Non-invasive hemoglobin detection method
By combining photoplethysmography with finger displacement and pressure data features, and using feature extraction and signal quality evaluation algorithms, the high cost and low accuracy of non-invasive hemoglobin detection have been solved, enabling rapid and reliable quantitative detection of hemoglobin, suitable for people of different genders.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUILIN UNIV OF ELECTRONIC TECH
- Filing Date
- 2023-10-18
- Publication Date
- 2026-06-16
AI Technical Summary
Existing non-invasive hemoglobin detection methods suffer from high detection costs, low accuracy, and failure to effectively assess interfering factors, thus affecting the accuracy of the test results.
Signals were acquired using the photoplethysmography method, and combined with finger displacement and pressure data features. The AdaCost-Xgboost model was used for quantitative detection of hemoglobin through feature extraction algorithms and signal quality evaluation.
It enables reliable and rapid non-invasive quantitative detection of hemoglobin, simplifies the operation, improves the stability and accuracy of the test, and allows for early prediction of hemoglobin-related diseases.
Smart Images
Figure CN117462121B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical testing technology, specifically to a non-invasive method for detecting hemoglobin. Background Technology
[0002] Currently, several methods exist for the quantitative detection of non-invasive hemoglobin, including near-infrared spectroscopy, imaging, Raman spectroscopy, conductivity analysis, photoacoustic spectroscopy, image analysis, and multi-wavelength photoplethysmography. While all these methods can perform non-invasive quantitative detection of hemoglobin, they all suffer from drawbacks such as high cost and low accuracy. For example, Chinese invention patent CN113749655 B provides a method for detecting blood oxygen saturation, comprising: acquiring the PPG detection signal of a user to be tested; processing the PPG detection signal based on a morphological filtering algorithm; and determining the blood oxygen saturation of the user based on the processed PPG detection signal. In this embodiment, after acquiring the PPG detection signal of the user to be tested, a morphological filtering algorithm is used to process the PPG detection signal. Morphological filtering can reduce the complexity of signal processing, improve signal processing effect, and thus improve the accuracy of blood oxygen saturation detection. However, the above method does not consider other interfering factors, and the effectiveness of the signal needs further improvement, affecting the final detection result. Summary of the Invention
[0003] To address the aforementioned problems, the purpose of this invention is to provide a non-invasive hemoglobin detection method that reliably extracts the main features of the pulse wave, enabling rapid quantitative detection of hemoglobin and achieving non-invasive detection.
[0004] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:
[0005] A non-invasive hemoglobin detection method includes the following steps:
[0006] (1) Collect user photoplethysmography pulse wave signal, finger displacement data characteristics, and finger pressure data characteristics;
[0007] (2) Use feature extraction algorithms to extract the morphological features of photoplethysmography (PPG) waves, and at the same time evaluate the signal quality of PPG wave signals, finger displacement data features, and finger pressure data features to determine whether the signal segment is usable.
[0008] (3) The morphological characteristics of photoplethysmography, finger displacement data and finger pressure data are averaged as the characteristics for hemoglobin quantification. The corresponding hemoglobin content can be obtained by conversion through a quantitative model.
[0009] Preferably, in step (2), the feature extraction algorithm includes
[0010] ① Feature point identification: Identify the inflection points of the pulse wave, i.e., the peaks and troughs, based on the morphological characteristics of the pulse wave; after identifying the feature points, record the size and position of each peak and trough in the first to fourth rows of the array respectively;
[0011] ② First feature classification: Subtract the identified peak from the next adjacent trough. Based on the pulse wave morphology, the height from the systolic blood pressure peak to the start point of the next cycle and the height from the diastolic blood pressure peak to the gap point can be obtained. Arrange them in the fifth row of the array according to their corresponding positions. Then, subtract the identified peak from the previous trough that is one trough away from it, and arrange them in the sixth row of the array according to their corresponding positions. There are a total of N peaks and troughs in the counting group. Then, sort the entire array from largest to smallest according to the size of the fifth row. The first half of the fifth row of the array, i.e. the larger part, consists of the heights from all systolic blood pressure peaks to the start point of the next cycle, and the second half of the array, i.e. the smaller part, consists of the heights from the diastolic blood pressure peak to the gap point.
[0012] ③ First feature selection: Starting from the array's starting point and midpoint, perform variance calculations on the N / 4 consecutive adjacent values in the first and second halves of the fifth row of the array, respectively. Reduce computation time by reusing the sliding variance recursive formula. Calculate the combination with the smallest variance for each part, and use the mean of the smallest variance combination for the first and second halves as features A1 and B, respectively, and save the mean used in the variance calculation. Then calculate the mean of the smallest variance combination for the first and second halves of the row containing the trough, respectively, to obtain features C and D. Subsequently, after re-sorting the entire array from largest to smallest according to the sixth row size, similarly extract the minimum variance and mean of the N / 4 consecutive adjacent values in the first half as feature A2.
[0013] ④ First feature calculation: The size of the dynamic part of PPG is calculated based on the size of features A1 and A2, and the variance is used as the weight. The weighted sum is calculated according to the following formula.
[0014]
[0015] in,
[0016] A: Size of the dynamic portion of the pulse wave;
[0017] A1: The mean of N / 4 consecutive values with the smallest variance in the height from the peak of the contraction pressure to the starting point of the next cycle;
[0018] R A1 The minimum variance of the height from the peak of contraction pressure to the starting point of the next cycle;
[0019] A2: The mean of N / 4 consecutive values with the smallest variance in height from the peak of systolic pressure to the starting point of this cycle;
[0020] RA2 The minimum variance of the height from the peak of the contraction pressure to the starting point of this cycle;
[0021] This completes the extraction of the main amplitude features of the pulse wave, which is also an important part of inferring blood components. These features are: A: size characteristics of the dynamic part of the pulse wave; B: amplitude from the diastolic pressure peak to the gap; C: size of the DC part of the pulse wave; and D: gap height.
[0022] ⑤ Second feature classification, second feature selection and second feature calculation: Calculate and extract the distance between the two systolic blood pressure peaks; the distance from the diastolic blood pressure peak to the gap; the distance between the systolic blood pressure peak and the starting point. Similarly, sort these three features from smallest to largest, and extract the average of the N / 4 consecutive values with the smallest variance according to the variance to obtain features E, F and G, thus obtaining the pulse wave time domain features.
[0023] Preferably, in step (2), the signal quality evaluation index is as follows: Since the features are selected based on variance, variance can effectively measure the degree of data dispersion. The minimum variance of each feature, as well as the information on the number of peaks and valleys, finger displacement data features, and finger pressure data features, are used as features to judge signal quality. A certain number of high-quality and low-quality signals are collected and manually labeled to determine whether they meet the requirements, forming a dataset for training the quality evaluation model. The AdaCost algorithm is used for modeling, and a decision tree is used as the base model. Finally, a binary classification model is trained. In addition, during the detection process, the finger displacement data features and finger pressure data features are monitored in real time. If the change is too large and exceeds the set threshold, it is judged as not meeting the requirements.
[0024] Preferably, before feature point recognition, the photoplethysmography (PPG) signal is first subjected to arithmetic mean filtering and FIR filtering.
[0025] Preferably, the quantitative model in step (3) is as follows: pulse wave features AG, finger displacement data and finger pressure data are used as features, the actual hemoglobin content is used as the result, and prediction models are trained according to gender, and the ADASYN algorithm is used to expand the imbalanced dataset.
[0026] Preferably, the prediction model uses AdaCost-Xgboost. First, the AdaCost algorithm is used to perform binary classification based on the minimum normal standard value of hemoglobin for men and women as the threshold to distinguish between normal people and anemic patients. Then, the Xgboost algorithm is used to perform regression on the classified datasets respectively, resulting in four non-invasive hemoglobin prediction models to adapt to different genders and populations.
[0027] Compared with the prior art, the present invention has the following advantages:
[0028] 1. This invention primarily uses photoplethysmography (PPG) for noninvasive hemoglobin detection. After analysis, two factors affecting the dynamics of the PPG were identified: the distance between the transmissive light source and the sensor during detection, and the pressure applied by the finger to the photodiode (hereinafter referred to as distance and pressure, respectively). These two factors are used as features for detection and incorporated into the noninvasive hemoglobin prediction model.
[0029] 2. This invention designs an automatic feature extraction algorithm and a signal quality assessment algorithm to extract the required features. It can efficiently and reliably extract the main features of the pulse wave, and can quickly realize the quantitative detection of hemoglobin. It is simple to operate, has a short detection time, good stability, and realizes non-invasive detection. It plays an important and positive role in the early prediction of the occurrence of hemoglobin-related diseases and patient monitoring. Attached Figure Description
[0030] Figure 1 This is a flowchart of the signal processing and feature extraction process of the present invention;
[0031] Figure 2 This is a block diagram of the detection system of the present invention;
[0032] Figure 3 This is a schematic diagram of the first feature classification.
[0033] Figure 4 This is a schematic diagram of the second feature classification. Detailed Implementation
[0034] The technical solutions provided by the present invention will be described in detail below with reference to specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.
[0035] This invention discloses a non-invasive method for detecting hemoglobin, comprising the following steps:
[0036] (1) Collect the user's fingertip photoplethysmography pulse wave signal, finger displacement data characteristics, and finger pressure data characteristics.
[0037] (2) Figure 2 As shown, the photoplethysmography (PPG) signal is first subjected to arithmetic mean filtering and FIR filtering. Then, the morphological features of the PPG are extracted using a feature extraction algorithm. At the same time, the signal quality of the PPG signal, finger displacement data features, and finger pressure data features is evaluated to determine whether the signal segment is usable.
[0038] The feature extraction algorithm includes feature point recognition, first feature classification, first selection classification, first feature calculation, second feature classification, second feature selection, and second feature calculation.
[0039] Feature point identification: Identify the inflection points of the pulse wave, i.e., the peaks and troughs, based on the morphological characteristics of the pulse wave. After identifying the feature points, record the size and position of each peak and trough in the first to fourth rows of the array.
[0040] First feature classification: Subtract the identified peak from its adjacent trough. Based on the pulse wave morphology, the height from the systolic blood pressure peak to the start point of the next cycle and the height from the diastolic blood pressure peak to the gap point can be obtained. These are then placed in the fifth row of the array according to their corresponding positions. Next, subtract the identified peak from the preceding trough that is one trough away from it, and place it in the sixth row of the array according to its corresponding position. There are a total of N peaks and troughs in the counting group.
[0041] The entire array is then sorted from largest to smallest according to the fifth row. Ignoring outliers, since the height from the systolic pressure peak to the start point of the next cycle is significantly greater than the height from the diastolic pressure peak to the gap point, the first half of the fifth row, the larger portion, consists of the heights from all systolic pressure peaks to the start point of the next cycle. The second half, the smaller portion, consists of the heights from the diastolic pressure peak to the gap point. Figure 3 As shown.
[0042] First feature selection: Starting from the array's beginning and midpoint, perform variance calculations on the N / 4 consecutive adjacent values in the first and second halves of the fifth row, respectively. Reduce computation time by reusing the moving average recursive formula. Calculate the combination with the minimum variance for each part, and use the mean of the minimum variance combinations for the first and second halves as features A1 and B, respectively, saving the mean used in the variance calculation. Then, calculate the mean of the minimum variance combinations for the first and second halves of the row containing the trough, obtaining features C and D. Subsequently, re-sort the entire array from largest to smallest according to the sixth row, following the same steps. Similarly, extract the minimum variance and mean of the N / 4 consecutive adjacent values in the first half as feature A2.
[0043] First feature calculation: The size of the dynamic part of PPG is calculated based on the size of features A1 and A2, and the variance is used as the weight. The results are then weighted and summed according to Formula 1.
[0044]
[0045] in,
[0046] A: Size of the dynamic portion of the pulse wave;
[0047] A1: The mean of N / 4 consecutive values with the smallest variance in the height from the peak of the contraction pressure to the starting point of the next cycle;
[0048] R A1 The minimum variance of the height from the peak of contraction pressure to the starting point of the next cycle;
[0049] A2: The mean of N / 4 consecutive values with the smallest variance in height from the peak of systolic pressure to the starting point of this cycle;
[0050] R A2 : The minimum variance of the height from the contraction pressure peak to the starting point of this cycle.
[0051] The above formula can effectively avoid the effects of pulse wave baseline drift and measurement errors.
[0052] This completes the extraction of the main amplitude characteristics of the pulse wave, a crucial part of inferring blood components. (The following are examples:)
[0053] A: Size characteristics of the dynamic portion of the pulse wave; B: Amplitude from the peak to the gap in diastolic blood pressure;
[0054] C: Size of the DC portion of the pulse wave; D: Height of the notch.
[0055] Second feature classification: Subtract the position of the trough from the position of its preceding adjacent peak, and re-fill the fifth row of the array. Subtract the position of the peak from the position of the peak that is separated from its preceding peak, and re-fill the sixth row of the array. For example... Figure 4 As shown.
[0056] Second feature selection: The fifth and sixth rows were reordered from largest to smallest based on their respective values. According to the pulse wave morphology, the first half of the fifth row consists of the distance from the systolic blood pressure peak to the starting point, and the second half consists of the distance from the diastolic blood pressure peak to the gap. In the sixth row, since the interval between diastolic and systolic blood pressure peaks is approximately equal, there is no need to segment the sixth row.
[0057] The second feature calculation: Based on variance, the average of the N / 4 consecutive values with the smallest variance in the first and second halves of the fifth row is extracted to obtain features E and F. The average of the N / 4 consecutive smallest variance values in the entire sixth row is then extracted and calculated to obtain feature G. Feature G can be converted to obtain heart rate data.
[0058] Thus, three time-domain characteristics of the pulse wave were obtained.
[0059] Signal quality evaluation metrics: Since features are selected based on variance, which effectively measures the dispersion of data, the minimum variance of each feature, along with information such as the number of peaks and troughs, and distance and pressure sensor data, are used as features to judge signal quality. A dataset is constructed by collecting a certain number of high-quality and low-quality signals and manually labeling them to determine whether they meet the requirements. This dataset is used to train the quality evaluation model. The AdaCost algorithm is used for modeling, and a decision tree is used as the base model, ultimately resulting in a binary classification model. Furthermore, during the detection process, distance and pressure sensor data are monitored in real time. If the changes are too large and exceed a set threshold, the data is judged as unacceptable.
[0060] (3) Hemoglobin prediction: Pulse wave features (AG), distance sensor data (finger displacement data), and pressure sensor data (finger displacement data) were used as features, and the actual hemoglobin content was used as the result. Prediction models were trained separately for each gender, and the ADASYN algorithm was used to augment the imbalanced dataset. The prediction model used AdaCost-Xgboost. First, the AdaCost algorithm was used to perform binary classification based on the minimum normal hemoglobin value for both men and women as a threshold, distinguishing between normal individuals and anemic patients. Then, the Xgboost algorithm was used to perform regression on the classified datasets. Four non-invasive hemoglobin prediction models were obtained to adapt to different genders and populations.
[0061] like Figure 2 As shown, the detection system of the present invention includes a feature acquisition module, a data processing module, a touch display unit, a storage unit, a signal processing and feature extraction unit, and a finger clamp mechanism, wherein the feature acquisition module is integrated into the finger clamp mechanism. The output end of the data processing module is connected to the touch display unit and the storage unit, and the input and output ends of the data processing module are both connected to the feature acquisition module, respectively realizing the functions of acquisition control, storage and display of photoplethysmography (PPG) signals.
[0062] The feature acquisition module includes a multi-wavelength photoplethysmography (PPG) acquisition unit, a distance acquisition unit, and a pressure acquisition unit. The input and output terminals of the multi-wavelength PPG acquisition unit are connected to the data processing module, and the output terminals of the distance and pressure acquisition units are connected to the data processing module.
[0063] The multi-wavelength photoplethysmography (PPG) acquisition unit includes four groups of eight LEDs with different wavelengths, an aluminum substrate, an analog front-end circuit, an LED drive signal multiplexing circuit, and photodiodes. The aluminum substrate is located below the LEDs, and the photodiodes can be directly illuminated by the LEDs. The output of the analog front-end circuit is connected to the LEDs via the LED drive signal multiplexing circuit, and the output of the photodiodes is connected to the analog front-end circuit. The data processing module uses a ZYNQ7020 as the core controller. The core controller sets the parameters of the analog front-end AFE4490 via the SPI protocol. The LED constant current drive signal emitted by the analog front-end drives the eight LEDs after being controlled by the analog switch TMUX1109. The eight LEDs illuminate the area above the fingertip in a sequential manner, and the photodiode below the fingertip receives the light transmitted through the fingertip and the filter, thus obtaining the multi-wavelength PPG signal.
[0064] The raw photoplethysmography (PPG) signals acquired by the multi-wavelength PPG acquisition unit are displayed in real time on the touchscreen and transmitted to the storage unit for storage after being processed by the data processing system. Incident light passes through a finger and a filter, then passes through a photoelectric sensor driving circuit to obtain PPG signals, which are ultimately returned to the main control chip for data analysis and hemoglobin prediction. The acquired feature data is then transmitted to the storage unit as a dataset for training the model.
[0065] Finger thickness characteristics are acquired by a KS8-10 displacement sensor. During detection, the displacement sensor data is monitored simultaneously; excessive fluctuations indicate significant finger tremor, which is displayed on the screen. Finger tip pressure characteristics are acquired by an FSR402 sensor. The analog signals acquired by both sensors are processed into digital signals by a 3PA1030 AD converter chip.
[0066] The data processing module uses the ZYNQ7020 as the main control chip, responsible for data acquisition and caching, storage, touch and display control, device timing control, etc. It acquires photoplethysmography (PPG) signals at a frequency of 500Hz, and then enters the signal processing and feature extraction unit. High-frequency noise is filtered out by an arithmetic mean filtering algorithm, and baseline drift is removed by a 0.8-10Hz FIR filter. The feature extraction algorithm designed in this invention identifies feature points to extract the morphological features of the PPG. At the same time, the signal quality evaluation model designed in this invention determines whether the signal segment is usable. The average morphological feature obtained after processing 400,000 sampling points for each wavelength is used as a feature. The average value of the data collected by the displacement sensor and the data collected by the pressure sensor is used as the feature for hemoglobin quantification. The corresponding hemoglobin content can then be obtained by conversion through the quantitative model.
[0067] The above embodiments are merely specific examples to further illustrate the purpose, technical solution, and beneficial effects of the present invention, and the present invention is not limited thereto. Any modifications, equivalent substitutions, improvements, etc., made within the scope of the disclosure of the present invention are included within the protection scope of the present invention.
Claims
1. A non-invasive method for detecting hemoglobin, characterized in that, Includes the following steps: (1) Collect user photoplethysmography pulse wave signal, finger displacement data characteristics, and finger pressure data characteristics; (2) Use feature extraction algorithms to extract morphological feature values of photoplethysmography (PPG), and at the same time evaluate the signal quality of PPG signal, finger displacement data features, and finger pressure data features to determine whether these signals are usable. (3) The morphological features of the photoplethysmography (PPG), finger displacement data, and finger pressure data are averaged as features for hemoglobin quantification. The corresponding hemoglobin content can then be obtained by conversion using a quantitative model. In step (2), the feature extraction algorithm includes... Feature point identification: Identify the inflection points of the pulse wave, i.e., the peaks and troughs, based on the morphological characteristics of the pulse wave; after identifying the feature points, record the size and position of each peak and trough in the first to fourth rows of the array respectively; First feature classification: Subtract the identified peak from the next adjacent trough. Based on the pulse wave morphology, the height from the systolic blood pressure peak to the start point of the next cycle and the height from the diastolic blood pressure peak to the gap point can be obtained. Arrange them in the fifth row of the array according to their corresponding positions. Then, subtract the identified peak from the previous trough that is one trough away from it, and arrange them in the sixth row of the array according to their corresponding positions. There are a total of N peaks and troughs in the counting group. Then, sort the entire array from largest to smallest according to the fifth row. The first half of the fifth row of the array, i.e. the larger part, consists of the heights from all systolic blood pressure peaks to the start point of the next cycle, and the second half of the array, i.e. the smaller part, consists of the heights from the diastolic blood pressure peak to the gap point. First feature selection: Starting from the array's starting point and midpoint, perform variance calculations on the N / 4 consecutive adjacent values in the first and second halves of the fifth row of the array, respectively, and reduce the computation time by reusing the sliding variance recursive formula; Calculate the combination with the smallest variance for each of the two parts, and use the mean of the combination with the smallest variance for the first half and the second half as features A1 and B respectively, and save the mean used in the variance calculation. Next, calculate the mean of the minimum variance combination of the first half and the second half of the row where the trough is located, and obtain features C and D. Then, after re-sorting the entire array from largest to smallest according to the size of the sixth row, extract the minimum variance and mean of the N / 4 consecutive adjacent values in the first half as feature A2. First feature calculation: The size of the dynamic part of PPG is calculated based on the size of features A1 and A2, with the variance as the weight, and a weighted sum is performed according to the following formula; ， in, A: Size of the dynamic portion of the pulse wave; The mean of N / 4 consecutive values with the smallest variance from the peak of the contraction pressure to the starting point of the next cycle; The minimum variance of the height from the peak of contraction pressure to the starting point of the next cycle; The mean of N / 4 consecutive values with the smallest variance from the peak of the contraction pressure to the starting point of this cycle; The minimum variance of the height from the peak of the contraction pressure to the starting point of this cycle; This completes the extraction of the main amplitude features of the pulse wave, which is also a crucial part for inferring blood components. These features are: A: Size characteristics of the dynamic portion of the pulse wave; B: Amplitude from the diastolic blood pressure peak to the gap; C: Size of the DC portion of the pulse wave; D: Gap height. Second feature classification, second feature selection and second feature calculation: Calculate and extract the distance from the peak of diastolic blood pressure to the gap; the distance from the peak of systolic blood pressure to the starting point; and the distance between two consecutive peaks. Sort these three features from smallest to largest, and extract the average of the N / 4 consecutive values with the smallest variance according to the variance to obtain features E, F and G, thus obtaining the pulse wave time domain features.
2. The non-invasive hemoglobin detection method according to claim 1, characterized in that: In step (2), the signal quality evaluation index is as follows: Since the features are selected based on variance, variance can effectively measure the degree of data dispersion. The minimum variance of each feature, as well as the information on the number of peaks and valleys, finger displacement data features, and finger pressure data features, are used as features to judge signal quality. A certain number of high-quality and low-quality signals are collected and manually labeled to determine whether they meet the requirements, forming a dataset for training the quality evaluation model. The AdaCost algorithm is used for modeling, and a decision tree is used as the base model. Finally, a binary classification model is trained. In addition, during the detection process, the finger displacement data features and finger pressure data features are monitored in real time. If the change is too large and exceeds the set threshold, it is judged as not meeting the requirements.
3. The non-invasive hemoglobin detection method according to claim 1, characterized in that: Before feature point identification, the photoplethysmography (PPG) signal is first subjected to arithmetic mean filtering and FIR filtering.
4. The non-invasive hemoglobin detection method according to claim 1, characterized in that, The quantitative model in step (3) is as follows: pulse wave features AG, finger displacement data and finger pressure data are used as features, the actual hemoglobin content is used as the result, and prediction models are trained according to gender. The ADASYN algorithm is used to expand the imbalanced dataset.
5. The non-invasive hemoglobin detection method according to claim 4, characterized in that: The prediction model uses AdaCost-Xgboost. First, the AdaCost algorithm is used to perform binary classification based on the minimum normal hemoglobin value for men and women as the threshold to distinguish between normal people and anemic patients. Then, the Xgboost algorithm is used to perform regression on the classified datasets to obtain four non-invasive hemoglobin prediction models to adapt to different genders and populations.