A diabetes detection system based on multi-sensor pulse features

Abstract

This application discloses a diabetes detection system based on multi-sensor pulse characteristics, relating to the field of medical signal processing. The system includes: a pulse wave data acquisition module that uses a pressure sensor and a photoelectric sensor to acquire the user's wrist pressure pulse wave and fingertip photoplethysmography (PPG) pulse wave; a preprocessing module to obtain usable single-cycle pulse wave data; a signal fusion module that converts the pressure pulse wave signal and PPG pulse wave signal into two-dimensional image signals through continuous wavelet transform to retain time-frequency domain information, and processes the two two-dimensional images using non-subsampled shear wave transform and a pulse-coupled neural network; and a feature extraction and calculation module that extracts pulse wave image features through a ResNet network and finally classifies them using a random forest. This detection system improves the fusion effect without increasing computational load, achieving non-invasive diabetes detection using multiple sensors and improving detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of medical signal processing, and more specifically, to a diabetes detection system based on multi-sensor pulse characteristics. Background Technology

[0002] With the development of technology and the improvement of social standards, the number of people suffering from diabetes in modern society is increasing. Diabetes is gradually becoming one of the three major chronic diseases in the world, alongside hypertension and cardiovascular disease. On May 19, 2023, the "World Health Statistics Report 2023" pointed out that diabetes caused 2 million deaths worldwide in just one year. According to data predicted by the authoritative medical journal "The Lancet" in 2023, the number of people with diabetes is expected to increase to 1.3 billion in the next 30 years, and every country will experience growth. Therefore, timely detection of diabetes is particularly important today.

[0003] Currently, the main methods for detecting diabetes include fasting blood glucose tests, oral glucose tolerance tests, glycated hemoglobin (HbA1c) tests, and minimally invasive portable blood glucose meters. Fasting blood glucose tests require patients to fast for more than eight hours before ingesting 237 ml of glucose solution, which can cause discomfort. Oral glucose tolerance tests require patients to ingest an aqueous solution containing 75g of anhydrous glucose, with blood samples taken at 0.5, 1.0, and 2.0 hours to measure blood glucose changes and observe the patient's glucose tolerance; this not only causes stomach discomfort but also requires three blood samples. HbA1c tests and minimally invasive portable blood glucose meters both require blood samples and are relatively expensive. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the problems of existing diabetes detection methods being invasive and painful, and typically focusing on a single signal for prediction, resulting in limited information acquisition and difficulty in comprehensively assessing the body's condition, this application provides a diabetes detection system based on multi-sensor pulse characteristics to overcome at least one of the shortcomings of the prior art, thus solving the problems mentioned in the background art.

[0006] This invention provides a diabetes detection system based on multi-sensor pulse characteristics, comprising:

[0007] S1, Pulse wave data acquisition module, uses array pressure sensor and photoelectric sensor to collect multi-channel pulse wave data;

[0008] S2, Signal preprocessing module, preprocesses multi-channel pulse wave data, including signal filtering, baseline drift removal and period segmentation, to obtain usable single-cycle pressure pulse wave data and photoplethysmography pulse wave data;

[0009] S3, the signal fusion module, converts the pulse wave data obtained in S2 into a two-dimensional image through continuous wavelet transform. The converted two-dimensional image is then decomposed into low-frequency and high-frequency sub-bands using non-subsampled shear wave transform (NSST). The low-frequency sub-bands are then fused using energy attribute weighted averaging, and the high-frequency sub-bands are fused using pulse coupled neural network (PCNN). Finally, the fused image is generated using inverse shear transform. The image fused using this method is superior in terms of detail preservation, contrast, and structural integrity.

[0010] S4, Feature Extraction and Calculation Module, uses a ResNet network to extract features from the fused image and uses a random forest for feature classification to achieve non-invasive and accurate detection of diabetes.

[0011] Furthermore, S1 employs an integrated multi-channel acquisition device for pulse wave signal acquisition in diabetic patients. This device uses a three-channel array pressure sensor to accurately acquire synchronous three-channel pressure pulse wave signals from the subject's wrist. Simultaneously, a high-performance photoelectric sensor is used to acquire the pulse wave signal from the tip of the subject's right index finger.

[0012] The preprocessing module includes signal filtering, baseline drift removal, and period segmentation. Specifically, the signal filtering method involves sequentially passing each group of three-channel pulse wave signals through an FIR low-pass filter with a cutoff frequency of fl to remove high-frequency noise. The amplitude and slope changes of each pulse wave are used to determine whether there is distortion or weak signal in each channel. The autocorrelation function is calculated to detect abrupt changes in the signal in order to eliminate abnormal signals.

[0013] If the signal is periodic, the autocorrelation function (ACF) will show significant peaks at certain delays, corresponding to the period of the signal. Conversely, it will not show obvious periodic peaks.

[0014] Furthermore, the ascending limb of the pulse wave is located during the systolic phase of the heart, where the fluctuations are most pronounced. Its first-order difference value is greater than zero throughout the entire pulse wave cycle. Therefore, by locating the point with the largest first-order difference value in the ascending limb of the pulse wave, which is easier to determine throughout the entire pulse cycle, we can then look to the left for the minimum value point of the waveform, i.e., the starting point.

[0015] Furthermore, the baseline drift is removed by using cubic spline curve fitting. The baseline is fitted based on the above starting point, and the overall signal fluctuation is corrected by calculating the difference between the signal and the baseline. Finally, periodic segmentation is performed based on the above starting point.

[0016] The signal fusion module performs two-dimensional conversion between the preprocessed pressure pulse wave signal and the photoplethysmography pulse wave signal using the continuous wavelet transform method.

[0017] Furthermore, in the signal fusion section of S3, two-dimensional images of pressure pulse wave and photoplethysmography pulse wave are first input, and noise removal and alignment are performed on the images to ensure spatial consistency between the images.

[0018] Non-subsampled shear wave transform (NSST) was applied to the preprocessed two-dimensional images of pressure pulse wave and photoplethysmography pulse wave, respectively, to decompose the images into low-pass subbands and high-pass subbands. The low-pass subbands contain the structural information of the image, while the high-pass subbands contain the detailed information of the image.

[0019] The formula is as follows:

[0020]

[0021] In the formula, j, k, and m are parameters in the shear wave transform, representing the decomposition level, direction, and spatial location, respectively.

[0022] Further, S3 calculates the intrinsic property (IP) and energy attribute (EA) of each low-frequency sub-band. IPA and IPB are intrinsic properties of the low-frequency sub-bands of the two-dimensional images of pressure pulse waves and photoplethysmography pulse waves, typically determined by the mean and median.

[0023] IPA = μA + μ′ A

[0024] IPB = μB + μ′ B

[0025] Where μ is the mean, μ′ is the median, and δ is a scaling factor.

[0026] Energy attributes are a method for measuring the energy of pixels in an image subband, reflecting the structural features of the image. For each pixel location (p, q), the energy attribute for both types of images is calculated using the following formula:

[0027] EAA(p,q)=e -δ|LA(p,q)-IPA|

[0028] EAB(p,q)=e -δ|LB(p,q)-IPB|

[0029] LA(p,q) and LB(p,q) are the pixel values ​​of the low-frequency subbands of the two images at position (p,q), respectively.

[0030] The low-frequency subbands are fused using the Energy Attribute Weighted Average method, as shown in the following formula:

[0031]

[0032] The low-frequency subbands of two-dimensional images of pressure pulse wave and photoplethysmography pulse wave are fused into a single low-frequency subband.

[0033] Furthermore, S3 uses a pulse-coupled neural network (PCNN) to process the high-frequency subband. The PCNN is used to determine the optimal fusion rule for the two pulse information to preserve important structural and texture information.

[0034] The formula is expressed as:

[0035]

[0036] In the formula, h[m] and g[m] are the coefficients of the low-pass filter and the high-pass filter, respectively.

[0037] Based on the output of PCNN, the high-frequency subbands of either pressure pulse wave or photoplethysmography pulse wave are selected as the fused high-frequency subbands.

[0038] Furthermore, S3 uses the inverse shearlet transform to recombine the sub-bands that have undergone feature extraction and fusion processing, reconstructing the fused image, restoring the integrity and consistency of the image, preserving the structural and functional information of the image, thereby providing more accurate details, while retaining the structural features of the input image, suppressing noise, and improving the overall quality of the fused image.

[0039] Furthermore, the generated fused image is input into a pre-trained ResNet network. Through layer-by-layer processing of convolutional layers, pooling layers, and residual modules, deep features of the image are extracted, as well as high-level semantic features, which improves the ability to extract deep features and helps to distinguish different image features.

[0040] Furthermore, the Random Forest (RF) algorithm is used to classify the extracted image features. Random Forest is an ensemble learning algorithm based on an improvement of the decision tree algorithm. It constructs multiple decision trees and combines their predictions to obtain the final classification result. When constructing a Random Forest, a subset of samples is first randomly selected from the training dataset as training data for the current decision tree, and a subset of features is randomly selected from the feature set to construct the decision tree. Then, using the selected training data and features, the decision tree is constructed by splitting nodes until a stopping condition is met. Finally, the deep image features extracted from the ResNet network are input into the Random Forest model.

[0041] Each decision tree in the random forest makes a classification decision based on the input features and outputs its own classification result. Through the above system, diabetes detection based on multi-sensor pulse features is realized.

[0042] Compared with the prior art, the present invention provides a diabetes detection system based on multi-sensor pulse characteristics, which has the following beneficial effects:

[0043] This invention provides a diabetes detection system based on multi-sensor pulse characteristics. It uses pressure sensors and photoelectric sensors to collect the user's pulse wave signals, eliminating the need for invasive procedures and avoiding the risks of pain and infection that may arise from traditional invasive detection methods. This improves user acceptance and convenience, solves the problems of invasiveness and pain in existing diabetes detection methods, and is suitable for users of different ages and physical conditions. It does not require professional operators, has broad applicability and promotional value, and is suitable for home screening of patients.

[0044] This invention employs continuous wavelet transform to convert pulse wave signals into two-dimensional image signals, which better preserves the time-frequency characteristics of the signal, providing a higher-quality data foundation for subsequent image processing and feature extraction. Simultaneously, by utilizing non-subsampled shear wave transform (NSST) and pulse-coupled neural network (PCNN) to process the two-dimensional image, the fusion effect is improved without significantly increasing computational load. This effectively extracts important feature information from the image, further enhancing signal recognizability and improving the reliability of the detection results.

[0045] This invention fuses pressure pulse wave (PPW) and photoplethysmography (PPG) signals, fully utilizing the advantages of both sensor signals and compensating for the limitations of a single sensor signal. PWP signals reflect changes in blood vessel wall pressure, while PPG signals reflect changes in blood volume. The fusion of these two signals provides a more comprehensive picture of the physiological state of human blood vessels and blood, thus offering richer information for diabetes detection and significantly improving accuracy. This contributes to higher early diagnosis rates and better treatment outcomes for diabetes, and is of great significance for diabetes prevention and control. Attached Figure Description

[0046] Figure 1 This is a block diagram of a diabetes detection system based on multi-sensor pulse characteristics according to an embodiment of the present invention;

[0047] Figure 2 This is a schematic diagram of the pulse image fusion process of the present invention;

[0048] Figure 3 This is a diagram of the ResNet network structure of the present invention;

[0049] Figure 4 This is a block diagram of the random forest structure of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Example

[0052] like Figure 1 As shown, a diabetes detection system based on multi-sensor pulse characteristics includes the following specific steps:

[0053] S1: Multi-channel pulse wave data are acquired using an array of pressure sensors and photoelectric sensors.

[0054] In this embodiment, an integrated three-channel pressure sensor is used to collect pressure pulse wave signals from three points on the user's wrist: cun, guan, and chi. A photoelectric sensor is used to collect photoplethysmography pulse waves from the user's right index finger.

[0055] S2: Preprocessing of multi-channel pulse wave data, including signal filtering, baseline drift removal, and period segmentation, can yield three-channel pressure pulse wave data and single-channel photoplethysmography pulse wave data that can be used in a single period.

[0056] The specific steps are as follows:

[0057] The preprocessing module includes signal filtering, baseline drift removal, and period segmentation. Specifically, the signal filtering method involves sequentially passing each group of three-channel pulse wave signals through an FIR low-pass filter with a cutoff frequency of fl to remove high-frequency noise. The pulse wave amplitude and slope changes are used to determine whether each channel of the pulse wave has distortion or weak signal. The autocorrelation function is calculated to detect whether there are abrupt changes in the signal, thereby eliminating abnormal signals.

[0058] If the signal is periodic, the autocorrelation function (ACF) will show significant peaks at certain delays, corresponding to the period of the signal. Conversely, it will not show obvious periodic peaks.

[0059] Furthermore, the ascending limb of the pulse wave is located during the systolic phase of the heart, where the fluctuations are most pronounced. Its first-order difference value is greater than zero throughout the entire pulse wave cycle. Therefore, by locating the point with the largest first-order difference value in the ascending limb of the pulse wave, which is easier to determine throughout the entire pulse cycle, we can then look to the left for the minimum value point of the waveform, i.e., the starting point.

[0060] Furthermore, the baseline drift is removed by using cubic spline curve fitting. The baseline is fitted based on the above starting point, and the overall signal fluctuation is corrected by calculating the difference between the signal and the baseline. Finally, periodic segmentation is performed based on the above starting point.

[0061] S3, the signal fusion module, fuses pulse signals from multiple sensors. The specific operation steps are as follows:

[0062] A1: The preprocessed pressure pulse wave signal and photoplethysmography pulse wave signal are converted into two dimensions using the continuous wavelet transform method.

[0063] A2: As Figure 2 As shown, non-subsampled shear wave transform (NSST) is applied to the preprocessed two-dimensional images of pressure pulse wave and photoplethysmography pulse wave, respectively, to decompose the images into low-pass subbands and high-pass subbands. The formula is as follows:

[0064]

[0065] In the formula, j, k, and m are parameters in the shear wave transform, representing the decomposition level, direction, and spatial location, respectively.

[0066] The low-frequency subband contains structural information of the image, while the high-frequency subband contains detailed information.

[0067] A2: Furthermore, since pressure pulse waves and photoplethysmography pulse waves have different information focuses, this example uses the Energy Attribute Weighted Average method to fuse low-frequency subbands and calculate the Intrinsic Property (IP) and Energy Attribute (EA) of each low-frequency subband.

[0068] IPA and IPB are intrinsic properties of the low-frequency subbands of two-dimensional images of pressure pulse waves and photoplethysmography pulse waves, typically determined by the mean and median, as shown in the following formula:

[0069] IPA = μ A +μ′ A

[0070] IPB = μ B +μ′ B

[0071] Where μ is the mean, μ′ is the median, and δ is a scaling factor used to control the range of the energy attribute. For each pixel location (p, q), the energy attribute of both images is calculated using the following formula:

[0072] EAA(p,q)=e -δ|LA(p,q)-IPA|

[0073] EAB(p,q)=e -δ|LB(p,q)-IPB|

[0074] LA(p,q) and LB(p,q) are the pixel values ​​of the low-frequency subbands of the two images at position (p,q), respectively.

[0075] The low-frequency subbands are fused using an energy attribute weighted average method, as shown in the following formula:

[0076]

[0077] Fusion is achieved by calculating the energy attribute (EA) of each low-frequency subband and then weighting the subbands based on these attributes.

[0078] A3: Use Pulse Coupled Neural Network (PCNN) to process high-frequency subbands while preserving detailed information.

[0079] The formula is expressed as:

[0080]

[0081] In the formula, h[m] and g[m] are the coefficients of the low-pass filter and the high-pass filter, respectively.

[0082] Based on the output of PCNN, the high-frequency subbands of either pressure pulse wave or photoplethysmography pulse wave are selected as the fused high-frequency subbands.

[0083] A4: The inverse shearlet transform is used to recombine the subbands that have undergone feature extraction and fusion to reconstruct the fused image, restore the integrity and consistency of the image, and preserve the structural and functional information of the image, thereby providing more accurate details.

[0084] S4, the feature extraction and classification module, extracts pulse wave image features from the fused image generated in S3 through a trained ResNet network and classifies them using a random forest, thus achieving the goal of non-invasive detection of diabetes.

[0085] A1: First, use a pre-trained ResNet network as a feature extractor on the prepared dataset. The ResNet network is as follows: Figure 3 As shown in Table 1, the network learns rich image feature representations through training, directly utilizing the weight parameters learned in the convolutional, pooling, and residual modules. The fused image obtained in S3 is input into the ResNet network, and deep-level features are extracted through layer-by-layer processing via convolutional, pooling, and residual modules. The ResNet network parameters are shown in Table 1.

[0086] Table 1 ResNet network parameters

[0087]

[0088] A2: As Figure 4 As shown, in this embodiment, the Random Forest (RF) algorithm is used to classify the extracted image features. When constructing the Random Forest, a subset of samples is randomly selected from the training dataset as training data for the current decision tree, and a subset of features is randomly selected from the feature set to construct the decision tree. Then, using the selected training data and features, the decision tree is constructed by splitting nodes until a stopping condition is met. The deep image features extracted from the ResNet network are then input into the Random Forest model. Each decision tree in the Random Forest makes a classification decision based on the input features and outputs its own classification result. This method enables diabetes detection based on multi-sensor pulse features.

[0089] In this example, the dataset is cross-validated 5-fold using MATLAB R2020a, with 80% of the data randomly allocated as the training set and the remaining 20% ​​as the test set. The model is retested five times, and the average of the predictions is calculated.

[0090] The performance evaluation metrics for this example are as follows:

[0091] This example uses a confusion matrix to reflect the classification results of the classification model. TP indicates that the diabetes sample was correctly diagnosed as diabetes by the classification model; FN indicates that the diabetes sample was incorrectly diagnosed as healthy by the classification model; FP indicates that the healthy sample was incorrectly diagnosed as diabetes by the classification model; and TN indicates that the healthy sample was correctly diagnosed as healthy by the classification model, as shown in Table 2.

[0092] Table 2 Confusion Matrix

[0093]

[0094] This study, based on the confusion matrix, further uses accuracy (A) CC Precision (P), recall (R), and F1 score are used as evaluation metrics to better assess the model's performance.

[0095]

[0096] Recall rate indicates the percentage of patients who actually have diabetes who are diagnosed. It reflects the rate of missed diagnosis. The higher the recall rate, the lower the rate of missed diagnosis.

[0097]

[0098] The F-score is a comprehensive indicator that takes into account both precision and recall.

[0099] When α = 1, it is the F1 score.

[0100]

[0101] To address the problems of existing diabetes detection methods being invasive and painful, and typically focusing on a single signal for prediction, resulting in limited information acquisition and difficulty in comprehensively assessing the body's condition, this application provides a diabetes detection system based on multi-sensor pulse characteristics to overcome at least one of the shortcomings of the prior art, thus solving the problems mentioned in the background art.

[0102] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A diabetes detection system based on multi-sensor pulse characteristics, characterized in that, include: The pulse wave data acquisition module uses an array of pressure sensors and photoelectric sensors to collect multi-channel pulse wave data. The signal preprocessing module preprocesses the multi-channel pulse wave data, including signal filtering, baseline drift removal, and period segmentation, to obtain usable single-cycle pressure pulse wave data and photoplethysmography pulse wave data. The signal fusion module converts the obtained pulse wave data into a two-dimensional image through continuous wavelet transform. The converted two-dimensional image is then decomposed into low-frequency and high-frequency sub-bands using non-subsampled shear wavelet transform. The low-frequency sub-bands are then fused using the energy attribute weighted average method, and the high-frequency sub-bands are fused using the pulse coupled neural network. Finally, the fused image is generated using inverse shear transform. The feature extraction and calculation module uses a ResNet network to extract features from the fused image and uses a random forest for feature classification, in order to achieve the goal of non-invasive and accurate detection of diabetes.

2. The diabetes detection system based on multi-sensor pulse characteristics according to claim 1, characterized in that: The pulse wave data acquisition module acquires the user's pulse wave signal using an integrated multi-channel acquisition device. This device accurately acquires synchronous three-channel pressure pulse wave signals from the subject's wrist using a three-channel array pressure sensor. Simultaneously, it acquires the subject's pulse wave signal using a high-performance photoelectric sensor.

3. The diabetes detection system based on multi-sensor pulse characteristics according to claim 1, characterized in that: The signal preprocessing module sequentially passes the three-channel pulse wave signals of each group through an FIR low-pass filter with a cutoff frequency of fl to remove high-frequency noise. It then determines whether each channel pulse wave is distorted or weak by measuring the amplitude and slope changes of the pulse wave. Finally, it detects whether the signal is non-periodic by calculating the autocorrelation function and eliminates abnormal signals.

4. The diabetes detection system based on multi-sensor pulse characteristics according to claim 1, characterized in that: The signal fusion module uses continuous wavelet transform to convert the preprocessed pressure pulse wave signal and photoplethysmography (PPG) signal into a two-dimensional image, so as to preserve the time-domain and frequency-domain information of the pressure pulse wave signal and PPG signal and facilitate the fusion of multi-sensor signals.

5. A diabetes detection system based on multi-sensor pulse characteristics according to claim 1, characterized in that: The signal fusion module uses non-subsampled shear wave transform to decompose the image into low-frequency subbands and high-frequency subbands. The low-frequency subband contains structural information of the image and is fused using an energy attribute weighted average method. The high-frequency subband contains detail information and is fused using a pulse-coupled neural network. Finally, an inverse shear transform is used to generate the fused image.

6. A diabetes detection system based on multi-sensor pulse characteristics according to claim 1, characterized in that: The feature extraction and calculation module performs feature extraction and calculation on the generated fused image for classification. Specifically, it extracts pulse wave image features from the fused image generated by the signal fusion module through a trained ResNet network and classifies them through a random forest, thus achieving the goal of non-invasive detection of diabetes.

Images