A multi-wavelength ppg non-invasive blood glucose estimation system based on spectral reconstruction

Abstract

This application relates to the field of non-invasive blood glucose detection technology and provides a multi-wavelength PPG non-invasive blood glucose estimation system based on spectral reconstruction. The system includes: an acquisition module for acquiring the emission spectral response curve of the light-emitting unit and the spectral response curve of the photoelectric receiving unit of a signal acquisition device; an acquisition module for acquiring demographic information of the subject, the subject's PPG signal, temperature and humidity signals, and time information; a determination module for determining the target absorption curve of the subject; a reconstruction module for spectrally reconstructing the subject's PPG signal to obtain the subject's spectral tensor; an encoding module for jointly encoding the spectral tensor, temperature and humidity signals, and time information to obtain fused features; and an estimation module for inputting the fused features into a blood glucose estimation model to obtain the blood glucose estimation result for the subject. This application can improve the accuracy and individual adaptability of non-invasive blood glucose estimation.

Description

Technical Field

[0001] This application belongs to the field of non-invasive blood glucose detection technology, and in particular relates to a multi-wavelength PPG non-invasive blood glucose estimation system based on spectral reconstruction. Background Technology

[0002] Photoplethysmography (PPG) is widely used for detecting parameters related to pulse, blood oxygen, blood pressure, and metabolic status due to its advantages of being non-invasive, portable, and low-cost. For blood glucose estimation, multi-wavelength PPG can reflect the differences in light absorption by blood and tissues at different wavelengths, and is therefore considered to have application potential.

[0003] However, existing technologies still have the following shortcomings. First, many methods directly use the PPG waveform acquired by the sensor as the model input, without explicitly considering the coupling effect between the broadband emission of the light-emitting device, tissue transmittance, and photoelectric receiving response. Since commercial LEDs typically have a wide emission bandwidth, the actual acquired PPG signal is essentially an integral response within a certain wavelength range, rather than a response at a single precise wavelength, leading to aliasing and weakening of fine-grained spectral information related to glucose. Second, existing methods often treat the multi-wavelength LED-photoelectric receiving path as a fixed black box, ignoring the emission spectral characteristics and receiving response characteristics of the device itself, making it difficult to fully extract spectral information. Third, there are significant differences in skin thickness, vascular distribution, and tissue composition among subjects, resulting in significant individual differences in tissue optical properties. However, most existing methods use a uniform modeling approach, failing to effectively characterize the stable spectral fingerprint of each individual, thus resulting in poor accuracy and individual adaptability of non-invasive blood glucose estimation. Summary of the Invention

[0004] This application provides a multi-wavelength PPG non-invasive blood glucose estimation system based on spectral reconstruction, which can solve the problems of poor accuracy and individual adaptability of non-invasive blood glucose estimation.

[0005] This application provides a multi-wavelength PPG non-invasive blood glucose estimation system based on spectral reconstruction, including: The acquisition module is used to acquire the emission spectral response curves of the light-emitting unit of the signal acquisition device under multiple wavelength channels and the spectral response curves of the photoelectric receiving unit of the signal acquisition device under multiple wavelength channels. The acquisition module is used to collect demographic information of the subjects under test. It uses a signal acquisition device to collect PPG signals of the subjects under test in multiple wavelength channels, as well as temperature and humidity signals corresponding to the PPG signals, and collects time information corresponding to the PPG signals. The determination module is used to determine the target absorption curve of the test subject in each wavelength channel based on the demographic information of the test subject, the demographic information of multiple training subjects, and the absorption curves of multiple training subjects in multiple wavelength channels. The absorption curves of the training subjects in multiple wavelength channels are obtained during the training process and stored in the absorption curve table. The absorption curve of the training subject in each wavelength channel is used to describe the individual characteristics of the training subject in that wavelength channel. The reconstruction module is used to perform spectral reconstruction of the PPG signal of the test subject in the wavelength channel for each of the multiple wavelength channels, based on the target absorption curve of the test subject in the wavelength channel, the emission spectral response curve of the light-emitting unit in the wavelength channel, and the spectral response curve of the photoelectric receiving unit in the wavelength channel, to obtain the spectral tensor of the test subject in the wavelength channel. The encoding module is used to perform joint feature encoding on the reconstructed spectral tensor, temperature and humidity signals, and time information to obtain fused features. The estimation module is used to input the fused features into the blood glucose estimation model for estimation, and obtain the blood glucose estimation results of the subject to be tested.

[0006] Optionally, determine the specific module used to perform the following steps: Based on the demographic information of the test subjects, obtain the demographic feature vectors of the test subjects; Based on the demographic information of each training subject, obtain the demographic feature vector of each training subject; For each training subject, calculate the Euclidean distance between the demographic feature vector of that training subject and the demographic feature vector of the test subject. Sort all calculated Euclidean distances in ascending order, and then sort the results. The training subjects corresponding to each Euclidean distance are used as One nearest neighbor training subject; For each of the multiple wavelength channels, The average of the absorption curves of the nearest neighbor training subjects in the wavelength channel is used as the target absorption curve of the subject to be tested in the wavelength channel.

[0007] Optionally, the refactoring module is specifically used to perform the following steps: The subject's score on the [number]th day was calculated using the following formula. Reconstructed spectral shape vector under each wavelength channel : ; The subject's score on the [number]th day was calculated using the following formula. spectral tensor under each wavelength channel : ; in, , Indicates the number of wavelength channels. Indicates the light-emitting unit in the first... Emission spectral response curves for each wavelength channel. Indicates the subject's time on Target absorption curves for each wavelength channel. Indicates that the photoelectric receiving unit is in the first Spectral response curves for each wavelength channel. Indicates the subject's time on PPG signal under each wavelength channel.

[0008] Optionally, the encoding module is specifically used to perform the following steps: The spectral tensor of the subject under each wavelength channel is normalized, and all normalized spectral tensors are spliced ​​together to obtain a multi-wavelength spectral tensor. The temperature and humidity signals are spliced ​​with the multi-wavelength spectral tensor along the channel dimension to obtain the multimodal splicing input; The encoder input features are obtained based on multimodal concatenation input and temporal information; The encoder input features are input into the FEDformer model for processing to obtain global features; The global features are input into the PatchTST model for processing to obtain local features; The demographic feature vector and local features of the test subjects are concatenated to obtain the fused features; the demographic feature vector of the test subjects is obtained based on the demographic information of the test subjects.

[0009] Optionally, the encoding module obtains the encoder input features by performing the following steps: Through formula Calculate the encoder input features ; in, Indicates value embedding function, This indicates a multimodal concatenated input. This indicates time information, including hour, day, month, and year. Indicates the time embedding function, This represents the final temporal embedding vector; ; , Indicates time information The result obtained by periodically encoding each time component; ; This represents a normalized time scalar representing any one of the time components: hour, day, month, or year. Representing time scalars The periodic encoding results.

[0010] Optionally, the absorption curves of the training subjects at multiple wavelength channels are generated by the absorption curve generation network during the training process.

[0011] Optionally, the absorption curve generation network includes an embedding layer, a fully connected layer, and a non-negative activation layer connected in sequence; The embedding layer is used to map the identification information of the training subjects into subject feature vectors in a continuous latent space, so as to transform the identification information of the training subjects into a learnable low-dimensional representation. The fully connected layer is used to perform a non-linear mapping on the subject feature vector output by the embedding layer to generate an absorption curve; The non-negative activation layer is used to impose a non-negative constraint on the absorption curve of the fully connected layer output, so that the absorption value of the absorption curve at each discrete wavelength point is greater than or equal to zero.

[0012] Optionally, multiple wavelength channels correspond one-to-one with multiple center wavelengths, including 665nm, 905nm, 1200nm, 1300nm, 1460nm and 1550nm.

[0013] The above-mentioned solution in this application has the following beneficial effects: In the embodiments of this application, a target absorption curve for the subject to be tested is determined based on the absorption curves of multiple training subjects. Based on this target absorption curve, the emission spectral response curve of the light-emitting unit of the PPG signal acquisition device, and the spectral response curve of the photoelectric receiving unit, the PPG signal of the subject to be tested is spectrally reconstructed. Then, the reconstructed spectral tensor, the temperature and humidity signals corresponding to the PPG signal of the subject to be tested, and time information are jointly feature-encoded. Finally, the encoded fusion features are processed using a blood glucose estimation model to obtain the blood glucose estimation result of the subject to be tested. Specifically, since this application no longer simply treats the PPG signal acquired by the broadband LED as a single waveform input, but reconstructs the PPG signal based on the emission spectral response curve, the absorption curve characterizing tissue transmittance, and the spectral response curve, this application can fully exploit spectral domain information and recover finer-grained wavelength-level information from the integral response, mitigating the spectral aliasing problem caused by broadband emission, thereby improving the accuracy and individual adaptability of non-invasive blood glucose estimation. Other beneficial effects of this application will be described in detail in the following detailed description section. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art 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.

[0015] Figure 1 This is a schematic diagram of the structure of a multi-wavelength PPG non-invasive blood glucose estimation system based on spectral reconstruction provided in an embodiment of this application; Figure 2 A schematic diagram of the spectral characteristics and signal formation in a six-wavelength PPG system; Figure 3 A visual representation of PPG signals and reconstructed values ​​across multiple wavelengths; Figure 4 A detailed visualization of PPG reconstructed signals across multiple wavelengths; Figure 5 This is a Parkes error grid analysis diagram in practice; Figure 6 This is a Bland-Altman analysis diagram in progress; Figure 7 For an implementation, embed dimension D and contrast weights ISO precision heatmap; Figure 8 A visualization of the embedded features of a subject during implementation; Figure 9 This is a standardized absorbance curve of a subject during the experiment. Detailed Implementation

[0016] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0017] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0018] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0019] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0020] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0021] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0022] To address the issues of poor accuracy and individual adaptability in current non-invasive blood glucose estimation methods, this application provides a multi-wavelength PPG non-invasive blood glucose estimation system based on spectral reconstruction. This system determines the target absorption curve for the test subject based on the absorption curves of multiple training subjects. Based on this target absorption curve, the emission spectral response curve of the light-emitting unit of the PPG signal acquisition device, and the spectral response curve of the photoelectric receiving unit, the PPG signal of the test subject is spectrally reconstructed. Then, the reconstructed spectral tensor, the temperature and humidity signals corresponding to the PPG signal of the test subject, and time information are jointly feature-encoded. Finally, the encoded fusion features are processed using a blood glucose estimation model to obtain the blood glucose estimation result for the test subject. Notably, this application does not simply treat the PPG signal acquired by the broadband LED as a single waveform input. Instead, it reconstructs the PPG signal based on the emission spectral response curve, the absorption curve characterizing tissue transmittance, and the spectral response curve. This allows the application to fully extract spectral domain information and recover finer-grained wavelength-level information from the integral response, mitigating the spectral aliasing problem caused by broadband emission, thereby improving the accuracy and individual adaptability of non-invasive blood glucose estimation.

[0023] The following describes the multi-wavelength PPG non-invasive blood glucose estimation system based on spectral reconstruction provided in this application by way of specific embodiments.

[0024] like Figure 1 As shown in the embodiments of this application, the multi-wavelength PPG non-invasive blood glucose estimation system based on spectral reconstruction includes an acquisition module, a collection module, a determination module, a reconstruction module, an encoding module, and an estimation module.

[0025] The acquisition module is used to acquire the emission spectral response curves of the light-emitting unit of the signal acquisition device under multiple wavelength channels and the spectral response curves of the photoelectric receiving unit of the signal acquisition device under multiple wavelength channels.

[0026] The aforementioned signal acquisition device is a PPG signal acquisition device, comprising: a light-emitting unit for generating light of various center wavelengths, a photoelectric receiving unit, and a temperature and humidity acquisition unit (specifically, a temperature and humidity sensor) for acquiring temperature and humidity signals. The photoelectric receiving unit can specifically include a transmissive photoelectric receiving unit and a reflective photoelectric receiving unit. The transmissive photoelectric receiving unit receives the light signal (emitted by the light-emitting unit) after it has passed through the subject's finger and converts the received light signal into a PPG signal. The reflective photoelectric receiving unit receives the light signal (emitted by the light-emitting unit) after it has been reflected by the subject's finger and converts the received light signal into a PPG signal. It should be noted that the basic structure of the signal acquisition device can adopt the signal acquisition device described in patent application number 202411530032.2. Furthermore, the number of temperature and humidity acquisition units, transmissive photoelectric receiving units, and reflective photoelectric receiving units in this application can be set according to actual conditions.

[0027] like Figure 2 As shown in (a) the relative spectral power distribution of the six LEDs; (b) the photodiode responsivity of two wavelength-dependent photodiodes; (c) two tissue transmittance curves; (d) the normalized element-wise spectral output of the integrated spectral signal), in a six-wavelength PPG system, the actual acquired PPG signal is not a response of a single precise wavelength, but rather an integral response formed by broadband light emitted by the emitting unit within a certain bandwidth, propagating through tissue, and then being received by the photodetector unit. This represents the relative spectral power distribution of the six LED channels. This indicates the responsivity of two photodiodes. This represents the tissue transmittance curve (measured by different instruments). This represents the normalized element-wise spectral output corresponding to the six LED channels. In some embodiments of this application, the original PPG signal... It can be represented as: ; in, and These represent the effective wavelength range boundaries of the corresponding light-emitting channels (i.e., wavelength channels). wavelength The overall spectral response at the location.

[0028] Furthermore, the overall spectral response is derived from the emission spectral response of the emitting unit. Organizational penetration rate and photoelectric receiving unit response Joint decision, namely: ; Wherein, ⊙ represents element-wise multiplication, and the tissue permeability and tissue absorption curve satisfy: ; in, Indicates the subject at wavelength The absorption curve at that location.

[0029] For ease of calculation, wavelengths can be selected uniformly within the wavelength range. Discrete wavelength points The sampling interval is Δ Then we have: ; This discretizes the broadband integral PPG signal into a superposition of multiple sub-band responses. This discretization provides the foundation for subsequent learning of individual absorption curves and spectral reconstruction. Wavelength point The overall spectral response at the location.

[0030] It should be noted that the emission spectral response curves of the luminescent unit under multiple wavelength channels and the spectral response curves of the photodetector unit under multiple wavelength channels can be obtained through calibration (if there are multiple photodetector units, the spectral response curves of each photodetector unit under multiple wavelength channels can be obtained through calibration) and are used as fixed prior quantities during training and inference; while the absorption curve (i.e., tissue absorption curve) is a subject-related quantity that varies with individual tissue thickness, vascular structure and tissue composition, and is therefore used as the core characterization object of individual spectral fingerprint.

[0031] The acquisition module is used to collect demographic information of the subjects under test. It uses a signal acquisition device to collect PPG signals of the subjects under test in multiple wavelength channels, as well as the temperature and humidity signals corresponding to the PPG signals, and collects the time information corresponding to the PPG signals.

[0032] The subjects mentioned above are individuals requiring blood glucose estimation. Their demographic information is presented as a four-dimensional vector, including age, gender, height, and weight. The various light emitted by the emitting unit corresponds one-to-one with multiple wavelength channels. The PPG signal for each wavelength channel refers to the PPG signal output by the photoelectric receiving unit when the emitting unit emits the light corresponding to that wavelength channel. It should be noted that if there are multiple photoelectric receiving units, the PPG signals output by the multiple photoelectric receiving units are weighted and averaged, and the weighted average result is used as the PPG signal for that wavelength channel.

[0033] In some embodiments of this application, multiple wavelength channels correspond one-to-one with multiple center wavelengths (i.e., the center wavelength of each of the above-mentioned light sources), including 665nm, 905nm, 1200nm, 1300nm, 1460nm and 1550nm.

[0034] The shape of the above PPG signal is as follows The temperature and humidity signal includes a temperature signal and a humidity signal, and the shapes of the temperature signal and the humidity signal are as follows: , Indicates the number of samples.

[0035] The temperature and humidity signals mentioned above refer to the ambient temperature and humidity of the environment where the signal acquisition device was located when the PPG signal was acquired. It should be noted that if there are multiple temperature and humidity acquisition units, the temperature and humidity signals output by the multiple units (some units output the temperature and humidity of the subject's skin surface, while others output the ambient temperature and humidity) are weighted and averaged, and the weighted average result is used as the aforementioned temperature and humidity signal. The temperature and humidity signal includes temperature and humidity information. The time information mentioned above refers to the time of PPG signal acquisition.

[0036] The determination module is used to determine the target absorption curve for the test subject in each wavelength channel based on the demographic information of the test subject, the demographic information of multiple training subjects, and the absorption curves of the multiple training subjects in multiple wavelength channels. The absorption curves of the training subjects in multiple wavelength channels are obtained during the training process and stored in the absorption curve table. The absorption curve of the training subject in each wavelength channel is used to describe the individual characteristics of the training subject in that wavelength channel, thereby characterizing the tissue transmittance of the training subject in that wavelength channel.

[0037] In some embodiments of this application, the determining module described above is specifically used to perform steps 11.1 to 11.5: Step 11.1: Based on the demographic information of the test subjects, obtain the demographic feature vector of the test subjects.

[0038] Specifically, the demographic information of the test subject can be input into the embedding layer for processing to obtain the demographic feature vector of the test subject.

[0039] Step 11.2: Based on the demographic information of each training subject, obtain the demographic feature vector of each training subject.

[0040] Specifically, the demographic information of the training subjects (including age, gender, height, weight, etc.) can be input into the embedding layer for processing to obtain the demographic feature vector of the training subjects.

[0041] Step 11.3: For each training subject, calculate the Euclidean distance between the demographic feature vector of the training subject and the demographic feature vector of the test subject.

[0042] Step 11.4: Sort all calculated Euclidean distances in ascending order, and then sort the results. The training subjects corresponding to each Euclidean distance are used as One nearest neighbor training subject.

[0043] Step 11.5, for each of the multiple wavelength channels, respectively, The average of the absorption curves of the nearest neighbor training subjects in the wavelength channel is used as the target absorption curve of the subject to be tested in the wavelength channel.

[0044] It is worth mentioning that by using a similar subject retrieval mechanism to estimate the target absorption curve for unknown subjects (i.e., test subjects), the model can still obtain individualized spectral reconstruction results when facing individuals who have not been seen, thereby enhancing the model's generalization ability and practicality.

[0045] The reconstruction module is used to perform spectral reconstruction of the PPG signal of the test subject in the wavelength channel for each of the multiple wavelength channels, based on the target absorption curve of the test subject in the wavelength channel, the emission spectral response curve of the emitting unit in the wavelength channel, and the spectral response curve of the photoelectric receiving unit in the wavelength channel, to obtain the spectral tensor of the test subject in the wavelength channel.

[0046] In some embodiments of this application, the above-described reconstruction module is specifically used to perform steps 12.1 to 12.2: Step 12.1, calculate the subject's score on the [number]th [day] using the following formula. Reconstructed spectral shape vector under each wavelength channel : ; Step 12.2, calculate the number of test subjects in the following formula at the first... spectral tensor under each wavelength channel : ; in, , Indicates the number of wavelength channels. Indicates the light-emitting unit in the first... Emission spectral response curves for each wavelength channel. Indicates the subject's time on Target absorption curves for each wavelength channel. Indicates that the photoelectric receiving unit is in the first Spectral response curves for each wavelength channel. Indicates the subject's time on PPG signal under each wavelength channel. This represents the transmittance obtained from the absorption curve (calculated according to Lambert-Beer's Law).

[0047] It should be noted that if the signal acquisition device includes multiple photoelectric receiving units, Multiple photoelectric receiving units can be used in the first The average or weighted average of the spectral response curves for each wavelength channel.

[0048] like Figure 3 As shown, visualizations of the original PPG signals and reconstructed values ​​across multiple wavelengths are presented ((a) original PPG signal; (b) reconstructed PPG signal, with different colors representing different wavelengths). The original PPG signal reflects the waveform changes of different wavelength channels over time; the reconstructed values ​​correspond to the two-dimensional or three-dimensional spectral representation formed in both the wavelength and time dimensions after expanding the broadband response. Compared to the original waveform, the reconstructed results not only preserve the temporal information of the pulse waveform but also introduce the discrete wavelength structure information within each channel, thus enabling a more granular characterization of the differences in tissue absorption and transmission of light at different wavelengths.

[0049] like Figure 4 The diagram shows a detailed visualization of the PPG reconstructed signal across multiple wavelengths ((a) PPG reconstructed signal corresponding to the 1200nm center wavelength; (b) PPG reconstructed signal corresponding to the 1300nm center wavelength; (c) PPG reconstructed signal corresponding to the 1460nm center wavelength; (d) PPG reconstructed signal corresponding to the 1550nm center wavelength; (e) PPG reconstructed signal corresponding to the 665nm center wavelength; (f) PPG reconstructed signal corresponding to the 905nm center wavelength). This diagram can be used to illustrate the differences in energy distribution across wavelength channels before and after reconstruction, as well as the detailed changes in wavelength dimensions. Through this reconstruction process, the fine-grained spectral information that was originally aliased in the broadband integral is unfolded, thus providing an input basis for the subsequent encoder to simultaneously model the "spectral domain relationship" and the "temporal domain relationship".

[0050] The encoding module is used to perform joint feature encoding on the reconstructed spectral tensor, temperature and humidity signals, and time information to obtain fused features.

[0051] In some embodiments of this application, the above-described encoding module is specifically used to perform steps 13.1 to 13.6: Step 13.1: Normalize the spectral tensor of the subject under each wavelength channel, and stitch together all the normalized spectral tensors to obtain a multi-wavelength spectral tensor.

[0052] Step 13.2: The temperature and humidity signals are spliced ​​with the multi-wavelength spectral tensor along the channel dimension to obtain the multimodal splicing input.

[0053] It should be noted that before stitching the temperature and humidity signals with the multi-wavelength spectral tensor, the temperature and humidity signals need to be interpolated first. Specifically, the temperature and humidity signals (i.e., temperature and humidity signals) are interpolated or upsampled along the time axis to the same time resolution as the PPG signal. Given the original sampling time index of the temperature and humidity signals. and PPG sampling time index Its linear interpolation function Written as: ; in, This represents the temperature or humidity signal to be interpolated. This represents the original sampling time index sequence of the temperature and humidity signal; This represents the sampling time index sequence of the PPG signal; Indicates the target interpolation time, and ; and express The middle is located at the target interpolation time. The adjacent sampling times on both sides, and satisfying ; and They represent signals respectively. At any moment and The value at that point. This formula represents the value obtained by using two adjacent sampling points at time. Linear estimation is performed on the signal value.

[0054] Interpolated temperature and humidity signal (i.e., combined temperature and humidity signal) It can be represented as: ; in, Represents a temperature signal. Indicates humidity signal, This indicates that the temperature and humidity signals are spliced ​​together according to channel dimensions. Indicates the batch size. It is kept consistent with the PPG signal in the time dimension to facilitate subsequent multimodal fusion.

[0055] After interpolation, the temperature and humidity signals The multi-wavelength spectral tensor is stitched together along the channel dimension to obtain the multi-modal stitched input. .

[0056] Step 13.3: Obtain encoder input features based on multimodal splicing input and time information.

[0057] Specifically, the encoding module obtains the encoder input features by performing the following steps: Through formula Calculate the encoder input features .

[0058] in, This indicates a value embedding function, used to... Projecting onto a unified latent space, its computational form is as follows: ; This indicates a multimodal concatenated input. This indicates time information, including hour, day, month, and year. This represents a time embedding function used to embed time information. Encoding as time features; This represents the final temporal embedding vector. Represents the learnable weights. This represents a learnable bias.

[0059] ; This is used to map the original periodic encoding result to the same feature dimension as the value embedding; Indicates time information The result is obtained by periodically encoding each time component.

[0060] ; This represents a normalized time scalar representing any one of the time components: hour, day, month, and year (i.e., In order, they are the time scalars after normalization for hours, days, months, and years. Representing time scalars The periodic encoding results. and These represent the periodic characteristics obtained based on the sine and cosine functions, respectively. Through the above mapping, time variables can be transformed into a periodic representation to express periodic patterns such as diurnal variations and seasonal changes.

[0061] Step 13.4: Input the encoder input features into the FEDformer model for processing to obtain global features.

[0062] Step 13.5: Input the global features into the PatchTST model for processing to obtain local features.

[0063] Step 13.6: The demographic feature vector and local features of the test subject are concatenated to obtain the fused features; the demographic feature vector of the test subject is obtained based on the demographic information of the test subject.

[0064] It is worth mentioning that the global features output by the FEDformer model mainly represent long-range spectral time dependence, while the PatchTST model further extracts short-term local waveform change features by performing local patch partitioning, position encoding, and Transformer encoding on the global features. Therefore, it can complement the aforementioned global time-series encoding module and jointly improve the accuracy and stability of blood glucose estimation.

[0065] The estimation module is used to input the fused features into the blood glucose estimation model for estimation, and obtain the blood glucose estimation results of the subject to be tested.

[0066] The blood glucose estimation model described above can be a fully connected layer. By inputting the fused features into this fully connected layer for processing, the blood glucose estimation result, i.e., the blood glucose value, can be obtained.

[0067] Overall, the system of this application is divided into three stages. In the first stage, a corresponding individual absorption curve is constructed for each subject, and the original PPG integral signal is reconstructed by expanding the absorption curve. In the second stage, the reconstructed spectral tensor is concatenated with the upsampled temperature and humidity signal, and then the time embedding is superimposed and input into a unified encoder. In the third stage, an embedding representation for contrastive learning is generated by the projection head, and the fused features are input into the regression head for blood glucose value prediction.

[0068] The absorption curve and the model training process are illustrated below.

[0069] The absorption curves of the training subjects at multiple wavelength channels were generated by an absorption curve generation network during the training process. This absorption curve generation network consists of an embedded layer, a fully connected layer, and a non-negative activation layer connected in sequence.

[0070] The embedding layer is used to map the identification information of the training subjects (such as unique identification numbers) into subject feature vectors in a continuous latent space, so as to transform the identification information of the training subjects into a learnable low-dimensional representation, enabling the model to characterize the stable tissue optical differences between different subjects.

[0071] The fully connected layer is used to perform a nonlinear mapping on the subject feature vector output by the embedding layer to generate absorption curves. Specifically, this fully connected layer performs a nonlinear mapping on the subject feature vector and, in conjunction with wavelength channel information, projects the relevant subject features onto the absorption spectrum space to generate absorption curve parameters for the corresponding wavelength channel.

[0072] The non-negative activation layer is used to impose non-negative constraints on the absorption curve output by the fully connected layer, ensuring that the absorption value at each discrete wavelength point is greater than or equal to zero. Specifically, this non-negative activation layer is used to impose non-negative constraints on the absorption curve parameters, ensuring that the absorption value at each generated discrete wavelength point is greater than or equal to zero, thereby satisfying the physical meaning of the absorption curve and improving the numerical stability of the training process.

[0073] It should be noted that the model training utilizes PPG signals and real blood glucose values ​​(used as labels) from multiple training subjects. The process of calculating blood glucose estimates during training is the same as that in the inference process described earlier (i.e., for each training subject, wavelength channels are separated, and based on the absorption curve, emission spectral response curve, and spectral response curve, the corresponding PPG signal is spectrally reconstructed to obtain a spectral tensor. Then, the spectral tensor, temperature and humidity signals, and time information are jointly coded as features. Finally, the blood glucose estimation model is used to process the fused features obtained by encoding to obtain the blood glucose estimate). The difference is that the absorption curves of the training subjects are generated using an absorption curve generation network during training, and this absorption curve generation network also participates in the training. That is, in each iteration of training, when the model parameters (including all learnable parameters in the blood glucose estimation process, such as learnable parameters in the FEDformer model, PatchTST model, and blood glucose estimation model) are adjusted based on the output blood glucose estimate and loss function, the model parameters of the absorption curve generation network are also adjusted accordingly.

[0074] In some embodiments of this application, supervised contrastive learning can be used to train the model, so that the sample representations of the same training subject are clustered in the feature space and the sample representations of different training subjects are separated in the feature space; and the supervised contrastive learning objective and the blood glucose estimation loss are used together as a joint training objective to train the model.

[0075] For example, model training employs a joint loss function, including regression loss and supervised contrast loss. The regression loss, used to constrain the error between the predicted and reference blood glucose values, can be the mean squared error. The supervised contrast loss, used to enhance aggregation between samples from the same subject (i.e., training subjects) and separation between samples from different subjects, can be calculated based on sample projection vectors and subject identifiers.

[0076] The effects of this application will be illustrated below with specific examples.

[0077] In this embodiment, a dataset was constructed using the data collection process corresponding to the oral glucose tolerance test, yielding 26,063 valid samples from 254 subjects, including 52 diabetic subjects. The data was divided into training, validation, and test sets in a 3:1:1 ratio, and further subject-level five-fold cross-validation was performed. Model inputs included six-channel PPG signals, three-channel temperature signals, three-channel humidity signals, time information, and demographic information. Evaluation metrics included root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), ISO 15197:2013 accuracy (ISO 15197:2013 is a standard number), and the percentage of Parkes error grid (PEG) A-zone.

[0078] To verify the effectiveness of the proposed SRFormer (i.e., the system provided in this application), this study selected several representative non-invasive blood glucose estimation methods for comparison, including traditional deep learning models, feature fusion methods, and state-of-the-art methods based on clustering or hierarchical modeling. As shown in Table 1, SRFormer achieves state-of-the-art performance across all evaluation metrics, with overall performance significantly outperforming existing non-invasive blood glucose estimation methods. In Table 1, CSS-DNN is a context-aware sparse deep neural network, MMG-NET is a multimodal blood glucose sensing network, VCC is a variable feature clustering correction module, DIROS is a dynamic insulin blood glucose optimization inference system, CBHFF is a cross-layer bidirectional hybrid feature fusion module, and DBSCAN-CNN is a density clustering fusion convolutional neural network.

[0079] Table 1 Performance comparison with existing studies

[0080] like Figure 5 The image shows a Parkes error grid analysis plot from one embodiment. Most of the scatter plots for the predicted and reference values ​​fall within region A, indicating that the method described in this application has good clinical acceptability. Figure 5 Zone A represents a clinically accurate prediction, meaning the prediction supports the correct treatment decision; Zone B indicates a possible error, but generally does not lead to significantly inappropriate clinical management; Zone C indicates the prediction error may lead to unnecessary treatment; Zone D indicates the prediction error may lead to the omission of necessary treatment; Zone E indicates the prediction error may lead to a treatment decision contrary to the correct approach. The clinical risk increases progressively from Zone A to Zone E.

[0081] like Figure 6The figure shown is a Bland-Altman consistency analysis diagram from one embodiment. The average deviation is small, and most sample points are within the 95% consistency limit, indicating that the predicted values ​​of the method in this application have good consistency with the reference blood glucose values.

[0082] As shown in Table 2, removing any module leads to significant degradation, and the magnitude of degradation is consistent across different metrics. This indicates that global-local joint temporal modeling constitutes the key structure for performance. Based on this, the SRB acts as an amplifier and regularizer on the spectral side, providing more discriminative input to the temporal backbone through individualized spectral reconstruction, thereby further reducing the error to an optimal level. In Table 2, FED represents the FEDformer model, PAT represents the PatchTST model, and SRB represents the reconstruction module. Indicates that it is disabled. Indicates activation.

[0083] Table 2 Module Ablation Experiment

[0084] like Figure 7 As shown, the embedding dimension D and the contrast weights are... The ISO accuracy heatmap is shown. In this embodiment, when the embedding dimension is 3 and the contrast weight is 0.1, the ISO accuracy reaches a relatively good level. This figure shows that there is a synergistic relationship between the spectral discrete dimension and the supervised contrast learning intensity used in this application, and that appropriate contrast constraints are beneficial to improving the accuracy of blood glucose detection.

[0085] like Figure 8 The figure shows a visualization of the embedding features of five random subjects ((a) Principal Component Analysis (PCA) projection, (b) t-distributed random nearest neighbor embedding (t-SNE) projection). The different subject samples exhibit good intra-class clustering and inter-class separation in the feature space, indicating that spectral reconstruction and supervised contrastive learning can effectively extract stable features relevant to the subjects. The numbers following each subject in the figure represent their identification information (i.e., their identification number).

[0086] like Figure 9 As shown, the normalized absorbance curves of five randomized subjects are presented ((a)–(f) represent the normalized spectral curves of LED channels 1 to 6). Significant differences exist in the absorption curves of different subjects across multiple wavelength channels, indicating that the individual absorption curves learned in this application can reflect individual differences in the optical properties of the subjects' tissues, thus providing a basis for personalized blood glucose testing.

[0087] In summary, this application constructs a spectral reconstruction mechanism based on individual absorption curves, expanding broadband integral multi-wavelength PPG signals into multi-wavelength spectral tensors containing individual spectral fingerprints. This is then combined with temperature, humidity, time, and demographic information for unified encoding and blood glucose regression, effectively improving the accuracy, stability, and individual adaptability of non-invasive blood glucose testing. Compared to schemes that directly model only the original PPG waveform, this application can more fully recover and utilize spectral domain information and explicitly consider cross-individual tissue optical differences, thus making it more suitable for blood glucose testing tasks in real-world applications.

[0088] Specifically, this application has at least the following beneficial effects: Firstly, this application no longer simply regards the PPG signal collected by broadband LED as a single waveform input. Instead, it establishes a discrete spectral model based on the emission spectral response, tissue transmittance, and photoelectric receiving response. This model can recover finer-grained wavelength-level information from the integral response, thereby mitigating the spectral aliasing problem caused by broadband luminescence.

[0089] Secondly, this application uses individual absorption curves and absorption curves to explicitly model the stable tissue optical differences of subjects, and introduces individual spectral fingerprints into the multi-wavelength PPG reconstruction process, which helps to improve the model's ability to adapt to cross-individual differences.

[0090] Third, this application estimates the target absorption curve for unknown subjects through a similar subject retrieval mechanism, enabling the model to obtain individualized spectral reconstruction results even when facing unseen individuals, thereby enhancing the model's generalization ability and practicality.

[0091] Fourth, this application, by jointly modeling the reconstructed multi-wavelength spectral tensor with temperature and humidity signals, time information and demographic information, and combining global temporal coding, local temporal coding and supervised contrastive learning, can simultaneously characterize long-range spectral time dependence, local waveform changes and individual identity-related features, thereby further improving the accuracy, stability and robustness of blood glucose estimation.

[0092] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A multi-wavelength PPG non-invasive blood glucose estimation system based on spectral reconstruction, characterized in that, include: The acquisition module is used to acquire the emission spectral response curves of the light-emitting unit of the signal acquisition device under multiple wavelength channels and the spectral response curves of the photoelectric receiving unit of the signal acquisition device under multiple wavelength channels; The acquisition module is used to acquire demographic information of the subject to be tested, acquire PPG signals of the subject to be tested in multiple wavelength channels using the signal acquisition device, as well as temperature and humidity signals corresponding to the PPG signals, and acquire time information corresponding to the PPG signals. The determination module is used to determine the target absorption curve of the test subject in each wavelength channel based on the demographic information of the test subject, the demographic information of multiple training subjects, and the absorption curves of multiple training subjects in multiple wavelength channels. The absorption curves of the training subjects in multiple wavelength channels are obtained during the training process and stored in the absorption curve table. The absorption curve of the training subject in each wavelength channel is used to describe the individual characteristics of the training subject in that wavelength channel. The reconstruction module is used to perform spectral reconstruction of the PPG signal of the subject under the wavelength channel for each of the multiple wavelength channels, based on the target absorption curve of the subject under the wavelength channel, the emission spectral response curve of the luminescent unit under the wavelength channel, and the spectral response curve of the photoelectric receiving unit under the wavelength channel, to obtain the spectral tensor of the subject under the wavelength channel. The encoding module is used to perform joint feature encoding on the reconstructed spectral tensor, the temperature and humidity signal, and the time information to obtain fused features; The estimation module is used to input the fused features into the blood glucose estimation model for estimation, so as to obtain the blood glucose estimation result of the subject to be tested.

2. The multi-wavelength PPG non-invasive blood glucose estimation system according to claim 1, characterized in that, The determining module is specifically used to perform the following steps: Based on the demographic information of the test subject, obtain the demographic feature vector of the test subject; Based on the demographic information of each training subject, obtain the demographic feature vector of each training subject; For each training subject, calculate the Euclidean distance between the demographic feature vector of that training subject and the demographic feature vector of the subject to be tested. Sort all calculated Euclidean distances in ascending order, and then sort the results. The training subjects corresponding to each Euclidean distance are used as One nearest neighbor training subject; For each of the multiple wavelength channels, The average of the absorption curves of the nearest neighboring training subjects in the wavelength channel is taken as the target absorption curve of the subject under test in the wavelength channel.

3. The multi-wavelength PPG non-invasive blood glucose estimation system according to claim 1, characterized in that, The reconstruction module is specifically used to perform the following steps: The subject of the test was calculated using the following formula at the [number]th [time]. Reconstructed spectral shape vector under each wavelength channel : ； The subject of the test was calculated using the following formula at the [number]th [time]. spectral tensor under each wavelength channel : ； in, , Indicates the number of wavelength channels. Indicates that the light-emitting unit is in the first... Emission spectral response curves for each wavelength channel. This indicates that the subject under test was on the [number]th [day / month]. Target absorption curves for each wavelength channel. Indicates that the photoelectric receiving unit is in the first... Spectral response curves for each wavelength channel. This indicates that the subject under test was on the [number]th [day / month]. PPG signal under each wavelength channel.

4. The multi-wavelength PPG non-invasive blood glucose estimation system according to claim 1, characterized in that, The encoding module is specifically used to perform the following steps: The spectral tensor of the subject under test in each wavelength channel is normalized, and all normalized spectral tensors are spliced ​​together to obtain a multi-wavelength spectral tensor. The temperature and humidity signals are spliced ​​with the multi-wavelength spectral tensor along the channel dimension to obtain a multimodal splicing input; The encoder input features are obtained based on the multimodal splicing input and the time information; The encoder input features are input into the FEDformer model for processing to obtain global features; The global features are input into the PatchTST model for processing to obtain local features; The demographic feature vector of the test subject and the local features are concatenated to obtain the fused feature; the demographic feature vector of the test subject is obtained based on the demographic information of the test subject.

5. The multi-wavelength PPG non-invasive blood glucose estimation system according to claim 4, characterized in that, The encoding module obtains the encoder input features by performing the following steps: Through formula Calculate the encoder input features ; in, Indicates value embedding function, This indicates the multimodal splicing input. This indicates time information, including hour, day, month, and year. Indicates the time embedding function, This represents the final temporal embedding vector; ； Indicates a fully connected projection layer. Indicates time information The result obtained by periodically encoding each time component; ； This represents a normalized time scalar representing any one of the time components among hours, days, months, and years. Representing time scalars The periodic encoding results.

6. The multi-wavelength PPG non-invasive blood glucose estimation system according to claim 1, characterized in that, The absorption curves of the training subjects at multiple wavelength channels are generated by the absorption curve generation network during the training process.

7. The multi-wavelength PPG non-invasive blood glucose estimation system according to claim 6, characterized in that, The absorption curve generation network includes an embedded layer, a fully connected layer, and a non-negative activation layer connected in sequence. The embedding layer is used to map the identification information of the training subjects into subject feature vectors in a continuous latent space, so as to transform the identification information of the training subjects into a learnable low-dimensional representation. The fully connected layer is used to perform nonlinear mapping on the subject feature vector output by the embedding layer to generate an absorption curve; The non-negative activation layer is used to impose a non-negative constraint on the absorption curve output by the fully connected layer, so that the absorption value of the absorption curve at each discrete wavelength point is greater than or equal to zero.

8. The multi-wavelength PPG non-invasive blood glucose estimation system according to claim 1, characterized in that, The multiple wavelength channels correspond one-to-one with multiple center wavelengths, including 665nm, 905nm, 1200nm, 1300nm, 1460nm and 1550nm.

Images