A non-invasive method, device and equipment for detecting multiple biochemical indicators in the human body
By constructing a pure baseline vector space and a dual-path weighted fusion mechanism to eliminate abnormal signals, high-precision non-invasive detection of blood glucose, blood oxygen saturation, and hemoglobin content is achieved. This solves the problems of poor stability of PPG signal acquisition and insufficient model generalization ability, making it suitable for home health monitoring.
Patent Information
- Application Number
- CN202611596226.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-15
- Publication Date
- 2026-11-17
AI Technical Summary
Existing PPG signal acquisition has poor stability, and the abnormal signal screening mechanism has shortcomings. It cannot achieve accurate synchronous prediction of multiple parameters, and the existing detection model has insufficient generalization ability, making it difficult to meet the high-precision and high-stability detection needs of home users.
A pure baseline vector space is constructed for hierarchical screening of dual indicators. Combined with dual-path weighted fusion and residual compensation mechanism, abnormal signals are eliminated to achieve high-precision prediction of various biochemical indicators.
It enables simultaneous non-invasive detection of blood glucose, blood oxygen saturation, and hemoglobin levels. The hardware is low-cost, easy to miniaturize, suitable for long-term home monitoring, simple to operate, accurately identifies PPG signal distortion, and improves the accuracy of multi-parameter detection.
Smart Images

Figure CN123101136A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of biomedical signal acquisition and processing, and in particular to a non-invasive method, device and equipment for detecting multiple biochemical indicators in the human body. Background Technology
[0002] Metabolic and cardiovascular chronic diseases are among the core chronic diseases threatening human health and increasing the social medical burden globally. Abnormal blood glucose, disordered blood oxygen saturation, and imbalanced hemoglobin levels are three core physiological indicators reflecting blood circulation, metabolism, respiratory oxygen supply, and hematopoietic function, and are also key indicators for clinical screening, diagnosis, and chronic disease management. Multiple clinical studies have shown that blood glucose, blood oxygen saturation, and hemoglobin levels are interconnected and mutually influential: high blood glucose damages vascular endothelial function, reduces blood oxygen-carrying and metabolic capacity, and easily leads to insufficient blood oxygen supply; while blood oxygen deficiency and abnormal hemoglobin levels interfere with the body's glucose metabolism process, exacerbating blood glucose fluctuations. Long-term abnormalities in all three indicators can induce various complications such as cardiovascular and cerebrovascular diseases, metabolic syndrome, anemia, and microcirculatory disorders, seriously threatening human health. Therefore, simultaneous, continuous, and dynamic monitoring of blood glucose, blood oxygen saturation, and hemoglobin levels is a core prerequisite for achieving routine assessment of human physiological health and early screening and control of chronic diseases. Both the World Health Organization and clinical guidelines clearly state that routine dynamic monitoring of physiological parameters and joint analysis of multiple indicators can effectively improve the accuracy of disease screening and the comprehensiveness of disease assessment, providing reliable data support for personalized health interventions and adjustments to clinical treatment plans. Therefore, home-use non-invasive testing devices capable of simultaneously detecting multiple physiological parameters possess significant application value and clinical significance that cannot be replaced by single testing devices in various scenarios such as home health monitoring, daily management of chronic diseases, and screening for sub-health conditions.
[0003] Currently, in clinical and home settings, the mainstream methods for detecting blood glucose, blood oxygen saturation, and hemoglobin levels are mostly single invasive or specialized wired testing methods. Blood glucose testing commonly uses finger-prick blood sampling, an invasive technique that collects a small amount of blood from the fingertip and relies on a biochemical reaction on a test strip to obtain a numerical value. This technology is accurate and mature, but it carries risks of pain during blood collection, wound infection, and high-frequency testing can easily cause patient resistance, making long-term continuous monitoring difficult. Traditional methods for detecting blood oxygen saturation and hemoglobin levels rely heavily on specialized medical equipment. Blood oxygen detection often uses fixed medical pulse oximeters, which have limited portability; hemoglobin level detection usually requires venous blood collection and laboratory biochemical analysis, making the process cumbersome and time-consuming, and unsuitable for the real-time, rapid testing needs of home and mobile settings. Meanwhile, most existing testing equipment operates on a single-parameter independent testing mode, requiring multiple devices to complete the testing of three indicators separately. This is cumbersome, inefficient, and the timing and environment of testing differ between devices, making it impossible to achieve synchronous matching and monitoring of multiple parameters. This makes it difficult to meet the needs of joint assessment of the human body's multi-dimensional physiological state, greatly limiting the convenience and comprehensiveness of routine home health monitoring.
[0004] Currently, non-invasive physiological parameter detection technology is developing rapidly, mainly divided into two major technical systems: optical detection and non-optical detection. Traditional non-invasive optical detection technologies, such as near-infrared spectroscopy and ultrasound, have significant shortcomings, generally suffering from weak effective signals, strong interference from individual differences and environmental factors, high equipment costs, large size, and poor portability, making them unsuitable for the needs of home-based, miniaturized, low-cost, and real-time detection. In contrast, photoplethysmography (PPG) technology relies on changes in blood volume to collect physiological signals, offering significant advantages such as convenient signal acquisition, low hardware costs, high equipment miniaturization, and real-time continuous acquisition. Furthermore, PPG signals contain rich information related to blood metabolism, microcirculation oxygenation, and blood components, showing a high correlation with blood glucose levels, blood oxygen saturation, and hemoglobin levels. This provides core data support for the simultaneous detection of multiple physiological parameters, making it highly suitable for the research and development and implementation of home-based non-invasive multi-parameter detection devices, and thus becoming a current research hotspot in the field of non-invasive physiological detection.
[0005] However, practical applications and existing research have revealed that current non-invasive multi-parameter detection technologies based on PPG signals still have significant technical shortcomings, making it difficult to meet the high-precision and high-stability detection needs of home use. These problems can be broadly categorized into two main aspects: First, PPG signal acquisition suffers from poor stability, and the abnormal signal screening mechanism has significant shortcomings. PPG signals are weak optical-electrical signals, highly susceptible to multiple factors during acquisition, including individual physiological differences, limb motion artifacts, ambient light interference, and changes in body position. This can easily lead to waveform distortion, baseline shift, and signal-noise superposition. Currently, most signal preprocessing methods, both domestically and internationally, employ fixed thresholds and single-index evaluation methods, which cannot adapt to the differentiated PPG waveform characteristics of different populations and physiological states. This easily results in the misjudgment and rejection of valid feature waveforms and the omission of abnormal interference waveforms, leading to distortion in subsequent multi-parameter feature extraction and significantly reducing the overall accuracy and stability of blood glucose, blood oxygen, and hemoglobin detection.
[0006] Second, existing detection models lack generalization ability and cannot achieve accurate synchronous prediction of multiple parameters. Current PPG detection algorithms are mostly designed for single physiological parameters, only mining single-dimensional feature information in the signal, lacking in-depth mining of multi-parameter correlation and complementary features; a few multi-parameter detection models mostly use a single algorithm or a simple fixed-weight fusion strategy, failing to fully utilize the physiological correlation of blood glucose, blood oxygen, and hemoglobin indicators and the complementary information of PPG multi-feature space. When faced with complex and variable PPG signals from different populations and under different physiological states, the models cannot balance global generalization ability and local prediction stability, making it difficult to achieve high-precision detection of the three physiological parameters simultaneously. The prediction errors of each indicator cannot stably reach the clinical-grade home monitoring standards, greatly limiting the practical application of PPG multi-parameter non-invasive detection technology. Summary of the Invention
[0007] The purpose of this application is to provide a non-invasive detection method, device, and equipment for multiple biochemical indicators in the human body. By constructing a pure benchmark vector space for dual-indicator hierarchical screening, and combining dual-path weighted fusion and residual compensation mechanisms, it can achieve accurate removal of abnormal signals and high-precision prediction of multiple biochemical indicators.
[0008] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a non-invasive method for detecting multiple biochemical indicators in the human body, including: Acquire photoplethysmography (PPG) signals; The photoplethysmography (PPG) signal is subjected to anomaly removal to obtain the target pulse wave signal. The anomaly removal includes: acquiring multiple normal pulse wave samples and constructing a pure reference vector space based on these samples; mapping the feature vector of the pulse wave signal to be tested to the pure reference vector space to obtain a low-dimensional projection vector and a reconstructed fitting vector; calculating the residual deviation between the feature vector of the pulse wave signal to be tested and the reconstructed fitting vector, as well as the spatial distance between the low-dimensional projection vector and the center vector of the pure reference vector space; calculating a comprehensive anomaly score based on the residual deviation and the spatial distance; and performing stratified screening of the pulse wave signal to be tested using an adaptive dynamic threshold to remove anomaly signals. The feature vector of the target pulse wave signal is extracted and input into a pre-trained prediction model to output predicted values of blood glucose level, blood oxygen saturation and hemoglobin level; the prediction model includes: blood glucose level prediction model, blood oxygen saturation prediction model and hemoglobin level prediction model. The blood glucose level prediction model, blood oxygen saturation prediction model, and hemoglobin level prediction model all include multiple independent basic calculation units. The output predicted values of blood glucose level, blood oxygen saturation, and hemoglobin level include: The feature vector of the target pulse wave signal is input into multiple sets of the basic computing units to obtain multiple predicted values and then concatenate them into a fusion matrix. The fusion weights of each basic computational unit in the fusion matrix are calculated based on a two-factor adaptive composite distance function to obtain the predicted mean of the fusion matrix path; the two factors include a global bias factor and a local stability factor. Calculate the residual compensation correction term in the original feature space; The predicted mean is combined with the residual compensation correction term to obtain the final predicted values of blood glucose level, blood oxygen saturation and hemoglobin level.
[0009] Optionally, the construction of a pure reference vector space based on normal pulse wave samples includes: Multiple raw pulse wave samples that passed the initial manual screening were obtained; The original pulse wave sample is amplified by signal amplification, which includes at least one of time flipping, amplitude flipping, and time-amplitude joint flipping, to obtain an amplified training sample set, and each training sample is assigned a corresponding category label. A multi-layer continuous computation network is constructed, which includes multiple cascaded two-layer basic computation modules, a fixed-length transformation layer, and a classification output layer. Using the amplified training sample set as input and minimizing the overall classification error as the optimization objective, the multi-layer continuous computation network is trained. After training, the network parameters are fixed to obtain a fixed feature extraction mapping relationship. The depth feature vectors of all original pulse wave samples are extracted using the fixed feature extraction mapping relationship to obtain a set of depth feature vectors; Outlier samples are removed from the deep feature vector set through multiple iterations. Each iteration includes: calculating the average reference vector and offset vector of the deep feature vector set, constructing an initial offset matrix and performing feature decomposition to obtain the directional reference vector and component weight values. The depth feature vector is projected onto the directional reference vector, outlier vectors whose projection length deviates from the preset interval are iteratively removed, the average reference vector and feature decomposition are recalculated, until the projection values of the remaining vectors are concentrated in the preset interval, and a clean normal sample set is obtained. According to the component weight values in descending order, directional reference vectors whose cumulative weights reach a preset ratio threshold are selected and combined to generate a projection basis matrix, which constitutes the pure reference vector space.
[0010] Optionally, the step of performing stratified screening of the pulse wave signal to be tested using an adaptive dynamic threshold to remove abnormal signals specifically includes: Calculate the comprehensive abnormality score of all samples in the normal pulse wave sample, sort them, and select the score value corresponding to the high quantile as the adaptive dynamic threshold. First-level coarse screening: When the comprehensive abnormal score of the pulse wave signal to be tested is greater than the adaptive dynamic threshold, it is marked as a suspected abnormal waveform and enters the second-level fine discrimination; when it is less than or equal to, it is judged as a normal waveform. The second layer of fine discrimination: For suspected abnormal waveforms, the number of components that exceed the normal residual range and the maximum deviation of a single component are counted. When the proportion of abnormal components exceeds the preset ratio or the maximum deviation of a single component exceeds the preset multiple of the normal range, it is judged as an abnormal waveform and removed; otherwise, it is judged as a normal waveform.
[0011] Optionally, the covariance function of the composite distance function is in the form of: ; In the formula, and To fuse the two sets of sample vectors within the input matrix, Indicates the first The sample at the th Predicted values on the base unit; Indicates the first The sample at the th The predicted values on the basic unit; D=M is the total number of columns in the fusion matrix, that is, the total number of basic units; For signal amplitude parameters; The variance of the signal; For random noise amplitude parameters; For noise variance; Let Kronecker discriminant function be used. This refers to the scale parameter specific to the d-th column of the fusion matrix; the scale parameter specific to the d-th column is jointly determined by the global bias factor and the local stability factor, and It is positively correlated with the global deviation factor and negatively correlated with the local stability factor.
[0012] Optionally, the residual compensation correction term is calculated using the following formula: ; In the formula: This is a residual compensation correction term; The total number of training samples; Similarity in the original feature space; The sample to be predicted; For training samples; For the first The average prediction bias of all basic units on each training sample; The predicted final blood glucose level, blood oxygen saturation, and hemoglobin level are determined using the following formula: ; in, These are the predicted values for the final biochemical parameters. The predicted mean of the fusion matrix path, This is the residual compensation correction term.
[0013] Optionally, the global deviation factor is calculated using the following formula: ; in, For the first The global mean square error of the basic unit; For the first The set of basic units for the first Predicted values for each sample; The total number of training samples; For the first The actual biochemical parameter values of each sample.
[0014] Optionally, the local stability factor is calculated using the following formula: ; in, For the first Set of basic units in the sample to be predicted The local stability factor within the neighborhood ranges from 1 to 1. ; Number of nearest neighbors; For the neighboring region The mean of the predicted values of the basic unit; For the neighboring region Variance of the predicted values of the base unit; Indicates the first The sample at the th Predicted values on the base unit; for the sample to be predicted Find it in the fusion space Let the nearest neighbor training samples be denoted as the neighborhood set. ; The sample to be predicted.
[0015] Optionally, the feature vector of the target pulse wave signal is extracted, including: Extract the R coefficient features and spectral entropy SE of the target pulse wave signal at multiple specific wavelengths. Complexity, permutation entropy (PE), fractal dimension (FD); The characteristics of R coefficients and spectral entropy SE at multiple specific wavelengths The complexity, permutation entropy (PE), and fractal dimension (FD) are concatenated to obtain the feature vector.
[0016] Secondly, this application provides a non-invasive detection device for multiple biochemical indicators in the human body, including: The signal acquisition module is used to acquire photoplethysmography (PPG) signals. An abnormal signal removal module is used to remove abnormal signals from the photoplethysmography (PPG) signal to obtain a target pulse wave signal. The abnormal signal removal includes: acquiring multiple normal pulse wave samples and constructing a pure reference vector space based on the normal pulse wave samples; mapping the feature vector of the pulse wave signal to be tested to the pure reference vector space to obtain a low-dimensional projection vector and a reconstructed fitting vector; calculating the residual deviation between the feature vector of the pulse wave signal to be tested and the reconstructed fitting vector, as well as the spatial distance between the low-dimensional projection vector and the center vector of the pure reference vector space; calculating a comprehensive abnormality score based on the residual deviation and the spatial distance; and performing stratified screening of the pulse wave signal to be tested using an adaptive dynamic threshold to remove abnormal signals. A multi-parameter biochemical index prediction module is used to extract the feature vector of the target pulse wave signal, input the feature vector of the target pulse wave signal into a pre-trained prediction model, and output the predicted values of blood glucose level, blood oxygen saturation and hemoglobin level; the prediction model includes: blood glucose level prediction model, blood oxygen saturation prediction model and hemoglobin level prediction model. The blood glucose level prediction model, blood oxygen saturation prediction model, and hemoglobin level prediction model all include multiple independent basic calculation units. The output predicted values of blood glucose level, blood oxygen saturation, and hemoglobin level include: The feature vector of the target pulse wave signal is input into multiple sets of the basic computing units to obtain multiple predicted values and then concatenate them into a fusion matrix. The fusion weights of each basic computational unit in the fusion matrix are calculated based on a two-factor adaptive composite distance function to obtain the predicted mean of the fusion matrix path; the two factors include a global bias factor and a local stability factor. Calculate the residual compensation correction term in the original feature space; The predicted mean is combined with the residual compensation correction term to obtain the final predicted values of blood glucose level, blood oxygen saturation and hemoglobin level.
[0017] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the non-invasive detection method for multiple biochemical indicators of the human body as described above.
[0018] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a non-invasive method, device, and equipment for detecting multiple biochemical indicators in the human body, which has the following beneficial effects: 1. This invention enables simultaneous, non-invasive detection of multiple biochemical indicators in the human body, including blood glucose, blood oxygen saturation, and hemoglobin levels. On one hand, it eliminates the need for multiple blood draws, avoiding the pain and infection risks associated with invasive testing. The detection operation is simple and suitable for long-term continuous home monitoring. On the other hand, by utilizing a PPG signal source to simultaneously extract multiple blood physiological indicators, it allows for a comprehensive assessment of the body's metabolic and microcirculatory status by combining the physiological correlations between these indicators. This overcomes the shortcomings of traditional single-parameter detection methods, such as limited information and inconsistent timing of multi-device step-by-step detection. Furthermore, this solution boasts low hardware costs, ease of miniaturization, and a balance between model generalization ability and predictive stability, improving the accuracy of multi-parameter detection and meeting the clinical needs of home health monitoring.
[0019] 2. This application proposes a hierarchical screening mechanism for abnormal signals based on a clean reference vector space. By constructing a standard numerical space unaffected by abnormal waveforms, and relying on both residual bias (characterizing local point distortion) and spatial distance (characterizing overall baseline shift) for dual numerical evaluation, combined with an adaptive dynamic threshold that automatically updates following the distribution of normal samples, hierarchical screening is performed. This scheme can effectively remove interfering samples, adapt to the unique pulse waveform characteristics of different testers, significantly reduce the probability of qualified waveforms being misjudged as abnormal samples, accurately identify distorted PPG signals, and provide a high-quality data foundation for subsequent prediction.
[0020] 3. This application innovatively proposes a prediction model for multiple human biochemical indicators based on multi-layer units and dual-path weighted fusion. When fusing the contributions of each computational unit, the scaling coefficients of each dimension are simultaneously correlated with both global prediction bias and local prediction stability, enabling the contribution weights to be adaptively allocated according to the sample location, overcoming the poor generalization ability of traditional fixed-weight fusion. Simultaneously, a residual compensation correction term from the original feature space is introduced and combined with the weighted mean of the low-dimensional spliced matrix. The two prediction paths operate in different feature spaces, complementing each other's information and fully exploiting the complementary information in the fusion matrix and the original features, achieving high-precision calculation of multiple human biochemical indicators from PPG signals.
[0021] 4. The detection system provided in this application has a high degree of hardware adaptability. It uses LEDs of a specific wavelength as a time-division multiplexing light source, which is rich in information on various biochemical indicators of the human body and has low hardware cost and small size. The matching clamp unit adopts a spring structure combined with a pressure sensor, which can maintain the consistency of measurement position and pressure, ensure the optimal detection pressure, effectively eliminate pressure interference, further improve prediction accuracy, and meet the application requirements of low cost, miniaturization, and portability for home devices. Attached Figure Description
[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0023] Figure 1 A schematic diagram illustrating the principle of a non-invasive multi-parameter biochemical index detector for human body provided in an embodiment of this application; Figure 2 This is a schematic diagram of a clamping unit structure provided in an embodiment of this application; Figure 3 A flowchart illustrating a non-invasive detection method for multiple biochemical indicators in the human body, provided as an embodiment of this application; Figure 4This is a schematic diagram of the Clarke error grid, a blood glucose evaluation index provided in an embodiment of this application. Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation
[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0025] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0026] Example 1 like Figure 1 As shown in the figure, this embodiment provides a schematic diagram of the principle of a non-invasive detection instrument for multiple biochemical indicators in the human body. The non-invasive detection instrument includes a clamp unit, a data acquisition unit, a data processing unit, and a display unit.
[0027] The fixture unit mainly includes a pressure sensor, a temperature sensor, and a first analog-to-digital (A / D) converter. The pressure sensor adjusts and collects the pressure at the measured site, converting the measured pressure into an electrical signal via the first A / D converter. This signal is then transmitted to the MCU minimum system to determine if the pressure requirement is met. If it is, the signal is sent to the display unit for further signal acquisition. If not, the display unit shows a text prompt to the subject to adjust the pressure until the requirement is met. The temperature sensor senses the subject's body temperature, converts it into an electrical signal via the first A / D converter, and transmits this signal to the MCU minimum system.
[0028] The data acquisition unit is used to acquire the photoplethysmography (PPG) signal of the measured part, and then transmit the signal to the data processing unit for further processing.
[0029] In this embodiment, the core component of the data acquisition light source is an LED lamp assembly, which provides a stable lighting foundation for the downstream photodetector to acquire light signals. Different biochemical indicators use different wavelengths: blood oxygen saturation detection uses wavelengths of 660nm and 905nm; hemoglobin content detection uses wavelengths of 660nm, 810nm, and 940nm; and blood glucose content detection uses wavelengths of 880nm, 940nm, 1200nm, and 1550nm. The photodetector is responsible for picking up the photoplethysmography (PPG) signal and needs to be compatible with the reflected light signals formed after different wavelengths of light are absorbed by the human fingertip tissue. Therefore, this device is used in combination with a silicon (Si) photodetector and an indium gallium arsenide (InGaAs) photodetector.
[0030] The data processing unit coordinates the entire testing process of the equipment, undertaking functions such as light source drive control, data acquisition and processing, raw data storage, image display signal transmission, and outputting a stable constant current drive power supply to the front-end data acquisition module.
[0031] The data processing unit consists of a signal conditioning circuit, an analog-to-digital converter module, and an MCU minimum control system. The raw signal acquired from the front end is first sent to the signal conditioning circuit, where it undergoes current-to-voltage conversion, signal amplification, and multi-stage filtering to optimize the waveform and reduce noise, thereby improving signal integrity. The optimized analog signal is then converted to a digital signal by the A / D module and sent to the MCU minimum system. The MCU minimum system performs waveform feature extraction and quantitative regression modeling on the input digital signal, ultimately calculating the predicted values for blood glucose level, blood oxygen saturation, and hemoglobin level, and transmitting the results to the display module. In addition, the touchscreen on the display module can receive various operation commands from the user and send them back to the MCU minimum system. The main control system then starts and stops the optical signal acquisition operation according to the commands, and simultaneously completes the corresponding data parsing and calculation according to the preset program logic.
[0032] The display module is responsible for outputting real-time predicted results of blood glucose levels, blood oxygen saturation, and hemoglobin levels, as well as the subject's body surface temperature detection data. In this embodiment, an LCD screen is selected as the visualization output carrier.
[0033] The fixture unit described above, except for Figure 1 In addition to the pressure sensor and the first A / D converter shown, the system also includes components such as a finger fixing slot, a data acquisition end circuit board fixing slot, and a spring fixing cylinder. Figure 2 As shown, the structure and function of the clamping unit are described below: like Figure 2As shown, 1 is the LED, temperature sensor, and detector hole; 2 is the finger placement area, used to fix the finger and ensure optimal detection; 3 is the area for placing the acquisition end circuit board; 4 is the screw fixing hole; 5 is the spring hole, for placing a spring to adjust the pressure and ensure that the detection meets the pressure conditions; 7 is the pressure sensor placement area; 6 is the Velcro fixing bracket; 8 is the spring fixing cylinder; and 9 is the housing limit bracket.
[0034] In summary, the clamp unit connects the pressure sensor to the finger being measured via a spring, fixes the LED and detector via LED and detector holes, and fixes the position of the entire clamp module via a limit bracket.
[0035] Example 2 like Figure 3 As shown, a non-invasive method for detecting multiple biochemical indicators in the human body is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method includes the following steps: Step 101: Acquire the photoplethysmography (PPG) signal.
[0036] Step 102: Perform abnormal signal removal on the photoplethysmography (PPG) signal to obtain the target pulse wave signal; the abnormal signal removal includes: acquiring multiple normal pulse wave samples, constructing a pure reference vector space based on the normal pulse wave samples; mapping the feature vector of the pulse wave signal to be tested to the pure reference vector space to obtain a low-dimensional projection vector and a reconstructed fitting vector; calculating the residual deviation between the feature vector of the pulse wave signal to be tested and the reconstructed fitting vector, and the spatial distance between the low-dimensional projection vector and the center vector of the pure reference vector space; calculating a comprehensive abnormality score based on the residual deviation and the spatial distance, and performing hierarchical screening of the pulse wave signal to be tested using an adaptive dynamic threshold to remove abnormal signals; Step 103: Extract the feature vector of the target pulse wave signal, input the feature vector of the target pulse wave signal into the pre-trained prediction model of the corresponding biochemical index, and output the predicted values of blood glucose content, blood oxygen saturation and hemoglobin content. The quantitative regression prediction model includes multiple independent basic calculation units, and the predicted output values for blood glucose level, blood oxygen saturation, and hemoglobin level include: The feature vector of the target pulse wave signal is input into multiple sets of the basic computing units to obtain multiple predicted values and then concatenate them into a fusion matrix. The fusion weights of each basic computational unit in the fusion matrix are calculated based on a two-factor adaptive composite distance function to obtain the predicted mean of the fusion matrix path; the two factors include a global bias factor and a local stability factor. Calculate the residual compensation correction term in the original feature space; The predicted mean is combined with the residual compensation correction term to obtain the final predicted values of blood glucose level, blood oxygen saturation, and hemoglobin level.
[0037] In practical operation, the non-invasive detection process for multiple biochemical indicators in the human body includes: Turn on the power to supply power to all units of the system and its own functional modules.
[0038] The system executes the initialization program, completes the parameter configuration for each hardware module, assigns initial operating parameters to each component, and puts the whole machine into a ready-to-test state.
[0039] After the equipment completes initialization, the operator uses the adjustable clamping mechanism to fix the subject's finger and precisely adjusts the clamping pressure to the standard detection range. This optimal pressure value was determined through prior simulation modeling and multiple sets of control experiments. First, the finger is inserted into a special positioning groove to complete the limiting fixation. Then, the clamping pressure is increased or decreased using the elastic spring structure, and the display screen is observed to monitor the feedback value until the pressure parameter reaches the detection standard. At this point, the system automatically starts the PPG pulse wave signal acquisition process.
[0040] The entire PPG signal acquisition process is controlled by a constant current drive circuit, which sequentially illuminates multiple LED light sources in a time-division manner. The conduction time of a single pulse for each light source is set to the hundreds of microseconds. Multi-wavelength signal sampling is completed through cyclic time-division illumination, with a total acquisition time of 12 seconds for a complete full-wavelength acquisition cycle. During blood glucose testing, the constant current drive circuit selects four LED light sources: 880nm, 940nm, 1200nm, and 1550nm. The signals from the 880nm and 940nm sources, after reflection from the fingertip tissue, are captured by a silicon-based (Si) photodetector. The 1200nm and 1550nm long-wave near-infrared LEDs emit stable beams; these beams are reflected after penetrating the fingertip tissue and are received by an InGaAs photodetector adapted to the long-wave near-infrared band. During blood oxygen saturation testing, the constant current drive circuit selects two LED sources: 660nm red light and 905nm near-infrared light. The light source outputs beams in a time-division manner. The light enters the fingertip tissue and is absorbed and scattered by the tissue to form a reflected light signal. The reflected light is received by a silicon-based (Si) photodetector. Two wavelength-corresponding photoplethysmography (PPG) signals are simultaneously acquired to extract the light absorption difference between oxyhemoglobin and deoxyhemoglobin in arterial blood, thereby calculating blood oxygen saturation. When detecting hemoglobin content, the constant current drive circuit selects three LED light sources (660nm, 810nm, and 940nm) to operate in a time-division manner, with 810nm being the isoabsorption point wavelength of hemoglobin. All of the above wavelengths are within the response range of the silicon-based detector. The fingertip reflected light is captured by the silicon-based (Si) photodetector. By combining the pulse wave amplitude difference between the isoabsorption point wavelength and the auxiliary wavelength, hemoglobin-related characteristic information is obtained.
[0041] After receiving the light reflected from the fingertip, the photodetector generates a weak raw current signal. To eliminate noise and adapt to the back-end acquisition circuit, the weak current needs to be converted from current to voltage (IV) and then processed through multiple stages of signal amplification to obtain an optimized complete pulse waveform signal. The processed waveform signal is then input to the signal conditioning circuit for noise reduction, filtering, and shaping. The raw PPG data corresponding to different wavelengths is stored separately and finally transmitted to the MCU minimum control system for subsequent multi-biochemical index calculation.
[0042] Abnormal spectral rejection was performed on the above signals, followed by feature extraction. The extracted features were then substituted into the established model to calculate the blood glucose level, blood oxygen saturation level, and hemoglobin level of the subject.
[0043] To quantitatively evaluate the accuracy and detection error of this non-invasive multi-parameter detection system in measuring blood glucose, blood oxygen saturation, and hemoglobin levels, the measured results of commercial high-precision detection equipment were used as the true values for each physiological parameter. Specifically, blood glucose values were selected from the Sinocare iCanCGM continuous glucose monitoring system, while blood oxygen saturation and hemoglobin levels were based on real-time measurements from a medical-grade non-invasive blood oxygen and blood parameter detector. These combined to construct a multi-parameter synchronous control experimental baseline. The dataset for this experiment originated from a multi-day non-invasive multi-parameter control trial conducted in October 2025, with adult healthy men as the subjects. Before the formal testing, all subjects were required to fast for 12 hours to eliminate dietary metabolic interference, and the optical path, pressure, and data parameter calibrations of the entire detection equipment were completed. The equipment established stable data communication with the host computer via serial port to ensure synchronous transmission and storage of multi-channel signals. Sample collection covered different time periods throughout the day (morning, noon, and evening), with multiple sets of time-series data collected at various time points before and after meals to enrich the sample distribution scenarios and improve the model's applicability.
[0044] Before formal signal acquisition, the system relies on the pressure sensor of the clamp to provide real-time feedback on the fingertip clamping pressure. After stabilizing the pressure to the optimal detection range, the system initiates a multi-wavelength PPG pulse wave signal synchronous acquisition process, simultaneously acquiring multi-band optical raw signals adapted for blood glucose, blood oxygen saturation, and hemoglobin detection. Throughout the acquisition process, the subject maintains physical and mental stability and limb fixation to prevent motion artifacts caused by large-amplitude movements. Simultaneously, the integrated limiting clamping structure of the device ensures tight contact between the fingertip skin and the optical sensing element, maintaining a stable detection distance and effectively guaranteeing the acquisition quality, data consistency, and experimental repeatability of multiple sets of raw signals. At the same moment the system completes the acquisition of multi-wavelength PPG raw signals, the subject's true blood glucose, blood oxygen saturation, and hemoglobin levels are simultaneously acquired using a commercial reference device. This strictly controls the time difference between the non-invasive detection data and the true data, minimizing control errors caused by time lag and providing reliable and aligned experimental data support for subsequent multi-parameter model training, feature calibration, and quantitative evaluation of detection accuracy.
[0045] After acquiring the PPG signal, abnormal spectral detection is first performed. This application proposes a waveform vector space calculation model, which relies on a clean reference space, dual-parameter numerical scoring, and adaptive hierarchical discrimination to adapt to individual signal differences and accurately detect distorted PPG signal waveforms. Then, spatiotemporal dual-modal feature extraction is performed. Finally, a quantitative regression model is used to predict blood glucose levels, blood oxygen saturation, and hemoglobin levels. In the prediction module, an innovative prediction model for multiple human biochemical indicators based on multi-layer units and dual-path weighted fusion is proposed. This model simultaneously correlates the scaling coefficients of each calculation dimension with the overall bias and local stability, allowing the contribution weights to be adaptively allocated according to the sample location. The final predicted value is expressed as the sum of the weighted mean of the low-dimensional spliced matrix and the residual correction term of the original feature space, achieving high-precision calculation of multiple human biochemical indicators from PPG signals.
[0046] The system control and signal processing module performs abnormal spectral removal of PPG signals within the MCU. The algorithm flow is as follows: Step 1: The original single-acquisition PPG signal has a fixed sampling length of 3000 sampling points. The original discrete signal sequence is denoted as... , The signal sequence after baseline drift cancellation is denoted as , In the formula This is a low-frequency offset correction function. Its function is to extract the slowly changing reference offset of the signal and perform difference cancellation to eliminate the overall waveform rise and fall problems. After baseline correction, high-frequency fine noise is filtered out to obtain a smooth and clean waveform sequence. The filtered signal sequence is denoted as... The formula is In the formula To smoothly truncate the computation function and filter out minute, irregular fluctuations, a clean signal sequence is obtained. After preprocessing, manual initial screening is performed to remove inferior signals with dense waveform spikes, excessively large peak-to-peak differences, and irregular pulsation intervals. Qualified waveforms are continuously collected and the total number of high-quality samples is controlled to meet the required standards. To form a high-quality sample set To enrich the sample system and adapt to subsequent four-dimensional feature mapping training, each high-quality original waveform is expanded using three types of global flip transformations, resulting in four groups of effective samples for each original waveform. The specific transformation rules and corresponding identifiers are as follows: For the original signal , Sample identification The original signal is time-flipped to obtain... Sample identification The original signal is obtained by amplitude inversion. Sample identification The original signal was obtained by flipping both its time and amplitude. Sample identification The four types of samples correspond to the following identifiers. A single original signal is used to expand four sets of training data.
[0047] After signal transformation, severely distorted and invalid data were manually removed, retaining only high-quality and valid signals to construct a standardized labeled dataset. The total number of normal samples after manual screening was [number missing]. N Uniform length parameter for a single signal T , No. i The sample signal vector is ,in Representing the i Sample No. T The signal amplitude at each sampling point ;No. i Classification identifier parameters corresponding to each sample The rules for determining the identifier are specified. When the signal is the original, untransformed signal, ;when When it is a time-flip signal, ;when When it is an amplitude inversion signal, ;when When it is a time-amplitude combined flip signal, The complete sample set is The corresponding set of identifiers is .
[0048] This step ultimately outputs a standardized sample dataset with accurate four-class classification labels. (That is, the final output is a standardized clean waveform dataset with four-dimensional labels of 0 / 1 / 2 / 3), with all samples having uniform dimensions and one-to-one labels, fully covering normal original signals and three types of transformed signals. This dataset is the sole training input data source for the feature extraction module of the multi-layer continuous operation in step 2.
[0049] Step 2: Based on the four-dimensional sample dataset constructed in Step 1, deep and subtle features of the waveform are mined through multiple sets of concatenated numerical transformation modules. A mapping structure is independently constructed, consisting of 5 identical concatenated operation modules, 1 set of 64-dimensional fixed-length transformation layers, and 1 set of four-dimensional normalized classification output layers, which are sequentially concatenated and spliced together.
[0050] The single-module operation has two layers of basic numerical conversion operations and one layer of interval extreme value filtering operations. The specific operation expression of the single module is as follows: ; ; ; In the formula:k This is the module number, with a value ranging from 1 to 5. It is an exponential linear transformation function, responsible for amplifying minute differences in waveform characteristics. For interval extreme value filtering operations, the data length is compressed while retaining core fluctuation information; Input data for the current module. The first module's input is the original waveform sequence, and subsequent modules' input is the output result of the previous module. It is the first k The weight matrices of the first and second layers of each module are used to mine signal feature correlations; It is the first k The bias vectors of the first and second layers of each module are trainable parameters of the model, used to correct feature offsets; For the first k The intermediate feature vectors output from the first layer of operations in each module serve as the input for the second layer of operations. For the first k The depth intermediate feature vector output from the second layer operation of each module; is the final output feature vector of the k-th module, and is the input of the next level module or transformation layer.
[0051] Will n A series of two-layer basic modules are stacked in series to complete multi-level deep feature extraction of the signal. The output of the stacked modules is fed into a fixed-length transformation layer to standardize the feature dimensions and output a fixed-dimensional basic feature vector. The transformation layer formula is as follows: ; This is the output feature vector of the nth (last) concatenated module, and the input of the transformation layer; The weight matrix of the fixed-length transformation layer has trainable parameters and is used for feature dimension mapping. A fixed-length transformation layer bias vector, trainable parameters, used for feature calibration; This is a one-sided truncated linear transformation function (self-developed dimensionality calibration function), which is used to truncate invalid feature intervals, constrain the feature output range, and ensure the stability of feature vectors. Let m be the m-dimensional fundamental feature vector corresponding to the i-th sample signal, which represents the core deep feature representation of the signal. .
[0052] In this embodiment, n is set to 5. After the five groups of modules are sequentially processed, a 64-dimensional fixed-length transformation layer is connected to achieve unified standardization of feature dimensions and output a 64-dimensional deep feature vector. Based on standardized fundamental feature vectors The system performs four-class classification through normalization mapping operations, outputting a four-dimensional probability vector corresponding to the classification probabilities of the original signal and the three transformed signals. The classification operation formula is as follows: ; This is the classification layer weight matrix, which contains trainable parameters used for the association mapping between features and classification labels. This is the classification layer bias vector, which can be trained parameters used to correct the classification results. The normalized interval mapping operation maps the output to the 0-1 probability interval, resulting in a four-dimensional probability vector. This is a four-dimensional probability output vector, with the four dimensions corresponding to the classification probabilities of labels 0, 1, 2, and 3, respectively. The label corresponding to the maximum probability is the final classification result of the signal.
[0053] This step ultimately outputs the m-dimensional deep fundamental feature vectors of all training samples. The sample four-class classification probability prediction value. The output classification prediction value is used for the overall error calculation in step three, and the output feature vector. Used for constructing the covariance matrix in step four.
[0054] The predicted identifier is obtained by performing a four-dimensional normalization operation on the feature vector. The parameter iterative update is achieved using global error constraints. The global error formula is as follows: ; The total global error loss value is the value for all samples. The smaller the loss value, the higher the model classification accuracy and the more accurate the feature extraction. N is the total number of training samples, which is consistent with the parameter definition in step 1. i is the sample sequence number, and all N samples are iterated over; c is the classification label category number, which takes values of 0, 1, 2, and 3, corresponding to the four types of signal samples; The core function of the indicator function (binary function) is to select samples of the corresponding category to participate in error calculation: when the sample's true label... When equal to category c, It participates in error accumulation; when When not equal to c, It is not included in error calculation; The predicted probability value of the i-th sample corresponding to the c-th class is taken from the normalized four-dimensional probability output vector.
[0055] The iteration objective is to minimize the overall error value, continuously adjusting all transformation weights and offset parameters; when the overall error fluctuation value is less than 1e after 100 consecutive iterations. -5 Once fully converged and stabilized, all parameters of the entire mapping structure are locked and solidified, forming a fixed and unchanging feature extraction operation relationship. .
[0056] Any input compliant PPG waveform sequence All of them can stably output a unique 64-dimensional standardized feature vector through fixed operational relationships, which are fixed as follows: .
[0057] This step solidifies the 64-dimensional feature extraction operation rules, providing a unified and stable source of feature values for the construction of the pure standard feature space in step 3 and the anomaly detection of the waveform under test in step 4.
[0058] Step 3: Through multiple rounds of iterative screening and elimination of a small number of slightly distorted outlier vectors hidden in the high-quality sample set, the interference of inferior samples on the standard space is eliminated, and a high-purity normal waveform feature benchmark space is constructed, thereby reducing the probability of misjudging normal waveforms from the root.
[0059] The original high-quality waveform that passed all tests in step 1 ( Enter the feature extraction rules solidified in step 2 one by one. This yields a set of 64-dimensional standard feature vectors for all samples in batches. .
[0060] Step 1: Solve for the global average baseline vector to eliminate overall numerical bias. In the formula: It is a 64-dimensional average center vector, representing the standard characteristic center position of all high-quality waveforms.
[0061] Step 2: Solve for the deviation vector of a single sample relative to the center. Construct the initial correlation matrix C. , In the formula: matrix C is used to quantify the numerical correlation between 64-dimensional features, and T is the matrix transpose operation.
[0062] Step 3: Matrix decomposition and multi-round iterative sieving of outlier vectors, performing basic decomposition operations on the incidence matrix: In the formula These are the component weights obtained after decomposition; the larger the weight value, the more it represents the main distribution pattern of the normal waveform vector carried in that direction; matrix Each column corresponds to a directional reference vector, resulting in multiple sets of characteristic directional vectors and corresponding weight coefficients. The larger the weight coefficient, the better that direction can characterize the core distribution pattern of the normal waveform.
[0063] Project all sample feature vectors onto each standard direction and calculate the projection offset of a single sample in the core direction; set concrete screening rules: remove extreme outlier vectors whose projection offset exceeds the mean of all samples ± 1.96 times the standard deviation.
[0064] After a single round of elimination, the remaining clean samples are retained. The new average center vector is recalculated, the correlation matrix is reconstructed, and the direction vector is re-decomposed and updated. The above iterative process is repeated until the number of eliminated samples is 0 for two consecutive rounds and all sample projection offsets are concentrated in the normal range.
[0065] Step 4: Construct the final pure baseline projection matrix. After iterative purification, select the core direction vectors with the top 85% weighting and combine them to form a dedicated pure feature space projection transformation matrix W. This matrix completely eliminates outlier interference and retains only the standard distribution pattern of normal PPG waveforms. After screening outliers and removing abnormal waveforms from the initial high-quality samples, the total number of pure standard waveform samples remaining is M. The value is determined autonomously by the data distribution of the original samples. <N。
[0066] This step outputs a purified sample set, a standard feature center vector, and a highly reliable projection transformation matrix W, providing a standard benchmark for the dual-index scoring, adaptive thresholding, and hierarchical detection in step 4.
[0067] Step 4: First, perform the same baseline correction and high-frequency filtering as in Step 1 on the newly acquired raw signal to obtain a clean waveform to be measured. Prioritize verifying the periodicity of PPG, directly eliminate extreme abnormal waveforms with disordered or irregular pulsation cycles, and retain periodic and compliant waveforms for feature detection.
[0068] Step 1: Extract the 64-dimensional feature vector of the waveform under test, and extract features using the fixed operation rules established in Step 2: This yields the original 64-dimensional feature vector of the waveform to be tested.
[0069] Step 2: Standard spatial projection and feature reconstruction operations. Using the purified standard projection matrix W from step 3, we complete the dimensionality reduction mapping and spatial restoration reconstruction. ; ; In the formula: To simplify the core features for dimensionality reduction, To reconstruct and fit feature vectors that conform to the standard normal space.
[0070] Step 3: Quantifying two types of waveform anomalies using dual indicators Indicator 1: Residual Squared Deviation (characterizing local waveform point distortion) Calculate the sum of squared differences between the original features and the fitted features point by point to quantify the degree of local waveform disorder: ; Indicator 2: Standardized spatial distance (characterizing the overall characteristic offset of the waveform) Based on the core feature centers of pure, normal samples, the spatial deviation between the feature to be tested and the standard cluster is calculated, weakening the interference of the principal direction weights and accurately reflecting the overall offset. ; In the formula: The average center of the dimensionality reduction features of pure, normal samples. These are weighting coefficients for each direction, enabling overall offset quantization.
[0071] Step 4: Dual-weighted joint anomaly scoring By fusing two indicators with fixed weights, a unique comprehensive anomaly score is obtained, and the weighting coefficients are concretely fixed. , Prioritize detection of local waveform distortion: Step 5: Adaptive Dynamic Threshold Solution Substitute each of the clean, normal samples from step 3 into the scoring process described above to obtain the standard score distribution set of all normal samples; calculate the mean of this set. with standard deviation Generate a dynamic judgment threshold that adapts to the individual waveform characteristics of the current subject, with concrete threshold rules: .
[0072] The threshold is automatically updated according to the waveform distribution of different subjects, which completely solves the problems of poor adaptability of fixed thresholds and high individual false detection rate.
[0073] Step 6: Two-level screening and judgment (coarse and fine) First layer: Main space coarse screening (quickly filtering qualified waveforms) like If the waveform shows no obvious abnormalities, it is directly admitted and proceeds to the subsequent feature extraction and regression prediction process. like : Marked as a suspected abnormal waveform, and enter the residual space for fine discrimination.
[0074] Second layer: Residual space fine-tuning (precisely eliminating false anomalies) For suspected abnormal waveforms, the distribution of 64-dimensional feature residual components is checked one by one: the number of components exceeding the normal residual range and the maximum deviation of a single component are counted; only when the proportion of abnormal components exceeds 15% or the maximum deviation of a single component exceeds 3 times the normal range are they judged as real abnormal waveforms and the removal operation is performed. If there are only a few minor residual fluctuations and no substantial waveform distortion, it is determined to be a valid normal waveform and is allowed to proceed to the subsequent prediction process.
[0075] Feature extraction of multiple PPG signals was performed in the MCU, with independent feature analysis conducted on the raw PPG waveforms of the blood glucose, blood oxygen saturation, and hemoglobin levels detection channels. Extracted feature types included R-coefficient features, spectral entropy (SE), and complexity. Permutation entropy (PE) and fractal dimension (FD). For blood glucose level detection, feature extraction is performed based on PPG signals acquired from four LEDs at 880nm, 1550nm, 940nm, and 1200nm. The blood glucose feature vector is [R]. 1 / 3 ,R 1 / 4 ,R 2 / 3 ,R 2 / 4 SE1,C 01 ,FD1,PE1,SE2,C 02 ,FD2,PE2,SE3,C 03 ,FD3,PE3,SE4,C 04 [FD4,PE4], where subscripts 1, 2, 3, and 4 represent the features of PPG signals acquired by LEDs with wavelengths of 880nm, 1550nm, 940nm, and 1200nm, respectively; for blood oxygen saturation detection, feature extraction is performed based on the PPG signals acquired by two LEDs at 660nm and 905nm: in addition to calculating the spectral entropy SE1 and SE2 corresponding to the two waveforms respectively, the complexity C 01 C 02 In addition to permutation entropy PE1 and PE2, and fractal dimension FD1 and FD2, the focus is on extracting the blood oxygen-specific R coefficient feature R. 1 / 2 The R 1 / 2 The ratio of the AC component amplitude of the 660nm red PPG to the AC component amplitude of the 905nm near-infrared PPG is used to characterize the oxygenation status of arterial blood. The blood oxygenation feature vector is [R]. 1 / 2 SE1, SE2, C 01 C 02 [PE1,PE2,FD1,FD2], where subscripts 1 and 2 correspond to 660nm and 905nm respectively; for hemoglobin content detection, feature extraction is performed based on PPG signals acquired from three LEDs at 660nm, 810nm, and 940nm; where 810nm is the wavelength of the isoabsorption point of hemoglobin, and the spectral entropy SE1, SE2, and SE3 of each of the three waveforms are calculated, with a complexity of C. 01 C 02 C 03 Permutation entropy PE1, PE2, PE3; fractal dimension FD1, FD2, FD3; simultaneously extract two sets of R ratio coefficients based on the 810nm isoabsorption point: R 1 / 2 R represents the PPG AC amplitude ratio between 660nm and 810nm. 3 / 2The ratio of PPG AC amplitude at 940nm to that at 810nm is given by the hemoglobin eigenvector [R]. 1 / 2 ,R 3 / 2 SE1, SE2, SE3, C 01 C 02 C 03 [PE1,PE2,PE3,FD1,FD2,FD3], where subscripts 1, 2, and 3 correspond to 660nm, 810nm, and 940nm, respectively.
[0076] The blood glucose feature vector, blood oxygen feature vector, and hemoglobin feature vector are sequentially fed into the corresponding quantitative prediction models. The blood glucose prediction model, blood oxygen saturation prediction model, and hemoglobin level prediction model all employ the same model training process. The only difference between the three lies in the input feature set used during the training phase and the corresponding biochemical ground truth labels. This unified model training process is as follows: Step 1: Let the original feature matrix be X, and the true label vector be Y, where , n The total number of training samples is N, and the total number of features extracted from the original PPG signal is N. When the input is a blood glucose feature vector, N is 20; when the input is a blood oxygen feature vector, N is 9; and when the input is a hemoglobin feature vector, N is 14. ,in For the first i The reference values of biochemical indicators for each sample, and the overall structural relationship of the training data are as follows: Three independent basic computing units were constructed. Each basic unit was trained independently based on the same original feature matrix X, and each generated a set of predicted values. The output expression of the first basic unit is: Using the feature matrix as input, the algorithm fits the mapping relationship from features to biochemical indicators through kernel function mapping and ridge regression regularization. It then generates preliminary predicted values for the biochemical indicators for each sample, outputting the first set of one-dimensional prediction vectors. The output expression for the second set of basic units is: Output via Gaussian process regression The output expression for the third set of basic units is: Output via lightweight gradient boosting unit The one-dimensional prediction vectors output by the three basic units are concatenated and integrated column-wise to generate a new matrix specifically for top-level fusion calculations. .
[0077] Step 2: Use the fusion matrix Z output from Step 1 as the model input. For any two sets of sample vectors within the fusion input matrix and The basic form of the covariance function corresponding to the composite distance is: ; In the formula, and To fuse two sets of sample vectors within the input matrix, where Indicates the first i The sample at the th d Predicted values on the base unit; The signal amplitude parameter controls the overall output amplitude of the covariance function; The variance of the signal; The random noise amplitude parameter characterizes the level of uninterpretable random noise in the training data. For noise variance; Let Kronecker discriminant function be defined as follows: If and only if i = j ,otherwise Its function is to add noise terms only on the diagonal to ensure the positive definiteness of the covariance matrix. The first dimension is a dedicated scaling parameter, matched with a kernel ridge regression unit; The second-dimensional proprietary scaling parameter is matched to the Gaussian process regression unit; It uses 3D-specific scale parameters and is matched with a lightweight gradient boosting unit.
[0078] Calculate the first d Global prediction bias of the base unit across all training samples: , d In equations 1, 2, and 3: For the first d The global mean square error of a basic unit is the average prediction bias of that unit across the entire training set. For the first d The set of basic units for the first i Predicted values for each sample; n This represents the total number of training samples. A larger value indicates lower global prediction accuracy for that unit. Actual calculations show that the lightweight gradient boosting unit has the smallest global bias (…). Minimum), Gaussian process regression unit centered ( Medium), the nuclear ridge regression unit has the largest bias ( maximum).
[0079] For the sample to be predicted Find it in the fusion space k In this embodiment, the nearest neighbor training samples are used. k Let 10 be the neighborhood set, and denote it as: The neighborhood relationship is determined by the Euclidean distance in the fusion space: If and only if Ranked in the top 10 of all training samples. Within this neighborhood, the first... d Local predictive stability of the basic unit: ; In the formula: For the first d Set of basic units in the sample to be predicted The local stability factor within the neighborhood ranges from 1 to 1. The closer the value is to 1, the more stable and reliable the predicted value of the cell is in the neighborhood. For the neighboring region d The mean of the predicted values of the basic unit; For the neighboring region d The smaller the variance of the predicted values of the base unit, the more consistent the local predictions are and the higher the stability.
[0080] This means that even if a basic unit has a large global bias, if its predicted value is very stable (small variance) in the 10 nearest neighbors of the sample to be predicted, it indicates that the unit has a reliable judgment on this local area and should be given a higher local weight.
[0081] By combining the global deviation factor and the local stability factor, the first... d Adaptive scaling parameters for dimension: ; In the formula: For the first d Set of basic units in the sample to be predicted The adaptive scaling parameter varies with the location of the sample to be predicted, rather than being a globally fixed value; The baseline scale parameter is initially set to 1.0; This is the global bias sensitivity coefficient, which controls the intensity of the influence of global prediction bias on scale parameters. It is initially set to 1.0. This is the local stability sensitivity coefficient, which controls the intensity of the influence of local stability on the scale parameters. It is initially set to 1.0. Will Substituting back into the composite distance function, analysis shows that when the global deviation... Small, local stability When the scale parameter is close to 1, The weight will be very small, but the fusion weight will be very large and dominant.
[0082] The fusion weight of each basic unit can be understood as... Positive correlation: ; This formula clearly shows that the weight decreases exponentially with global deviation and increases exponentially with local stability. Due to scale parameters As the samples to be predicted change, a global bias factor (let) is used during the training phase. Alternatively, construct the training sample covariance matrix by taking the average stability of the training set: The prediction phase uses the complete two-factor scaling parameters. Calculate the association vector between the sample to be predicted and the training samples.
[0083] During the training phase, a global bias factor is used to construct the covariance matrix of the training samples (where local stability is set to the average level of the training set or zero). The prediction phase uses the complete two-factor scaling parameters. Calculate the association vector between the sample to be predicted and the training samples.
[0084] This step outputs the two-factor adaptive composite distance function. Global deviation vector Scale parameters to be optimized Signal amplitude Noise amplitude Training sample covariance matrix .
[0085] Step 3: Part 1: Parameter Optimization and Fusion Matrix Path Prediction Define the objective function for parameter optimization: denote all parameters to be optimized as a vector. In the formula The residual compensation coefficient is included in the optimization. With maximizing the data matching degree as the optimization direction, a log-marginal likelihood objective function is constructed: ; In the formula, To optimize the objective function value, a larger value indicates better model parameters. The higher the degree of matching with the training data; K is the training sample covariance matrix generated by the composite distance function, with dimension 1. Its calculation depends on (Use a global bias factor during the training phase); n Y represents the total number of training samples; Y represents the true label vector. First item The first term is the data fitting term, which measures the deviation between the model's predicted values and the true labels; the smaller the deviation, the larger this term is. The second term is the model complexity penalty term, where... The first term is the determinant of the covariance matrix, used to prevent overfitting of the model; the third term is a normalization constant, independent of the parameters.
[0086] The optimization objective is: ; The gradient descent numerical optimization algorithm is used to iteratively update the parameter set. The specific process for all variables is as follows: (1) Initialization: Set the initial values of the parameters , , , , , Learning rate Convergence threshold Maximum number of iterations T = 500.
[0087] (2) Iterative update: in the first t In each iteration, the gradient of the objective function with respect to each parameter is calculated, and the parameters are updated along the gradient ascent direction.
[0088] (3) Convergence Judgment: Continuously calculate the logarithmic matching index of the objective function and continuously adjust the parameter values towards the maximum value of the index. When the magnitude of the change in the objective function value between two adjacent rounds meets the threshold or reaches the maximum number of iterations, convergence is determined. When the parameters converge, it is determined that the parameters have converged.
[0089] (4) Convergence result: The global optimal parameter values are obtained after convergence. .
[0090] Substituting the optimal parameters yields the optimal covariance matrix. .
[0091] For the sample to be predicted, its fusion vector is first generated by the basic computing units in step 1: ,in This is the original feature vector of the sample to be predicted.
[0092] The distance correlation vector between the sample to be predicted and all training samples is calculated using a two-factor adaptive scaling parameter: In the formula This is the distance correlation vector between the sample to be predicted and the training samples. Each element represents the covariance between the new sample and the corresponding training sample under a two-factor composite distance function. Note the scale parameter used here. That is, it includes local stability factors.
[0093] The predicted mean of the fusion matrix path is: In the formula, The predicted mean of biochemical indicators for the fusion matrix path; Y is the covariance matrix of the training samples after substituting the optimal parameters; Y is the true label vector of the training samples.
[0094] Part Two: Residual Compensation Correction Items For each training sample i and each basic unit d Calculate the predicted residuals: In the formula For the first d The set of basic units for the first i The prediction residual of a sample is the deviation between the actual value and the predicted value. A positive residual indicates that the biochemical indicators were underestimated in that unit, while a negative residual indicates that they were overestimated.
[0095] Construct the residual matrix: ,in Let be the residual vector of the d-th unit. For each sample, take the mean of the residuals of all units to obtain the sample-level residual vector: , Indicates the first i The average prediction bias of all basic units on a training sample reflects the systematic shift direction of the fusion prediction within that sample region.
[0096] In the original feature space, define the sample to be predicted. With training samples Similarity between them: In the formula: The original feature space similarity has a value range of 1. The closer the value is to 1, the more similar the two samples are in the original feature space. The square of the Euclidean distance in the original feature space; The bandwidth parameter of the original feature space is taken as the median of the pairwise distances of the training samples in this embodiment (approximately the typical value of the feature space scale) to ensure that the similarity function covers a reasonable range.
[0097] This similarity function is independent of the fusion matrix path and measures the proximity between samples directly in the original feature space, thus capturing original feature information not included in the fusion matrix.
[0098] Based on the original feature space similarity, the average residuals of the training samples are weighted and summed to obtain the residual compensation correction term for the sample to be predicted: ; In the formula: This is the residual compensation correction term, which is the weighted average of the average prediction bias of each basic unit on training samples similar to the sample to be predicted in the original feature space. If similar samples are generally underestimated... ,but The fusion prediction value should be adjusted upwards if necessary, and downwards if it is downwards. The denominator is a normalization factor to ensure that the dimensions of the correction term are consistent with the biochemical index value.
[0099] The predicted mean of the fusion matrix path is adaptively combined with the residual compensation correction term to obtain the final predicted values of each biochemical indicator: ; In the formula, These are the predicted values for the final biochemical indicators; This is the predicted mean of the fusion matrix path (first part of the output); This is the residual compensation correction term (second part of the output); The optimal residual compensation coefficient is determined by the parameter optimization process, controlling the contribution ratio of the residual compensation path in the final prediction. When the residual information is valid, the path prediction degenerates into a pure fusion matrix; when the residual information is valid, the path prediction degenerates into a pure fusion matrix. Take a positive value to correct for systematic bias.
[0100] The two paths operate in different feature spaces and complement each other: the fusion matrix path is responsible for the main prediction, and the residual compensation path is responsible for correcting systematic biases.
[0101] This step outputs the optimal parameter set. ; and the final predicted values of biochemical indicators .
[0102] Step 4: After the biochemical indicator prediction calculation is completed, the final predicted values will be... The data is fed into the driving circuit of the LCD display. After digital-to-analog conversion and display refresh, the biochemical index values of the subject are displayed on the screen in digital form, including blood glucose level (unit: mmol / L), blood oxygen saturation (unit: %), and hemoglobin level (unit: g / dL).
[0103] To systematically evaluate the multi-parameter measurement accuracy of this detector, six control experiments were conducted. The blood glucose level measured by the Sinocare iCanCGM continuous glucose monitoring system was used as the standard reference, while the results of a commercial medical-grade fingertip pulse oximeter and venous blood routine test were used as standard references for blood oxygen saturation and hemoglobin levels, respectively. This application introduces Clarke Error Grid Analysis based on the root mean square error of prediction to further evaluate the system's ability to detect blood glucose levels. The system test results are shown in the table below: Table 1. Blood Glucose Content Error Assessment
[0104] Table 2. Blood Oxygen Saturation Error Assessment
[0105] Table 3. Error Assessment of Hemoglobin Content
[0106] The six control experiments in this application showed that the MARD values in the blood glucose level error evaluation index were all below 15% (mean 13.10%), and the proportion of samples falling into zones A and B reached 100%. Furthermore, the statistical analysis of the percentage of samples with an absolute relative deviation (ARD) of less than or equal to 20% showed that the mean of the six experiments reached 80%, with three experiments exceeding 80% in compliance rate. The mean RMSE in the blood oxygen saturation evaluation index was 1.68%, and the mean RMSE in the hemoglobin level evaluation index reached 1.20 g / dL. In summary, the system demonstrates good performance in predicting multiple parameters including blood glucose level, blood oxygen saturation, and hemoglobin level, and its predictive accuracy basically meets practical measurement requirements.
[0107] Based on the same inventive concept, this application also provides a non-invasive detection device for multiple human biochemical indicators to implement the aforementioned non-invasive detection method for multiple human biochemical indicators. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations in the embodiments of the non-invasive detection device for multiple human biochemical indicators provided below can be found in the limitations of the non-invasive detection method for multiple human biochemical indicators described above, and will not be repeated here.
[0108] Example 3: In one exemplary embodiment, a non-invasive detection device for multiple human biochemical indicators is provided, comprising: The signal acquisition module is used to acquire photoplethysmography (PPG) signals. An abnormal signal removal module is used to remove abnormal signals from the photoplethysmography (PPG) signal to obtain a target pulse wave signal. The abnormal signal removal includes: acquiring multiple normal pulse wave samples and constructing a pure reference vector space based on the normal pulse wave samples; mapping the feature vector of the pulse wave signal to be tested to the pure reference vector space to obtain a low-dimensional projection vector and a reconstructed fitting vector; calculating the residual deviation between the feature vector of the pulse wave signal to be tested and the reconstructed fitting vector, as well as the spatial distance between the low-dimensional projection vector and the center vector of the pure reference vector space; calculating a comprehensive abnormality score based on the residual deviation and the spatial distance; and performing stratified screening of the pulse wave signal to be tested using an adaptive dynamic threshold to remove abnormal signals. A multi-parameter biochemical index prediction module is used to extract the feature vector of the target pulse wave signal, input the feature vector of the target pulse wave signal into a pre-trained prediction model, and output the predicted values of blood glucose level, blood oxygen saturation and hemoglobin level; the prediction model includes: blood glucose level prediction model, blood oxygen saturation prediction model and hemoglobin level prediction model. The blood glucose level prediction model, blood oxygen saturation prediction model, and hemoglobin level prediction model all include multiple independent basic calculation units. The output predicted values of blood glucose level, blood oxygen saturation, and hemoglobin level include: The feature vector of the target pulse wave signal is input into multiple sets of the basic computing units to obtain multiple predicted values and then concatenate them into a fusion matrix. The fusion weights of each basic computational unit in the fusion matrix are calculated based on a two-factor adaptive composite distance function to obtain the predicted mean of the fusion matrix path; the two factors include a global bias factor and a local stability factor. Calculate the residual compensation correction term in the original feature space; The predicted mean is combined with the residual compensation correction term to obtain the final predicted values of blood glucose level, blood oxygen saturation and hemoglobin level.
[0109] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 5 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores data on non-invasive detection methods for various human biochemical indicators. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a non-invasive detection method for various human biochemical indicators.
[0110] Those skilled in the art will understand that Figure 5 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.
[0111] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.
[0112] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).
[0113] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.
[0114] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0115] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A non-invasive method for detecting multiple biochemical indicators in the human body, characterized in that, The non-invasive detection methods for multiple human biochemical indicators include: Acquire photoplethysmography (PPG) signals; The photoplethysmography (PPG) signal is subjected to anomaly removal to obtain the target pulse wave signal. The anomaly removal includes: acquiring multiple normal pulse wave samples and constructing a pure reference vector space based on these samples; mapping the feature vector of the pulse wave signal to be tested to the pure reference vector space to obtain a low-dimensional projection vector and a reconstructed fitting vector; calculating the residual deviation between the feature vector of the pulse wave signal to be tested and the reconstructed fitting vector, as well as the spatial distance between the low-dimensional projection vector and the center vector of the pure reference vector space; calculating a comprehensive anomaly score based on the residual deviation and the spatial distance; and performing stratified screening of the pulse wave signal to be tested using an adaptive dynamic threshold to remove anomaly signals. The feature vector of the target pulse wave signal is extracted and input into a pre-trained prediction model to output predicted values of blood glucose level, blood oxygen saturation and hemoglobin level; the prediction model includes: blood glucose level prediction model, blood oxygen saturation prediction model and hemoglobin level prediction model. The blood glucose level prediction model, blood oxygen saturation prediction model, and hemoglobin level prediction model all include multiple independent basic calculation units. The output predicted values of blood glucose level, blood oxygen saturation, and hemoglobin level include: The feature vector of the target pulse wave signal is input into multiple sets of the basic computing units to obtain multiple predicted values and then concatenate them into a fusion matrix. The fusion weights of each basic computational unit in the fusion matrix are calculated based on a two-factor adaptive composite distance function to obtain the predicted mean of the fusion matrix path; the two factors include a global bias factor and a local stability factor. Calculate the residual compensation correction term in the original feature space; The predicted mean is combined with the residual compensation correction term to obtain the final predicted values of blood glucose level, blood oxygen saturation and hemoglobin level.
2. The non-invasive detection method for multiple biochemical indicators in the human body according to claim 1, characterized in that, The construction of a pure baseline vector space based on normal pulse wave samples includes: Multiple raw pulse wave samples that passed the initial manual screening were obtained; The original pulse wave sample is amplified by signal amplification, which includes at least one of time flipping, amplitude flipping, and time-amplitude joint flipping, to obtain an amplified training sample set, and each training sample is assigned a corresponding category label. A multi-layer continuous computation network is constructed, which includes multiple cascaded two-layer basic computation modules, a fixed-length transformation layer, and a classification output layer. Using the amplified training sample set as input and minimizing the overall classification error as the optimization objective, the multi-layer continuous computation network is trained. After training, the network parameters are fixed to obtain a fixed feature extraction mapping relationship. The depth feature vectors of all original pulse wave samples are extracted using the fixed feature extraction mapping relationship to obtain a set of depth feature vectors; Outlier samples are removed from the deep feature vector set through multiple iterations. Each iteration includes: calculating the average reference vector and offset vector of the deep feature vector set, constructing an initial offset matrix and performing feature decomposition to obtain the directional reference vector and component weight values. The depth feature vector is projected onto the directional reference vector, outlier vectors whose projection length deviates from the preset interval are iteratively removed, the average reference vector and feature decomposition are recalculated, until the projection values of the remaining vectors are concentrated in the preset interval, and a clean normal sample set is obtained. According to the component weight values in descending order, directional reference vectors whose cumulative weights reach a preset ratio threshold are selected and combined to generate a projection basis matrix, which constitutes the pure reference vector space.
3. The non-invasive detection method for multiple biochemical indicators in the human body according to claim 1, characterized in that, The step of performing stratified screening of the pulse wave signal to be tested, combined with an adaptive dynamic threshold, to remove abnormal signals, specifically includes: Calculate the comprehensive abnormality score of all samples in the normal pulse wave sample, sort them, and select the score value corresponding to the high quantile as the adaptive dynamic threshold. First-level coarse screening: When the comprehensive abnormal score of the pulse wave signal to be tested is greater than the adaptive dynamic threshold, it is marked as a suspected abnormal waveform and enters the second-level fine discrimination; when it is less than or equal to, it is judged as a normal waveform. The second layer of fine discrimination: For suspected abnormal waveforms, the number of components that exceed the normal residual range and the maximum deviation of a single component are counted. When the proportion of abnormal components exceeds the preset ratio or the maximum deviation of a single component exceeds the preset multiple of the normal range, it is judged as an abnormal waveform and removed; otherwise, it is judged as a normal waveform.
4. The non-invasive detection method for multiple biochemical indicators in the human body according to claim 1, characterized in that, The covariance function of the composite distance function is in the form of: ; In the formula, and To fuse the two sets of sample vectors within the input matrix, Indicates the first The sample at the th Predicted values on the base unit; Indicates the first The sample at the th The predicted values on the basic unit; D=M is the total number of columns in the fusion matrix, that is, the total number of basic units; For signal amplitude parameters; The variance of the signal; For random noise amplitude parameters; For noise variance; Let Kronecker discriminant function be used. This refers to the scale parameter specific to the d-th column of the fusion matrix; the scale parameter specific to the d-th column is jointly determined by the global bias factor and the local stability factor, and It is positively correlated with the global deviation factor and negatively correlated with the local stability factor.
5. The non-invasive detection method for multiple biochemical indicators in the human body according to claim 1, characterized in that, The residual compensation correction term is calculated using the following formula: ; In the formula: This is a residual compensation correction term; The total number of training samples; Similarity in the original feature space; The sample to be predicted; For training samples; For the first The average prediction bias of all basic units on each training sample; The predicted final blood glucose level, blood oxygen saturation, and hemoglobin level are determined using the following formula: ; in, These are the predicted values for the final biochemical indicators. The predicted mean of the fusion matrix path, This is the residual compensation correction term.
6. The non-invasive detection method for multiple biochemical indicators in the human body according to claim 1, characterized in that, The global deviation factor is calculated using the following formula: ; in, For the first The global mean square error of the basic unit; For the first The set of basic units for the first Predicted values for each sample; The total number of training samples; For the first The actual biochemical parameter values of each sample.
7. The non-invasive detection method for multiple biochemical indicators in the human body according to claim 1, characterized in that, The local stability factor is calculated using the following formula: ; in, For the first Set of basic units in the sample to be predicted The local stability factor within the neighborhood ranges from 1 to 1. ; Number of nearest neighbors; For the neighboring region The mean of the predicted values of the basic unit; For the neighboring region Variance of the predicted values of the base unit; Indicates the first The sample at the th Predicted values on the base unit; for the sample to be predicted Find it in the fusion space Let the nearest neighbor training samples be denoted as the neighborhood set. ; The sample to be predicted.
8. The non-invasive detection method for multiple biochemical indicators in the human body according to claim 1, characterized in that, Extracting the feature vector of the target pulse wave signal includes: Extract the R coefficient features and spectral entropy SE of the target pulse wave signal at multiple specific wavelengths. Complexity, permutation entropy (PE), fractal dimension (FD); The characteristics of R coefficients and spectral entropy SE at multiple specific wavelengths The complexity, permutation entropy (PE), and fractal dimension (FD) are concatenated to obtain the feature vector.
9. A non-invasive detection device for multiple biochemical indicators in the human body, characterized in that, The non-invasive detection device for multiple human biochemical indicators includes: The signal acquisition module is used to acquire photoplethysmography (PPG) signals. An abnormal signal removal module is used to remove abnormal signals from the photoplethysmography (PPG) signal to obtain a target pulse wave signal. The abnormal signal removal includes: acquiring multiple normal pulse wave samples and constructing a clean reference vector space based on the normal pulse wave samples; mapping the feature vector of the pulse wave signal to be tested to the clean reference vector space to obtain a low-dimensional projection vector and a reconstructed fitting vector; calculating the residual deviation between the feature vector of the pulse wave signal to be tested and the reconstructed fitting vector, as well as the spatial distance between the low-dimensional projection vector and the center vector of the clean reference vector space; calculating a comprehensive abnormality score based on the residual deviation and the spatial distance; and performing stratified screening of the pulse wave signal to be tested using an adaptive dynamic threshold to remove abnormal signals. A multi-parameter biochemical index prediction module is used to extract the feature vector of the target pulse wave signal, input the feature vector of the target pulse wave signal into a pre-trained prediction model, and output the predicted values of blood glucose level, blood oxygen saturation and hemoglobin level; the prediction model includes: blood glucose level prediction model, blood oxygen saturation prediction model and hemoglobin level prediction model. The blood glucose level prediction model, blood oxygen saturation prediction model, and hemoglobin level prediction model all include multiple independent basic calculation units. The output predicted values of blood glucose level, blood oxygen saturation, and hemoglobin level include: The feature vector of the target pulse wave signal is input into multiple sets of the basic computing units to obtain multiple predicted values and then concatenate them into a fusion matrix. The fusion weights of each basic computational unit in the fusion matrix are calculated based on a two-factor adaptive composite distance function to obtain the predicted mean of the fusion matrix path; the two factors include a global bias factor and a local stability factor. Calculate the residual compensation correction term in the original feature space; The predicted mean is combined with the residual compensation correction term to obtain the final predicted values of blood glucose level, blood oxygen saturation and hemoglobin level.
10. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the non-invasive detection method for multiple biochemical indicators of the human body as described in any one of claims 1-8.
