An AI-assisted diagnosis and treatment method for renal hypoplasia

Abstract

The application relates to the field of AI-assisted diagnosis and treatment, and discloses an AI-assisted diagnosis and treatment method for renal dysplasia. First, the unstructured case information of a renal dysplasia patient is converted into structured medical record information, and a disease recognition model for renal dysplasia is established. Then, patient recognition is performed to obtain a congenital renal dysplasia sample, and a disease feature extraction model is established to quantize and extract structured features from the patient medical record text, thereby obtaining key disease feature indexes of the patient. A fusion feature vector corresponding to each patient is constructed, and the necessity of disease surgical intervention is judged. The application solves the problems in the prior art, such as the large number of disease types, the complex pathology, the lack of specificity in clinical manifestations, the misdiagnosis, the missed diagnosis or the delayed treatment, the poor data fusion capability, the poor model generalization capability and the poor clinical interpretability of the existing AI system, and the inability to support more accurate and intelligent diagnosis and treatment decisions.

Description

Technical Field

[0001] This invention relates to the field of AI-assisted diagnosis and treatment, and in particular to an AI-assisted diagnosis and treatment method for renal dysplasia. Background Technology

[0002] In recent years, the application of artificial intelligence in the medical field has been continuously expanding, especially in disease prediction, risk assessment, and assisted diagnosis, demonstrating great potential. However, for renal dysplasia, especially urinary tract obstruction, existing AI-assisted diagnostic systems still face many challenges. First, these diseases are diverse, with complex pathologies, and clinical manifestations often lack specificity, easily leading to misdiagnosis, missed diagnosis, or delayed treatment. Second, obtaining high-quality medical data on renal dysplasia is difficult, and labeled data is scarce, resulting in insufficient training data for existing AI models, making it difficult to support the generalization and stable application of models in diverse cases.

[0003] Furthermore, current AI models often focus on modeling single-modal data, making it difficult to fully integrate multi-source information from clinical, imaging, and laboratory sources, thus limiting their comprehensive characterization of disease progression pathways and treatment outcomes. In the diagnosis and treatment of complex structural birth defects, existing systems generally lack the ability to dynamically predict natural outcomes, continuously evaluate the effects of surgical interventions, and provide individualized postoperative management. Therefore, there is an urgent need to make breakthroughs in improving the data fusion capabilities, model generalization abilities, and clinical interpretability of AI systems to support more accurate and intelligent diagnostic and treatment decisions. Summary of the Invention

[0004] This invention aims to provide an AI-assisted diagnosis and treatment method for renal dysplasia, which solves the problem that existing AI systems have poor data fusion capabilities, model generalization capabilities, and clinical interpretability, which are unable to support more accurate and intelligent diagnosis and treatment decisions due to the large number of types of this disease, complex pathology, and lack of specific clinical manifestations.

[0005] To achieve the above objectives, the present invention provides the following method:

[0006] This invention provides an AI-assisted diagnostic and treatment method for renal dysplasia:

[0007] S1: Convert unstructured case information of patients with renal dysplasia into structured medical record information and establish a disease identification model for renal dysplasia;

[0008] S2: Patient identification is performed using the renal dysplasia disease identification model to obtain congenital renal dysplasia samples. Based on the congenital renal dysplasia samples, a disease feature extraction model is established to perform feature quantification and structured feature extraction on the patient's medical record text to obtain the patient's key disease feature indicators.

[0009] S3: Based on the extraction of key disease characteristic indicators of the patients, construct a fusion feature vector corresponding to each patient to determine the necessity of surgical intervention for the disease.

[0010] Preferably, the steps of identifying patients using the renal dysplasia disease identification model to obtain congenital renal dysplasia samples, establishing a disease feature extraction model based on the congenital renal dysplasia samples, and performing feature quantification and structured feature extraction on the patient's medical record text to obtain the patient's key disease feature indicators include: inputting structured text fields from a MySQL data table containing text fields related to surgical judgment; defining a dictionary `data_sources` to map table names and their key fields and disease keywords in the MySQL data table, filtering out all tables containing the target fields, sequentially reading the text information in the corresponding fields, and for each record, concatenating non-empty fields into a complete text sentence to obtain an analysis sample; preprocessing the analysis sample to obtain a preprocessed text result; calculating the TF value and IDF value based on the preprocessed text result, calculating the TF-IDF value using the TF value and IDF value, and outputting the TF-IDF feature matrix; filtering the TF-IDF feature matrix based on a set of medical keywords for the preprocessed text result, performing feature quantification and structured feature extraction on the patient's medical record text to obtain the patient's key disease feature indicators.

[0011] Preferably, the step of preprocessing the analysis sample to obtain the preprocessed text result includes:

[0012] Punctuation marks and non-Chinese / English characters were removed from the analyzed samples:

[0013] ;

[0014] Where CleanedText represents the cleaned text, Regex is the regular expression, and Filter is the original text.

[0015] Jieba word segmentation and stop word removal

[0016] ;

[0017] Among them, TokenizedText represents the text result after word segmentation, that is, the set after splitting the original text into independent words, wᵢ represents the i-th word obtained after word segmentation, jiebacut refers to "Jieba word segmentation", CleanedText refers to the clean text after preprocessing above, which is the input for word segmentation operation, and StopWords represents stop words, that is, common words without actual semantics that need to be filtered.

[0018] Preferably, the steps of calculating the TF value and IDF value based on the preprocessed text result, calculating the TF-IDF value using the TF value and IDF value, and outputting the TF-IDF feature matrix include:

[0019] Based on the preprocessed text results, calculate the TF and IDF values:

[0020] Term Frequency (TF): The frequency of a word in a text, calculated using the following formula:

[0021] ;

[0022] in, The total number of words is the number of times word w appears in text d, and total_words(d) is the total number of words in text d.

[0023] Inverse Document Frequency (IDF): Measures the general importance of a word; the formula is as follows:

[0024] ;

[0025] Where N is the total number of texts in the corpus, The number of texts containing the word w;

[0026] The formula for calculating TF-IDF value is:

[0027] TF-IDF(w,d)=TF(w,d)×IDF(w);

[0028] The output is the TF-IDF feature matrix:

[0029] ;

[0030] Where n is the total number of texts and m is the size of the vocabulary.

[0031] Preferably, the step of filtering the TF-IDF feature matrix based on the medical keyword set for the preprocessed text results includes: extracting the TF-IDF values ​​corresponding to the keywords defined in the MySQL data table to form a preliminary feature vector; further filtering by retaining features with variance greater than a threshold using variance selection, and finally outputting the standardized feature vector for each text segment.

[0032] vd=[TF-IDF(w1,d),TF-IDF(w2,d),...,TF-IDF(wk,d)];

[0033] Where k is the number of keywords after filtering, the output vd vector can be used as a structured feature variable for subsequent classification model training, patient cluster analysis or feature visualization, and can be bound to the original database records to realize the structured writing back of information.

[0034] Preferably, the steps for constructing a fusion feature vector for each patient based on the extracted key disease characteristic indicators, and for determining the necessity of surgical intervention, include: inputting multi-type clinical characteristic information; constructing a BERT model structure, adding a classification layer after the [CLS] vector output by the BERT model structure to achieve binary prediction of natural outcome trend / necessity of surgical intervention; evaluating the BERT model structure through accuracy, confusion matrix, classification report, and area under the receiver operating characteristic curve; and introducing the SHAP value analysis method to rank the key features affecting natural outcome and surgical judgment, identify the main influencing factors, and improve the interpretability of the prediction model.

[0035] Preferably, the step of inputting multiple types of clinical feature information includes: the multiple types of clinical feature information include: symptom-related multi-hot coding features, structural examination results, laboratory test indicators, and medical history symptom intensity indicators; among all types of clinical feature information, textual features are converted into word embeddings through BERT's tokenizer, and combined with positional embeddings and segment embeddings to form an input sequence; numerical features are processed through standardization ( The text is then concatenated with the embedded text and used as the final input to the BERT model structure.

[0036] Preferably, the step of constructing the BERT model structure and adding a classification layer after the [CLS] vector output by the BERT model structure to achieve binary classification prediction of natural outcome trend / necessity of surgical intervention includes:

[0037] Model output: For sample i, the classification layer output is:

[0038] yi = softmax(W⋅CLSi+b);

[0039] Where CLSi is the BERT output vector at the [CLS] position of sample i, W is the classification layer weight matrix, and b is the bias term.

[0040] Preferably, the step of training the loss function of the BERT model structure includes:

[0041] Loss function: Cross-entropy loss is used, and the formula is as follows:

[0042] ;

[0043] Where N is the number of samples and C is the number of categories (C=2 here). For the true label of sample i, Predict probabilities for the model; training parameters: learning rate 5e-5, batch size=16, epoch=10, using the Adam optimizer.

[0044] Preferably, the steps of introducing the SHAP value analysis method to rank the key features affecting natural outcomes and surgical decisions by importance, identify the main influencing factors, and improve the interpretability of the prediction model include:

[0045] The SHAP method is used to calculate the impact of each input feature on the model output;

[0046] For sample x, the SHAP value of the j-th feature is:

[0047] ;

[0048] Where N: the entire set of features; S is a subset of features, excluding j; This is the model output when only features from a subset of SSS are used;

[0049] The SHAP value satisfies the principles of additivity and fairness:

[0050] ;

[0051] The top features influencing the prediction model's predictions are shown using SHAP bar charts, including ultrasound findings, laboratory indicators, and TF-IDF high-weight keywords.

[0052] The beneficial effects of this invention are reflected in the following aspects: This invention establishes a precise identification model for renal dysplasia based on structured case information of patients with renal dysplasia. Methodologically, it integrates the diagnostic experience and medical knowledge of clinical experts, constructing a disease screening process that combines rule-driven and data-driven approaches. Specifically, it employs regularized rule extraction and knowledge graph mapping to extract key clinical features, and automatically identifies suspected cases of renal dysplasia through algorithms, thereby improving the accuracy and efficiency of screening. A BERT pre-trained model is used for fine-tuning, accurately determining the natural course of the disease and the necessity of surgical intervention. To enhance the interpretability of the model, the SHAP value analysis method is introduced to rank the importance of key features affecting the natural course and surgical decisions, clarifying the main influencing factors, thus providing quantifiable and interpretable evidence for clinical decision-making. Attached Figure Description

[0053] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0054] Figure 1 This is a flowchart illustrating an AI-assisted diagnosis and treatment method for renal dysplasia provided in an embodiment of the present invention. Detailed Implementation

[0055] To enable those skilled in the art to better understand the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0057] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0058] In recent years, the application of artificial intelligence in the medical field has been continuously expanding, especially in disease prediction, risk assessment, and assisted diagnosis, demonstrating great potential. However, for renal dysplasia, especially urinary tract obstruction, existing AI-assisted diagnostic systems still face many challenges. First, these diseases are diverse, with complex pathologies, and clinical manifestations often lack specificity, easily leading to misdiagnosis, missed diagnosis, or delayed treatment. Second, obtaining high-quality medical data on renal dysplasia is difficult, and labeled data is scarce, resulting in insufficient training data for existing AI models, making it difficult to support the generalization and stable application of models in diverse cases.

[0059] Furthermore, current AI models often focus on modeling single-modal data, making it difficult to fully integrate multi-source information from clinical, imaging, and laboratory sources, thus limiting their comprehensive characterization of disease progression pathways and treatment outcomes. In the diagnosis and treatment of complex structural birth defects, existing systems generally lack the ability to dynamically predict natural outcomes, continuously evaluate the effects of surgical interventions, and provide individualized postoperative management. Therefore, there is an urgent need to make breakthroughs in improving the data fusion capabilities, model generalization abilities, and clinical interpretability of AI systems to support more accurate and intelligent diagnostic and treatment decisions.

[0060] This invention aims to provide an AI-assisted diagnosis and treatment method for renal dysplasia, which solves the problem that existing AI systems have poor data fusion capabilities, model generalization capabilities, and clinical interpretability, which are unable to support more accurate and intelligent diagnosis and treatment decisions due to the large number of types of this disease, complex pathology, and lack of specific clinical manifestations.

[0061] like Figure 1 As shown in the figure, a specific embodiment of the present invention provides an AI-assisted diagnosis and treatment method for renal dysplasia, comprising the following steps:

[0062] S1: Convert unstructured case information of patients with renal dysplasia into structured medical record information to establish a disease identification model for renal dysplasia.

[0063] In this embodiment of the invention, a precise identification model for renal dysplasia is established based on structured case information of patients. Methodologically, the diagnostic experience and medical knowledge of clinical experts are integrated to construct a disease screening process that combines rule-driven and data-driven approaches. Specifically, regularized rule extraction and knowledge graph mapping are used to extract key clinical features, and algorithms are used to automatically identify suspected cases of renal dysplasia, thereby improving the accuracy and efficiency of screening.

[0064] S2: Patients are identified using a disease identification model for renal dysplasia to obtain samples of congenital renal dysplasia. Based on these samples, a disease feature extraction model is established. Feature quantification and structured feature extraction are performed on the patient's medical record text to obtain the patient's key disease feature indicators.

[0065] In this embodiment of the invention, a structured text field from a MySQL data table containing text fields related to surgical judgment is input; a dictionary `data_sources` is defined to map table names in the MySQL data table to their key fields and disease keywords, filtering out all tables containing the target fields, and sequentially reading the text information in the corresponding fields. For each record, non-empty fields are concatenated into a complete text sentence to obtain an analysis sample; the analysis sample is preprocessed to obtain the preprocessed text result; punctuation and non-Chinese / English characters are removed from the analysis sample.

[0066] ;

[0067] Where CleanedText represents the cleaned text, Regex is the regular expression, and Filter is the original text.

[0068] Jieba word segmentation and stop word removal

[0069] ;

[0070] Wherein, TokenizedText represents the text result after tokenization, that is, the set of independent words after splitting the original text; wᵢ represents the i-th word obtained after tokenization; jiebacut refers to "Jieba Tokenization"; CleanedText refers to the clean text after preprocessing, which is the input for the tokenization operation; StopWords represents stop words, that is, common words without actual semantics that need to be filtered; based on the preprocessed text result, TF and IDF values ​​are calculated.

[0071] Term Frequency (TF): The frequency of a word in a text, calculated using the following formula:

[0072] ;

[0073] in, The total number of words is the number of times word w appears in text d, and total_words(d) is the total number of words in text d.

[0074] Inverse Document Frequency (IDF): Measures the general importance of a word; the formula is as follows:

[0075] ;

[0076] Where N is the total number of texts in the corpus, The number of texts containing word w; calculate the TF-IDF value using the TF and IDF values, and output the TF-IDF feature matrix; the formula for calculating the TF-IDF value is:

[0077] TF-IDF(w,d)=TF(w,d)×IDF(w);

[0078] The output is the TF-IDF feature matrix:

[0079] ;

[0080] Where n is the total number of texts and m is the vocabulary size; for the preprocessed text results, the TF-IDF feature matrix is ​​filtered based on the medical keyword set, and feature quantization and structured feature extraction are performed on the patient's medical record text to obtain the patient's key disease feature indicators; the TF-IDF values ​​corresponding to the keywords defined in the MySQL data table are extracted to form a preliminary feature vector; through the variance selection method, features with variance greater than a threshold are retained for further filtering, and finally the standardized feature vector of each text segment is output:

[0081] vd=[TF-IDF(w1,d),TF-IDF(w2,d),...,TF-IDF(wk,d)];

[0082] Where k is the number of keywords after filtering.

[0083] The output vd vector can be used as a structured feature variable for subsequent classification model training, patient cluster analysis, or feature visualization, and can be bound to the original database records to achieve structured information write-back.

[0084] S3: Based on the extraction of key disease characteristic indicators of the patients, construct a fusion feature vector corresponding to each patient to determine the necessity of surgical intervention for the disease.

[0085] In this embodiment of the invention, multiple types of clinical feature information are input; these include: symptom-related multi-hot encoded features, structural examination results, laboratory test indicators, and medical history symptom intensity indicators; among all types of clinical feature information, textual features are converted into word embeddings using BERT's tokenizer, and combined with positional embeddings and segment embeddings to form an input sequence; numerical features are processed through standardization (…). The text is then concatenated with the embedded text and used as the final input to the BERT model structure. The BERT model structure is constructed by adding a classification layer after the [CLS] vector output by the BERT model structure to achieve binary classification prediction of natural outcome trend / necessity of surgical intervention. Model output: For sample i, the classification layer output is:

[0086] yi = softmax(W⋅CLSi+b);

[0087] Where CLSi is the BERT output vector at the [CLS] position of sample i, W is the classification layer weight matrix, and b is the bias term; the loss function is the cross-entropy loss, and the formula is:

[0088] ;

[0089] Where N is the number of samples and C is the number of categories (C=2 here). For the true label of sample i, The model predicts probabilities; training parameters: learning rate 5e-5, batch size=16, epoch=10, using the Adam optimizer; the BERT model structure is evaluated using accuracy, confusion matrix, classification report, and area under the receiver operating characteristic curve; the SHAP value analysis method is introduced to rank the importance of key features affecting natural outcomes and surgical decisions, identify the main influencing factors, and improve the interpretability of the prediction model; the SHAP method is used to calculate the impact of each input feature on the model output;

[0090] For sample x, the SHAP value of the j-th feature is:

[0091] ;

[0092] Where N: the entire set of features; S is a subset of features, excluding j; This is the model output when only features from a subset of SSS are used;

[0093] The SHAP value satisfies the principles of additivity and fairness:

[0094] ;

[0095] The SHAP bar chart illustrates the top features influencing the prediction of the prediction model, including ultrasound findings, laboratory indicators, and TF-IDF high-weight keywords.

[0096] The beneficial effects of this invention are reflected in the following aspects: This invention establishes a precise identification model for renal dysplasia based on structured case information of patients with renal dysplasia. Methodologically, it integrates the diagnostic experience and medical knowledge of clinical experts, constructing a disease screening process that combines rule-driven and data-driven approaches. Specifically, it employs regularized rule extraction and knowledge graph mapping to extract key clinical features, and automatically identifies suspected cases of renal dysplasia through algorithms, thereby improving the accuracy and efficiency of screening. A BERT pre-trained model is used for fine-tuning, accurately determining the natural course of the disease and the necessity of surgical intervention. To enhance the interpretability of the model, the SHAP value analysis method is introduced to rank the importance of key features affecting the natural course and surgical decisions, clarifying the main influencing factors, thus providing quantifiable and interpretable evidence for clinical decision-making.

[0097] The above descriptions are merely embodiments of the present invention. Commonly known technical solutions or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. An AI-assisted diagnostic and treatment method for renal dysplasia, characterized in that, The method includes: S1: Convert unstructured case information of patients with renal dysplasia into structured medical record information and establish a disease identification model for renal dysplasia; S2: Patient identification is performed using the renal dysplasia disease identification model to obtain congenital renal dysplasia samples. Based on the congenital renal dysplasia samples, a disease feature extraction model is established to perform feature quantification and structured feature extraction on the patient's medical record text to obtain the patient's key disease feature indicators. S3: Based on the extraction of key disease characteristic indicators of the patients, construct a fusion feature vector corresponding to each patient to determine the necessity of surgical intervention for the disease.

2. The AI-assisted diagnosis and treatment method for renal dysplasia according to claim 1, characterized in that, The steps of identifying patients using the renal dysplasia disease identification model to obtain congenital renal dysplasia samples, establishing a disease feature extraction model based on the congenital renal dysplasia samples, and performing feature quantification and structured feature extraction on the patient's medical record text to obtain the patient's key disease feature indicators include: Input a structured text field from a MySQL table containing text fields related to surgical decisions; Define a dictionary `data_sources` to map the table names of the MySQL data tables to their key fields and disease keywords. Filter out all tables containing the target fields, and read the text information in the corresponding fields in sequence. For each record, concatenate the non-empty fields into a complete text sentence to obtain the analysis sample. The analyzed sample is preprocessed to obtain the preprocessed text result; Based on the preprocessed text results, the TF value and IDF value are calculated, the TF-IDF value is calculated using the TF value and IDF value, and the TF-IDF feature matrix is ​​output. Based on the preprocessed text results, the TF-IDF feature matrix is ​​filtered according to the medical keyword set, and the patient's medical record text is subjected to feature quantification and structured feature extraction to obtain the patient's key disease feature indicators.

3. The AI-assisted diagnosis and treatment method for renal dysplasia according to claim 2, characterized in that, The step of preprocessing the analysis sample to obtain the preprocessed text result includes: Punctuation marks and non-Chinese / English characters were removed from the analyzed samples: ； Where CleanedText represents the cleaned text, Regex is the regular expression, and Filter is the original text. Jieba word segmentation and stop word removal ； Among them, TokenizedText represents the text result after word segmentation, that is, the set after splitting the original text into independent words, wᵢ represents the i-th word obtained after word segmentation, jiebacut refers to "Jieba word segmentation", CleanedText refers to the clean text after preprocessing above, which is the input for word segmentation operation, and StopWords represents stop words, that is, common words without actual semantics that need to be filtered.

4. The AI-assisted diagnosis and treatment method for renal dysplasia according to claim 2, characterized in that, The steps of calculating the TF and IDF values ​​based on the preprocessed text results, calculating the TF-IDF value using the TF and IDF values, and outputting the TF-IDF feature matrix include: Based on the preprocessed text results, calculate the TF and IDF values: Term Frequency (TF): The frequency of a word in a text, calculated using the following formula: ； in, The total number of words is the number of times word w appears in text d, and total_words(d) is the total number of words in text d. Inverse Document Frequency (IDF): Measures the general importance of a word; the formula is as follows: ； Where N is the total number of texts in the corpus, The number of texts containing the word w; The formula for calculating TF-IDF value is: TF-IDF(w,d)=TF(w,d)×IDF(w); The output is the TF-IDF feature matrix: ； Where n is the total number of texts and m is the size of the vocabulary.

5. The AI-assisted diagnosis and treatment method for renal dysplasia according to claim 2, characterized in that, The step of filtering the TF-IDF feature matrix based on the medical keyword set for the preprocessed text results includes: Extract the TF-IDF values ​​corresponding to the keywords defined in the MySQL data table to form a preliminary feature vector; By using variance selection, features with variance greater than a threshold are retained for further filtering, ultimately outputting a standardized feature vector for each text segment: vd=[TF-IDF(w1,d),TF-IDF(w2,d),...,TF-IDF(wk,d)]; Where k is the number of keywords after filtering. The output vd vector can be used as a structured feature variable for subsequent classification model training, patient cluster analysis, or feature visualization, and can be bound to the original database records to achieve structured information write-back.

6. The AI-assisted diagnosis and treatment method for renal dysplasia according to claim 1, characterized in that, The step of constructing a fusion feature vector for each patient based on the extracted key disease feature indicators, and determining the necessity of surgical intervention for the disease, includes: Input multiple types of clinical feature information; Construct a BERT model structure, and add a classification layer after the [CLS] vector output by the BERT model structure to achieve binary classification prediction of natural outcome trend / necessity of surgical intervention; The BERT model architecture was evaluated using accuracy, confusion matrix, classification report, and area under the receiver operating characteristic curve. The SHAP value analysis method is introduced to rank the key features affecting natural outcomes and surgical decisions by importance, identify the main influencing factors, and improve the interpretability of the prediction model.

7. The AI-assisted diagnosis and treatment method for renal dysplasia according to claim 6, characterized in that, The step of inputting multiple types of clinical feature information includes: The multi-type clinical feature information includes: symptom-related multi-thermal coding features, structural examination results, laboratory test indicators, and medical history symptom intensity indicators; Among all types of clinical feature information, text features are converted into word embeddings by BERT's tokenizer, and then combined with position embeddings and segment embeddings to form the input sequence. Numerical features are processed through standardization ( The text is then concatenated with the embedded text and used as the final input to the BERT model structure.

8. The AI-assisted diagnosis and treatment method for renal dysplasia according to claim 6, characterized in that, The steps of constructing the BERT model structure and adding a classification layer after the [CLS] vector output by the BERT model structure to achieve binary classification prediction of natural outcome trend / necessity of surgical intervention include: Model output: For sample i, the classification layer output is: yi = softmax(W⋅CLSi+b); Where CLSi is the BERT output vector at the [CLS] position of sample i, W is the classification layer weight matrix, and b is the bias term.

9. The AI-assisted diagnosis and treatment method for renal dysplasia according to claim 8, characterized in that, The steps for training the loss function of the BERT model structure include: Loss function: Cross-entropy loss is used, and the formula is as follows: ； Where N is the number of samples and C is the number of categories (C=2 here). For the true label of sample i, Predict probabilities for the model; Training parameters: learning rate 5e-5, batch size=16, epoch=10, using Adam optimizer.

10. The AI-assisted diagnosis and treatment method for renal dysplasia according to claim 9, characterized in that, The steps of introducing the SHAP value analysis method to rank the key features affecting natural outcomes and surgical decisions, identify the main influencing factors, and improve the interpretability of the prediction model include: The SHAP method is used to calculate the impact of each input feature on the model output; For sample x, the SHAP value of the j-th feature is: ； Where N: the entire set of features; S is a subset of features, excluding j; This is the model output when only features from a subset of SSS are used; The SHAP value satisfies the principles of additivity and fairness: ； The top features influencing the prediction model's predictions are shown using SHAP bar charts, including ultrasound findings, laboratory indicators, and TF-IDF high-weight keywords.

Images