Training method of inflammatory bowel disease diagnosis and prediction model based on data driving

Abstract

The invention provides a training method and system of an inflammatory bowel disease diagnosis and prediction model based on data driving, a storage medium and electronic equipment, and relates to the technical field of model training. According to the method, the BP feedforward neural network is adopted to process and analyze the standard data of each inspection index extracted from the ultrasonic diagnosis report text, so that the rehabilitation cycle of the patient with the inflammatory bowel disease is predicted. Through training and verification, the model shows high accuracy and reliability, and provides a powerful decision support tool for doctors. In addition, the dynamic adjustment and updating capability of the model ensures that the model can adapt to continuously changing clinical data in practical application, and the application value in personalized treatment planning and disease management is further enhanced.

Description

Technical Field

[0001] The present invention relates to the field of model training technology, and in particular to a training method, system, storage medium and electronic device for a data-driven inflammatory bowel disease diagnosis and prediction model. Background Art

[0002] With the improvement of economic and technological levels, people pay more and more attention to their physical health. In particular, the diagnosis of inflammatory bowel disease has gradually become one of the focuses of people's attention. Timely and accurate detection and diagnosis are directly related to the hospitalization and rehabilitation cycle of patients with inflammatory bowel disease, thereby saving a lot of manpower and material resources.

[0003] In the modern medical system, with the rapid development of medical imaging technology, the role of imaging in clinical disease diagnosis has become increasingly important. In particular, ultrasound imaging technology plays a key role in the diagnosis of many diseases. However, the effectiveness of ultrasound diagnosis depends largely on the professional skills and experience of the operating physician. At the same time, there are significant differences in the diagnostic ability among ultrasound physicians, and it takes a high cost and a long period to train a high-level ultrasound physician. In addition, since the diagnostic results are affected by the subjective judgment of the physician, different doctors will give different diagnostic opinions on the same case. Therefore, limited by the physician's professional knowledge and diagnostic experience, there is a certain possibility of misdiagnosis in the ultrasound diagnosis of inflammatory bowel disease, which directly affects the consistency and accuracy principles of medical diagnosis.

[0004] The auxiliary diagnosis system with computational data processing technology as the core provides a breakthrough for the above problems and also marks an important trend in the future development of precision medicine. At the same time, with the development of the new generation of information technology and its widespread application in clinical practice, the large-scale accumulation of clinical diagnostic data has become possible, providing a data basis for the use of machine learning methods for auxiliary diagnosis. Combining computer-aided diagnosis technology for the diagnosis of inflammatory bowel disease not only helps to construct quantitative diagnostic indicators, but also ensures the objectivity and accuracy of the diagnosis. Therefore, studying how to quickly and accurately predict the recovery period of inflammatory bowel disease is a meaningful task with great application prospects and economic benefits. Summary of the invention

[0005] 1. Technical issues to be resolved

[0006] In view of the deficiencies in the prior art, the present invention provides a training method, system, storage medium and electronic device for a data-driven inflammatory bowel disease diagnostic prediction model, which solves the technical problem of being unable to quickly and accurately predict the recovery period of inflammatory bowel disease.

[0007] (II) Technical solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0009] A data-driven training method for a diagnostic prediction model of inflammatory bowel disease, comprising:

[0010] Acquire historical inflammatory bowel disease data; wherein the historical inflammatory bowel disease data includes a number of ultrasound diagnosis report texts covering multiple examination indicators and recovery cycle labels;

[0011] Each of the ultrasound diagnosis report texts is sequentially segmented, keyword information is extracted, and standardized to obtain a standardized inflammatory bowel disease feature matrix;

[0012] All standardized inflammatory bowel disease feature matrices were divided into training sets and validation sets;

[0013] The training set is used as the input of the BP feedforward neural network, and combined with the recovery cycle label, the network is iteratively optimized until the preset termination condition is met, so as to build the inflammatory bowel disease diagnosis and prediction model;

[0014] The validation set is used as input of the inflammatory bowel disease diagnosis and prediction model to obtain validation results to evaluate the performance of the model.

[0015] Preferably, the inspection indicators include quantitative indicators and non-quantitative indicators; wherein:

[0016] The quantitative indicators include intestinal wall thickness, blood flow Limberg grade, omental thickening and hypoechoic nodules, and intestinal wall elasticity value;

[0017] The non-quantitative indicators include local intestinal stenosis and dilatation and CDFI blood flow signals.

[0018] Preferably, the step of sequentially performing word segmentation, keyword information extraction and standardization processing on each of the ultrasonic diagnostic report texts comprises:

[0019] The ultrasound diagnosis report text is segmented by using forward maximum matching and reverse maximum matching respectively, and the segmentation results of the two different methods are compared. If the segmentation results are consistent, one of them is selected as the final word segmentation result; if the segmentation results are inconsistent, it is classified as an ambiguous segmentation result, and then the ambiguity is eliminated according to specific rules;

[0020] Identify and extract keyword information from the word segmentation results through a preset feature extraction framework, and convert the extracted features and entity relationships into a structured data format;

[0021] For the quantitative indicators of intestinal wall thickness, blood flow Limberg grade, omental thickening and hypoechoic nodules, and intestinal wall elasticity, the following standardized processing formula was used:

[0022]

[0023] Among them, x is the original data value of any quantitative indicator, μ is the mean of the original data, and σ is the standard deviation of the original data;

[0024] Non-quantitative indicators such as local intestinal stenosis and dilatation and CDFI blood flow signals are divided into two levels: 0-1.

[0025] Preferably, the training set and the validation set are divided using a K-fold cross-validation method.

[0026] Preferably, the training process of the BP feedforward neural network includes:

[0027] Step 10, Network initialization:

[0028] Assign a non-zero random number between (-1, 1) to the connection weight of each layer, and set the threshold a of each neuron in the hidden layer j and the threshold b of each neuron in the output layer k ;

[0029] Step 20, input signal forward propagation:

[0030] The standardized inflammatory bowel disease feature matrix is ​​set to X = (X 1 , X 2 , ..., X m ) is a variable, W ij is the weight between the i-th neuron in the input layer and the j-th neuron in the hidden layer, W jk is the weight between the kth neuron in the hidden layer and the kth neuron in the output layer; the calculation formula of the input H of the kth neuron in the hidden layer is:

[0031]

[0032] Among them, X 1 , X 2 , ..., X m They represent the standardized values ​​of different inspection index data, m represents the total number of inspection indexes; a j is the threshold of the jth neuron in the hidden layer; q represents the total number of neurons in the hidden layer;

[0033] The output of the kth neuron in the hidden layer is I j The calculation formula is:

[0034]

[0035] Where f is the activation function;

[0036] Similarly, the output value Y of the kth neuron in the output layer is calculated k for:

[0037]

[0038] Among them, q represents the total number of neurons in the output layer;

[0039] Step 30, error signal back propagation:

[0040] The predicted output Y of the kth neuron in the output layer k and the corresponding rehabilitation cycle label T k The error formula is:

[0041] Err k =Y k (1-Y k )(T k -Y k ), k = 1, 2, ..., l

[0042] The error calculation formula of the jth neuron in the hidden layer is:

[0043]

[0044] Step 40, update weights and thresholds:

[0045] The weight calculation formula of the new j-th neuron in the hidden layer is:

[0046] W ij =W ij +ΔW ij =W ij +LErr j X i

[0047] The threshold update formula for the jth neuron in the hidden layer is:

[0048] a j =a j +Δa j =a j +LErr j

[0049] Where, ΔW ij represents the weight change from the kth neuron in the input layer to the jth neuron in the hidden layer; Δa j represents the threshold change of the jth neuron in the hidden layer; L is the learning rate;

[0050] Similarly, update and calculate the weight and threshold of the kth neuron in the output layer;

[0051] Step 50, determine whether the expected error or maximum number of iterations is reached; if the expected error or maximum number of iterations is reached, end; otherwise return to Step 20.

[0052] Preferably, the accuracy, ROC curve and AUC value are used as evaluation indicators of model performance.

[0053] A data-driven inflammatory bowel disease diagnostic prediction model training system, comprising:

[0054] An acquisition module, used to acquire historical inflammatory bowel disease data; wherein the historical inflammatory bowel disease data includes a number of ultrasound diagnosis report texts covering multiple examination indicators and recovery cycle labels;

[0055] A processing module, used for sequentially performing word segmentation, keyword information extraction and standardization processing on each of the ultrasonic diagnostic report texts to obtain a standardized inflammatory bowel disease feature matrix;

[0056] A partitioning module is used to divide all standardized inflammatory bowel disease feature matrices into a training set and a validation set;

[0057] A training module, used to use the training set as the input of the BP feedforward neural network, and in combination with the rehabilitation cycle label, iteratively optimize the network until a preset termination condition is met, so as to build the inflammatory bowel disease diagnosis and prediction model;

[0058] The verification module is used to use the verification set as the input of the inflammatory bowel disease diagnosis and prediction model to obtain verification results to evaluate the performance of the model.

[0059] A storage medium stores a computer program for training a data-driven inflammatory bowel disease diagnostic prediction model, wherein the computer program enables a computer to execute the training method described above.

[0060] An electronic device, comprising:

[0061] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, the programs including instructions for executing the training method as described above.

[0062] (III) Beneficial effects

[0063] The present invention provides a training method, system, storage medium and electronic device for a data-driven inflammatory bowel disease diagnosis and prediction model. Compared with the prior art, it has the following beneficial effects:

[0064] In the present invention, the standard data of various examination indicators extracted from the ultrasound diagnosis report text are processed and analyzed by BP feedforward neural network to predict the recovery period of patients with inflammatory bowel disease. After training and verification, the model shows a high degree of accuracy and reliability, providing doctors with a powerful decision support tool. In addition, the dynamic adjustment and update capabilities of the model ensure that it can adapt to the ever-changing clinical data in practical applications, further enhancing its application value in personalized treatment planning and disease management. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0066] Figure 1 A block diagram of a data-driven inflammatory bowel disease diagnosis and prediction model training method provided in an embodiment of the present invention;

[0067] Figure 2 A schematic diagram of a BP neural network-based inflammatory bowel disease recovery cycle prediction model provided in an embodiment of the present invention;

[0068] Figure 3 A schematic diagram of a confusion matrix provided by an embodiment of the present invention;

[0069] Figure 4 An ROC curve and AUC value based on BP neural network prediction are provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0071] The embodiments of the present application provide a training method, system, storage medium and electronic device based on a data-driven inflammatory bowel disease diagnostic prediction model, thereby solving the technical problem of being unable to quickly and accurately predict the recovery period of inflammatory bowel disease. It avoids the technical drawbacks of related technologies that rely on the doctor's experience and professional judgment, work around inflammatory bowel disease images, and are difficult to systematically analyze and integrate the patient's complete information.

[0072] The technical solution in the embodiment of the present application is to solve the above technical problems, and the overall idea is as follows:

[0073] In view of the shortcomings of the prior art, the embodiment of the present invention proposes a set of standardized processing schemes for inflammatory bowel disease diagnosis text data, and on this basis, constructs an auxiliary diagnosis model for inflammatory bowel disease based on BP neural network. This model can make full use of the professional judgment and clinical experience of doctors, while analyzing the text data in the report to mine implicit patterns and associations in the disease progression and healing process. This model analyzes the sentence characteristics of pathological text data and combines Chinese word segmentation technology to extract the characteristics of ultrasound examinations of inflammatory bowel disease, thereby completing the transformation of diagnostic reports from non-standard free text to structured data. It not only provides strong decision-making support for the treatment of patients with chronic intestinal diseases, but also opens up a new path for machine learning research using medical text data, demonstrating the huge potential and application prospects in the field of medical artificial intelligence.

[0074] In addition, the embodiment of the present invention further regards the patient's recovery period (discharge status) as a binary classification problem based on the BP neural network, and calculates the prediction accuracy of the model based on the difference between the patient's predicted recovery period and the actual recovery period. At the same time, the commonly used performance evaluation criteria for classification problems in the medical clinical field are mainly used, including accuracy, confusion matrix, ROC curve and AUC value to verify the prediction results of the model.

[0075] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0076] Embodiment 1:

[0077] like Figure 1 As shown, an embodiment of the present invention provides a data-driven training method for an inflammatory bowel disease diagnosis and prediction model, comprising:

[0078] S1. Obtaining historical inflammatory bowel disease data; wherein the historical inflammatory bowel disease data includes a number of ultrasound diagnosis report texts covering multiple examination indicators and their recovery cycle labels;

[0079] S2. performing word segmentation, keyword information extraction and standardization processing on each of the ultrasound diagnosis report texts in turn to obtain a standardized inflammatory bowel disease feature matrix;

[0080] S3. Divide all standardized inflammatory bowel disease feature matrices into training sets and validation sets;

[0081] S4, using the training set as the input of the BP feedforward neural network, and combining the rehabilitation cycle label, iteratively optimizing the network until a preset termination condition is met, so as to build the inflammatory bowel disease diagnosis and prediction model;

[0082] S5. Use the validation set as input of the inflammatory bowel disease diagnosis and prediction model to obtain validation results to evaluate the performance of the model.

[0083] In an embodiment of the present invention, by analyzing inflammatory bowel disease medical text data through machine learning methods, the patient's recovery period (i.e., whether the patient meets the discharge conditions) can be predicted. It can not only reduce human errors and improve the objectivity and consistency of predictions, but also mine deep information from a large amount of text data to achieve dynamic monitoring and comprehensive evaluation of disease progression. By combining clinical experience and data analysis, machine learning provides a more accurate and efficient prediction tool for the treatment and management of chronic intestinal diseases, which helps to achieve personalized treatment and optimize the allocation of medical resources.

[0084] Next, we will introduce the steps of the above scheme in detail:

[0085] In step S1, historical inflammatory bowel disease data is obtained; wherein the historical inflammatory bowel disease data includes a number of ultrasound diagnosis report texts covering multiple examination indicators and recovery cycle labels.

[0086] In this step, relevant historical data with short-term rehabilitation and long-term rehabilitation labels are collected and appropriately deleted, and basic information such as doctors' diagnostic opinions, pathological results, and examination data are summarized and integrated. Examination indicators of the disease are extracted from the collected information based on the characteristics of historical data and medical expertise.

[0087] The inspection indicators include quantitative indicators and non-quantitative indicators; among which:

[0088] The quantitative indicators include intestinal wall thickness, blood flow Limberg grade, omental thickening and hypoechoic nodules, and intestinal wall elasticity value;

[0089] The non-quantitative indicators include local intestinal stenosis and dilatation and CDFI blood flow signals.

[0090] Afterwards, the original data were integrated and graded according to the six examination indicators. The quantitative indicators of intestinal wall thickness, blood flow Limberg grade, omental thickening and hypoechoic nodules, and intestinal wall elasticity were divided into six levels; the non-quantitative indicators of intestinal local stenosis and dilatation, and CDFI blood flow signal were divided into two levels.

[0091] In step S2, each of the ultrasound diagnosis report texts is sequentially segmented, keyword information is extracted, and standardized to obtain a standardized inflammatory bowel disease feature matrix.

[0092] The step of performing word segmentation, keyword information extraction and standardization on each ultrasonic diagnostic report text in turn includes:

[0093] The ultrasound diagnostic report text is segmented using forward maximum matching and reverse maximum matching (bidirectional maximum matching algorithm) respectively, and the segmentation results of the two different methods are compared. If the segmentation results are consistent, one of them is selected as the final word segmentation result; if the segmentation results are inconsistent, they are classified as ambiguous segmentation results, and then the ambiguity is eliminated according to specific rules.

[0094] By using Chinese word segmentation technology and natural language processing tools, key information can be effectively extracted from the text, and the preset feature extraction framework can be used to identify and extract keyword information in the text, such as disease names and ultrasound indicator features. Finally, by converting the extracted features and entity relationships into a structured data format, the transformation from non-standard free text to structured data is completed.

[0095] For the quantitative indicators of intestinal wall thickness, blood flow Limberg grade, omental thickening and hypoechoic nodules, and intestinal wall elasticity, since the maximum and minimum values ​​in the original data are unknown, the conventional min-max standardization method is not applicable. Therefore, the following standardized processing formula is used:

[0096]

[0097] Among them, x is the original data value of any quantitative indicator, μ is the mean of the original data, and σ is the standard deviation of the original data. This formula standardizes the original data by subtracting the mean and dividing by the standard deviation to make them have the same dimension, which helps to improve the performance and stability of the model in subsequent data analysis and modeling.

[0098] Non-quantitative indicators such as local intestinal stenosis and dilatation and CDFI blood flow signals are divided into two levels: 0-1.

[0099] Afterwards, the standardized data are summarized to obtain a standardized inflammatory bowel disease feature matrix corresponding to each ultrasound diagnosis report text.

[0100] In step S3, all standardized inflammatory bowel disease feature matrices are divided into a training set and a validation set.

[0101] Exemplarily, in this step, a K-fold cross-validation method is used to divide the training set and the validation set.

[0102] The standardized inflammatory bowel disease feature matrix is ​​divided into K parts by the K-fold cross-validation method. Each time, K-1 parts are selected as training sets, and the remaining 1 part is used as a validation set. K experiments are performed and the average of the validation results is taken. This process is repeated K times, so that each subset is used as a validation set once, and the remaining K-1 subsets are used as training sets. Finally, the validation results of K experiments are averaged as the performance evaluation indicator of the model.

[0103] In step S4, the training set is used as Figure 2 The BP feed-forward neural network shown is inputted, and combined with the recovery cycle label, the network is iteratively optimized until the preset termination condition is met to build the inflammatory bowel disease diagnosis and prediction model.

[0104] BP feedforward neural network is a feedforward neural network that uses the error back propagation algorithm. It consists of an input layer, a hidden layer, and an output layer. This neural network is designed to handle a binary classification problem. The hidden layer has one or more layers, the neurons between layers are fully connected, and the neurons within the layer are independent of each other. The training steps of the BP neural network are as follows:

[0105] Step 10, Network initialization:

[0106] Assign a non-zero random number between (-1, 1) to the connection weight of each layer, and set the threshold a of each neuron in the hidden layer j and the threshold b of each neuron in the output layer k ;

[0107] Step 20, input signal forward propagation:

[0108] The standardized inflammatory bowel disease feature matrix is ​​set to X = (X 1 , X 2 , ..., X m ) is a variable, W ij is the weight between the i-th neuron in the input layer and the j-th neuron in the hidden layer, W jk is the weight between the jth neuron in the hidden layer and the kth neuron in the output layer; the input H of the kth neuron in the hidden layer is j The calculation formula is:

[0109]

[0110] Among them, X 1 , X 2 , ..., X m They represent the standardized values ​​of different inspection index data, m represents the total number of inspection indexes; a j is the threshold of the jth neuron in the hidden layer; q represents the total number of neurons in the hidden layer;

[0111] The output of the kth neuron in the hidden layer is I j The calculation formula is:

[0112]

[0113] Where f is the activation function;

[0114] Similarly, the output value Y of the kth neuron in the output layer is calculated k for:

[0115]

[0116] Among them, q represents the total number of neurons in the output layer;

[0117] Step 30, error signal back propagation:

[0118] The predicted output Y of the kth neuron in the output layer k and the corresponding rehabilitation cycle label T k The error formula is:

[0119] Err k =Y k (1-Y k )(T k -Y k ), k = 1, 2, ..., l

[0120] The error calculation formula of the jth neuron in the hidden layer is:

[0121]

[0122] Step 40, update weights and thresholds:

[0123] The weight calculation formula of the new j-th neuron in the hidden layer is:

[0124] W ij =W ij +ΔW ij =W ij +LErr j X i

[0125] The threshold update formula for the jth neuron in the hidden layer is:

[0126] a j =a j +Δa j =a j +LErr j

[0127] Where, ΔWij represents the weight change from the kth neuron in the input layer to the jth neuron in the hidden layer; Δa j represents the threshold change of the jth neuron in the hidden layer; L is the learning rate;

[0128] Similarly, update and calculate the weight and threshold of the kth neuron in the output layer;

[0129] Step 50, determine whether the expected error or maximum number of iterations is reached; if the expected error or maximum number of iterations is reached, end; otherwise return to Step 20.

[0130] In step S5, the validation set is used as the input of the inflammatory bowel disease diagnosis and prediction model to obtain validation results to evaluate the performance of the model.

[0131] Exemplarily, in this step, the accuracy, ROC curve and AUC value are used as evaluation indicators of model performance to provide a method for comprehensively measuring the model's ability to distinguish different categories (specifically short-term rehabilitation and long-term rehabilitation).

[0132] In order to help understand the advantages of the embodiments of the present invention, the following specific examples are provided for illustration:

[0133] (1) Example Background

[0134] Taking 500 pieces of inflammatory bowel disease data collected by a tertiary hospital as an example, the presence or absence of repeated rules is used as the criterion for judging long-term recovery and short-term recovery. The specific data sample is shown in Table 1.

[0135] Table 1 Example of disease description data

[0136]

[0137]

[0138] (2) Text information extraction and data standardization

[0139] The data were in the form of Excel tables. Based on six characteristic indicators (as shown in Table 2), such as intestinal wall thickness, local intestinal stenosis and dilatation, blood flow Limberg grade, omental thickening and hypoechoic nodules, intestinal wall elasticity, and CDFI blood flow signal, a two-way maximum matching algorithm was used to extract text information.

[0140] Table 2 Characteristic indicators of inflammatory bowel disease and their descriptions

[0141]

[0142] The comprehensive analysis of the above indicators can provide important basis for the diagnosis and evaluation of chronic intestinal diseases. These indicators reflect the typical ultrasound manifestations of inflammatory bowel disease, including structural changes, blood flow dynamics, and the response of surrounding tissues, which are of great significance for clinicians to formulate treatment plans and monitor disease progression.

[0143] Taking Table 2 as an example, after automatic text extraction and data standardization, the diagnostic vector generated is (1, 1, 1, 1, 0, 1).

[0144] (3) Data selection

[0145] In order to ensure more effective training and learning of the neural network model, in this case, the K-fold cross-validation method is used to divide the training set (450 data) and the validation set (50 data), and then the constructed feedforward neural network prediction model is used to conduct experiments to obtain prediction results. The number of training samples of four types of quantitative features on six levels and the number of training samples of two types of non-quantitative features on two levels are shown in Table 3, and the number of test samples is shown in Table 4.

[0146] The quantitative indicators of intestinal wall thickness, blood flow Limberg grade, omental thickening and hypoechoic nodules / intestinal wall elasticity are divided into six levels. The non-quantitative indicators of intestinal local stenosis and dilatation, CDFI blood flow signal are divided into two levels.

[0147] Table 3 Number of training samples for six types of features

[0148] Features\Classification 0 0.2 0.4 0.6 0.8 1 total Thickening of the intestinal wall 66 54 71 121 78 65 445 Limberg blood flow classification 19 31 36 154 53 28 321 Omental thickening and hypoechoic nodules 24 31 97 54 131 42 379 Intestinal wall elasticity 36 50 77 124 96 38 421 Local intestinal stenosis and dilatation 217 184 401 CDFI blood flow signal 259 128 387

[0149] Table 4 Number of validation samples for six types of features

[0150]

[0151]

[0152] As can be seen from the table, the dataset is somewhat unbalanced, so it places high demands on the feature extraction capability of the model.

[0153] (4) Model selection and training

[0154] Based on a structured data set of 500 historical ultrasound examinations of inflammatory bowel disease, the BP neural network was used to predict the patient's recovery period. The experiment was carried out under the CPU Intel Core 2.8 GHz, 8G memory, win10 operating system and MATLAB R2018a environment.

[0155] As shown in Table 6, when constructing the BP neural network model, a neuron is selected in the output layer. Since the output range of the sigmoid function is (0,1), it is very suitable for binary classification problems, so the sigmoid function is used as the activation function. The model output can be interpreted as the probability of belonging to a certain category.

[0156] When compiling the model, the Adam optimizer is selected. This optimizer combines the advantages of the Momentum and RMSprop optimizers, effectively minimizes the loss function through adaptive learning rate adjustment, and is suitable for processing a variety of different deep learning tasks. The loss function uses binary_crossentropy, which can effectively measure the difference between the predicted probability and the actual label when dealing with binary classification problems. At the same time, this patent uses accuracy as the evaluation criterion of the model to intuitively measure the model performance.

[0157] The challenges encountered when adjusting the learning rate and setting the expected error include: if the learning rate is set too high, it may cause instability in the network training process; if it is set too low, it may increase the time required for training and slow down the convergence. Similarly, if the expected error is set too high, it may reduce the effect of training; if it is set too low, it may be difficult to control the training time. Therefore, after multiple rounds of experimental adjustments and comprehensive consideration of key factors such as training accuracy and network stability, the expected error is set to 0.001, the learning rate is set to 0.01, and the maximum number of iterations is adjusted to 100 times to achieve the best training effect and network performance.

[0158] Table 6 BP neural network structure related parameters

[0159] Network Layers Number of neurons in the input layer Number of neurons in the hidden layer Number of neurons in the output layer 3 layers 6 10 1 Activation functions of each layer Expected Error Learning Rate Maximum number of iterations Sigmoid 0.001 0.01 100

[0160] After 100 training cycles, the model achieved an accuracy of 93.06% on the training set. At the same time, the training loss dropped to 0.2097, indicating that the difference between the model's prediction and the actual label is small, indicating that the model is able to fit the training data well.

[0161] In the validation phase, the model also demonstrated its generalization ability on data independent of the training set and the test set, with a validation accuracy of 88.89% and a validation loss of 0.2854. This result shows that the model can not only perform well on the training data, but also maintain high accuracy and low error levels on new and unknown data, reducing the risk of overfitting and ensuring the generalization performance of the model.

[0162] Finally, the performance on the test set further verified the effectiveness of the model, with a test accuracy of 86.67% and a test loss of 0.2808. This result highlights the model's good predictive ability for unseen data, indicating that the model has high practical value and accuracy in practical application scenarios, especially in the prediction of rehabilitation cycles.

[0163] In summary, it can be seen that the model performed stably throughout the experiment, maintaining high accuracy and low loss values ​​at all stages, proving its efficiency and reliability in predicting the recovery period. This provides a solid foundation for further research and model optimization, and also opens up new possibilities for its application in the field of medical rehabilitation.

[0164] Furthermore, the embodiments of the present invention use the accuracy rate, ROC curve and AUC value as important indicators for evaluating model performance, and provide a method for comprehensively measuring the model's ability to distinguish different categories. The ROC curve provides an intuitive view of model performance by plotting the false positive rate (FPR) and true positive rate (TPR) at multiple classification thresholds. Here, FPR is the rate at which the model mistakenly judges benign samples as malignant, while TPR (also known as sensitivity) reflects the model's ability to correctly identify malignant samples. Together, these two indicators reveal the model's balanced effect when dealing with positive and negative categories. The specific definitions of these concepts are as follows:

[0165] True positives (TP): The number of short-term rehabilitation samples predicted correctly. That is, the patient's actual rehabilitation result is short-term rehabilitation, and the prediction result of the BP / RBF neural network model is also short-term rehabilitation. False positives (FP): The number of long-term rehabilitation samples predicted incorrectly. That is, the patient's actual rehabilitation result is long-term rehabilitation, but the prediction result of the BP / RBF neural network model is short-term rehabilitation. True negatives (TN): The number of long-term rehabilitation samples predicted correctly. That is, the patient's actual rehabilitation result is long-term rehabilitation, and the prediction result of the BP / RBF neural network model is also long-term rehabilitation. False negatives (FN): The number of short-term rehabilitation samples predicted incorrectly. That is, the patient's actual rehabilitation result is short-term rehabilitation, but the prediction result of the BP / RBF neural network model is long-term rehabilitation. The above TP, FP, TN, and FN can form a matrix that can reflect the classification of the BP / RBF neural network model, that is, the confusion matrix. According to the confusion matrix, the number of samples with all rehabilitation results of short-term rehabilitation P = TP + FN, and the number of samples with all rehabilitation results of long-term rehabilitation N = FP + TN. In this way, the calculation formula for the prediction accuracy of the rehabilitation attribute prediction model based on BP / RBF neural network is:

[0166]

[0167] ROC curve and AUC value: In order to better judge the quality of the model, this paper introduces the concepts of false positive rate (FPR) and true positive rate (TPR) in combination with the confusion matrix. FPR represents the ratio of incorrect predictions in samples with long-term rehabilitation results, and TPR represents the ratio of correct predictions in samples with short-term rehabilitation results. The calculation formula is as follows:

[0168]

[0169]

[0170] The ROC curve is a curve drawn with FPR as the horizontal axis and TPR as the vertical axis, also called the sensitivity curve. The AUC value is the area under the ROC curve. Obviously, the horizontal axis FPR and the vertical axis TPR of the ROC curve are between 0 and 1, and the AUC value is also between 0 and 1. The larger the AUC value, the better the prediction effect of the model.

[0171] Combining the confusion matrix to calculate FPR and TPR enables the model performance to be evaluated from a more detailed perspective, especially in medical diagnosis scenarios where the prediction results have a greater impact on subsequent decisions. In this way, not only can the overall performance of the model be evaluated, but also the weaknesses of the model under specific conditions can be identified and improved, such as reducing the error of judging benign as malignant, thereby improving specificity while maintaining high sensitivity. This comprehensive and detailed evaluation method provides a solid foundation for the optimization and application of the model, ensuring that higher accuracy and reliability can be achieved in practical applications.

[0172] The confusion matrix of the inflammatory bowel disease prediction model based on BP neural network is as follows Figure 3 As shown, the ROC curve and AUC value are as follows Figure 4 shown.

[0173] Embodiment 2:

[0174] The embodiment of the present invention provides a training system for a data-driven inflammatory bowel disease diagnosis and prediction model, comprising:

[0175] An acquisition module, used to acquire historical inflammatory bowel disease data; wherein the historical inflammatory bowel disease data includes a number of ultrasound diagnosis report texts covering multiple examination indicators and recovery cycle labels;

[0176] A processing module, used for sequentially performing word segmentation, keyword information extraction and standardization processing on each of the ultrasonic diagnostic report texts to obtain a standardized inflammatory bowel disease feature matrix;

[0177] A partitioning module is used to divide all standardized inflammatory bowel disease feature matrices into a training set and a validation set;

[0178] A training module, used to use the training set as the input of the BP feedforward neural network, and in combination with the rehabilitation cycle label, iteratively optimize the network until a preset termination condition is met, so as to build the inflammatory bowel disease diagnosis and prediction model;

[0179] The verification module is used to use the verification set as the input of the inflammatory bowel disease diagnosis and prediction model to obtain verification results to evaluate the performance of the model.

[0180] Embodiment 3:

[0181] An embodiment of the present invention provides a storage medium storing a computer program for training a data-driven inflammatory bowel disease diagnostic prediction model, wherein the computer program enables a computer to execute the training method described in Example 1.

[0182] Embodiment 4:

[0183] An embodiment of the present invention provides an electronic device, including:

[0184] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the programs include instructions for executing the training method as described in Example 1.

[0185] It is understandable that the training system, storage medium and electronic device based on the data-driven inflammatory bowel disease diagnostic prediction model provided in the embodiments of the present invention correspond to the training method based on the data-driven inflammatory bowel disease diagnostic prediction model provided in the embodiments of the present invention, and the explanations, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the training method and will not be repeated here.

[0186] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0187] 1. By deeply analyzing the detailed information in the doctor's diagnosis report, the embodiment of the present invention can accurately extract disease characteristics and clinical indicators, and combine with the machine learning model to significantly improve the accuracy of the prediction of the recovery period of chronic intestinal diseases.

[0188] 2. The embodiments of the present invention can provide each patient with a personalized recovery period prediction, helping doctors to develop more accurate treatment and rehabilitation plans for patients, thereby improving treatment efficiency and patient satisfaction.

[0189] 3. By regularly analyzing and updating diagnostic reports, the embodiments of the present invention can dynamically monitor disease progression and treatment response, adjust rehabilitation plans in a timely manner, and achieve real-time and dynamic disease management.

[0190] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0191] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data-driven training method for a diagnostic prediction model for inflammatory bowel disease, characterized in that: include: Acquire historical inflammatory bowel disease data; wherein the historical inflammatory bowel disease data includes a number of ultrasound diagnosis report texts covering multiple examination indicators and recovery cycle labels; Each of the ultrasound diagnosis report texts is sequentially segmented, keyword information is extracted, and standardized to obtain a standardized inflammatory bowel disease feature matrix; All standardized inflammatory bowel disease feature matrices were divided into training sets and validation sets; The training set is used as the input of the BP feedforward neural network, and combined with the recovery cycle label, the network is iteratively optimized until the preset termination condition is met, so as to build the inflammatory bowel disease diagnosis and prediction model; The validation set is used as input of the inflammatory bowel disease diagnosis and prediction model to obtain validation results to evaluate the performance of the model.

2. The training method according to claim 1, characterized in that: The inspection indicators include quantitative indicators and non-quantitative indicators; among which: The quantitative indicators include intestinal wall thickness, blood flow Limberg grade, omental thickening and hypoechoic nodules, and intestinal wall elasticity value; The non-quantitative indicators include local intestinal stenosis and dilatation and CDFI blood flow signals.

3. The training method according to claim 2, characterized in that: The step of sequentially performing word segmentation, keyword information extraction and standardization processing on each ultrasonic diagnostic report text comprises: The ultrasound diagnosis report text is segmented by using forward maximum matching and reverse maximum matching respectively, and the segmentation results of the two different methods are compared. If the segmentation results are consistent, one of them is selected as the final word segmentation result; if the segmentation results are inconsistent, it is classified as an ambiguous segmentation result, and then the ambiguity is eliminated according to specific rules; Identify and extract keyword information from the word segmentation results through a preset feature extraction framework, and convert the extracted features and entity relationships into a structured data format; For the quantitative indicators of intestinal wall thickness, blood flow Limberg grade, omental thickening and hypoechoic nodules, and intestinal wall elasticity, the following standardized processing formula was used: Among them, x is the original data value of any quantitative indicator, μ is the mean of the original data, and σ is the standard deviation of the original data; Non-quantitative indicators such as local intestinal stenosis and dilatation and CDFI blood flow signals are divided into two levels: 0-1.

4. The training method according to claim 1, characterized in that: The training set and the validation set are divided into the training set and the validation set using the K-fold cross validation method.

5. The training method according to claim 1, characterized in that: The training process of the BP feedforward neural network includes: Step 10, Network initialization: Assign a non-zero random number between (-1, 1) to the connection weight of each layer, and set the threshold a of each neuron in the hidden layer j and the threshold b of each neuron in the output layer k ; Step 20, input signal forward propagation: The standardized inflammatory bowel disease feature matrix is ​​set to X = (X1, X2, ..., X m ) is a variable, W ij is the weight between the i-th neuron in the input layer and the j-th neuron in the hidden layer, W jk is the weight between the jth neuron in the hidden layer and the kth neuron in the output layer; the input H of the kth neuron in the hidden layer is j The calculation formula is: Among them, X1, X2, ..., X m They represent the standardized values ​​of different inspection index data, m represents the total number of inspection indexes; a j is the threshold of the jth neuron in the hidden layer; q represents the total number of neurons in the hidden layer; The output of the kth neuron in the hidden layer is I j The calculation formula is: Where f is the activation function; Similarly, the output value Y of the kth neuron in the output layer is calculated k for: Among them, q represents the total number of neurons in the output layer; Step 30, error signal back propagation: The predicted output Y of the kth neuron in the output layer k and the corresponding rehabilitation cycle label T k The error formula is: Err k =And k (1-Y k )(T k -AND k ),k,1,2,...,l The error calculation formula of the jth neuron in the hidden layer is: Step 40, update weights and thresholds: The weight calculation formula of the new j-th neuron in the hidden layer is: IN ij =In ij +ΔW ij =In ij +LErr j X i The threshold update formula for the jth neuron in the hidden layer is: in j = yes j +Δa j = yes j +LErr j Where, ΔW ij represents the weight change from the kth neuron in the input layer to the jth neuron in the hidden layer; Δa j represents the threshold change of the jth neuron in the hidden layer; L is the learning rate; Similarly, update and calculate the weight and threshold of the kth neuron in the output layer; Step 50, determine whether the expected error or maximum number of iterations is reached; if the expected error or maximum number of iterations is reached, end; otherwise return to Step 20.

6. The training method according to claim 1, characterized in that: The accuracy, ROC curve and AUC value are used as evaluation indicators of model performance.

7. A training system for a data-driven inflammatory bowel disease diagnostic prediction model, characterized in that: include: An acquisition module, used to acquire historical inflammatory bowel disease data; wherein the historical inflammatory bowel disease data includes a number of ultrasound diagnosis report texts covering multiple examination indicators and recovery cycle labels; A processing module, used for sequentially performing word segmentation, keyword information extraction and standardization processing on each of the ultrasonic diagnostic report texts to obtain a standardized inflammatory bowel disease feature matrix; A partitioning module is used to divide all standardized inflammatory bowel disease feature matrices into a training set and a validation set; A training module, used to use the training set as the input of the BP feedforward neural network, and in combination with the rehabilitation cycle label, iteratively optimize the network until a preset termination condition is met, so as to build the inflammatory bowel disease diagnosis and prediction model; The verification module is used to use the verification set as the input of the inflammatory bowel disease diagnosis and prediction model to obtain verification results to evaluate the performance of the model.

8. A storage medium, characterized in that: The computer program for training a data-driven inflammatory bowel disease diagnosis prediction model is stored therein, wherein the computer program enables a computer to execute the training method according to any one of claims 1 to 6.

9. An electronic device, characterized in that: include: one or more processors; Memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs comprising instructions for executing the training method according to any one of claims 1 to 6.

Images