Non-invasive measurement method of intracranial pressure based on artificial intelligence technology and near infrared technology

Through artificial intelligence technology and near-infrared technology, an AI model was established and non-invasive intracranial pressure monitoring was used to monitor non-invasive intracranial pressure, which solved the problem of trauma risk and insufficient accuracy of invasive monitoring, and achieved high-accuracy non-invasive intracranial intracranial Pressure measurement ensures patient safety and reduces costs.

CN114795171BActive Publication Date: 2024-11-26SUZHOU XUNYI TECH CO LTD

Patent Information

Application Number
CN202210110367.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-29
Publication Date
2024-11-26
Estimated Expiration
2042-01-29

AI Technical Summary

Technical Problem

The existing intracranial pressure monitoring methods mainly rely on invasive technology, which has problems with insufficient trauma risk and accuracy, especially for patients who have not undergone surgery, it is impossible to monitor. Although near-infrared technology is non-invasive, its accuracy is not high, and it is difficult to be clinically Widely used.

Method used

Using non-invasive measurement methods based on artificial intelligence and near-infrared technology, a large data set is established by collecting near-infrared and invasive intracranial pressure data from postoperative patients, training AI models to predict intracranial pressure in non-invasive patients, and improving monitoring Accuracy and reliability.

Benefits of technology

Fast and accurate non-invasive measurement of intracranial pressure is achieved, with a prediction error of less than 1%, which improves clinical adaptability and reliability of predicted results, avoids trauma to patients, and reduces monitoring costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114795171B_ABST
    Figure CN114795171B_ABST
Patent Text Reader

Abstract

A non-invasive intracranial pressure measurement method based on artificial intelligence technology and near-infrared technology, including: 1. Collecting intracranial near-infrared data and invasive intracranial pressure data of multiple postoperative patients; performing multiple interval collections on each postoperative patient, and the collection of near-infrared data is greater than or equal to 100 frames; through collection, a large data set including array near-infrared data and array invasive intracranial pressure data is formed; 2. Establishing an AI model through the large data set in an AI computing workstation; 3. Deploying the AI ​​model on the AI ​​computing workstation; when measuring the intracranial pressure of a new non-invasive patient, first collect the intracranial near-infrared data of the new patient and input it into the AI ​​computing workstation to calculate the predicted intracranial pressure value of the new patient. The present invention makes it possible to accurately grasp the trend of intracranial pressure changes of patients in real time in a non-invasive way, ensures the life safety of patients, and reduces application costs.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of biomedical engineering technology, and in particular to a non-invasive intracranial pressure measurement method based on artificial intelligence technology and near infrared technology. Background Art

[0002] The detection of intracranial pressure (ICP) is very important for the management of patients with intracranial brain injury, especially for patients with cerebral edema and cerebral hemorrhage. The changing trend of intracranial pressure has great guiding significance for intracranial injury. Maintaining a stable intracranial pressure helps to obtain a better prognosis and avoid craniocerebral injury or even death. The "Chinese Expert Consensus on Intracranial Pressure Monitoring in Craniocerebral Trauma" recommends that patients with acute severe craniocerebral trauma (GCS 3-8 points) should undergo timely intracranial pressure testing and use intracranial pressure monitoring data as one of the main reference data for their clinical monitoring.

[0003] At present, the clinical monitoring method of intracranial pressure mainly adopts invasive monitoring method, but this method is invasive and often causes infection, bleeding, tissue damage and other injuries to patients. At the same time, since surgery is required to place invasive monitoring equipment, for patients who need to monitor intracranial pressure but have not yet undergone surgical treatment or have no surgical indications, intracranial pressure data cannot be measured by this method.

[0004] Near infrared technology (NIRS) is a commonly used biological imaging and detection technology. The principle of this technology is to use the characteristics of multiple scattering of near infrared light in tissues. At the same time, light interacts with human tissues (such as brain tissue, etc.). Near infrared light follows the Beer-Lambert law, so that after a series of absorption and scattering in the tissue, the attenuation of light and other information are captured by sensors such as CCD, and then used to analyze tissue components. The absorption of near infrared light is mainly related to oxygenated hemoglobin HbO2 and deoxygenated hemoglobin Hb. The infrared absorption peaks of the two hemoglobins are concentrated in the two wavelength ranges of 850nm and 760nm. Therefore, this device is often used to detect the blood oxygen content in the skull, and derive the cerebral blood oxygen saturation RSO2 through Hb and HbO2. There is a certain correlation between cerebral blood oxygen saturation and intracranial pressure. Existing technologies usually use near infrared equipment to measure Hb and HbO2, and directly derive intracranial pressure ICP based on empirical formulas. Although this method solves the problem of non-invasiveness, the empirical formula usually has a large error, so it cannot solve the problem of intracranial pressure detection accuracy.

[0005] The existing patent 1 (201911014649.8) mentions using a near-infrared device to measure the cerebral blood flow heartbeat pulsatility index PI, and establishing a correlation between PI and ICP to deduce the value of ICP. However, based on the correlation image described in the patent (such as Figure 1) It can be seen that although there is a correlation between PI and ICP, their linear relationship is not strong. Therefore, the patent's method of simply estimating ICP through the linear correlation ICP=K*PI+b has a very large error, and this method is difficult to promote and apply in clinical practice.

[0006] The existing patent 2 (201810913909.4) mentions the use of near-infrared equipment to measure cerebral blood oxygen saturation RSO2, and the use of mean arterial blood pressure MAP to establish the empirical formula ICP = (MAP-1.267*RSO2) to derive the ICP value. Based on the data provided in the existing patent 2, compared with the existing patent 1, this method can improve the accuracy of ICP calculation due to the addition of MAP sensing equipment and key blood pressure data. However, there are still problems such as large errors. In addition, the added blood pressure monitoring equipment also further increases the cost and difficulty of using the equipment in clinical practice.

[0007] In summary, the currently commonly used method for monitoring intracranial pressure in clinical practice is mainly invasive monitoring, which is invasive monitoring and causes certain harm to patients, and can only be used for patients after surgery. Near-infrared technology can be used for non-invasive monitoring of intracranial pressure, but this method has insufficient accuracy, resulting in limited clinical application.

[0008] Therefore, how to solve the above-mentioned deficiencies in the prior art has become a subject to be studied and solved by the present invention. Summary of the invention

[0009] The purpose of the present invention is to provide a non-invasive measurement method of intracranial pressure based on artificial intelligence technology and near infrared technology.

[0010] In order to achieve the above object, the technical solution adopted by the present invention is:

[0011] A non-invasive measurement method for intracranial pressure based on artificial intelligence technology and near-infrared technology.

[0012] Preparation stage: positioning a collection end of a near-infrared collection device on the head of a patient with postoperative craniocerebral disease, and simultaneously connecting the near-infrared collection device and an invasive intracranial pressure monitoring device to a first AI computing workstation;

[0013] The measuring method comprises:

[0014] Step 1: In clinical practice, near-infrared data of the patient's brain after surgery are collected by the near-infrared acquisition device, and at the same time, invasive intracranial pressure data of the patient after surgery are collected by the invasive intracranial pressure monitoring device;

[0015] The above data are collected from a plurality of postoperative patients, and the number of patients collected is greater than or equal to 100;

[0016] The above data are collected multiple times for each postoperative patient, and the near-infrared data and the invasive intracranial pressure data are recorded simultaneously at least every half an hour; wherein the collection of the near-infrared data is required to be greater than or equal to 100 frames;

[0017] Through the above acquisition, a large data set including the array of near-infrared data and the array of invasive intracranial pressure data is formed;

[0018] Step 2: Establishing an AI model using the large data set in the first AI computing workstation;

[0019] Step 3: After the AI ​​model is established, deploy the AI ​​model in the first AI computing workstation or the second AI computing workstation;

[0020] When measuring the intracranial pressure of a new non-invasive patient, the intracranial near-infrared data of the new patient is collected and input into the first or second AI computing workstation where the AI ​​model is deployed, and the predicted intracranial pressure value of the new patient is calculated and output by the first or second AI computing workstation.

[0021] The relevant contents in the above technical solution are explained as follows:

[0022] 1. In the above scheme, the near-infrared acquisition device is a multi-channel acquisition device, and the number of near-infrared data points collected on the patient's head is greater than or equal to four, and the points at least cover the frontal lobe and temporal lobe areas of the patient's head.

[0023] 2. In the above scheme, the near infrared data in the large data set needs to be preprocessed, and the preprocessing operation includes filtering the near infrared data;

[0024] The data preprocessing method comprises the following steps:

[0025] 1. The sampling frequency of the near-infrared acquisition device is set to be greater than or equal to 10 Hz, and the distance between the photocells is set to be greater than or equal to 20 mm; each channel collects at least four light attenuations of different wavelengths, and the collection wavelength range is 720 to 1000 nm, and the interval between the collection wavelengths is greater than or equal to 25 nm;

[0026] The light attenuation is calculated as follows:

[0027] The incident light intensity is I0, the outgoing light intensity is I1, and the attenuation OD is

[0028]

[0029] 3. In the above scheme, the wavelet filtering is divided into the following steps:

[0030] 1. Perform orthogonal wavelet transform on the measured signal, select an orthogonal wavelet and decomposition layer number N, N is a positive integer, and perform N-layer wavelet decomposition on the original signal;

[0031] Second, nonlinear threshold processing is performed on the wavelet transform coefficients of the measured signal. The high-frequency coefficients of each layer from 1 to N are processed by the threshold, and the low-frequency coefficients are not processed;

[0032] 3. Reconstruct the processed wavelet coefficients, reconstruct the signal of the Nth layer low-frequency coefficients after decomposition and the processed first to Nth layer high-frequency coefficients to obtain the filtered signal.

[0033] 4. In the above scheme, the process of establishing an AI model includes the following steps:

[0034] 1. Randomly selecting patients from the large data set and dividing them into a training group, a validation group and a test group, wherein the ratio of the training group to the validation group is at least 2:1, and the ratio of the validation group to the test group is 1:1;

[0035] 2. Build an AI model, use the data of the training group to perform multiple rounds of training on the AI ​​model, and use the data of the validation group after each round of training to test the accuracy of the intracranial pressure prediction value output by the AI ​​model. The testing method is as follows:

[0036] The near infrared data is input into the AI ​​model, and the error comparison method is shown in the following formula:

[0037] Error = abs(ICPval – ICPpre) / ICPval

[0038] ICPpre is the predicted intracranial pressure value, and ICPval is the invasive intracranial pressure data;

[0039] 3. Record the average Error value of the validation group and compare it with the last training result. If the prediction accuracy of the AI ​​model is improved, continue training. If the accuracy of the AI ​​model is not significantly improved, the training is considered to be over.

[0040] 4. After the AI ​​model training is completed, the accuracy of the AI ​​model is tested using the test group data, and the error values ​​of all patients are recorded, and the maximum error and average error are calculated as the basis for evaluating the clinical adaptability of the AI ​​model.

[0041] 5. In the above solution, the AI ​​model is a neural network model.

[0042] 6. In the above scheme, for each craniocerebral disease, a corresponding AI model is established;

[0043] In step three, when measuring the intracranial pressure of a new non-invasive patient, first determine what kind of intracranial disease the patient has, and collect the intracranial near-infrared data of the new patient. Then, input the data into an AI computing workstation that is equipped with an AI model corresponding to the intracranial disease, and calculate and output the predicted intracranial pressure of the new patient through the AI ​​computing workstation.

[0044] The working principle and advantages of the present invention are as follows:

[0045] The present invention provides a non-invasive intracranial pressure measurement method based on artificial intelligence technology and near-infrared technology, comprising: 1. collecting intracranial near-infrared data and invasive intracranial pressure data of multiple postoperative patients; performing multiple interval collection on each postoperative patient, and the collection of near-infrared data is greater than or equal to 100 frames; through the collection, a large data set including an array of near-infrared data and an array of invasive intracranial pressure data is formed; 2. An AI model is established in an AI computing workstation through the large data set; 3. The AI ​​model is deployed on the AI ​​computing workstation; when measuring the intracranial pressure of a new non-invasive patient, the intracranial near-infrared data of the new patient is first collected and input into the AI ​​computing workstation to calculate the predicted intracranial pressure value of the new patient.

[0046] Compared with the existing technology, the present invention collects data from existing invasive patients and uses artificial intelligence technology to establish an AI model. The AI ​​model is used to realize the rapid and accurate non-invasive prediction of intracranial pressure based on near-infrared technology. The error between the predicted intracranial pressure and the actual intracranial pressure value monitored by invasive equipment can be less than 1%, which greatly improves the clinical adaptability of near-infrared monitoring of intracranial pressure and the reliability of the prediction results.

[0047] On the one hand, the present invention can avoid the trauma caused to the patient in order to obtain accurate intracranial pressure. On the other hand, the obtained intracranial pressure prediction value is reliable and accurate, making it possible to accurately grasp the patient's intracranial pressure change trend in real time in a non-invasive manner, thereby ensuring the patient's life safety and reducing application costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Attached Figure 1 It is a relationship diagram between PI and ICP in the prior art;

[0049] Attached Figure 2 is a flowchart of an embodiment of the present invention;

[0050] Attached Figure 3 Another flowchart of the embodiment of the present invention for different types of diseases;

[0051] Attached Figure 4 A flowchart of establishing an AI model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0053] Embodiment: The present invention will be clearly described below with diagrams and detailed descriptions. After understanding the embodiments of the present invention, any person skilled in the art can make changes and modifications based on the techniques taught by the present invention without departing from the spirit and scope of the present invention.

[0054] The terms used herein are only for describing specific embodiments and are not intended to be limiting of the present invention. Singular forms such as "a", "this", "here", "this" and "the" as used herein also include plural forms.

[0055] The terms “first”, “second”, etc. used in this document do not specifically refer to an order or sequence, nor are they used to limit the present case. They are only used to distinguish components or operations described with the same technical terms.

[0056] The terms “include,” “including,” “have,” etc. used in this document are open-ended terms, meaning including but not limited to.

[0057] The terms used in this document generally have the ordinary meaning of each term used in this field, in the context of the case and in the specific context, unless otherwise noted. Certain terms used to describe the present invention will be discussed below or elsewhere in this specification to provide additional guidance to those skilled in the art in describing the present invention.

[0058] See attached Figure 2 Shown is a non-invasive method for measuring intracranial pressure based on artificial intelligence technology and near-infrared technology.

[0059] Preparation stage: The acquisition end of a near-infrared acquisition device is fixed in the form of a headgear on the head of a patient with postoperative craniocerebral diseases, such as cerebral hemorrhage, cerebral edema, cerebral infarction, brain tumor, etc. The near-infrared acquisition device and an invasive intracranial pressure monitoring device are simultaneously connected to a first AI computing workstation for communication.

[0060] The measuring method comprises:

[0061] Step 1: In clinical practice, the near-infrared data of the patient's brain after surgery is collected by the near-infrared acquisition device, and at the same time, the precise invasive intracranial pressure data of the patient after surgery is collected by the invasive intracranial pressure monitoring device.

[0062] The above data are collected from a plurality of the postoperative patients, and the number of patients collected is greater than or equal to 100. For a certain type of craniocerebral disease, such as cerebral hemorrhage and cerebral edema, the number of patients collected is preferably greater than 150.

[0063] The above data are collected multiple times for each postoperative patient, and the near-infrared data and the invasive intracranial pressure data are recorded at least every half an hour (preferably every hour); the collection of near-infrared data is required to be greater than or equal to 100 frames. For example: for a device with a sampling frequency of 50Hz, the data collection should be no less than 2s (50*2=100); for a device with a sampling frequency of 20Hz, the data collection should be no less than 5s (20*5=100).

[0064] Through the above acquisition, a large data set including the array of near-infrared data and the array of invasive intracranial pressure data is formed.

[0065] Step 2: Establish an AI model using the large data set in the first AI computing workstation.

[0066] Step 3: After the AI ​​model is established, the AI ​​model is deployed in the first AI computing workstation or the second AI computing workstation. When measuring the intracranial pressure of a new non-invasive patient, the intracranial near-infrared data of the new patient is collected and input into the AI ​​computing workstation where the AI ​​model is deployed, and the predicted intracranial pressure value of the new patient is calculated and output by the AI ​​computing workstation.

[0067] Among them, the near-infrared acquisition device is a multi-channel acquisition device, and the near-infrared data points collected on the patient's head are greater than or equal to four (preferably greater than or equal to eight), and the points at least cover the frontal lobe and temporal lobe areas of the patient's head.

[0068] Preferably, the near infrared data in step 1 and step 3 need to be preprocessed, and the preprocessing operation includes filtering the near infrared data; the data preprocessing method includes the following steps:

[0069] 1. Set the sampling frequency of the near-infrared acquisition device to be greater than or equal to 10 Hz (preferably 50 Hz), and the photoelectrode spacing to be greater than or equal to 20 mm (preferably 30 mm); each channel collects at least four different wavelengths of light attenuation, and the collection wavelength range is 720~1000 nm. The spacing of the collected wavelengths is greater than or equal to 25 nm, so that the mutual interference between the light attenuation is small. Preferably, the light attenuation of four wavelengths of 735 nm, 760 nm, 805 nm and 850 nm is collected. The number of channels corresponds to the number of collection points. Taking an eight-channel device as an example, the eight-channel device can collect 32 data per frame at a frequency of 50 Hz. When the sampling time is 10 s, a total of 4*8*10*50=16000 data can be obtained.

[0070] The light attenuation is calculated as follows:

[0071] The incident light intensity is I0, the outgoing light intensity is I1, and the attenuation OD is:

[0072]

[0073] Second, for each set of data in each channel, use a 0.001-10Hz bandpass filter to filter it, and then perform wavelet filtering on the data.

[0074] For each set of data in each channel (different wavelengths), wavelet transform has the ability to characterize the local characteristics of the signal in the time and frequency domain. Wavelet transform does not require denoising post-processing and has a fast computing speed, which has great advantages in infrared spectral analysis.

[0075] Wherein, the wavelet filtering is divided into the following steps:

[0076] 1. Perform orthogonal wavelet transform on the measured signal, select an orthogonal wavelet and decomposition layer number N, N is a positive integer, and perform N-layer wavelet decomposition on the original signal;

[0077] Second, nonlinear threshold processing is performed on the wavelet transform coefficients of the measured signal. The high-frequency coefficients of each layer from 1 to N are processed by the threshold, and the low-frequency coefficients are not processed;

[0078] 3. Reconstruct the processed wavelet coefficients, reconstruct the signal of the Nth layer low-frequency coefficients after decomposition and the processed first to Nth layer high-frequency coefficients to obtain the filtered signal.

[0079] The wavelet filtering method preferably uses Haar wavelet and Daubechies wavelet, and the preferred decomposition layer is 4 layers, and the original data is subjected to 4-layer wavelet decomposition. The signal-to-noise ratio of the data is improved by wavelet transformation, baseline drift is eliminated, and the authenticity of the data is ensured.

[0080] The invasive intracranial pressure data does not require preprocessing.

[0081] Among them, Figure 4 As shown, in step 2, the process of establishing an AI model includes the following steps:

[0082] 1. Randomly selecting patients from the large data set and dividing them into a training group, a validation group and a test group, wherein the ratio of the training group to the validation group is at least 2:1, and the ratio of the validation group to the test group is 1:1, so as to ensure that the validation group and the test group have at least 25 to 100 patients;

[0083] Each patient has the corresponding near-infrared data and the invasive intracranial pressure data of multiple time periods.

[0084] The more patient data there is, the higher the ratio of the training group is, and the lower the ratio of the validation group and the test group is. Taking 150 to 300 patients as an example, the ratio of the training group, validation group, and test group can be 4:1:1; if the number of patients is greater than 10,000, the ratio of the training group, validation group, and test group can be 18:1:1. Ensure that there are at least 25 to 100 patients in the test group and validation group to effectively test the model. At the same time, the more patients in the training group, the more accurate the trained AI model is.

[0085] 2. Build an AI model, use the data of the training group to perform multiple rounds of training on the AI ​​model, and use the data of the validation group after each round of training to test the accuracy of the intracranial pressure prediction value output by the AI ​​model. The testing method is as follows:

[0086] The near infrared data is input into the AI ​​model, and the error comparison method is shown in the following formula:

[0087] Error = abs(ICPval – ICPpre) / ICPval

[0088] ICPpre is the predicted intracranial pressure value, and ICPval is the invasive intracranial pressure data;

[0089] 3. Record the average Error value of the validation group and compare it with the last training result. If the prediction accuracy of the AI ​​model improves, continue training. If the accuracy of the AI ​​model does not improve significantly after multiple rounds of training, the training is considered to be over.

[0090] Among them, the smaller the Error value, the higher the prediction accuracy of the AI ​​model, and the larger the Error value, the lower the prediction accuracy of the AI ​​model.

[0091] 4. After the AI ​​model training is completed, the accuracy of the AI ​​model is tested using the test group data, and the error values ​​of all patients are recorded, and the maximum error and average error are calculated as the basis for evaluating the clinical adaptability of the AI ​​model.

[0092] For example, when the maximum error is less than 10% and the average error is less than 5%, it can be judged to have clinical adaptability. However, for different craniocerebral diseases, the reference standards for the maximum error and average error may be different. At the same time, the minimum error can also be calculated as the basis for evaluation.

[0093] Among them, when collecting the near-infrared data in step one and step three, the number of channels, collection position, sampling frequency, photoelectrode spacing, collection wavelength and preprocessing method are all required to be consistent.

[0094] Among them, the input of the AI ​​model is near-infrared data (preferably filtered near-infrared data), and the output is intracranial pressure data monitored by an invasive intracranial pressure device. The best intracranial pressure prediction effect is achieved by training and tuning the AI ​​model.

[0095] The AI ​​model is preferably a convolutional neural network model. The number of layers of the neural network model can be further adjusted according to the amount of data in the large data set to achieve the most accurate prediction value.

[0096] Preferably, the present invention can establish an AI model corresponding to each craniocerebral disease, and then construct a set including several AI models; the process of establishing each AI model is carried out according to the preparation stage, step one and step two.

[0097] In step three, when measuring the intracranial pressure of a new non-invasive patient, the intracranial near-infrared data of the new patient is first collected, and at the same time, it is determined what kind of cranial and brain disease the patient has. The data is then input into an AI computing workstation that is deployed with an AI model corresponding to the cranial and brain disease. The AI ​​computing workstation calculates and outputs the predicted intracranial pressure of the new patient.

[0098] In clinical practice, the correlation between near-infrared data and invasive intracranial pressure data of different intracranial diseases, such as cerebral hemorrhage, hydrocephalus, cerebral infarction, brain tumor, etc., is different. Therefore, a unified AI model for predicting the intracranial pressure values ​​of various intracranial diseases may have a certain degree of error. In order to further improve the accuracy of intracranial pressure value prediction and minimize the error, the technical solution of the present invention can establish respective AI models for different types of diseases. During the use of the equipment, the clinician selects the corresponding AI model according to the patient's disease, thereby achieving more accurate intracranial pressure prediction. The specific flow chart is as follows Figure 3 shown.

[0099] The "preparation stage" and "step one" of the present invention are only for postoperative patients, and the postoperative patients are already in an invasive state, that is, the present invention will not perform invasive operations on originally non-invasive patients in order to perform "step two" and "step three", and the collection of invasive intracranial pressure data of postoperative patients by invasive intracranial pressure monitoring equipment is also a prior art. The technical solution of the present invention does not include the operation step of performing surgical opening on the patient's head.

[0100] The near-infrared acquisition device, the invasive intracranial pressure monitoring device, the first AI computing workstation and the second AI computing workstation used in the present invention are all existing equipment, and their specific structural configurations or selections are not limited to the technical solution of the present invention.

[0101] The first AI computing workstation is used to analyze the data collected by the near-infrared acquisition device and the invasive intracranial pressure monitoring device and to establish an AI model. The second AI computing workstation is a workstation used by the hospital in specific use. Since it does not require data analysis and processing, nor does it require the establishment of an AI model, its hardware requirements are lower than those of the first AI computing workstation, so as to achieve the purpose of reducing the threshold and cost of use.

[0102] The present embodiment is described below with 5 specific cases:

[0103] Case 1:

[0104] The eight-channel OctaMon+ near-infrared acquisition headgear device provided by the American Cortech company was used to collect intracranial near-infrared data of patients with postoperative trauma. At the same time, the American Ennoga Camino MPM1 multi-parameter intracranial pressure monitor was used to collect the patients' invasive intracranial pressure data.

[0105] The number of patients collected data exceeded 300 (including at least 150 patients with cerebral hematoma and 150 patients with cerebral edema). Near-infrared data and invasive intracranial pressure data were collected from each patient once an hour. Each collection time was 10 seconds, and the total collection time was 48 hours. Therefore, each patient contained at least 48 sets of data, and the total data volume of all patients exceeded 14,400 sets of data.

[0106] Wavelet transform is used to filter the 8-channel near-infrared data collected each time to obtain filtered data.

[0107] The 300 patients were divided into a training group, a test group and a validation group. The training group contained all the data of 200 patients, the test group contained all the data of 50 patients, and the validation group contained all the data of 50 patients.

[0108] Build a convolutional neural network as an AI model, use L2-Loss and Smooth L1-Loss as the loss function for AI model training, perform multiple rounds of training on the AI ​​model, and use test data to find the AI ​​model with the highest accuracy.

[0109] Using all the data of the test group patients, the near-infrared data was input into the AI ​​model, and the output result was the intracranial pressure prediction value ICPpre. The intracranial pressure prediction value ICPpre was compared with the measured invasive intracranial pressure data ICPval. The error comparison method is shown in the following formula:

[0110] Error = abs(ICPval – ICPpre) / ICPval

[0111] Through the error calculation method described above, the errors of the test group were all less than 5%, with a minimum value of 0.4% and an average of 1.9%.

[0112] Case 2:

[0113] The eight-channel OctaMon+ near-infrared acquisition headgear device provided by the American Cortech company was used to collect intracranial near-infrared data of patients with postoperative trauma. At the same time, the American Ennoga Camino MPM1 multi-parameter intracranial pressure monitor was used to collect the patients' invasive intracranial pressure data.

[0114] The number of patients collected data exceeded 150, all of whom were patients who underwent cerebral hematoma puncture. Near-infrared data and invasive intracranial pressure data were collected from each patient once an hour. Each collection time was 10 seconds, and the total collection time was 48 hours. Therefore, each patient contained at least 48 sets of data, and the total data volume of all patients exceeded 7,200 sets of data.

[0115] Wavelet transform is used to filter the 8-channel near-infrared data collected each time to obtain filtered data.

[0116] The 150 patients were divided into a training group, a test group and a validation group. The training group contained all the data of 100 patients, the test group contained all the data of 25 patients, and the validation group contained all the data of 25 patients.

[0117] Build a convolutional neural network as an AI model, use L2-Loss and Smooth L1-Loss as the loss function for AI model training, perform multiple rounds of training on the AI ​​model, and use test data to find the AI ​​model with the highest accuracy.

[0118] Using all the data of the test group patients, the near-infrared data was input into the AI ​​model, and the output result was the intracranial pressure prediction value ICPpre. The intracranial pressure prediction value ICPpre was compared with the measured invasive intracranial pressure ICPval. The error comparison method is shown in the following formula:

[0119] Error = abs(ICPval – ICPpre) / ICPval

[0120] Using the error calculation method described above, the errors of the test group were all less than 4.2%, with a minimum value of 0.38% and an average of 1.7%.

[0121] Case 3:

[0122] The eight-channel OctaMon+ near-infrared acquisition headgear device provided by the American Cortech company was used to collect intracranial near-infrared data of patients with postoperative trauma. At the same time, the American Ennoga Camino MPM1 multi-parameter intracranial pressure monitor was used to collect the patients' invasive intracranial pressure data.

[0123] The number of patients collected data exceeded 150, all of whom were patients who underwent brain edema puncture surgery. Near-infrared data and invasive intracranial pressure data were collected from each patient once. Each collection time was 10 seconds, and the total collection time was 48 hours. Therefore, each patient contained at least 48 sets of data, and the total data volume of all patients exceeded 7,200 sets of data.

[0124] Wavelet transform is used to filter the 8-channel near-infrared data collected each time to obtain filtered data.

[0125] The 150 patients were divided into a training group, a test group and a validation group. The training group contained all the data of 100 patients, the test group contained all the data of 25 patients, and the validation group contained all the data of 25 patients.

[0126] Build a convolutional neural network as an AI model, use L2-Loss and Smooth L1-Loss as the loss function for AI model training, perform multiple rounds of training on the AI ​​model, and use test data to find the AI ​​model with the highest accuracy.

[0127] Using all the data of the test group patients, the near-infrared data was input into the AI ​​model, and the output result was the predicted intracranial pressure value ICPpre. The predicted intracranial pressure value ICPpre was compared with the measured invasive intracranial pressure ICPval. The error comparison method is shown in the following formula:

[0128] Error = abs(ICPval – ICPpre) / ICPval

[0129] Through the error calculation method described above, the errors of the test group were all less than 4.0%, the minimum value was 0.39%, and the average value was 1.8%.

[0130] Case 4:

[0131] The eight-channel OctaMon+ near-infrared acquisition headgear device provided by the American Cortech company was used to collect intracranial near-infrared data of patients with postoperative trauma. At the same time, the American Ennoga Camino MPM1 multi-parameter intracranial pressure monitor was used to collect the patients' invasive intracranial pressure data.

[0132] The number of patients collected data exceeded 300 (including at least 150 patients with cerebral hematoma and 150 patients with cerebral edema). Near-infrared data and invasive intracranial pressure data were collected from each patient once. Each collection time was 10 seconds, and the total collection time was 48 hours. Therefore, each patient contained at least 48 sets of data, and the total data volume of all patients exceeded 14,400 sets of data.

[0133] Wavelet transform is used to filter the 4-channel near-infrared data collected each time to obtain filtered data.

[0134] The 300 patients were divided into a training group, a test group and a validation group. The training group contained all the data of 200 patients, the test group contained all the data of 50 patients, and the validation group contained all the data of 50 patients.

[0135] Build a convolutional neural network as an AI model, use L2-Loss and Smooth L1-Loss as the loss function for AI model training, perform multiple rounds of training on the AI ​​model, and use test data to find the AI ​​model with the highest accuracy.

[0136] Using all the data of the test group patients, the near-infrared data was input into the AI ​​model, and the output result was the predicted intracranial pressure value ICPpre. The predicted intracranial pressure value ICPpre was compared with the measured invasive intracranial pressure ICPval. The error comparison method is shown in the following formula:

[0137] Error = abs(ICPval – ICPpre) / ICPval

[0138] Through the error calculation method described above, the errors of the test group were all less than 7%, with a minimum value of 0.69% and an average of 3.1%.

[0139] Test group patient number Disease types of patients in the test group 8-channel near-infrared data / All patients with cerebral hematoma and cerebral edema participated in the establishment of AI models 8 channels of near-infrared data / only patients with cerebral hematoma participated in the establishment of AI model 8 channels of near-infrared data / only patients with cerebral edema participated in the establishment of AI model 4-channel near-infrared data / All patients with cerebral hematoma and cerebral edema participated in the establishment of AI models 1 Brain hematoma 0.400 1.037 4.562 2 Brain hematoma 0.436 0.388 0.895 3 Brain hematoma 2.732 1.570 0.704 4 Brain hematoma 0.722 0.631 0.687 5 Brain hematoma 4.457 4.186 1.475 6 Brain hematoma 2.788 1.319 2.479 7 Brain hematoma 2.905 0.868 3.336 8 Brain hematoma 1.909 0.678 2.451 9 Brain hematoma 1.727 2.623 4.850 10 Brain hematoma 1.472 1.517 0.831 11 Brain hematoma 2.791 0.548 1.086 12 Brain hematoma 2.545 2.280 4.211 13 Brain hematoma 3.130 2.601 3.076 14 Brain hematoma 0.416 1.712 1.942 15 Brain hematoma 0.810 1.674 0.929 16 Brain hematoma 2.574 1.465 1.394 17 Brain hematoma 0.849 2.134 5.269 18 Brain hematoma 0.939 2.003 1.860 19 Brain hematoma 1.516 1.645 2.463 20 Brain hematoma 0.943 2.827 4.748 21 Brain hematoma 0.603 1.577 4.861 22 Brain hematoma 1.920 1.109 4.109 23 Brain hematoma 1.144 2.320 3.315 24 Brain hematoma 1.780 1.088 5.884 25 Brain hematoma 2.909 2.917 1.899 26 Brain edema 1.936 1.738 4.360 27 Brain edema 1.615 1.285 5.927 28 Brain edema 0.438 0.390 1.853 29 Brain edema 3.501 2.727 1.901 30 Brain edema 2.995 1.477 1.936 31 Brain edema 2.529 1.389 4.709 32 Brain edema 1.892 1.013 4.083 33 Brain edema 3.058 3.271 0.740 34 Brain edema 1.132 1.631 4.087 35 Brain edema 2.535 3.110 7.938 36 Brain edema 0.413 0.583 1.009 37 Brain edema 2.626 2.898 5.180 38 Brain edema 1.255 1.656 1.308 39 Brain edema 1.027 0.742 4.253 40 Brain edema 1.541 1.193 3.260 41 Brain edema 0.698 1.990 2.174 42 Brain edema 0.665 1.697 2.086 43 Brain edema 2.168 1.391 2.619 44 Brain edema 0.771 1.351 5.031 45 Brain edema 4.775 2.063 2.924 46 Brain edema 2.085 2.382 1.777 47 Brain edema 2.591 1.907 3.870 48 Brain edema 2.721 1.717 5.766 49 Brain edema 2.699 1.316 0.720 50 Brain edema 3.153 4.009 6.930 average value 1.905 1.709 1.797 3.115

[0140] Table 1

[0141] Table 1 shows that the average error of the above four cases is less than 5%, and the maximum error is less than 10%, which meets clinical adaptability. At the same time, from the above cases 1, 2, and 3, we can see that for different diseases, we can select data to train the AI ​​model in a targeted manner, so as to further improve the accuracy of the AI ​​model in predicting ICP. From cases 1 and 4, we can see that increasing the number of infrared acquisition channels can improve the accuracy of prediction.

[0142] Compared with the existing technology, the present invention collects data from existing invasive patients and uses artificial intelligence technology to establish an AI model. The AI ​​model is used to realize the rapid and accurate non-invasive prediction of intracranial pressure based on near-infrared technology. The error between the predicted intracranial pressure and the actual intracranial pressure value monitored by invasive equipment can be less than 1%, which greatly improves the clinical adaptability of near-infrared monitoring of intracranial pressure and the reliability of the prediction results.

[0143] On the one hand, the present invention can avoid the trauma caused to the patient in order to obtain accurate intracranial pressure. On the other hand, the obtained intracranial pressure prediction value is reliable and accurate, making it possible to accurately grasp the patient's intracranial pressure change trend in real time in a non-invasive manner, thereby ensuring the patient's life safety and reducing application costs.

[0144] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A non-invasive intracranial pressure measurement method based on artificial intelligence technology and near infrared technology, characterized in that: Preparation stage: positioning a collection end of a near-infrared collection device on the head of a patient with postoperative craniocerebral disease, and simultaneously connecting the near-infrared collection device and an invasive intracranial pressure monitoring device to a first AI computing workstation; The measuring method comprises: Step 1: In clinical practice, near-infrared data of the patient's brain after surgery are collected by the near-infrared acquisition device, and at the same time, invasive intracranial pressure data of the patient after surgery are collected by the invasive intracranial pressure monitoring device; The above data are collected from a plurality of postoperative patients, and the number of patients collected is greater than or equal to 100; The above data are collected multiple times for each postoperative patient, and the near-infrared data and the invasive intracranial pressure data are recorded simultaneously at least every half an hour; wherein the collection of the near-infrared data is required to be greater than or equal to 100 frames; Through the above acquisition, a large data set including the array of near-infrared data and the array of invasive intracranial pressure data is formed; Step 2: Establishing an AI model using the large data set in the first AI computing workstation; Step 3: After the AI ​​model is established, deploy the AI ​​model in the first AI computing workstation or the second AI computing workstation; When measuring intracranial pressure of a new non-invasively measured patient, by collecting intracranial near-infrared data of the new non-invasively measured patient, the data is input into a first or second AI computing workstation where the AI ​​model is deployed, and a predicted intracranial pressure value of the new non-invasively measured patient is calculated and output by the first or second AI computing workstation; The near infrared data in step 1 and step 3 need to be preprocessed, and the preprocessing operation includes performing wavelet filtering on the near infrared data; The wavelet filtering is divided into the following steps: (1) Perform an orthogonal wavelet transform on the measured signal, select an orthogonal wavelet and the number of decomposition layers N, where N is a positive integer, and perform N-layer wavelet decomposition on the original signal; (2) Perform nonlinear threshold processing on the wavelet transform coefficients of the measured signal. The high-frequency coefficients of each layer from 1 to N are processed by the threshold, and the low-frequency coefficients are not processed; (3) Reconstruct the processed wavelet coefficients, reconstruct the signal of the Nth layer low-frequency coefficients after decomposition and the processed first to Nth layer high-frequency coefficients to obtain the filtered signal; For each type of intracranial disease, a corresponding AI model is established respectively; in step three, when measuring the intracranial pressure of a new non-invasively measured patient, the intracranial near-infrared data of the new non-invasively measured patient is collected, and then the data is input into an AI computing workstation that is equipped with an AI model corresponding to the intracranial disease suffered by the new non-invasively measured patient, and the predicted intracranial pressure value of the new non-invasively measured patient is calculated and output by the AI ​​computing workstation.

2. The method for non-invasive measurement of intracranial pressure according to claim 1, characterized in that: The near-infrared acquisition device is a multi-channel acquisition device, and the number of near-infrared data points collected on the patient's head is greater than or equal to four, and the points at least cover the frontal lobe and temporal lobe areas of the patient's head.

3. The method for non-invasive measurement of intracranial pressure according to claim 2, characterized in that: The near infrared data preprocessing method in step 1 and step 3 comprises the following steps: (1) The sampling frequency of the near-infrared acquisition device is set to be greater than or equal to 10 Hz, and the distance between the photocells is set to be greater than or equal to 20 mm; each channel collects at least four light attenuations of different wavelengths, and the collection wavelength range is 720~1000nm, and the interval between the collection wavelengths is greater than or equal to 25nm; The light attenuation is calculated as follows: The incident light intensity is I0, the outgoing light intensity is I1, and the attenuation OD is: ; (2) For each set of data in each channel, a 0.001-10 Hz bandpass filter is used to filter it, and then a wavelet filter is performed on the data.

4. The method for non-invasive measurement of intracranial pressure according to claim 1 or 3, characterized in that: The process of building an AI model includes the following steps: (1) Randomly selecting patients from the large data set and dividing them into a training group, a validation group and a test group, wherein the ratio of the training group to the validation group is at least 2:1, and the ratio of the validation group to the test group is 1:1; (2) Build an AI model, use the data of the training group to perform multiple rounds of training on the AI ​​model, and use the data of the validation group after each round of training to test the accuracy of the intracranial pressure prediction value output by the AI ​​model. The testing method is as follows: The near infrared data is input into the AI ​​model, and the error comparison method is shown in the following formula: Error = abs(ICPval – ICPpre) / ICPval ICPpre is the predicted intracranial pressure value, and ICPval is the invasive intracranial pressure data; (3) Record the average Error value of the validation group and compare it with the last training result. If the prediction accuracy of the AI ​​model is improved, continue training. If the accuracy of the AI ​​model is not significantly improved, the training is considered to be over. (4) After the AI ​​model training is completed, the accuracy of the AI ​​model is tested using the test group data, and the error values ​​of all patients are recorded. The maximum error and average error are calculated as the basis for evaluating the clinical adaptability of the AI ​​model.

5. The method for non-invasive measurement of intracranial pressure according to claim 1, characterized in that: The AI ​​model is a neural network model.

Citation Information

Patent Citations

  • Determining intracranial pressure non-invasively by acoustic transducer

    CN101150989A

  • Multi-parameter-based intracranial pressure noninvasive detection method and device

    CN102429651A

  • Noninvasive intracranial pressure measuring method and noninvasive intracranial pressure analyzer using method

    CN103654760A

  • Method for monitoring intracranial pressure based on convolutional neural network algorithm

    CN106983504A

  • System and method for non-invasively monitoring intracranial pressure

    CN107613858A

Cited By

  • Continuous prediction method for intracranial pressure of traumatic brain injury patient based on LSTM (Long Short Term Memory) model

    CN121637194A