Inplausible machine learning disease auxiliary classification method
In the differential diagnosis of Crohn's disease and intestinal tuberculosis, max-pooling operation and multiple classifiers combined with the patient's clinical and endoscopic characteristics are solved, and auxiliary diagnosis of high accuracy and interpretability is achieved.
Patent Information
- Application Number
- CN202510670885.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-23
- Publication Date
- 2025-08-15
AI Technical Summary
The prior art has problems with poor classification results due to insufficient model interpretation and small training data sets in the differential diagnosis of Crohn's disease (CD) and intestinal tuberculosis (ITB), especially in deep learning models, abnormal signals are easily weakened and masked.
The max-pooling operation was used to maximize the probability of lesion characteristics of multiple endoscopic images. Combined with the clinical characteristics of the patients, classifiers such as logistic regression, random forest, support vector machine, two-layer MLP and XGBoost were used to build auxiliary classification models, and interpretability was enhanced through feature attribution methods.
It improves the accuracy of differential diagnosis between CD and ITB, achieves multi-center generalization and robustness comparable to that of experts, and reduces the risk of misdiagnosis.
Smart Images
Figure CN120495780A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a disease classification method, and in particular to an explainable machine learning disease auxiliary classification method. Background Art
[0002] The diagnosis of highly similar diseases currently presents challenges, such as the differential diagnosis of Crohn's disease (CD) and intestinal tuberculosis (ITB). The two diseases have a high degree of overlap in clinical symptoms, such as abdominal pain, diarrhea, weight loss, and fever; there is also some overlap in endoscopic and imaging manifestations. For example, longitudinal ulcers and a cobblestone appearance can be seen in CD, while transverse ulcers, ileocecal valve involvement, or tuberculous granulomas are common in ITB. However, many lesions may also show similar inflammatory manifestations at the pathological level or imaging. In addition, there is currently a lack of biomarkers or serological indicators with sufficient sensitivity and specificity, making it particularly difficult to diagnose early or atypical cases. Because the treatment options for CD and ITB are very different, misdiagnosis or delayed diagnosis can seriously affect the effectiveness of subsequent treatment and increase the burden on patients.
[0003] In recent years, artificial intelligence (AI), particularly deep learning technology, has made initial progress in assisted diagnosis through gastrointestinal endoscopy. Some studies have utilized convolutional neural networks (CNNs) to automatically identify colonoscopic images, reporting diagnostic accuracy rates of 78% to 85% or even higher. Other approaches have employed few-shot learning, contextual learning, or fusion of multidimensional features from electronic medical records to increase the accuracy of differentiating CD from ITB to over 80%.
[0004] However, current research still has the following shortcomings:
[0005] (1) Insufficient model interpretability: Deep models are often regarded as “black boxes”. Clinicians have higher requirements for the transparency and visualization of their interpretation process. The lack of credibility limits their practical application.
[0006] (2) Due to the small number of training data sets, when using neural network prediction to perform averaging or summing operations, abnormal signals are easily weakened and masked, resulting in poor classification results of the trained classification model. Summary of the Invention
[0007] The purpose of the present invention is to provide an explainable machine learning disease auxiliary classification method in order to overcome the problem that when using neural network prediction to perform averaging or summing operations, abnormal signals are easily weakened and masked, resulting in poor classification effect of the trained classification model.
[0008] The purpose of the present invention can be achieved by the following technical solutions:
[0009] An interpretable machine learning disease auxiliary classification method, the method comprising the following steps:
[0010] S1. Collect case training data and extract initial features from the case training data;
[0011] S2, screening clinical features from initial features;
[0012] S3, obtain the final clinical diagnosis from the case training data as label data;
[0013] S4. Obtain endoscopic training images, use a priori deep learning models to obtain lesion features in the training images, and obtain the probability corresponding to each lesion feature. For multiple images, take the maximum value of multiple probabilities of the same lesion feature dimension as a single scalar of the lesion feature;
[0014] S5. Concatenate the clinical features and lesion features to obtain an N-dimensional feature vector, input it into the classifier for training and verification, and obtain an auxiliary classification model;
[0015] S6. Obtain actual endoscopic images and actual case data, input them into the auxiliary classification model to obtain the auxiliary classification results and the corresponding features with the greatest contribution.
[0016] Furthermore, the auxiliary classification model includes one or more of a logistic regression classifier, a random forest classifier, a support vector machine classifier, a two-layer MLP classifier and an XGBoost classifier.
[0017] Furthermore, the training process of the support vector machine classifier is specifically as follows:
[0018] A preliminary fit is performed on the training set, and then the support vector machine parameter grid is defined. The parameters are optimized using grid search combined with 10-fold stratified cross-validation to obtain the support vector machine classifier.
[0019] Furthermore, the support vector machine parameter grid includes regularization parameters and kernel function parameters.
[0020] Furthermore, the training process of the XGBoost classifier is as follows:
[0021] Set the XGBoost classifier parameter grid, use grid search combined with 10-fold stratified cross-validation to optimize the parameters, and obtain the XGBoost classifier.
[0022] Furthermore, the classifier parameter grid includes the maximum depth, number of estimators, subsampling ratio, and column sampling ratio.
[0023] Furthermore, the training process of the two-layer MLP classifier is specifically as follows:
[0024] Define a two-layer neural network model, use the cross entropy loss function and Adam optimizer for optimization, and obtain a two-layer MLP classifier.
[0025] Furthermore, during the training process of the logistic regression classifier, the logistic regression classifier uses a feature attribution method for the auxiliary classification results, and outputs the top k features of each patient that contribute most to the output of the logistic regression classifier.
[0026] Furthermore, when the auxiliary classification model includes multiple ones of a logistic regression classifier, a random forest classifier, a support vector machine classifier, a two-layer MLP classifier, and an XGBoost classifier, the auxiliary classification result is obtained based on a weighted average of the outputs of each classifier.
[0027] Furthermore, during the verification process, multiple different data sources are used.
[0028] Compared with the prior art, the present invention has the following beneficial effects:
[0029] The present invention uses a max-pooling operation in the classification model. For multiple images of each patient, the multiple probabilities of the same feature dimension are maximized to obtain a single scalar of lesion characteristics, reducing the difference in training and test distributions. The max-pooling object of the present invention is special, which is the lesion prediction probability vector of multiple images. That is, max-pooling is used here to solve the aggregation problem of a probability vector - traditional averaging or summing will weaken and mask abnormal signals, while max-pooling can well preserve abnormal signals, ensuring that the final important lesion information flows to the final symptom classification, thereby improving the accuracy of symptom classification. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 is a flow chart of the present invention;
[0031] Figure 2 Schematic diagram of max-pooling of endoscopic lesion features. DETAILED DESCRIPTION
[0032] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.
[0033] The present invention proposes an interpretable machine learning disease auxiliary classification method, the flow chart is as follows Figure 1 As shown, the method includes the following steps:
[0034] S1. Collect case training data and extract initial features from the case training data;
[0035] S2, screening clinical features from initial features;
[0036] S3, obtain the final clinical diagnosis from the case training data as label data;
[0037] S4. Obtain endoscopic training images, use a priori deep learning models to obtain lesion features in the training images, and obtain the probability corresponding to each lesion feature. For multiple images, take the maximum value of multiple probabilities of the same lesion feature dimension as a single scalar of the lesion feature;
[0038] S5. Concatenate the clinical features and lesion features to obtain an N-dimensional feature vector, input it into the classifier for training and verification, and obtain an auxiliary classification model;
[0039] S6. Obtain actual endoscopic images and actual case data, input them into the auxiliary classification model to obtain the auxiliary classification results and the corresponding features with the greatest contribution.
[0040] For example, the present invention can be applied to a comprehensive diagnostic method and system for Crohn's disease (CD) and tuberculous bowel disease (ITB) based on machine learning. By combining the patient's clinical diagnostic characteristics with the endoscopic lesion characteristics, the overall lesion probability information of the patient's endoscopic images is obtained using processing methods such as max-pooling, and then the multi-dimensional features are further spliced and normalized, and classifiers such as logistic regression, random forest, support vector machine, two-layer MLP and XGBoost are trained respectively to achieve differential diagnosis of CD and ITB. In internal and external validation, the present invention can achieve an accuracy comparable to or even better than that of experts, and has certain interpretability, multi-center scalability and good robustness, which can effectively reduce the risk of misdiagnosis and has significant clinical application value.
[0041] The present invention at this moment comprises the following steps:
[0042] Obtain and screen the patient's clinical diagnostic characteristics, eliminating non-specific features and features with extremely unbalanced categories;
[0043] Max-pooling operation is performed on the lesion prediction values of multiple endoscopic images of the same patient to obtain the maximum prediction probability of each lesion feature;
[0044] The clinical diagnostic features obtained by screening are spliced and normalized with the endoscopic image lesion features processed by max-pooling; the schematic diagram of endoscopic lesion feature max-pooling is shown in Figure 2 shown.
[0045] The concatenated feature vectors were input into multiple classification models for training to obtain a diagnostic model for distinguishing CD from ITB.
[0046] During application, the patient's clinical characteristics and endoscopic images are processed in the same manner, and a diagnosis conclusion of CD or ITB is outputted through the diagnostic model.
[0047] The original dataset was divided into a training set and a test set in a ratio of 8:2, where each patient had 3-5 typical endoscopic images. Then, an internal validation set and an external validation set were introduced. The internal validation set and the original dataset were of the same origin, and the number of endoscopic images of each patient was about 40. The external validation set was of a different origin from the original dataset, and the number of endoscopic images of each patient was still about 40.
[0048] Various classification models include logistic regression, random forest, support vector machine (Gaussian kernel configuration), two-layer MLP (with dropout), and XGBoost.
[0049] In the feature screening stage, 20 effective clinical features were retained, and non-specific features and features with extremely unbalanced categories were eliminated;
[0050] Eight endoscopic lesion characteristics of the patients were retained, including longitudinal ulcer, fish-mouth ileocecal valve, annular ulcer, aphthous ulcer, irregular ulcer, cobblestone change, no lesion and erosion;
[0051] The two are concatenated and then normalized to obtain the final classification features for each patient.
[0052] For support vector machines, XGBoost models all use the ten-fold cross-validation method to find the optimal parameters;
[0053] For the two-layer MLP classifier, the performance of the model on the test set is used to find the best training epoch.
[0054] The training results of the classifier are comprehensively judged using the maximum class probability method.
[0055] By using the feature attribution method, we can obtain the features that contribute most to the diagnosis of each patient to enhance the interpretability of the diagnosis results. The features with the highest contribution are highly consistent with the doctor's diagnostic experience, which can improve the reliability and interpretability of the diagnosis results.
[0056] To address these challenges, the present invention utilizes a combination of machine learning and deep learning classifiers, incorporating multi-dimensional information such as the patient's clinical diagnostic features and endoscopic lesion characteristics, to construct a comprehensive diagnostic system or method for distinguishing CD from ITB. This system aims to improve diagnostic accuracy and scalability across multiple centers and sample types. Through data cleaning, feature screening, and max-pooling, the present invention improves the understanding of the patient's overall characteristics in real-world scenarios while maintaining model complexity. This approach achieves diagnostic performance comparable to or even superior to that of experts, while also providing excellent interpretability.
[0057] The main purpose of this invention is to provide an AI-based comprehensive diagnostic system / method for CD and ITB. This system / method utilizes a combination of a patient's clinical diagnostic features and endoscopic lesion characteristics to train multiple classifiers for differential diagnosis. The system overcomes the shortcomings of existing technologies in terms of data quality, model interpretability, and multi-dimensional feature fusion by:
[0058] Multi-dimensional information fusion: combining the patient's clinical characteristics (such as test indicators, symptom characteristics, medical history, etc.) with endoscopic lesion characteristics to enhance the model's ability to distinguish;
[0059] Feature screening and distribution consistency strategy: The data is rigorously cleaned and screened. In practical applications, when a single patient has multiple endoscopic images, operations such as max-pooling are introduced to reduce the distribution difference between training and testing.
[0060] Multi-classifier ensemble: Multiple models, including logistic regression (LR), random forest (RF), support vector machine (SVM, using Gaussian kernel), two-layer MLP (with dropout), and XGBoost, are used for differential diagnosis separately or in combination;
[0061] Multi-center and external validation: Even when there are a large number of difficult cases in the internal validation set, a high level of accuracy can be maintained, and the model can be tested in an external validation set (different data sources) to ensure that the model has certain generalizability and robustness.
[0062] Further detailed steps are described below:
[0063] Data collection: A total of 342 patients diagnosed with CD or ITB in a hospital were collected. Basic information (age, sex, etc.), main clinical symptoms (fever, weight loss, etc.), laboratory tests (PPD, IGRAs, TB-AB (IgG), etc.), and other characteristics were obtained, totaling about 30 items.
[0064] Feature screening: Based on expert experience, statistical analysis, and ablation experiments, the 20 features most relevant to differential diagnosis and with an acceptable missing rate are retained. Features are sourced from multiple sources and are not guaranteed to be complete, including 1) clinical symptoms; 2) endoscopic findings; and 3) laboratory parameters. Some features may be missing. Therefore, the classifier proposed in this patent can still make robust predictions based on some features even in the absence of features.
[0065] Label confirmation: The final clinical diagnosis (CD / ITB) is used as the classification label;
[0066] Endoscopic image acquisition: For training data, 3 to 5 representative images of lesions were collected from each patient; for validation scenarios, up to 40 images were collected from each patient.
[0067] Pre-condition disease discriminator: Utilizes a priori deep learning models to output corresponding probabilities for several pathological features in the image (such as longitudinal ulcers, annular ulcers, etc.);
[0068] Max-pooling: For each patient's multiple images, the multiple probabilities for the same feature dimension are maximized to obtain a single scalar lesion feature, reducing the difference between training and testing distributions. Traditional max-pooling is often used to normalize image features, but this approach is specialized in that it targets lesion prediction probability vectors across multiple images. In other words, max-pooling solves the problem of aggregating probability vectors. Traditional averaging or summing can weaken or mask abnormal signals, while max-pooling effectively preserves these abnormal signals, ensuring that important lesion information flows to the final symptom prediction.
[0069] Model selection: logistic regression (LR), random forest (RF), SVM (Gaussian kernel), two-layer MLP (with dropout) and XGBoost;
[0070] A support vector machine (SVM) classifier was first preliminarily fitted on the training set. A parameter grid was then defined, including the regularization parameter (C:[0.1,1,3,6]) and kernel function parameter (gamma:["scale","auto",0.01,0.1,1]). Grid search combined with 10-fold stratified cross-validation was used to optimize the training set using accuracy as the evaluation metric to find the optimal parameter combination and train the optimal model.
[0071] The XGBoost classifier defines a parameter grid, including the maximum depth (max_depth: [6, 9]), the number of estimators (n_estimators:
[100] ), the subsampling ratio (subsample: [0.8, 1.0]), and the column sampling ratio (colsample_bytree: [0.3, 0.5, 0.6, 0.8]). Then, grid search is combined with 10-fold stratified cross-validation to fit the training set with accuracy as the evaluation metric to find the best parameter combination and train the best model.
[0072] A simple two-layer neural network model is defined using a multi-layer perceptron (MLP) and optimized using the cross-entropy loss function and the Adam optimizer (lr = 0.001). During training, the model iterates for up to 1000 epochs, with each epoch using forward propagation to compute output, calculate loss, and then backpropagate and update parameters. An early stopping mechanism (patience = 200) is also introduced to evaluate the test set accuracy at each epoch. If the accuracy exceeds the historical best, the current model state is saved. If there is no improvement for 200 consecutive epochs, the model is stopped early. Finally, the best model state is loaded and the accuracy is calculated on the test set, internal validation set, and external validation set.
[0073] Classification process: Clinical features and lesion features are concatenated to obtain an N-dimensional feature vector, which is input into the classifier and output as a predicted label of "CD" or "ITB";
[0074] Internal validation: Testing was performed on validation data from the same source as the training set, which contained approximately 40 endoscopic images per patient.
[0075] External validation: Evaluate data from different hospitals (but with similar test indicators) to observe cross-center generalization;
[0076] Enhanced interpretability: For the logistic regression classifier, we use feature attribution to identify the k features that contribute most to the diagnosis for each patient, enhancing the interpretability of the diagnosis. The top contributing features highly overlap with the doctor's diagnostic experience, improving the reliability and interpretability of the diagnosis.
[0077] Result analysis: The model is significantly superior to single features or traditional diagnostic methods in terms of accuracy, AUC, F1 score, etc., and reaches the level of top human experts.
[0078] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.
Claims
1. An interpretable machine learning disease auxiliary classification method, characterized in that: The method comprises the following steps: S1. Collect case training data and extract initial features from the case training data; S2, screening clinical features from initial features; S3, obtain the final clinical diagnosis from the case training data as label data; S4. Obtain endoscopic training images, use a priori deep learning models to obtain lesion features in the training images, and obtain the probability corresponding to each lesion feature. For multiple images, take the maximum value of multiple probabilities of the same lesion feature dimension as a single scalar of the lesion feature; S5. Concatenate the clinical features and lesion features to obtain an N-dimensional feature vector, input it into the classifier for training and verification, and obtain an auxiliary classification model; S6. Obtain actual endoscopic images and actual case data, input them into the auxiliary classification model to obtain the auxiliary classification results and the corresponding features with the greatest contribution.
2. The interpretable machine learning disease auxiliary classification method according to claim 1, characterized in that: The auxiliary classification model includes one or more of a logistic regression classifier, a random forest classifier, a support vector machine classifier, a two-layer MLP classifier and an XGBoost classifier.
3. The interpretable machine learning disease auxiliary classification method according to claim 2, characterized in that: The training process of the support vector machine classifier is specifically as follows: A preliminary fit is performed on the training set, and then the support vector machine parameter grid is defined. The parameters are optimized using grid search combined with 10-fold stratified cross-validation to obtain the support vector machine classifier.
4. The interpretable machine learning disease auxiliary classification method according to claim 3, characterized in that: The support vector machine parameter grid includes regularization parameters and kernel function parameters.
5. The interpretable machine learning disease auxiliary classification method according to claim 2, characterized in that: The training process of the XGBoost classifier is as follows: Set the XGBoost classifier parameter grid, use grid search combined with 10-fold stratified cross-validation to optimize the parameters, and obtain the XGBoost classifier.
6. The interpretable machine learning disease auxiliary classification method according to claim 5, characterized in that: The classifier parameter grid includes the maximum depth, number of estimators, subsampling ratio, and column sampling ratio.
7. The interpretable machine learning disease auxiliary classification method according to claim 2, characterized in that: The training process of the two-layer MLP classifier is specifically as follows: Define a two-layer neural network model, use the cross entropy loss function and Adam optimizer for optimization, and obtain a two-layer MLP classifier.
8. The interpretable machine learning disease auxiliary classification method according to claim 2, characterized in that: During the training process of the logistic regression classifier, the logistic regression classifier uses a feature attribution method for the auxiliary classification results and outputs the top k features of each patient that contribute most to the output of the logistic regression classifier.
9. The interpretable machine learning disease auxiliary classification method according to claim 2, characterized in that: When the auxiliary classification models include multiple types of logistic regression classifiers, random forest classifiers, support vector machine classifiers, two-layer MLP classifiers and XGBoost classifiers, the auxiliary classification results are obtained based on the weighted average of the outputs of each classifier.
10. The interpretable machine learning disease auxiliary classification method according to claim 1, characterized in that: During the verification process, several different data sources are used.