Gastroesophageal reflux AI auxiliary typing method based on esophageal function and morphological characteristics

Abstract

The invention discloses a gastroesophageal reflux AI auxiliary typing method based on esophageal function and morphological characteristics, and belongs to the technical field of gastroesophageal reflux, a first index parameter is acquired, esophageal functional characteristics are evaluated by using a pre-training model according to the first index parameter, and a functional score is calculated; wherein the first index parameter comprises a high-resolution esophageal power index parameter and an esophageal pH-impedance combined index parameter; obtaining esophageal endoscopic image index parameters, and according to the esophageal endoscopic image index parameters, evaluating the severity of the foraminal hernia defect and cardia relaxation by using the deep learning convolutional neural network prediction model, and calculating a morphological score; comprehensively evaluating the severity of the gastroesophageal reflux according to the functional score and the morphological score, and completing the precise typing of the gastroesophageal reflux. The severity of esophageal dysfunction and morphological defect is effectively evaluated based on a comprehensive index evaluation system, and precise typing of gastroesophageal reflux disease is completed.

Description

Technical Field

[0001] This invention belongs to the field of gastroesophageal reflux technology, specifically relating to an AI-assisted classification method for gastroesophageal reflux based on esophageal function and morphological characteristics. Background Technology

[0002] Gastroesophageal reflux disease (GERD) is a condition characterized by the reflux of stomach contents, causing discomfort and / or complications. Clinically, GERD is generally classified into esophageal symptom syndrome and extraesophageal symptom syndrome. Acid reflux and heartburn are typical clinical symptoms, while atypical symptoms include chest pain, back pain, and abdominal pain. Long-term erosion of the esophagus by refluxed material can also lead to complications such as esophageal stricture, bleeding, cancer, reflux laryngitis, cough, and asthma. The diverse clinical manifestations make accurate diagnosis and treatment selection for this disease quite challenging.

[0003] Currently, there is no strict grading standard for gastroesophageal reflux disease (GERD). Detailed classification helps in assessing adjustments to anti-reflux regimens and predicting prognosis. However, due to the complex and varied clinical manifestations of GERD, there is currently no clear and definitive disease classification standard. Reflux specialists typically assess the severity of GERD and the appropriate treatment plan based on a combined analysis of the degree of lower esophageal sphincter relaxation within the gastroesophageal valve and the patient's clinical symptoms. For example, milder cases of GERD can be treated with medications that inhibit gastric acid secretion and proton pump inhibitors; while when the patient has severe gastroesophageal valve relaxation and is accompanied by pathological manifestations and symptoms of reflux esophagitis, anti-reflux surgery should be considered.

[0004] While the diagnosis, classification, and severity assessment of GERD can currently be conducted using methods such as gastroscopy, esophageal impedance-pH monitoring, and esophageal manometry, the lack of clear, well-defined, and easily applicable classification criteria means that effectively utilizing the data from these examinations to objectively evaluate GERD severity remains highly dependent on the physician's personal experience. In some primary care hospitals lacking experienced physicians, the lack of clinical experience or the use of single testing methods may lead to misclassification of the disease, resulting in inappropriate clinical intervention plans. Therefore, there is an urgent need to develop an auxiliary diagnostic method that can accurately assess the classification and severity of GERD to effectively support clinical decision-making.

[0005] Currently, the diagnosis, classification, and severity assessment of GERD can be conducted using methods such as gastroscopy, combined esophageal impedance-pH monitoring, and esophageal manometry. Gastroscopy provides direct imaging of lesions from the lower esophagus to the esophageal mucosa (classified into grades A to D based on the Los Angeles classification) and the degree of lower esophageal sphincter relaxation. This method can reflect the severity of gastroesophageal valve function in the etiology of GERD to some extent and assess the size of hiatal hernia from a certain perspective (however, it is greatly influenced by the endoscopist's experience). It is one of the commonly used examinations in reflux surgery to assess whether GERD has anatomical abnormalities. However, because endoscopic imaging data cannot directly reflect the gastroesophageal motility parameters and reflux load, the sensitivity and specificity of this method in the diagnosis of GERD are limited, and it cannot directly determine whether a GERD patient requires surgery.

[0006] Esophageal impedance-pH combined monitoring and high-resolution esophageal motility testing are important diagnostic methods for GERD that have been increasingly used in clinical practice in recent years. Among them, pH-impedance combined monitoring can effectively monitor gastroesophageal reflux load, providing comprehensive reflux parameters. Therefore, it can clarify the correlation between reflux events and symptom events, effectively compensating for the limitations of endoscopic diagnosis of GERD. Secondly, high-resolution esophageal motility testing has high sensitivity and specificity for the dynamic parameters of hiatal hernia, allowing for accurate assessment of the integrity and function of the gastroesophageal reflux barrier. However, in clinical practice, due to the high correlation and mutual influence between the morphology of hiatal hernia, reflux load, and motility parameters, it is difficult to effectively assess the severity and surgical indications of gastroesophageal reflux disease based on a single detection method or independent indicator. To effectively improve the accuracy of GERD classification and severity assessment, machine learning technology can be used, employing convolutional neural networks (CNNs) in deep learning to comprehensively analyze combined detection data from esophageal functional and morphological studies. This allows for precise assessment of gastroesophageal reflux disease classification and severity, and the model can be used to determine whether a patient has surgical indications. In recent years, a series of machine learning methods have emerged for the diagnosis or classification of GERD. These methods employ different mathematical models and algorithms to assess the severity of GERD. Some of these methods analyze gastroscopy data based on an improved model of the random forest algorithm to classify GERD according to its severity; others use general CNN algorithms to assess the condition of the cardia in gastroscopy images; and still others use structured state-space sequence analysis to analyze reflux data in pH-impedance combined detection.

[0007] However, analysis and testing of the aforementioned auxiliary diagnostic models based on different machine learning algorithms reveal that these algorithms can only perform modeling and analysis based on single-source test data, failing to comprehensively analyze combined esophageal functional and gastroscopy data. This results in limited accuracy in disease classification and severity assessment. Furthermore, since most auxiliary diagnostic methods rely primarily on gastroscopy images, they cannot effectively assess esophageal functional and physiological indicators. For GERD patients without hiatal hernia morphological changes, these models have significant limitations and cannot make accurate diagnoses based solely on gastroscopy images. Summary of the Invention

[0008] In view of this, the present invention provides an AI-assisted classification method for gastroesophageal reflux based on esophageal function and morphological characteristics, which is used to solve the problems existing in the prior art. It is an AI-assisted decision-making method that accurately assesses the classification and severity of gastroesophageal reflux disease based on the joint detection data of esophageal function and endoscopic images.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A method for AI-assisted classification of gastroesophageal reflux based on esophageal function and morphological characteristics includes: S1, acquiring first indicator parameters, and using a pre-trained model to evaluate esophageal functional characteristics and calculate functional scores based on the first indicator parameters; wherein, the first indicator parameters include high-resolution esophageal motility parameters and esophageal pH-resistance joint parameters; S2, acquiring esophageal endoscopic image indicator parameters, and using a deep learning convolutional neural network prediction model to evaluate the severity of foramen hernia defects and cardia relaxation based on the esophageal endoscopic image indicator parameters, and calculate morphological scores; S3, comprehensively assessing the severity of gastroesophageal reflux based on functional and morphological scores to complete accurate classification of gastroesophageal reflux.

[0011] Furthermore, high-resolution esophageal motility parameters include: contraction front velocity, lower esophageal sphincter resting pressure, lower esophageal sphincter diastolic pressure, distal latency, distal contraction integral, lower esophageal sphincter length, bolus pressure, hiatal hernia, peristaltic contractions, synchronous contractions, ineffective peristalsis, multiple rapid swallowing DCI value, and the difference between the upper and lower borders of the esophageal sphincter.

[0012] Furthermore, the combined esophageal pH-resistance parameters include: exposure duration, number of acid refluxes, duration of acid exposure, proportion of acid exposure when upright, proportion of acid exposure when lying down, total proportion of acid exposure, average acid clearance time, longest acid exposure duration, Demeester score, number of resistive refluxes - liquid - acidic, number of resistive refluxes - liquid - weakly acidic, number of resistive refluxes - liquid - alkaline, number of resistive refluxes - liquid - all, number of resistive refluxes - mixture - acidic, number of resistive refluxes - mixture - weakly acidic, number of resistive refluxes - mixture - alkaline, number of resistive refluxes - mixture - all, number of resistive refluxes - total - acidic, number of resistive refluxes - total - weakly acidic, number of resistive refluxes - total - alkaline, and number of resistive refluxes - total - all.

[0013] Further, based on the first indicator parameter, the pre-trained model is used to evaluate esophageal functional characteristics and calculate the functional score, including: Step 1, obtaining the reflux severity label and the first indicator parameter of the training samples, calculating the Pearson correlation coefficient between each indicator parameter and the reflux severity label of the training samples based on the first calculation formula, and selecting the indicator parameters corresponding to the Pearson correlation coefficient that meet the preset first standard to construct the first indicator group; Step 2, calculating the correlation between each indicator parameter and the reflux severity label of the training samples based on the recursive feature elimination method, and selecting the indicator parameters that meet the preset second standard to construct the second indicator group; Step 3, obtaining the pre-selected key indicator parameters to construct the third indicator group; Step 4, constructing models using the random forest method based on the first, second, and third indicator groups respectively, and calculating the esophageal functional score; Step 5, calculating the average of the esophageal functional scores of the models constructed in Step 4 to obtain the functional score.

[0014] Furthermore, the first calculation formula includes:

[0015]

[0016] Where x represents an indicator of the training sample, y represents the backflow severity label of the training sample, n represents the number of training samples, and the subscript i represents the i-th training sample. i Let y represent the metric of the i-th training sample. i This represents the label indicating the severity of backflow for the i-th training sample;

[0017] In step 4, a model is constructed using the random forest method, and the esophageal functional score is calculated, including:

[0018]

[0019] Among them, h j (x) is the j-th decision tree, w jThe corresponding weights are T, where T is the total number of decision trees, and H(x) is a random forest model containing T decision trees.

[0020] In a random forest, each decision tree randomly selects m indicators and k samples; based on the selected indicators and samples, the Gini coefficient is calculated.

[0021]

[0022] Where Y is the number of sample types, D is the sample dataset, and p a This represents the proportion of samples in category a to the total number of samples.

[0023] The formula for calculating the functional score in step 5 includes:

[0024]

[0025] Among them, FS PCCs The esophageal functional score, FS, is derived from the index selected by the Pearson correlation coefficient. RFE FS represents the esophageal functional score obtained from the indicators selected by the recursive feature elimination method. Exert Esophageal functional score derived from indicators selected by specialists.

[0026] Further, step 2 includes: S121, inputting all the first indicator parameters into the pre-trained SVM model to obtain the performance evaluation of the SVM model and generating the performance evaluation index ranking F1; S122, removing the indicator parameter f* with the lowest performance evaluation index ranking in F1 to obtain new first indicator parameters; S123, repeating steps S121 and S122 until the performance evaluation of the SVM model reaches the preset expectation, and obtaining the latest first indicator parameters to construct the second indicator group.

[0027] Furthermore, the esophageal endoscopy image parameters include: a preset number of color images of the esophageal position and a preset number of color images of the cardia position; wherein, the color images of the esophageal position are acquired at the location where the esophageal mucosa is observed at the lower end of the esophagus, and the color images of the cardia position are acquired in the direction of observation from the stomach cavity towards the esophagus. At the same time, the image resolution of both the color images of the esophageal position and the color images of the cardia position is not less than 256 pixels * 256 pixels.

[0028] Furthermore, the construction of the deep learning convolutional neural network prediction model in S2 includes: acquiring training samples, dividing the training samples into training and test sets according to a preset partitioning ratio; wherein, the esophageal endoscopy image index parameters of each sample in the training set are rotated and mirrored; acquiring the initial DenseNet network; inputting the training set into the initial DenseNet network for training, obtaining the weights for classifying the degree of hernia defects and esophageal laxity at each orifice, and calculating the probability of classifying the severity of hernia defects and esophageal laxity at each orifice based on the weights; wherein, the activation function of the activation layer of the DenseNet network is f(x) = max(0,x); the loss function is the cross-entropy loss function: H(p,q) = -∑p i log(q i ), where p is the true distribution, q is the predicted distribution, and the subscript i represents the i-th sample; when the initial training result of the DenseNet network reaches the preset training conditions, the DenseNet network forms a deep learning convolutional neural network prediction model by establishing dense connections between all previous layers and subsequent layers; the morphological score calculation method in S2 includes: using the severity of foramen hernia defects and cardia relaxation analyzed by the DenseNet network on the parameters of esophageal endoscopy images as the morphological score.

[0029] Furthermore, the validation method for the deep learning convolutional neural network prediction model includes: obtaining a test set, scaling the esophageal endoscopy image parameters of the samples in the test set to 224*224 size, and obtaining multiple 56*56 matrices based on convolution and pooling; after one processing step of DenseBlock and Transition, a matrix of 28*28 size is obtained; the network structure of DenseNet consists of DenseBlock and Transition; through repeated processing of DenseBlock until the matrix is ​​7*7 size, the matrix is ​​pooled to obtain a 1*1 size matrix; the 1*1 size matrix is ​​input to a fully connected layer to obtain the weights for classifying each type of foramen hernia defect and esophageal laxity, which are used to validate the deep learning convolutional neural network prediction model.

[0030] Furthermore, based on cross-validation, the threshold for the pre-trained model at 95% specificity was analyzed and set to 0.42; the AG score was 2 points when the functional score of the sample data was greater than the threshold, and 1 point when it was less than the threshold.

[0031] Based on cross-validation, the threshold for the deep learning convolutional neural network prediction model at 95% specificity was set to 0.39; the AG score was 2 points when the morphological score of the sample data was greater than the threshold, and 1 point when it was less than the threshold.

[0032] The overall score is the sum of the AG score of the functional score and the AG score of the morphological score. If the total score is less than or equal to 2, a non-surgical treatment plan is recommended; if the total score is greater than 2, a surgical treatment plan is recommended.

[0033] The beneficial effects of this invention are as follows:

[0034] This invention can correct diagnostic errors caused by a single detection method, and can effectively assess the severity of esophageal dysfunction and morphological defects based on a comprehensive index evaluation system, complete the accurate classification of gastroesophageal reflux disease, and improve the accuracy of assessment of the severity grading and surgical indications of gastroesophageal reflux disease.

[0035] Other advantages, objectives, and features of the invention will be set forth in the following description and will be apparent to those skilled in the art in some respects, or may be learned by practice of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0036] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0037] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0038] Figure 1 This is a flowchart of an AI-assisted classification method for gastroesophageal reflux based on esophageal function and morphological characteristics, as described in an embodiment of the present invention.

[0039] Figure 2 This is a flowchart illustrating the specific technical implementation of an AI-assisted classification method for gastroesophageal reflux based on esophageal function and morphological characteristics in an embodiment of the present invention.

[0040] Figure 3 This is a schematic diagram of a specific embodiment of an AI-assisted classification method for gastroesophageal reflux based on esophageal function and morphological characteristics, as described in this invention. Detailed Implementation

[0041] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0042] like Figure 1As shown, this invention provides an AI-assisted classification method for gastroesophageal reflux based on esophageal function and morphological characteristics, comprising: S1, acquiring first index parameters, and using a pre-trained model to evaluate esophageal functional characteristics and calculate a functional score FS based on the first index parameters; wherein, the first index parameters include high-resolution esophageal motility index parameters and esophageal pH-resistance joint index parameters; S2, acquiring esophageal endoscopic image index parameters, and using a deep learning convolutional neural network prediction model to evaluate the severity of foramen hernia defects and cardia relaxation based on the esophageal endoscopic image index parameters and calculate a morphological score MS; S3, comprehensively assessing the severity of gastroesophageal reflux based on the functional score FS and the morphological score MS to complete the accurate classification of gastroesophageal reflux.

[0043] The working principle and beneficial effects of the above technical solution are as follows: Please refer to [link / reference]. Figure 2 This invention utilizes the AutoGERD algorithm, an artificial intelligence pre-trained model based on machine learning and convolutional neural network theory, to digitally normalize and artificially quantify key functional and morphological parameters of the esophagus for gastroesophageal reflux disease (GERD). Combined with modern medical diagnostic methods such as high-resolution esophageal motility testing, esophageal pH-impedance combined detection, and endoscopy, it quantitatively analyzes and provides real-time assessment of the disease classification and severity of GERD patients. Based on the output of the AutoGERD algorithm, it provides professional suggestions and plans for further treatment options, thus achieving AI-assisted diagnosis and decision support for GERD. The invented technical solution differs significantly from traditional gastroesophageal reflux disease (GERD) assessment and classification schemes. It can accurately assess the severity of GERD patients without the involvement of senior physicians. In particular, it can significantly improve the problem of inadequate diagnosis of GERD patients without hiatal hernia by imaging examinations, while simultaneously achieving accurate GERD classification. The technical solution protected by this invention provides a novel AI-assisted approach based on the GERD assessment system, which can greatly improve the diagnostic accuracy of primary healthcare institutions lacking senior medical personnel. It can also enable more GERD patients to receive professional AI-assisted diagnosis and personalized treatment plan assessment services.

[0044] In one embodiment, high-resolution esophageal motility parameters include: contraction front velocity, lower esophageal sphincter resting pressure, lower esophageal sphincter diastolic pressure, distal latency, distal contraction integral, lower esophageal sphincter length, bolus pressure, hiatal hernia, peristaltic contraction, synchronous contraction, ineffective peristalsis, multiple rapid swallowing DCI value, and difference between the upper and lower borders of the esophageal sphincter.

[0045] The working principle and beneficial effects of the above technical solution are as follows: Based on the sample data to be collected, the specific values ​​of 13 dynamic indicators output by the examination equipment are acquired and stored using the self-developed AutoGERD system. Among them, the sample data to be collected is preferably obtained from any sample data collected after high-resolution esophageal motility examination according to medical orders in the medical database. The 13 high-resolution esophageal motility indicators are: contraction front velocity (CFV), lower esophageal sphincter resting pressure (LES-P), lower esophageal sphincter diastolic pressure (LES-IRP), distal latency (DL), distal contraction integral (DCI), lower esophageal sphincter length (LES Length), intrabolus pressure (IBP), hiatal hernia (HH), peristaltic contraction (PC), synchronous contraction (SC), ineffective peristalsis (IEC), multiple rapid swallowing DCI value (MRS-DCI), and difference between the upper and lower edges of the esophageal sphincter (DUL). The acquisition of the above data is beneficial to improving the accuracy of subsequent classification.

[0046] In one embodiment, the combined esophageal pH-resistance parameters include: exposure duration, number of acid refluxes, duration of acid exposure, proportion of acid exposure while upright, proportion of acid exposure while lying down, total proportion of acid exposure, average acid clearance time, longest acid exposure duration, Demeester score, number of resistive refluxes - liquid - acidic, number of resistive refluxes - liquid - weakly acidic, number of resistive refluxes - liquid - alkaline, number of resistive refluxes - liquid - all, number of resistive refluxes - mixture - acidic, number of resistive refluxes - mixture - weakly acidic, number of resistive refluxes - mixture - alkaline, number of resistive refluxes - mixture - all, number of resistive refluxes - total - acidic, number of resistive refluxes - total - weakly acidic, number of resistive refluxes - total - alkaline, and number of resistive refluxes - total - all;

[0047] The working principle and beneficial effects of the above technical solution are as follows: Based on the second sample data to be collected, the specific values ​​of 21 kinetic indicators output by the examination equipment are acquired and stored using the self-developed AutoGERD system. Preferably, the second sample data to be collected is obtained from any second sample data collected from a medical database after 24-hour esophageal pH-impedance combined testing according to medical orders. The 21 esophageal pH-impedance combined indicator parameters are: exposure duration (WMT), acid reflux frequency (RE), acid exposure duration (TRE), acid exposure ratio when upright (PUT), acid exposure ratio when lying flat (PST), total acid exposure ratio (PTT), mean acid clearance time (TRAC), longest acid exposure duration (LRE), Demeester score (DMS), and impedance reflux frequency-liquid- The following data were collected: Acidity (RE-L-Ac), Number of Reverse Flows - Liquid - Weak Acid (RE-L-Wa), Number of Reverse Flows - Liquid - Alkaline (RE-L-Al), Number of Reverse Flows - Liquid - All (RE-L-Ak), Number of Reverse Flows - Mixture - Acidity (RE-M-Ac), Number of Reverse Flows - Mixture - Weak Acidity (RE-M-Wa), Number of Reverse Flows - Mixture - Alkaline (RE-M-Al), Number of Reverse Flows - Mixture - All (RE-M-Ak), Number of Reverse Flows - Total - Acidity (RE-T-Ac), Number of Reverse Flows - Total - Weak Acidity (RE-T-Wa), Number of Reverse Flows - Total - Alkaline (RE-T-Al), and Number of Reverse Flows - Total - All (RE-T-Ak). Collecting these data helps improve the accuracy of subsequent classification.

[0048] In one embodiment, based on the first indicator parameter, a pre-trained model is used to evaluate esophageal functional characteristics and calculate a functional score, including: Step 1, obtaining the reflux severity label and the first indicator parameter of the training samples; calculating the Pearson correlation coefficient between each indicator parameter and the reflux severity label of the training samples based on the first calculation formula; and selecting the indicator parameters corresponding to the Pearson correlation coefficient that meet the preset first standard to construct a first indicator group; Step 2, calculating the correlation between each indicator parameter and the reflux severity label of the training samples based on the recursive feature elimination method, and selecting the indicator parameters that meet the preset second standard to construct a second indicator group; Step 3, obtaining the pre-selected key indicator parameters to construct a third indicator group; Step 4, constructing models using the random forest method based on the first, second, and third indicator groups to calculate the esophageal functional score; Step 5, calculating the average of the esophageal functional scores of the models constructed in Step 4 to obtain the functional score, including:

[0049] The first calculation formula includes:

[0050]

[0051] Where x represents an indicator of the training sample, y represents the backflow severity label of the training sample, n represents the number of training samples, and the subscript i represents the i-th training sample. i Let y represent the metric of the i-th training sample. i This represents the label indicating the severity of backflow for the i-th training sample;

[0052] In step 4, a model is constructed using the random forest method, and the esophageal functional score is calculated, including:

[0053]

[0054] Among them, h j (x) is the j-th decision tree, w j The corresponding weights are T, where T is the total number of decision trees, and H(x) is a random forest model containing T decision trees.

[0055] In a random forest, each decision tree randomly selects m indicators and k samples; based on the selected indicators and samples, the Gini coefficient is calculated.

[0056]

[0057] Where Y is the number of sample types, D is the sample dataset, and p a This represents the proportion of samples in category a to the total number of samples.

[0058] The formula for calculating the functional score in step 5 includes:

[0059]

[0060] Among them, FS PCCs The esophageal functional score, FS, is derived from the index selected by the Pearson correlation coefficient. RFE FS represents the esophageal functional score obtained from the indicators selected by the recursive feature elimination method. Expert Esophageal functional score derived from indicators selected by specialist physicians;

[0061] Step 2 includes: S121, inputting all the first index parameters into the pre-trained SVM model to obtain the performance evaluation of the SVM model and generating the performance evaluation index ranking F1; S122, removing the index parameter f with the lowest performance evaluation index ranking from F1. * S123. Repeat steps S121 and S122 until the performance evaluation of the SVM model reaches the preset expectation, and obtain the latest first indicator parameters to construct the second indicator group.

[0062] The working principle and beneficial effects of the above technical solution are as follows: Select effective high-resolution esophageal motility indicators and esophageal pH-resistance joint index parameters from the sample data and send them into the self-developed AutoGERD dynamic characteristic analysis module. Use the pre-trained model to evaluate the esophageal functional characteristics and calculate the functional score.

[0063] Furthermore, the pre-trained model construction and scoring method is as follows:

[0064] A. Calculate the Pearson correlation coefficient between each indicator and the disease severity score of the training set samples; the Pearson correlation coefficient between two variables is defined as the quotient of the covariance and standard deviation between the two variables, and the specific calculation formula is as follows:

[0065]

[0066] Where x represents an indicator of the training sample, y represents the backflow severity label of the training sample, n represents the number of training samples, and the subscript i represents the i-th training sample. i Let y represent the metric of the i-th training sample. i This represents the label indicating the severity of backflow for the i-th training sample;

[0067] The indicators selected were DMS, HH, LES-IRP, LES-Length, LESP, LRE, PC, PTT, PUT, RE, TRAC, and TRE, with an absolute correlation coefficient greater than 0.1 and an absolute correlation coefficient less than 0.9.

[0068] B. A recursive feature elimination method is used to calculate the correlation between each indicator and the disease severity score of the training set samples. The specific steps are as follows: input the features of all indicators into the model. It is worth noting that "all indicators" here refers to the 34 indicators in the first indicator parameter; obtain the model's performance evaluation and derive the F1 ranking of the performance evaluation indicators; typically, this method uses an SVM model, which aims to find a hyperplane to separate the two classes of data while maximizing the geometric margin; therefore, this method is transformed into... The constraint is y i (w T x i +b)≥1. Where x is the indicator, y is the label, w is the weight vector of the indicator, and ||w|| represents the norm of w. This algorithm uses the L2 norm, i.e. b is a constant, the subscript i represents the i-th sample, and the subscript m represents the feature of the m-th indicator;

[0069] Select the feature with the lowest performance evaluation index. * ;

[0070] Remove the last eigenvalue f * This yields a new sequence dataset F←Ff * ;

[0071] The model is trained again to obtain a new set of performance evaluation metrics, F2.

[0072] Repeat the previous steps until the model's performance evaluation meets expectations, or the model's performance reaches its optimal level across all subsets of metrics.

[0073] The metrics for selecting the optimal number of features are: CFV, DCI, DL, DMS, HH, IBP, LES-IRP, LES-Length, LESP, MRS-DCI, TRAC, TRE, and WMT.

[0074] C. Based on the experience of senior experts, six indicators that are of great clinical importance were selected: DMS, HH, LES-IRP, LES-Length, LESP, and PTT.

[0075] D. For the above three combinations of indicators, models are constructed using the random forest method to calculate the esophageal functional score; the specific formula is as follows:

[0076]

[0077] Among them, h j (x) is the j-th decision tree, w j For the corresponding weights, T is the total number of decision trees, and H(x) is a random forest model containing T decision trees; each decision tree in the random forest randomly selects n indicators and k samples; based on the selected indicators and samples, the Gini coefficient is calculated:

[0078]

[0079] Where Y is the number of sample types, D is the sample dataset, and p a This represents the proportion of samples in category a to the total number of samples.

[0080] E. The final esophageal functional score is obtained by averaging the scores from the three different models; the specific formula is as follows:

[0081]

[0082] Among them, FS PCCs The esophageal functional score, FS, is derived from the index selected by the Pearson correlation coefficient. RFE FS represents the esophageal functional score obtained from the indicators selected by the recursive feature elimination method. Expert Esophageal functional score derived from indicators selected by specialist physicians;

[0083] It is worth noting that the three esophageal functional scores in step E are derived from different indicators (obtained from steps A, B, and C), and the three different esophageal functional scores are calculated using the method in step D.

[0084] The Gini coefficient is the basis for constructing the decision tree in the D step of the random forest. The attribute with the smallest Gini coefficient is used as the optimal splitting attribute. That is, the smaller the Gini coefficient, the more suitable it is for splitting labels.

[0085] In one embodiment, the esophageal endoscopy image parameters include: a preset number of color images of the esophageal position and a preset number of color images of the cardia position; wherein, the color images of the esophageal position are acquired at the location where the esophageal mucosa is observed at the lower end of the esophagus, and the color images of the cardia position are acquired in the direction of observation from the stomach cavity towards the esophagus; at the same time, the image resolution of both the color images of the esophageal position and the color images of the cardia position is not less than 256 pixels * 256 pixels.

[0086] The working principle and beneficial effects of the above technical solution are as follows: Based on the third sample data to be collected, the self-developed AutoGERD system is used to extract and store at least one color image each of the esophagus and cardia from the gastroscopy report of the third sample data; the esophageal image needs to show the position of the esophageal mucosa at the lower end of the esophagus, and the image resolution should not be less than 256 pixels * 256 pixels; the cardia image needs to be observed from the stomach cavity towards the esophagus, and the image resolution should not be less than 256 pixels * 256 pixels; the collection of the above data is beneficial to improving the accuracy of subsequent typing.

[0087] In one embodiment, the construction of the deep learning convolutional neural network prediction model in S2 includes: acquiring training samples, dividing the training samples into a training set and a test set according to a preset division ratio; wherein, the esophageal endoscopy image index parameters of each sample in the training set are rotated and mirrored; acquiring an initial DenseNet network; inputting the training set into the initial DenseNet network for training, obtaining the weights for classifying the degree of hernia defects and esophageal laxity at each orifice, and calculating the probability of classifying the severity of hernia defects and esophageal laxity at each orifice based on the weights; wherein, the activation function of the activation layer used by the DenseNet network is f(x) = max(0,x); the loss function is the cross-entropy loss function: H(p,q) = -∑p i log(q i), where p is the true distribution, q is the predicted distribution, and the subscript i represents the i-th sample; when the initial training results of the DenseNet network reach the preset training conditions, the DenseNet network forms a deep learning convolutional neural network prediction model by establishing dense connections between all preceding layers and subsequent layers; the morphological score calculation method in S2 includes: using the severity of foramen hernia defects and cardia relaxation analyzed by the DenseNet network on the esophageal endoscopy image index parameters as the morphological score; the verification method of the deep learning convolutional neural network prediction model includes: obtaining the test set, scaling the esophageal endoscopy image index parameters of the samples in the test set to 2 The 24*224 matrix is ​​processed by convolution and pooling to obtain multiple 56*56 matrices. After one pass of DenseBlock and Transition, a 28*28 matrix is ​​obtained. The DenseNet network structure consists of DenseBlock and Transition. Through repeated processing with DenseBlock until the matrix reaches a 7*7 size, pooling is performed to obtain a 1*1 matrix. The 1*1 matrix is ​​then input to a fully connected layer to obtain weights for classifying each type of foramen hernia defect and esophageal laxity, which are used to validate the deep learning convolutional neural network prediction model.

[0088] The working principle and beneficial effects of the above technical solution are as follows: The deep learning convolutional neural network prediction model construction and scoring method are as follows: The training images are divided into training and test sets in an 8:2 ratio; each image in the training set is rotated and mirrored to become four new images; each image is passed through the DenseNet network to obtain the weights of each category and calculate the probability of each category. The classification or typing here is a typology of the severity of gastroesophageal reflux disease (GERD). Preferably, the specific typology result is a percentage between 0 and 1, which represents the current severity of GERD. The classification result of the weights is obtained in the same way as the independent validation described below. The validation involves feeding independent data into the model, and the model calculates the score based on the formula and weights obtained from the training set. Similar results; DenseNet forms a deep CNN neural network by establishing dense connections between all preceding layers and subsequent layers, achieving feature reuse in the channel dimension, mitigating the gradient vanishing phenomenon, and reducing computational cost. It's worth noting that since this CNN neural network includes a DenseNet network, subsequent data processing directly uses the DenseNet network for data processing in its representation. However, in practical applications, it uses the DenseNet network within the CNN neural network for data prediction. The DenseNet network structure mainly consists of DenseBlocks and Transitions.

[0089] DenseBlock: DenseBlock is a basic module of DenseNet. In this module, the feature maps of each layer are of the same size and can be connected along the channel dimension. The non-linear combination function in DenseBlock uses a structure of normalization layer, activation layer and 3×3 convolutional layer. In order to reduce the amount of computation, a structure containing normalization layer, activation layer and 1×1 convolutional layer is added before the combination function in DenseBlock.

[0090] The Transition layer primarily connects two adjacent DenseBlocks and reduces the feature map size; it consists of a normalization layer, an activation layer, a 1×1 convolutional layer, and a 2×2 average pooling layer.

[0091] The activation function used in DenseNet's activation layer is f(x) = max(0,x);

[0092] The loss function used is the cross-entropy loss function: H(p,q)=-∑p i log(q i ), where p is the true distribution, q is the predicted distribution, and the subscript i indicates the i-th sample;

[0093] During verification, the input image is scaled to 224×244, and after convolution and pooling, it becomes multiple matrices of size 56×56; after DenseBlock(1) and Transition(1), the matrix size becomes 28×28; this process is repeated until DenseBlock(4) is completed, resulting in a 7×7 matrix, which is then pooled into a 1×1 matrix; finally, after passing through a fully connected layer, the weights for each category are obtained.

[0094] Finally, morphological scores were calculated based on the severity of the parameters of the esophageal endoscopy images analyzed using DenseNet.

[0095] In one embodiment, the severity of gastroesophageal reflux is comprehensively assessed based on functional and morphological scores, and information on whether surgical indications are present is output. This information is then integrated with the AutoGERD pre-set medical record database to achieve real-time recording and updating of case information.

[0096] Furthermore, the comprehensive scoring method is as follows:

[0097] Based on cross-validation, the threshold for the pre-trained model at 95% specificity was set to 0.42; when the functional score (FS) of the sample data is greater than the threshold, the AG score is 2 points, and when it is less than the threshold, the AG score is 1 point.

[0098] Based on cross-validation, the threshold for the deep learning convolutional neural network prediction model at 95% specificity was set to 0.39; when the morphological score (MS) of the sample data was greater than the threshold, the AG score was 2 points, and when it was less than the threshold, the AG score was 1 point.

[0099] The composite score is the sum of the AG score of the functional score (FS) and the AG score of the morphological score (MS). If the total score is less than or equal to 2, a non-surgical treatment plan is recommended; if the total score is greater than 2, a surgical treatment plan is recommended.

[0100] The gastroesophageal reflux classification results include recommended surgical treatment and recommended non-surgical treatment options.

[0101] Compared with existing technologies, this invention utilizes a self-developed AutoGERD scalable open-source framework, combined with currently accepted GERD disease classification theories and esophageal specialist medical practice experience, to intelligently analyze real-time parameters of esophageal function and morphology. Without requiring the personal experience of senior physicians, it comprehensively assesses the degree of changes in the morphology and function of esophageal foramen hernia, automatically completing the accurate classification of gastroesophageal reflux disease (GERD), improving the accuracy of assessing the severity of GERD and whether surgical indications are necessary. It is worth noting that the results of the technical solution protected by this invention are mainly in the form of software programs. These software algorithms and executable code are integrated into the supporting software of relevant medical devices or hospital information management systems. The system is simple and reusable to deploy, and can conveniently serve a large number of patients with gastroesophageal reflux disease (GERD), generating continuous economic and social benefits. Simultaneously, it is at a forward-looking preliminary research stage in the field of accurate diagnosis of GERD both domestically and internationally, possessing a technological first-mover advantage and unique advantages in AI medical applications. In particular, it has a positive effect on promoting the large-scale implementation of AI in digital healthcare. AI-assisted diagnostic applications have a significant promoting effect on alleviating the strain on social medical resources. The technical solution protected by this invention further enhances the feasibility and credibility of AI medical rehabilitation technology in the field of GERD assessment, building upon the significant technological breakthroughs achieved in existing artificial intelligence and deep learning technologies.

[0102] To more clearly illustrate the technical solutions of the embodiments of the present invention, the technical solutions of the present invention will be described below in conjunction with specific embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0103] like Figure 3 As shown, the medical records of two randomly selected suspected gastroesophageal reflux disease patients (0001 and 0002) were analyzed according to the specific operating procedures:

[0104] Patients 1.0001 and 0002 underwent high-resolution esophageal motility testing, esophageal pH-impedance combined testing, and gastroscopy as prescribed by their doctors.

[0105] Table 1 shows the 13 kinetic parameters and 21 impedance parameters of patients 2.0001 and 0002:

[0106] Table 1

[0107]

[0108]

[0109] For patient 3,0001, 12 indicators were selected for analysis of the PCC pre-training model, resulting in a FS (PCCs) score of 0.278; 13 indicators were selected for analysis of the RFE pre-training model, resulting in a FS (RFE) score of 0.169; and 6 indicators were selected for analysis of the Expert pre-training model, resulting in a FS (Expert) score of 0.360. The average of these three scores was used to obtain the functional score (FS) of the gastroesophageal system, which was 0.269.

[0110] 4. The images of the cardia and esophagus of patient 0001 were input into the DenseNet convolutional neural network model. After processing such as Convolution, Pooling, and Transition, several DenseBlocks were formed, and finally a one-dimensional vector value was obtained. After comprehensive processing, the morphological score (MS) of the gastroscopy image was obtained as 0.2948.

[0111] 5. The functional score (FS) of the gastroesophageal system of patient 0001 was converted to the AG score. Since 0.269 < 0.42, the AG score was 1. The morphological score (MS) of the gastroscopy image was converted to the AG score. Since 0.2948 < 0.39, the AG score was 1.

[0112] 6. The combined AG score is 2; the patient's gastroesophageal reflux disease is considered to be of mild severity, and a non-surgical treatment plan is recommended; for patient 0002, repeat steps 3-6;

[0113] For patient 7.0002, 12 indicators were selected for analysis of the PCC pre-training model, resulting in a FS (PCCs) score of 0.839; 13 indicators were selected for analysis of the RFE pre-training model, resulting in a FS (RFE) score of 0.666; and 6 indicators were selected for analysis of the Expert pre-training model, resulting in a FS (Expert) score of 0.710. The average of these three scores was used to obtain the functional score (FS) of the gastroesophageal system, which was 0.738.

[0114] 8. The images of the cardia and esophagus of patient 0002 were input into the DenseNet convolutional neural network model. After processing such as Convolution, Pooling, and Transition, several DenseBlocks were formed, and finally a one-dimensional vector value was obtained. After comprehensive processing, the morphological score (MS) of the gastroscopy image was obtained as 0.7272.

[0115] 9. The functional score (FS) of the gastroesophageal system of patient 0002 was converted to the AG score. Since 0.738 > 0.42, the AG score was 2. The morphological score (MS) of the gastroscopy image was converted to the AG score. Since 0.7272 > 0.39, the AG score was 2.

[0116] 10. The combined AG score is 4; the patient's gastroesophageal reflux disease is considered to be severe, and surgical treatment is recommended.

[0117] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.

Claims

1. An AI-assisted classification method for gastroesophageal reflux based on esophageal function and morphological characteristics, characterized in that, include: S1. Obtain the first indicator parameters. Based on the first indicator parameters, use a pre-trained model to evaluate the esophageal functional characteristics and calculate the functional score. The first indicator parameters include high-resolution esophageal motility parameters and esophageal pH-resistance joint parameters. S2. Obtain the esophageal endoscopy image indicator parameters. Based on the esophageal endoscopy image indicator parameters, use a deep learning convolutional neural network prediction model to evaluate the severity of foramen hernia defects and cardia relaxation and calculate the morphological score. S3. Based on the functional score and morphological score, comprehensively assess the severity of gastroesophageal reflux and complete the accurate classification of gastroesophageal reflux.

2. The AI-assisted classification method for gastroesophageal reflux based on esophageal function and morphological characteristics according to claim 1, characterized in that, High-resolution esophageal motility parameters include: contraction front velocity, lower esophageal sphincter resting pressure, lower esophageal sphincter diastolic pressure, distal latency, distal contraction integral, lower esophageal sphincter length, bolus pressure, hiatal hernia, peristaltic contractions, synchronous contractions, ineffective peristalsis, multiple rapid swallowing DCI value, and the difference between the upper and lower borders of the esophageal sphincter.

3. The AI-assisted classification method for gastroesophageal reflux based on esophageal function and morphological characteristics according to claim 1, characterized in that, The combined esophageal pH-resistance parameters include: exposure duration, number of acid refluxes, duration of acid exposure, proportion of acid exposure when upright, proportion of acid exposure when lying down, total proportion of acid exposure, average acid clearance time, longest acid exposure duration, Demeester score, number of impedance refluxes - liquid - acidic, number of impedance refluxes - liquid - weakly acidic, number of impedance refluxes - liquid - alkaline, number of impedance refluxes - liquid - all, number of impedance refluxes - mixture - acidic, number of impedance refluxes - mixture - weakly acidic, number of impedance refluxes - mixture - alkaline, number of impedance refluxes - mixture - all, number of impedance refluxes - total - acidic, number of impedance refluxes - total - weakly acidic, number of impedance refluxes - total - alkaline, and number of impedance refluxes - total - all.

4. The AI-assisted classification method for gastroesophageal reflux based on esophageal function and morphological characteristics according to claim 1, characterized in that, Based on the first indicator parameter, the esophageal functional characteristics are evaluated using a pre-trained model, and the functional score is calculated, including: Step 1, obtaining the reflux severity label and the first indicator parameter of the training samples, calculating the Pearson correlation coefficient between each indicator parameter and the reflux severity label of the training samples based on the first calculation formula, and selecting the indicator parameters corresponding to the Pearson correlation coefficient that meet the preset first standard to construct the first indicator group; Step 2, calculating the correlation between each indicator parameter and the reflux severity label of the training samples based on the recursive feature elimination method, and selecting the indicator parameters that meet the preset second standard to construct the second indicator group; Step 3, obtaining the pre-selected key indicator parameters to construct the third indicator group; Step 4, constructing models using the random forest method based on the first, second, and third indicator groups, and calculating the esophageal functional score; Step 5, calculating the average of the esophageal functional scores of the models constructed in Step 4 to obtain the functional score.

5. The AI-assisted classification method for gastroesophageal reflux based on esophageal function and morphological characteristics according to claim 4, characterized in that, First calculation formula include: Where x represents an indicator of the training sample, y represents the backflow severity label of the training sample, n represents the number of training samples, and the subscript i represents the i-th training sample. i Let y represent the metric of the i-th training sample. i This represents the label indicating the severity of backflow for the i-th training sample; In step 4, a model is constructed using the random forest method, and the esophageal functional score is calculated, including: Among them, h j (x) is the j-th decision tree, w j The corresponding weights are T, where T is the total number of decision trees, and H(x) is a random forest model containing T decision trees. In a random forest, each decision tree randomly selects m indicators and k samples; based on the selected indicators and samples, the Gini coefficient is calculated. Where Y is the number of sample types, D is the sample dataset, and p a This represents the proportion of samples in category a to the total number of samples. The formula for calculating the functional score in step 5 includes: Among them, FS PCCs The esophageal functional score, FS, is derived from the index selected by the Pearson correlation coefficient. RFE FS represents the esophageal functional score obtained from the indicators selected by the recursive feature elimination method. Expert Esophageal functional score derived from indicators selected by specialists.

6. The AI-assisted classification method for gastroesophageal reflux based on esophageal function and morphological characteristics according to claim 4, characterized in that, Step 2 includes: S121, inputting all the first index parameters into the pre-trained SVM model to obtain the performance evaluation of the SVM model and generating the performance evaluation index ranking F1; S122, removing the index parameter f with the lowest performance evaluation index ranking from F1. * S123. Repeat steps S121 and S122 until the performance evaluation of the SVM model reaches the preset expectation, and obtain the latest first indicator parameters to construct the second indicator group.

7. The AI-assisted classification method for gastroesophageal reflux based on esophageal function and morphological characteristics according to claim 1, characterized in that, The esophageal endoscopy image parameters include: a preset number of color images of the esophageal position and a preset number of color images of the cardia position; wherein, the color images of the esophageal position are acquired at the position where the esophageal mucosa is observed at the lower end of the esophagus, and the color images of the cardia position are acquired in the direction of observation from the stomach cavity towards the esophagus. At the same time, the image resolution of both the color images of the esophageal position and the color images of the cardia position is not less than 256 pixels * 256 pixels.

8. The AI-assisted classification method for gastroesophageal reflux based on esophageal function and morphological characteristics according to claim 1, characterized in that, The construction of the deep learning convolutional neural network prediction model in S2 includes: acquiring training samples, dividing the training samples into training and test sets according to a preset partitioning ratio; wherein, the esophageal endoscopy image parameters of each sample in the training set are rotated and mirrored; acquiring the initial DenseNet network; inputting the training set into the initial DenseNet network for training, obtaining the weights for classifying the degree of hernia defects and esophageal laxity at each foramen, and calculating the probability of classifying the severity of hernia defects and esophageal laxity at each foramen based on the weights; wherein, the activation function of the activation layer of the DenseNet network is f(x) = max(0, x); the loss function is the cross-entropy loss function: H(p, q) = -∑p i log(q i ), where p is the true distribution, q is the predicted distribution, and the subscript i represents the i-th sample; when the initial training result of the DenseNet network reaches the preset training conditions, the DenseNet network forms a deep learning convolutional neural network prediction model by establishing dense connections between all previous layers and subsequent layers; the morphological score calculation method in S2 includes: using the severity of foramen hernia defects and cardia relaxation analyzed by the DenseNet network on the parameters of esophageal endoscopy images as the morphological score.

9. The AI-assisted classification method for gastroesophageal reflux based on esophageal function and morphological characteristics according to claim 8, characterized in that, The validation method for the deep learning convolutional neural network prediction model includes: obtaining a test set, scaling the parameters of the esophageal endoscopy images in the test set to 224*224, and obtaining multiple 56*56 matrices through convolution and pooling; after one processing step of DenseBlock and Transition, a 28*28 matrix is ​​obtained; the DenseNet network structure consists of DenseBlock and Transition; through repeated processing of DenseBlock until the matrix is ​​7*7, the matrix is ​​pooled to obtain a 1*1 matrix; the 1*1 matrix is ​​input to a fully connected layer to obtain the weights for classifying each type of foramen hernia defect and esophageal laxity, which are used to validate the deep learning convolutional neural network prediction model.

10. The AI-assisted classification method for gastroesophageal reflux based on esophageal function and morphological characteristics according to claim 1, characterized in that, The severity of gastroesophageal reflux is assessed by comprehensively evaluating functional and morphological scores, and a precise classification of gastroesophageal reflux is completed, including: Based on cross-validation, the threshold for the pre-trained model at 95% specificity was analyzed and set to 0.42; the AG score was 2 points when the functional score of the sample data was greater than the threshold, and 1 point when it was less than the threshold. Based on cross-validation, the threshold for the deep learning convolutional neural network prediction model at 95% specificity was set to 0.39; the AG score was 2 points when the morphological score of the sample data was greater than the threshold, and 1 point when it was less than the threshold. The overall score is the sum of the AG score of the functional score and the AG score of the morphological score. If the total score is less than or equal to 2, a non-surgical treatment plan is recommended; if the total score is greater than 2, a surgical treatment plan is recommended.

Images