Otological disease auxiliary diagnosis large model construction method and large model system

Abstract

The invention provides an otological disease auxiliary diagnosis large model construction method and a large model system, and relates to the technical field of multi-modal learning modeling, and the method comprises the following steps: building four diagnosis sub-models, comprising a hearing loss evaluation model, an auditory meatus and tympanic membrane / middle ear state or lesion evaluation model, an auditory meatus and middle ear structure evaluation model and a medical history and medical record model. The four diagnosis sub-models are trained; the outputs of the four diagnosis sub-models are fused, and a large ear disease clinical diagnosis model is constructed; evaluating and optimizing the performance of the otological disease clinical diagnosis large model; according to the method, multiple sub diagnosis models are fused, multi-source heterogeneous data are used for training, the generalization ability of the model can be improved, the model can adapt to different application scenes, the decision process of the model can be better understood by analyzing the output of each sub diagnosis model, and the credibility of the model is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of multimodal learning modeling, and in particular to a method for constructing a large model for auxiliary diagnosis of ear diseases and a large model system. Background Art

[0002] In the diagnosis of ear diseases, the following methods are generally used: 1. Pure tone audiometry, which gives the nature of hearing loss, such as sensorineural, conductive, and mixed. If it is sensorineural, it generally does not support otitis media inflammation, while conductive and mixed types support otitis media inflammation or ossicular chain diseases such as otosclerosis. Therefore, a simple hearing test result can give one or more diagnostic suggestions. In the information integration link of this project, it not only verifies with other sub-diagnostic models in the diagnostic process, but also supplements treatment and prognosis information. For example, by calculating the air conduction difference, we can understand the activity of the ossicular chain, assist in whether to prepare for ossicular reconstruction before surgery, and predict postoperative treatment. Treatment effect; 2. Ear canal irrigation and endoscopic examination, generally performed by an endoscopist and a report will be issued. The endoscopic examination will give one or more diagnostic suggestions. The diagnostic level is generally high, but the disadvantage is the lack of understanding of deep structures, the range of lesions, and hearing; 3. Temporal bone CT plain scan, CT examination will give one or more diagnostic suggestions. CT can be used to make a preliminary diagnosis, and further understand the range of the lesion, assist in the decision-making of surgical methods, and play a role in both mutual verification and supplementation in the entire information integration process; 4. Medical history collection, asking the patient's main symptoms and duration, the clinician will interview and write the medical record, and give one or more diagnostic suggestions.

[0003] The above-mentioned various auxiliary diagnosis methods are mostly performed separately, using only a single algorithm or single modality data, resulting in insufficient identification ability for complex cases and insufficient diversity of training data, resulting in insufficient coverage of edge cases by the model and inability to fully capture disease characteristics, leading to missed diagnoses or misdiagnoses. Therefore, the present invention proposes a method for constructing a large model for auxiliary diagnosis of ear diseases and a large model system to solve the problems existing in the prior art. Summary of the Invention

[0004] In response to the above problems, the present invention proposes a method for constructing a large-scale model for auxiliary diagnosis of ear diseases and a large-scale model system. The method for constructing a large-scale model for auxiliary diagnosis of ear diseases and the large-scale model system effectively combine four sub-diagnostic models to construct a comprehensive large-scale model for clinical diagnosis of ear diseases. The large-scale model can comprehensively utilize multi-source data such as the patient's hearing test results, ear images, ear structure information, medical history and medical records to achieve accurate diagnosis of ear diseases.

[0005] To achieve the purpose of the present invention, the present invention is implemented through the following technical solutions:

[0006] The first aspect of the present invention provides a method for constructing a large auxiliary diagnosis model for ear diseases based on multiple sub-diagnostic models, comprising the following steps:

[0007] S1: Establish four sub-diagnostic models, including: hearing loss assessment model, ear canal and tympanic membrane / middle ear status or lesion assessment model, ear canal and middle ear structure assessment model, and medical history model;

[0008] S2: Train the above four sub-diagnostic models separately;

[0009] S3: Fuse the outputs of the four sub-diagnostic models to construct a large clinical diagnosis model for ear diseases;

[0010] S4: Evaluate and optimize the performance of large clinical diagnostic models for otolaryngology diseases.

[0011] A further improvement is that: S1 includes: collecting data, preprocessing data and building a model, wherein the collected data includes:

[0012] Collect pure tone audiometry, acoustic immittance testing, and otoacoustic emissions audiometry results to build a hearing loss assessment model;

[0013] Collect otoscope images and ear CT / MRI images to construct an evaluation model of the ear canal and tympanic membrane / middle ear status or pathology;

[0014] Collect ear CT / MRI images to construct an evaluation model of the ear canal and middle ear structure;

[0015] Medical history and medical records are collected, including basic patient information, medical history, symptoms, signs, previous diagnosis and treatment records, to build a medical history and medical record model.

[0016] A further improvement is that the data preprocessing includes the following steps:

[0017] Remove noise from hearing test results and standardize data formats;

[0018] Perform image enhancement and segmentation of lesion areas on ear endoscopy images and ear CT / MRI images;

[0019] Perform text cleaning, entity recognition, and relationship extraction on medical history records.

[0020] A further improvement is that in S2, training the hearing loss assessment model includes the following steps:

[0021] Input: hearing test results;

[0022] Output: type and degree of hearing loss;

[0023] Training process: Using a deep learning model, the hearing test results are analyzed to identify and extract features related to hearing loss;

[0024] The training of the model for assessing the status or lesions of the ear canal and tympanic membrane / middle ear includes the following steps:

[0025] Input: otoendoscopic images, ear CT / MRI images;

[0026] Output: Status and lesion type of ear canal and tympanic membrane / middle ear;

[0027] Training process: Use deep learning models to analyze otoendoscopic images and ear CT / MRI images to identify lesion areas and lesion types;

[0028] Training the ear canal and middle ear structure assessment model involves the following steps:

[0029] Input: CT / MRI images of the ear;

[0030] Output: structural information of the ear canal and middle ear;

[0031] Training process: Use deep learning models to analyze ear CT / MRI images and extract structural features of the ear canal and middle ear;

[0032] Training the medical history model involves the following steps:

[0033] Input: medical history text;

[0034] Output: patient's basic information, medical history, symptoms, signs, previous diagnosis and treatment records;

[0035] Training process: Use natural language processing technology to extract key information from medical history text.

[0036] A further improvement is that the outputs of the four sub-diagnosis models are fused in S3, including:

[0037] Feature splicing: splicing the output feature vectors of the four sub-diagnostic models to form a comprehensive feature vector;

[0038] Feature selection: Select the most effective features from the comprehensive feature vector to reduce feature dimensions and improve model efficiency;

[0039] Classifier training: Use the selected features to train a classifier for classification and diagnosis of ear diseases.

[0040] A further improvement is that the feature splicing includes the following steps:

[0041] Assume that the output feature vectors of the four sub-diagnostic models are:

[0042] Hearing loss assessment model:

[0043] ;

[0044] Model for assessing the status or lesions of the ear canal and tympanic membrane / middle ear:

[0045] ;

[0046] Model for evaluating the ear canal and middle ear structures:

[0047] ;

[0048] Medical history model:

[0049] ;

[0050] These four eigenvectors are concatenated to obtain a comprehensive eigenvector:

[0051] .

[0052] A further improvement is that the feature selection adopts any one of the three methods of L1 regularization, recursive feature elimination, and feature selection based on importance score, wherein:

[0053] The method of using L1 regularization includes the following steps: adding an L1 regularization term to the loss function so that the model automatically selects important features during training. The mathematical expression is:

[0054]

[0055] in, are model parameters, It is The comprehensive feature vector of samples, is the corresponding label, is the regularization coefficient, yes The L1 norm of

[0056] The method of recursive feature elimination includes the following steps: recursively training the model, removing the least important features in each iteration until the number of remaining features reaches a preset value;

[0057] Importance score-based feature selection includes the following steps: using all features to train the model and calculating the importance score of each feature; sorting the features from low to high according to their importance scores, and gradually removing the features with the lowest scores; repeating the above steps until the number of remaining features reaches a preset value.

[0058] A further improvement is that the classifier training adopts any one of the support vector machine, random forest, and gradient boosting tree methods, wherein:

[0059] The method of using support vector machines includes: finding a hyperplane that maximizes the interval between two types of samples. The mathematical expression is:

[0060]

[0061] The constraints are:

[0062]

[0063]

[0064]

[0065] in, is the normal vector of the hyperplane, is the bias term, is the penalty parameter, is the slack variable;

[0066] The random forest method involves building multiple decision trees and voting or averaging their predictions. The specific methods include:

[0067] Randomly extract n samples from the training set as the training set of a decision tree, randomly select s features as candidate features of the decision tree, repeat the above operation to build T decision trees, and use each decision tree to make predictions for new samples, and vote or average the prediction results to get the final result;

[0068] The method of using gradient boosting tree includes: gradually reducing the loss function of the model by iteratively training the decision tree. The mathematical expression is:

[0069]

[0070] in, It is After the model has been iterated, It is After the model has been iterated, is the learning rate, It is A decision tree.

[0071] A further improvement is that the step S4 includes the following steps:

[0072] The constructed clinical diagnosis model for ear diseases was evaluated using the test set. The evaluation indicators included accuracy, recall, and F1 value.

[0073] Optimize the model based on the evaluation results, including adjusting model parameters, increasing training data, and trying different model fusion methods.

[0074] A second aspect of the present invention provides a large-scale model system for auxiliary diagnosis of otological diseases based on multiple sub-diagnostic models, which is applied to the method for constructing a large-scale model for auxiliary diagnosis of otological diseases based on multiple sub-diagnostic models described in any one of the first aspects, and includes a hearing loss assessment model, an ear canal and tympanic membrane / middle ear status or lesion assessment model, an ear canal and middle ear structure assessment model, a medical history model, and a fusion output unit. The hearing loss assessment model is used to assess hearing loss; the ear canal and tympanic membrane / middle ear status or lesion assessment model is used to assess the ear canal and tympanic membrane, middle ear status or lesions; the ear canal and middle ear structure assessment model is used to extract the ear canal and middle ear structure; and the medical history model is used to identify medical history.

[0075] The fusion output unit is used to fuse the outputs of the hearing loss assessment model, the ear canal and tympanic membrane / middle ear status or lesion assessment model, the ear canal and middle ear structure assessment model, and the medical history model to construct a large clinical diagnosis model for ear diseases and perform auxiliary diagnosis of ear diseases.

[0076] The beneficial effects of the present invention are:

[0077] 1. By fusing multiple sub-diagnostic models, the present invention can effectively improve the diagnostic accuracy of ear diseases and reduce misdiagnosis and missed diagnosis. By using multi-source heterogeneous data for training, the generalization ability of the model can be improved, so that it can adapt to different application scenarios. By analyzing the output of each sub-diagnostic model, the decision-making process of the model can be better understood, the credibility of the model can be improved, and the comprehensive use of multi-source data can achieve accurate diagnosis of ear diseases, which is beneficial to assist doctors in treating ear diseases.

[0078] 2. Through the three steps of feature splicing, feature selection and classifier training, the present invention can effectively integrate the four sub-diagnostic models to construct a comprehensive clinical diagnosis model for ear diseases. This model can comprehensively utilize multi-source data to achieve accurate diagnosis of ear diseases and has important clinical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 is a flow chart of the method of the present invention;

[0080] Figure 2 It is a system composition diagram of the present invention. DETAILED DESCRIPTION

[0081] In order to deepen the understanding of the present invention, the present invention will be further described in detail below with reference to the examples. The examples are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0082] Example 1

[0083] according to Figure 1 、 2 As shown, this embodiment proposes a method for constructing a large auxiliary diagnosis model for ear diseases based on multiple sub-diagnosis models, including the following steps:

[0084] S1: Establish four sub-diagnostic models, including: hearing loss assessment model, ear canal and tympanic membrane / middle ear status or lesion assessment model, ear canal and middle ear structure assessment model and medical history model.

[0085] S2: Train the above four sub-diagnostic models separately.

[0086] S3: The outputs of the four sub-diagnostic models are integrated to construct a large clinical diagnosis model for ear diseases.

[0087] S4: Evaluate and optimize the performance of large clinical diagnostic models for otolaryngology diseases.

[0088] The S1 includes data collection, data preprocessing and model building.

[0089] The embodiments of the present invention are explained in detail below:

[0090] First, the data sources of the four sub-diagnostic models are explained:

[0091] Hearing loss assessment model: pure tone audiometry, acoustic immittance testing, otoacoustic emissions and other hearing test results.

[0092] Model for assessing the status or lesions of the ear canal and tympanic membrane / middle ear: otoendoscopic images, ear CT / MRI images.

[0093] Model for the evaluation of the ear canal and middle ear structures: CT / MRI images of the ear.

[0094] Medical history model: patient basic information, medical history, symptoms, signs, previous diagnosis and treatment records, etc.

[0095] The following describes the data preprocessing method:

[0096] Perform corresponding preprocessing on the data required by each sub-diagnosis model, for example:

[0097] Hearing test results: removing noise and standardizing data format.

[0098] Otoendoscopic images, ear CT / MRI images: image enhancement, segmentation of lesion areas.

[0099] Medical history and medical records: text cleaning, entity recognition, and relationship extraction.

[0100] The following describes the training method for each sub-diagnosis model:

[0101] Hearing loss assessment model:

[0102] Input: Hearing test results.

[0103] Output: Hearing loss type (e.g. conductive, sensorineural, mixed, etc.), degree (e.g. mild, moderate, severe, profound, etc.).

[0104] Model training process: Deep learning models, such as convolutional neural networks (CNN) or recurrent neural networks (RNN), can be used to analyze hearing test results and identify characteristics of hearing loss.

[0105] Model for assessing the status or lesions of the ear canal and tympanic membrane / middle ear:

[0106] Input: Otoendoscopic images, ear CT / MRI images.

[0107] Output: Status of the ear canal and tympanic membrane / middle ear (e.g. normal, inflammation, effusion, perforation, etc.) and type of lesion (e.g. otitis media, cholesteatoma, tympanic membrane perforation, etc.).

[0108] Model training process: Deep learning models, such as convolutional neural networks (CNNs), can be used to analyze otoendoscopic images and ear CT / MRI images to identify lesion areas and lesion types.

[0109] Model for evaluating the ear canal and middle ear structures:

[0110] Input: CT / MRI images of the ear.

[0111] Output: Structural information of the ear canal and middle ear (ear canal width, tympanic cavity size, ossicular chain integrity, etc.).

[0112] Model training process: Deep learning models, such as convolutional neural networks (CNNs), can be used to analyze ear CT / MRI images and extract the structural features of the ear canal and middle ear.

[0113] Medical history model:

[0114] Input: Medical history record text.

[0115] Output: patient's basic information, medical history, symptoms, signs, previous diagnosis and treatment records, etc.

[0116] Model training process: Natural language processing techniques, such as named entity recognition (NER) and relation extraction (RE), can be used to extract key information from medical history text.

[0117] Specifically, in step S3, the outputs of the four sub-diagnosis models are fused to construct a large clinical diagnosis model for ear diseases. The model fusion method can adopt the following steps:

[0118] Feature splicing: The output features of the four sub-diagnostic models are spliced ​​to form a comprehensive feature vector.

[0119] Feature selection: Select the most effective features from the comprehensive feature vector to reduce feature dimensions and improve model efficiency; examples include L1 regularization, recursive feature elimination (RFE), and feature selection based on importance scores.

[0120] Classifier training: Use the selected features to train a classifier, such as support vector machine (SVM), random forest (RF), gradient boosted tree (GBDT), to classify and diagnose ear diseases.

[0121] Specifically, the methods used in S4 to evaluate and optimize the performance of the large-scale clinical diagnosis model for ear diseases include:

[0122] Model evaluation: The constructed large model for clinical diagnosis of ear diseases is evaluated using a test set. Evaluation indicators include accuracy, recall rate, F1 value, etc.

[0123] Model optimization: Optimize the model based on the evaluation results, such as adjusting model parameters, increasing training data, and trying different model fusion methods.

[0124] The present invention achieves accurate diagnosis of ear diseases by comprehensively utilizing multi-source data and has important clinical application value.

[0125] Specifically, in this embodiment, the feature stitching includes the following steps:

[0126] Assume that the output feature vectors of the four sub-diagnostic models are:

[0127] Hearing loss assessment model:

[0128] ;

[0129] Model for assessing the status or lesions of the ear canal and tympanic membrane / middle ear:

[0130] ;

[0131] Model for evaluating the ear canal and middle ear structures:

[0132] ;

[0133] Medical history model:

[0134] ;

[0135] These four eigenvectors are concatenated to obtain a comprehensive eigenvector:

[0136] .

[0137] Specifically, in this embodiment, the feature selection adopts any one of L1 regularization (Lasso), recursive feature elimination (RFE), and feature selection based on importance score, wherein:

[0138] The specific method of L1 regularization (Lasso) is to add an L1 regularization term to the loss function so that the model automatically selects important features during training. The mathematical expression is:

[0139]

[0140] in, are model parameters, It is The comprehensive feature vector of samples, is the corresponding label, is the regularization coefficient, yes The L1 norm of .

[0141] The specific method of recursive feature elimination (RFE) is to recursively train the model and remove the least important features in each iteration until the preset number of features is reached.

[0142] RFE is an iterative elimination strategy that trains a model (such as SVM, logistic regression, random forest, etc.), ranks the importance of features, and then gradually removes the least important features and retrains the model until the set number of features is reached. The specific steps include:

[0143] (1) Training the model.

[0144] (2) Obtain the importance or weight of each feature.

[0145] (3) Remove the least important features.

[0146] (4) Retrain the model on the remaining features.

[0147] (5) Repeat steps (2)-(4) until the specified number of features are retained.

[0148] Importance score-based feature selection includes the following steps: using all features to train the model and calculating the importance score of each feature; sorting the features from low to high importance scores and gradually removing the features with the lowest scores; repeating the above steps until the number of remaining features reaches the preset value. The specific steps include:

[0149] a. Train an initial model using all features.

[0150] b. Calculate the importance score of each feature.

[0151] c. Remove the feature with the lowest importance score.

[0152] d. Repeat steps a to c until the number of remaining features meets the preset value.

[0153] Specifically, in this embodiment, the classifier training adopts any one of the support vector machine, random forest, and gradient boosting tree methods, wherein:

[0154] The support vector machine (SVM) method involves finding a hyperplane that maximizes the interval between two types of samples. The mathematical expression is:

[0155]

[0156] The constraints are:

[0157]

[0158]

[0159]

[0160] in, is the normal vector of the hyperplane, is the bias term, is the penalty parameter, is the slack variable.

[0161] The Random Forest (RF) method involves building multiple decision trees and voting or averaging their predictions. Specific methods include:

[0162] Randomly extract n samples from the training set as the training set of a decision tree, randomly select s features as candidate features of the decision tree, repeat the above operation, build T decision trees, and use each decision tree to make predictions for new samples. The prediction results are voted or averaged to obtain the final result.

[0163] The gradient boosting tree (GBDT) method involves iteratively training the decision tree to gradually reduce the model's loss function, which is mathematically expressed as:

[0164]

[0165] in, It is After the model has been iterated, It is After the model has been iterated, is the learning rate, It is A decision tree.

[0166] Through three steps, namely feature concatenation, feature selection, and classifier training, this paper effectively integrates four sub-diagnostic models to construct a comprehensive clinical diagnosis model for ear diseases. This model can comprehensively utilize multi-source data to achieve accurate diagnosis of ear diseases and has important clinical application value.

[0167] Example 2

[0168] according to Figure 1 、 2 As shown, this embodiment proposes a large model system for auxiliary diagnosis of ear diseases based on multiple sub-diagnostic models, which is applied to the large model construction method for auxiliary diagnosis of ear diseases based on multiple sub-diagnostic models described in any one of the first embodiments, including a hearing loss assessment model, an ear canal and tympanic membrane / middle ear status or lesion assessment model, an ear canal and middle ear structure assessment model, a medical history model and a fusion output unit. The hearing loss assessment model is used to assess hearing loss; the ear canal and tympanic membrane / middle ear status or lesion assessment model is used to assess the ear canal and tympanic membrane, middle ear status or lesions; the ear canal and middle ear structure assessment model is used to extract the ear canal and middle ear structure; and the medical history model is used to identify the medical history.

[0169] The fusion output unit is used to fuse the outputs of the hearing loss assessment model, the ear canal and tympanic membrane / middle ear status or lesion assessment model, the ear canal and middle ear structure assessment model, and the medical history model to construct a large clinical diagnosis model for ear diseases and perform auxiliary diagnosis of ear diseases.

[0170] By fusing multiple sub-diagnostic models, the present invention can effectively improve the diagnostic accuracy of ear diseases and reduce misdiagnosis and missed diagnosis. By using multi-source heterogeneous data for training, the generalization ability of the model can be improved, making it adaptable to different application scenarios. By analyzing the output of each sub-diagnostic model, the decision-making process of the model can be better understood, the credibility of the model can be improved, and the comprehensive utilization of multi-source data can achieve accurate diagnosis of ear diseases, which is helpful for assisting doctors in treating ear diseases. Moreover, through the three steps of feature splicing, feature selection and classifier training, the present invention can effectively fuse the four sub-diagnostic models to construct a comprehensive large model for clinical diagnosis of ear diseases. This model can comprehensively utilize multi-source data to achieve accurate diagnosis of ear diseases and has important clinical application value.

[0171] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a large auxiliary diagnosis model for ear diseases based on multiple sub-diagnostic models, characterized in that: The following steps are involved: S1: Establish four sub-diagnostic models, including: hearing loss assessment model, ear canal and tympanic membrane / middle ear status or lesion assessment model, ear canal and middle ear structure assessment model, and medical history model; S2: Train the above four sub-diagnostic models separately; S3: Fuse the outputs of the four sub-diagnostic models to construct a large clinical diagnosis model for ear diseases; S4: Evaluate and optimize the performance of large clinical diagnostic models for otolaryngology diseases.

2. The method for constructing a large-scale auxiliary diagnosis model for ear diseases based on multiple sub-diagnostic models according to claim 1, characterized in that: The S1 includes: collecting data, preprocessing data and building a model, wherein the collected data includes: Collect pure tone audiometry, acoustic immittance testing, and otoacoustic emissions audiometry results to build a hearing loss assessment model; Collect otoscope images and ear CT / MRI images to construct an evaluation model of the ear canal and tympanic membrane / middle ear status or pathology; Collect ear CT / MRI images to construct an evaluation model of the ear canal and middle ear structure; Medical history and medical records are collected, including basic patient information, medical history, symptoms, signs, previous diagnosis and treatment records, to build a medical history and medical record model.

3. The method for constructing a large-scale auxiliary diagnosis model for ear diseases based on multiple sub-diagnostic models according to claim 2, characterized in that: The data preprocessing comprises the following steps: Remove noise from hearing test results and standardize data formats; Perform image enhancement and segmentation of lesion areas on ear endoscopy images and ear CT / MRI images; Perform text cleaning, entity recognition, and relationship extraction on medical history records.

4. The method for constructing a large-scale auxiliary diagnosis model for ear diseases based on multiple sub-diagnostic models according to claim 1, characterized in that: In S2, training the hearing loss assessment model includes the following steps: Input: hearing test results; Output: type and degree of hearing loss; Training process: Using a deep learning model, the hearing test results are analyzed to identify and extract features related to hearing loss; The training of the model for assessing the status or lesions of the ear canal and tympanic membrane / middle ear includes the following steps: Input: otoendoscopic images, ear CT / MRI images; Output: Status and lesion type of ear canal and tympanic membrane / middle ear; Training process: Use deep learning models to analyze otoendoscopic images and ear CT / MRI images to identify lesion areas and lesion types; Training the ear canal and middle ear structure assessment model involves the following steps: Input: CT / MRI images of the ear; Output: structural information of the ear canal and middle ear; Training process: Use deep learning models to analyze ear CT / MRI images and extract structural features of the ear canal and middle ear; Training the medical history model involves the following steps: Input: medical history text; Output: patient's basic information, medical history, symptoms, signs, previous diagnosis and treatment records; Training process: Use natural language processing technology to extract key information from medical history text.

5. The method for constructing a large-scale auxiliary diagnosis model for ear diseases based on multiple sub-diagnostic models according to claim 1, characterized in that: The outputs of the four sub-diagnostic models are integrated in S3, including: Feature splicing: splicing the output feature vectors of the four sub-diagnostic models to form a comprehensive feature vector; Feature selection: Select the most effective features from the comprehensive feature vector to reduce feature dimensions and improve model efficiency; Classifier training: Use the selected features to train a classifier for classification and diagnosis of ear diseases.

6. The method for constructing a large-scale auxiliary diagnosis model for ear diseases based on multiple sub-diagnostic models according to claim 5, characterized in that: The feature splicing includes the following steps: Assume that the output feature vectors of the four sub-diagnostic models are: Hearing loss assessment model: ； Model for assessing the status or lesions of the ear canal and tympanic membrane / middle ear: ； Model for evaluation of ear canal and middle ear structures: ； Medical history model: ； These four eigenvectors are concatenated to obtain a comprehensive eigenvector: 。 7. The method for constructing a large-scale auxiliary diagnosis model for ear diseases based on multiple sub-diagnostic models according to claim 5, characterized in that: The feature selection adopts any one of the three methods: L1 regularization, recursive feature elimination, and feature selection based on importance score, wherein: The method of using L1 regularization includes the following steps: adding an L1 regularization term to the loss function so that the model automatically selects important features during training. The mathematical expression is: ； in, are model parameters, It is The comprehensive feature vector of samples, is the corresponding label, is the regularization coefficient, yes The L1 norm of The method of recursive feature elimination includes the following steps: recursively training the model, removing the least important features in each iteration until the number of remaining features reaches a preset value; Importance score-based feature selection includes the following steps: using all features to train the model and calculating the importance score of each feature; sorting the features from low to high according to their importance scores, and gradually removing the features with the lowest scores; repeating the above steps until the number of remaining features reaches a preset value.

8. The method for constructing a large-scale auxiliary diagnosis model for ear diseases based on multiple sub-diagnostic models according to claim 5, characterized in that: The classifier training adopts any one of the support vector machine, random forest, and gradient boosting tree methods, wherein: The method of using support vector machines includes: finding a hyperplane that maximizes the interval between two types of samples. The mathematical expression is: ； The constraints are: ； ； ； in, is the normal vector of the hyperplane, is the bias term, is the penalty parameter, is the slack variable; The random forest method involves building multiple decision trees and voting or averaging their predictions. The specific methods include: Randomly extract n samples from the training set as the training set of a decision tree, randomly select s features as candidate features of the decision tree, repeat the above operation to build T decision trees, and use each decision tree to make predictions for new samples, and vote or average the prediction results to get the final result; The method of using gradient boosting tree includes: gradually reducing the loss function of the model by iteratively training the decision tree. The mathematical expression is: in, It is After the model has been iterated, It is After the model has been iterated, is the learning rate, It is A decision tree.

9. The method for constructing a large-scale auxiliary diagnosis model for ear diseases based on multiple sub-diagnostic models according to claim 1, characterized in that: The S4 comprises the following steps: The constructed clinical diagnosis model for ear diseases was evaluated using the test set. The evaluation indicators included accuracy, recall, and F1 value. Optimize the model based on the evaluation results, including adjusting model parameters, increasing training data, and trying different model fusion methods.

10. A system for auxiliary diagnosis of otological diseases based on multiple sub-diagnostic models, applied to the method for constructing an auxiliary diagnosis of otological diseases based on multiple sub-diagnostic models as described in any one of claims 1 to 9, characterized in that: It includes a hearing loss assessment model, an ear canal and tympanic membrane / middle ear status or lesion assessment model, an ear canal and middle ear structure assessment model, a medical history model and a fusion output unit. The hearing loss assessment model is used to assess hearing loss; the ear canal and tympanic membrane / middle ear status or lesion assessment model is used to assess the ear canal and tympanic membrane, middle ear status or lesions; the ear canal and middle ear structure assessment model is used to extract the ear canal and middle ear structure; and the medical history model is used to identify medical history. The fusion output unit is used to fuse the outputs of the hearing loss assessment model, the ear canal and tympanic membrane / middle ear status or lesion assessment model, the ear canal and middle ear structure assessment model, and the medical history model to construct a large clinical diagnosis model for ear diseases and perform auxiliary diagnosis of ear diseases.

Images