CN115054209B

CN115054209B - Multi-parameter physiological information detection system and method based on intelligent mobile equipment

Info

Publication number: CN115054209B
Application number: CN202210393707.8A
Authority: CN
Inventors: 马喜波; 夏超然; 朱雅芳
Original assignee: Hangzhou Huashi Novi Medical Technology Co ltd
Current assignee: Hangzhou Huashi Novi Medical Technology Co ltd
Priority date: 2022-04-14
Filing date: 2022-04-14
Publication date: 2023-09-26
Anticipated expiration: 2042-04-14
Also published as: CN115054209A

Abstract

The invention provides a multi-parameter physiological information detection system and method based on intelligent mobile equipment, wherein the intelligent mobile equipment comprises illumination equipment and image pickup equipment which are positioned on the same side of the intelligent mobile equipment, and the detection system comprises: and the signal acquisition module and the signal processing module. The invention predicts heart rate, blood oxygen and blood pressure by utilizing the intelligent terminal camera and the finger tip video acquired by light. The system utilizes video to take frames, divides images into channels, and synthesizes PPG signals from continuous video frames at finger ends by pixel point weighted summation; and extracting characteristics by utilizing a Unet network on the basis of PPG, and respectively predicting heart rate, blood oxygen and blood pressure values by utilizing branch networks of a plurality of branches consisting of a multi-layer convolutional neural network.

Description

Multi-parameter physiological information detection system and method based on intelligent mobile equipment

Technical Field

The invention relates to the field of medical instruments, in particular to a physiological information detection system.

Background

In the face of population aging and high-intensity trend of daily work of office workers, monitoring of conventional physiological indexes (heart rate, blood oxygen, blood pressure and the like) at any time and any place becomes an urgent need.

In the above requirements, accurate measurement of blood pressure is a difficulty. Blood pressure (usually called systolic and diastolic blood pressure is the pressure of aortic blood on the wall of blood vessels during systole and diastole of a human body) is taken as an important index of physiological signals of the human body, and can reflect cardiovascular functions of the human body, and the biggest harm of abnormal blood pressure values to the human body is possible to cause damage to important organs such as heart, brain, kidney and the like, such as cerebral infarction, renal artery stenosis, myocardial infarction and the like. Therefore, the daily monitoring of the blood pressure can fully know the real-time condition and the dynamic change condition of the blood pressure of the individual, and has very important significance for the health monitoring of the individual.

1. The devices for detecting physiological indexes on the market are mostly independent in function, such as finger-clip type pulse oximeter for detecting blood oxygen and sphygmomanometer for detecting blood pressure. There are few integrated, one-stop, comprehensive devices for multiple physiological parameter sensing.

2. The existing blood pressure detection equipment in the market is divided into two major categories, namely mechanical type and electronic type.

The mechanical mercury sphygmomanometer is pressurized by a sleeve belt, and the stethoscope distinguishes pulse sounds to measure the blood pressure value. This method requires a professional to operate and this measurement may cause a white gown effect, i.e. the pressure of the cuff may cause discomfort to the subject when measured in the face of differences in doctor blood pressure and normal conditions.

The electronic type can be divided into an electronic sphygmomanometer and an intelligent watch with a blood pressure measuring function. The electronic sphygmomanometer is opposite to the mercury sphygmomanometer, and only the listening sound of a person replaced by the electronic listening sound is not influenced by a doctor white coat, and is more objective and stable. However, the cuff-based pressurizing mode is not changed, and the cuff-based pressurizing device is inconvenient to carry due to the volume problem. Although the existing blood pressure measuring watch integrated with the miniature air pump and the double-layer narrow air bag is heavy in overall weight, the air bag under the watchband is airtight, and the watch is not suitable for daily long-term wearing.

However, other smart watches with blood pressure measurement function, whether based on two PPG signals or a combination scheme of PPG and ECG signals, have high hardware requirements for measuring blood pressure, and must be mounted on the smart watch.

In summary, the existing blood pressure measurement scheme is based in part on the cuff pressurization measurement mode, which is not convenient enough on one hand and brings a certain uncomfortable feeling to the tested person on the other hand; partially utilizing the PPG optoelectronic signal, the method relies on hardware devices such as photosensors.

Disclosure of Invention

In view of the above, it is desirable to provide a method and system for detecting multi-parameter physiological information that does not require a conventional pressurizing device such as a cuff or a hardware device such as a separate photosensor.

Specifically, the invention provides a multi-parameter physiological information detection system based on an intelligent mobile device, which is characterized in that the intelligent mobile device comprises a lighting device and a camera device which are positioned on the same side of the intelligent mobile device, and the detection system comprises: the signal acquisition module and the signal processing module are connected with each other,

the illumination equipment is used for illuminating a specific part of the body of the individual to be tested;

the camera equipment is used for collecting video images of the irradiated part;

the signal acquisition module is used for acquiring multi-frame images within a preset time obtained by the camera equipment and extracting PPG signals based on the images;

the signal processing module comprises a UNet-ANN multi-branch feedback enhancement model, and the UNet-ANN multi-branch feedback enhancement model is used for carrying out multi-parameter physiological information detection based on the PPG signal.

Preferably, the system further comprises a data processing module for framing the acquired video images, weighted summing the R-channel pixel values of each frame of images, and finally combining into a continuous "PPG signal".

Preferably, the UNet-ANN multi-branch feedback reinforcement model includes an asymmetric UNet backbone network and a multi-branch ANN network, and each branch of the multi-branch ANN network is used for detecting a physiological information parameter.

Preferably, the number of channels of the input signal in the UNet backbone network is smaller than the number of sampling points, and the number of channels of the output signal is larger than the number of sampling points.

Preferably, the physiological information parameters include one or more of heart rate, blood oxygen, systolic pressure, and diastolic pressure.

Preferably, the UNet-ANN multi-branch feedback enhancement model performs parameter adjustment of the UNet main network based on differences between true values and detection values of heart rate and blood oxygen of the individual to be tested as feedback conditions.

Preferably, the image capturing apparatus is further configured to capture a face video of an individual to be tested, and perform sex determination of the individual to be tested based on the face video.

Preferably, the data processing module is used for performing 0.5-8HZ elliptical band-pass filtering processing on the finger tip PPG signal, detecting abnormal points and cleaning; windowing is carried out on the 'PPG signal' for 30 seconds, the window width is 256 sampling points, and the step length is 150 sampling points; judging whether the data subjected to the window cleaning has data segments meeting the data quality requirements or not, and extracting corresponding data segments.

In another aspect, the present invention provides a method for constructing a multi-parameter physiological information detection model, which is characterized in that the method includes:

step 1), irradiating the tail end of a body part of an individual to be detected by using a flash lamp of the intelligent mobile equipment;

step 2), acquiring multi-frame continuous images of an individual to be detected by utilizing a camera of the intelligent mobile equipment;

step 3), extracting an end PPG signal based on the image and processing the end PPG signal;

step 4), constructing a UNet-ANN multi-branch feedback reinforcement model;

step 5), training the reinforcement model by using the marked finger end PPG signal.

Preferably, the step of processing the terminal PPG signal comprises:

step 2.1, framing the obtained video, weighting and summing the R channel pixel values of each frame of image, and combining the R channel pixel values into a section of continuous PPG signal;

step 2.2, performing 0.5-8HZ elliptical bandpass filtering treatment on the PPG signal, detecting abnormal points and cleaning;

step 2.3, windowing PPG signals;

step 2.4, judging that the data subjected to the window cleaning is subjected to data segment extraction and checking according to the step 3, if the data is checked to be qualified, carrying out the next step, otherwise, re-acquiring the data;

step S3: extraction of PPG signals, the steps comprising:

step 3.1, searching peak points for the processed PPG signals, and finding peak points and valley points in the PPG signals;

step 3.2: d_peak representing the fluctuation condition of the peak point and d_valley representing the fluctuation condition of the valley point are calculated through the formula d= (max-min)/min, and if d_peak is less than 0.05 and d_valley is less than 0.08, the PPG data of the segment is reserved;

step 3.3: judging whether the number of cycles contained in the current signal window is less than a preset value or not based on the number of peak points and valley points, and adopting a second extraction scheme if the number of cycles contained in the current signal window is less than the preset value;

the second extraction scheme:

step 3.4 dividing the processed PPG Signal into windows of a Single period

Step 3.5: calculating the correlation coefficient r of adjacent periodic waveforms, and if r > is a correlation coefficient pre-threshold value, reserving the PPG signal;

step 3.6: if the PPG signal processed in step 2 does not meet both extraction criteria, the segment of PPG signal is directly deleted.

The term "body-specific part" and "body end" as used herein refer to a body part, such as a finger tip, an earlobe, or a toe tip, that is capable of blood-related information collection.

Advantageous effects

The invention creatively provides a method for acquiring video images of the tail end (such as finger end and ear end) of an individual to be detected by utilizing the combination of a camera of a mobile phone and an intelligent mobile terminal with a flat lamp and lamplight, extracting a photoplethysmography (PPG) signal from the video images, and carrying out weighted summation on the acquired video frames to convert the weighted summation into PPG signal sampling points. According to the invention, the acquired PPG signal is utilized to acquire multiple physiological parameter indexes by utilizing the multi-branch feedback reinforcement model, so that synchronous measurement of multiple physiological parameters based on a mobile phone and without professional equipment can be realized at the same time.

Compared with the finger-clip type pulse oximeter which is fixed at the finger tip, the finger-clip type pulse oximeter has the advantages that light (green light or red light and near infrared light in general) emitted in a fixed wavelength band and the PPG signals acquired by the corresponding photoelectric sensors are unstable, so that the existing equipment and method cannot effectively measure parameters such as blood oxygen, blood pressure and the like based on the PPG extracted from continuous video frames of the finger tip acquired by the rear camera of the mobile phone. However, the invention realizes the extraction and the utilization of PPG signals collected by a mobile phone through the data processing and the design adjustment of a model, and has higher accuracy through experimental verification.

The multi-branch feedback reinforcement model predicts heart rate, blood oxygen and blood pressure values jointly by the Unet backbone network and a plurality of ANN branches. In the pre-training stage of the model, the feedback of heart rate and blood oxygen branches is utilized to optimize the Unet backbone network. And then fixing a backbone network, and training a blood pressure prediction branch on the basis.

In addition, the invention corrects the model prediction result by using the sex and age signals of the portrait prediction acquired by the front camera

Note that: although blood pressure is a physiological index with large individual difference, the blood pressure has certain statistical significance with signals such as individual age, sex and the like under a large sample. And combining the individual information obtained from the portrait, and prompting the subject to perform secondary measurement if the model predicted physiological index and the statistical predicted result are excessively different. The three measurements still do not fluctuate too much. The system and method of the present invention avoids some occasional errors and extreme results to some extent.

Drawings

FIG. 1 is a flow chart of parameter detection by the system of the present invention

Fig. 2 is a schematic diagram of a process for reconstructing a finger PPG signal according to the present invention

FIG. 3 is a schematic view of the overall framework of the model of the present invention

FIG. 4 is a detailed frame diagram of a backbone network Unet of the model of the present invention

FIG. 5 is a detailed frame diagram of a branched network of the model in the present invention

Detailed Description

The multi-parameter physiological information detection system based on the intelligent mobile device in the embodiment is described by taking a smart phone as an example, and of course, other intelligent mobile devices such as a tablet computer and the like can be adopted to realize the system of the invention.

The intelligent mobile device needs to be provided with the lighting device and the camera device which are positioned on the same side, and the rear camera is mainly adopted at present, so that illumination and image acquisition can be realized on the same side. Of course, for special cases, it is also possible to use an auxiliary light source for illumination and then use the camera of the mobile device for image acquisition. The detection system of the present embodiment includes: and the signal acquisition module and the signal processing module.

In this embodiment, the information collecting hardware device includes: a mobile phone flash lamp and a mobile phone rear camera. The flash lamp is used for irradiating the finger end or the ear end of the human body; the camera is used for collecting video images of the irradiated part; the signal acquisition module is used for acquiring 20-40s, such as multi-frame images within 30s, obtained by the image pickup equipment and extracting a PPG signal based on the images; the signal processing module comprises a UNet-ANN multi-branch feedback enhancement model, and the UNet-ANN multi-branch feedback enhancement model is used for carrying out multi-parameter physiological information detection based on the PPG signal.

The signal processing module is loaded with an application program of the multi-branch feedback strengthening algorithm network and runs in a processor of the mobile device. The invention can utilize the mobile phone camera, the light and the application program to realize the measurement of human physiological indexes (heart rate, blood oxygen and blood pressure) and the preservation and analysis system. Specific model training and parameter measurement are as follows:

s1: data acquisition

Step 1.1, a person to be tested takes a rest slightly to ensure that each physiological index of the person to be tested is in a stable state;

and 1.2, setting the video recording of the mobile phone to 60fps or adjusting according to the precision.

And 1.3, acquiring and storing the face video by using a front-facing camera of the mobile phone (the step is only required to be executed once when any subject uses the function for the first time). It should be noted that, the steps 1.1-1.3 may be implemented by separately setting a login setting module in the system, which is an optional module, and is mainly used for performing personnel information login confirmation and mobile phone parameter adjustment.

And 1.4, pressing the finger on the lamplight of the mobile phone and the rear camera connected with the lamplight, keeping a quiet and stable state, and collecting continuous image frames of the finger end for 30 seconds.

S2: data processing

And 2.1, framing the finger end video obtained in the step 1.3, weighting and summing the R channel pixel values of each frame of image, and finally combining the R channel pixel values into a section of continuous finger end PPG signal.

And 2.2, carrying out 0.5-8HZ elliptical band-pass filtering treatment on the finger tip PPG signal, detecting abnormal points and cleaning.

Step 2.3, windowing is carried out in the 'PPG signal' for 30 seconds, the window width is 256 sampling points, and the step length is 150 sampling points.

Step 2.4, judging that the data after the window cleaning and dividing is subjected to data segment extraction and checking according to the step 3, if yes, carrying out the next step, otherwise, returning to 1.3 to re-acquire the data

S3: extraction of PPG signals

First extraction protocol

And 3.1, finding peak points and valley points in the PPG signal by the processed PPG signal through searching the peak points and the valley points.

Step 3.2: d_peak (representing the fluctuation of the peak point) and d_valley (representing the fluctuation of the valley point) are respectively calculated by the maximum value (max) of the peak point and the valley point found in the step 3.1, the minimum value (min) is calculated by the formula d= (max-min)/min, and if d_peak is less than 0.05 and d_valley is less than 0.08 (wherein 0.05,0.08 is an empirical value and can be different according to the change of the equipment), the PPG data of the segment is reserved;

step 3.3: and judging whether the number of the periods contained in the current signal window is less than a preset value (the number of the peak points or the estimated points is not more than 2) based on the number of the peak points and the valley points, and adopting a second extraction scheme.

The second extraction scheme:

3.4 divide the preprocessed PPG signal into a plurality of single-period windows (each segment of data is 256 samples in length, each period window corresponds to one heartbeat period of the data measurer).

Step 3.5: the correlation coefficient r of the adjacent periodic waveform is calculated (by adopting the existing correlation algorithm), and if r >0.85 (0.85 is an empirical value), the segment of PPG signal is reserved.

Step 3.6: if the PPG signal processed in step 2 does not meet the two extraction criteria, the PPG signal is directly deleted.

S4: model prediction

And 4.1, calling an openCV library to judge the gender and age information of the subject from the face video obtained in the step 1.2.

And 4.2, inputting the preprocessed finger tip PPG signal into a pre-trained UNet-ANN multi-branch feedback reinforcement model to obtain predicted heart rate, blood oxygen and blood pressure values. The specific structure of the model is discussed in further detail below.

And 4.3, adjusting the 3.2 model by combining the subject information acquired in the step 4.1.

S5: result preservation

Step 5.1 selecting whether to save the measured results

Step 5.2 if 5.1 chooses to save, save the result to personal data and conduct a statistical analysis.

Note that: detailed description of heart rate, blood oxygen and blood pressure prediction model

The UNet-ANN multi-branch feedback reinforcement model in this embodiment is divided into two major parts, namely a main network and a branch network.

The backbone network adopts a Unet network, the branch network adopts an ANN branch network, and in the embodiment, a single Unet network and the structural configuration of four ANN branch networks are adopted.

A unet network architecture:

the original Unet is of a symmetrical structure, and the Unet is used as a feature extraction module to adjust the structure, so that an asymmetrical structure is adopted. The Unet network was originally created to solve the problem of medical image segmentation, and the Unet network acquires context information and location information using a U-type network structure. In this embodiment, the asymmetric Unet backbone network is mainly divided into two left and right modules. The left module consists of four layers, each layer consisting essentially of 4 two-stage 3*1 convolution kernels plus 2*1 downsampling sub-modules (the number of convolution kernels and sub-modules for each layer being adjustable). The right module also comprises four layers, each layer consists of 4 convolution kernels of two stages 3*1 and 2*1 up-sampling submodules, each layer between the left module and the right module is connected through splicing, and output is carried out from the rightmost side. The asymmetry is reflected by the difference between the input layer, which is a PPG signal of 1 channel 256 sample point length, and the output layer, which is a multi-feature signal of 1024 channel 2 sample point length. The improved Unet structure replaces two-dimensional convolution of the image by one-dimensional convolution, and a plurality of output channels are used for replacing a single channel in the original Unet in the last layer (the uppermost layer of the right module) to obtain a plurality of characteristic values, and the characteristic values are respectively sent to each ANN branch network.

As shown in fig. 4, the PPG signal with the length of 1 channel 256 sampling points is taken as input, and is input into the main network Unet, and becomes a signal with the length of 64 channel 256 sampling points under the convolution network of two stages 3*1; the signal is changed into a signal with the length of 128 sampling points of 64 channels through a 2*1 downsampling network; a signal that becomes 128 channels 128 sample points long under a two-stage 3*1 convolutional network; a signal with the length of 128 channels and 64 sampling points is changed into a signal with the length of 128 channels through a 2*1 downsampling network; a signal that becomes 256 channels 64 sample points long under the convolution network of two stages 3*1; a signal with the length of 256 channels and 32 sampling points is changed into a signal with the length of 256 channels through a 2*1 downsampling network; a signal that becomes 512 channels 32 sample points long under the convolution network of two stages 3*1; a signal with the length of a sampling point of 512 channels and 16 is changed into a signal with the length of a sampling point of 512 channels through a 2*1 downsampling network; a signal which becomes 1024 channels with 16 sampling points length under the convolution network of two stages 3*1; a signal which is changed into a signal with the length of 512 channels 32 sampling points through a 2*1 up-sampling network; the signal with the length of the sampling point of 512 channels 32 and the signal with the length of the sampling point of 512 channels 32 on the left side are spliced to form a signal with the length of the sampling point of 1024 channels 32; a signal that becomes 512 channels 32 sample points long under the convolution network of two stages 3*1; a signal with a length of 256 channels and 64 sampling points is changed into a signal with a length of 256 channels through a 2*1 up-sampling network; the signal with the length of 256 channels 64 sampling points is spliced with the signal with the length of 256 channels 64 sampling points on the left side to form a signal with the length of 512 channels 64 sampling points; a signal that becomes 256 channels 64 sample points long under the convolution network of two stages 3*1; a signal that becomes 128 channels 128 sample points long through a 2*1 up-sampling network; the signal with 128 channel 128 sampling point length is spliced with the signal with 128 sampling point length of the left side 128 channel to form a signal with 256 channel 128 sampling point length; a signal that becomes 128 channels 128 sample points long under a two-stage 3*1 convolutional network; a signal with the length of 256 sampling points of 64 channels is changed into a signal with the length of 256 sampling points of 64 channels through a 2*1 up-sampling network; the signal with the length of 64 channels and 256 sampling points is spliced with the signal with the length of 64 channels and 256 sampling points on the left side to form a signal with the length of 128 channels and 256 sampling points; a signal with the length of 128 channels and 64 sampling points is changed into a signal with the length of 128 channels through a 4*1 downsampling network; at 5*1, a convolution network with a step size of 3 changes into a signal with a length of 256 channels 20 sampling points; in 5*1, the signal with the length of the sampling point of the 512 channels 6 is changed into a signal with the length of the sampling point of the 512 channels 6 under a convolution network with the step length of 3; finally, at 3*1, the step size is changed into a signal with 1024 channels and 2 sampling points under the convolution network with the step size of 2.

Ann branch network:

the processing of the branch network is mainly based on the multi-characteristic signal of the 1024 channel 2 sampling point length of the output layer of the main network to perform characteristic combination processing. For heart rate and blood oxygen saturation, the predicted value is obtained through the characteristic processing of the multi-layer convolution. The high-pressure and low-pressure values of the blood pressure values are associated data, and the data are respectively obtained by carrying out convolution processing on one dimension of the characteristic signal with the length of 1024 channel 2 sampling points to obtain predicted values.

The signal with 1024 channel 2 sampling point length is changed into the signal with 1024 channel 1 sampling point length through the convolution network of 2*1; the 1024-channel 1 sampling point length signal is changed into a 64-channel 1 sampling point length signal through a 1*1 convolution network; the 64-channel 1 sample length signal is converted to a 1-channel 1 sample length signal, i.e., a predicted heart rate value, via a 1*1 convolution network.

2.1Unet backbone network finally output 1024-channel 2-sampling-point length signal for predicting blood oxygen

The signal with 1024 channel 2 sampling point length is changed into the signal with 1024 channel 1 sampling point length through the convolution network of 2*1; the 1024-channel 1 sampling point length signal is changed into 256-channel 1 sampling point length signal through a 1*1 convolution network; the signal with the length of 256 channel 1 sampling points is changed into the signal with the length of 64 channel 1 sampling points through a convolution network of 1*1; the 64-channel 1 sample point length signal is changed into a 1-channel 1 sample point length signal, namely the predicted blood oxygen value, through a 1*1 convolution network.

2.2Unet backbone network finally output 1024-channel 2-sampling-point-length signal for predicting blood pressure

Taking a signal with the sampling point length of 1024 channels of first dimension data to be changed into a signal with the sampling point length of 256 channels 1 through a convolution network of 1*1; the signal with the length of 256 channel 1 sampling points is changed into the signal with the length of 64 channel 1 sampling points through a convolution network of 1*1; the 64-channel 1 sample length signal is converted to a 1-channel 1 sample length signal, i.e., the predicted systolic pressure, through the convolution network of 1*1.

Taking the signal with the sampling point length of 1024 channels of second dimension data to be changed into the signal with the sampling point length of 256 channels 1 through a convolution network of 1*1; the signal with the length of 256 channel 1 sampling points is changed into the signal with the length of 64 channel 1 sampling points through a convolution network of 1*1; the 64-channel 1 sample length signal is converted to a 1-channel 1 sample length signal, i.e., the predicted diastolic pressure, through the convolution network of 1*1.

It should be noted that, although the specific number of channels and sampling points is given above, those skilled in the art may adjust the number of channels and the number of sampling points according to the needs in practical use, and the present invention is not limited thereto.

The experimental test shows that the inventor actually tests by the model constructed by the invention, 100 people in the training set, 50 people in the verification set and 50 people in the test set are selected, and the final test effect proves that the method can be applied to practical application.

Heart rate accuracy: the probability of more than 90% falls within +/-3bpm

Accuracy of blood oxygen saturation: the probability of more than 90% falls within +/-2%

Blood pressure accuracy: the probability of over 85% high pressure SBP falls within +/-10bpm

The probability of over 95% low pressure DBP falls within +/-10 bpm.

Claims

1. A multi-parameter physiological information detection system based on an intelligent mobile device, wherein the intelligent mobile device comprises a lighting device and a camera device on the same side of the intelligent mobile device, the detection system comprising: the illumination equipment is used for irradiating a specific body part of an individual to be detected; the camera equipment is used for collecting video images of the irradiated part; the signal acquisition module is used for acquiring multi-frame images within a preset time obtained by the camera equipment and extracting PPG signals based on the images; the signal processing module comprises a UNet-ANN multi-branch feedback enhancement model, the UNet-ANN multi-branch feedback enhancement model is used for carrying out multi-parameter physiological information detection based on the PPG signal, wherein the UNet-ANN multi-branch feedback enhancement model adopts an asymmetric structure and comprises a single-Unet main network and four ANN branch networks, the asymmetric Unet main network is divided into a left module and a right module, each layer of the left module and the right module is connected through splicing, the right module is output from the rightmost side, and the four ANN branch networks are respectively used for detecting heart rate values, blood oxygen values, systolic pressure and diastolic pressure.

2. The multi-parameter physiological information detection system according to claim 1, further comprising a data processing module for framing the acquired video images, weighted summing the R-channel pixel values of each frame of images, and finally combining into a continuous PPG signal.

3. The system of claim 1, wherein the number of channels of the input signal is less than the number of samples and the number of channels of the output signal is greater than the number of samples in the UNet backbone network.

4. A multi-parameter physiological information detection system according to claim 3, wherein said UNet-ANN multi-branch feedback enhancement model performs parameter adjustment of UNet backbone network based on differences between true values and detection values of heart rate and blood oxygen of an individual to be detected as feedback conditions.

5. The multi-parameter physiological information detection system according to claim 1, wherein the image pickup device is further configured to capture a face video of the individual to be measured, and perform sex determination of the individual to be measured based on the face video.

6. The multi-parameter physiological information detection system according to claim 2, wherein,

the data processing module is used for carrying out 0.5-8HZ elliptical band-pass filtering processing on the finger tip PPG signal, detecting abnormal points and cleaning; windowing is carried out on the PPG signal for 30 seconds, the window width is 256 sampling points, and the step length is 150 sampling points; judging whether the data subjected to the window cleaning has data segments meeting the data quality requirements or not, and extracting corresponding data segments.

7. A method for constructing a multi-parameter physiological information detection model, the method comprising: step 1), irradiating the tail end of a body part of an individual to be detected by using a flash lamp of the intelligent mobile equipment; step 2), acquiring multi-frame continuous images of an individual to be detected by utilizing a camera of the intelligent mobile equipment; step 3), extracting an end PPG signal based on the image and processing the end PPG signal; step 4), constructing a UNet-ANN multi-branch feedback reinforcement model, wherein the UNet-ANN multi-branch feedback reinforcement model adopts an asymmetric structure and comprises a single-Unet main network and four ANN branch networks, the asymmetric Unet main network is divided into a left module and a right module, each layer of the left module and each layer of the right module are connected through splicing, the left module and the right module are output from the rightmost side, and the four ANN branch networks are respectively used for detecting heart rate values, blood oxygen values, systolic pressure and diastolic pressure; step 5), training the reinforcement model by using the marked finger end PPG signal.

8. The method of claim 7, wherein the step of processing the terminal PPG signal comprises: step 2.1, framing the obtained video, weighting and summing the R channel pixel values of each frame of image, and combining the R channel pixel values into a section of continuous PPG signal; step 2.2, performing 0.5-8HZ elliptical bandpass filtering treatment on the PPG signal, detecting abnormal points and cleaning; step 2.3, windowing PPG signals; step 2.4, judging that the data subjected to the window cleaning is subjected to data segment extraction and checking according to the step 3, if the data is checked to be qualified, carrying out the next step, otherwise, re-acquiring the data; step S3: extraction of PPG signals, the steps comprising: step 3.1, searching peak points for the processed PPG signals, and finding peak points and valley points in the PPG signals; step 3.2: d_peak representing the fluctuation condition of the peak point and d_valley representing the fluctuation condition of the valley point are calculated through the formula d= (max-min)/min, and if d_peak is less than 0.05 and d_valley is less than 0.08, the PPG data of the segment is reserved; step 3.3: judging whether the number of cycles contained in the current signal window is less than a preset value or not based on the number of peak points and valley points, and adopting a second extraction scheme if the number of cycles contained in the current signal window is less than the preset value; the second extraction scheme: step 3.4 window step 3.5 of dividing the processed PPG signal into individual periods: calculating the correlation coefficient r of adjacent periodic waveforms, and if r > is a correlation coefficient pre-threshold value, reserving the PPG signal; step 3.6: if the PPG signal processed in step 2 does not meet both extraction criteria, the segment of PPG signal is directly deleted.