carotid artery blood oxygen detection system according to claim 6, characterized in that, The pulse oximeter includes three sets of LEDs, which are arranged in a ring around the photodetector as the origin. The distances between the three sets of LEDs and the photodetector are arranged in a gradient, with the values ​​ranging from 2.0cm to 3.2cm.

8. The flexible wearable carotid artery blood oxygen detection system according to claim 6, characterized in that, The analog front-end unit needs to perform preprocessing before the input electrical signal. The preprocessing includes removing DC components, signal inversion, and moving average filtering.

Images