Calculus risk prediction method based on machine learning and electronic equipment

Abstract

The invention belongs to the technical field of calculus risk assessment, and provides a calculus risk prediction method based on machine learning and electronic equipment, and the method comprises the steps: to-be-detected original data collection and calculus risk prediction; the construction process of the stone risk prediction model comprises the steps of unmarked data collection, image feature extraction, index feature extraction, text semantic extraction, feature fusion, feature clustering analysis, iterative training, fine adjustment small sample data collection, meta-learning fine adjustment and cluster marking. According to the method, through a semi-supervised learning strategy of unmarked data clustering and small sample adjustment, the demand of marked data is greatly reduced; according to the method, clustering and unsupervised iteration are fused through DBSCAN and DPC double algorithms, so that pre-training under unmarked data is realized, and the reliability of unmarked training is improved; through combination of multi-modal data fusion, clustering iteration and meta-learning fine tuning, the robustness of the model is improved.

Description

Technical Field

[0001] This invention relates to the field of kidney stone risk assessment technology, and in particular to a kidney stone risk prediction method and electronic device based on machine learning. Background Technology

[0002] Existing stone prediction technologies are mainly divided into three categories: metabolomics-based prediction methods, image analysis-based prediction methods, and clinical indicator-based prediction methods. Metabolomics-based prediction method CN117238491B uses three techniques, including liquid chromatography-mass spectrometry, to collect urinary metabolite data. After preprocessing such as filtering, denoising, and normalization, features are extracted using a Markov model, combined with support vector machines and neural networks to achieve prediction. This method can reflect the biochemical mechanisms at the molecular level, but data acquisition costs are high and processing is complex. Image analysis-based prediction method CN120877920B integrates urinary metabolic data with microscopic images of urinary sediment crystals. Metabolic features are extracted using PCA+PLS-DA, and an improved Inception-ResNet-v2 network is used to identify crystal types. Prediction is then performed using XGBoost and BiLSTM models, providing intuitive morphological information, but it requires sophisticated equipment and relies on professional image interpretation. The prediction method based on clinical indicators (CN112053784B) focuses on calcium oxalate kidney stones. It combines three clinical indicators with four relative abundance values ​​of gut microbiota and uses a random forest algorithm for modeling. The AUC value reaches 0.91. The data acquisition is simple, but the applicability is limited.

[0003] While existing methods can detect kidney stones, several issues remain: 1) Strong dependence on labeled data: Most existing methods rely on supervised learning, requiring a large amount of high-quality labeled data. 2) Insufficient standardization in multimodal fusion: Heterogeneous data vary greatly in feature space and distribution, lacking a unified fusion standard, resulting in high computational resource requirements and poor interpretability. 3) Lack of algorithm robustness: Prone to overfitting when faced with data noise, outliers, and class imbalance, exhibiting weak stability and generalization ability. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the purpose of this invention is to provide a method and electronic device for predicting the risk of kidney stones based on machine learning, which solves the problems of strong dependence on labeled data, insufficient standardization of multimodal fusion, and poor robustness in the prior art.

[0005] To achieve the above objectives, the present invention provides the following solution: A machine learning-based method for predicting kidney stone risk includes: Collect imaging data and indicator data of the patients to be tested to obtain the raw data to be tested; The raw data to be detected is input into a pre-trained stone risk prediction model for identification to obtain stone risk prediction results; the stone risk prediction results include: stone presence determination status and stone risk level; the construction process of the stone risk prediction model includes: Extract unlabeled image data, indicator data, and text data from historical patients in the historical database; The image data is subjected to multimodal image standardization, morphological and texture feature extraction, feature fusion and dimensionality reduction processing in sequence to obtain image feature vectors; The indicator data is sequentially cleaned, standardized and encoded, and feature filtered to obtain the indicator feature vector; The text data is subjected to vector transformation and semantic enhancement to obtain a text semantic vector; A cross-modal residual attention network is used to fuse the image feature vector, the index feature vector, and the text semantic vector to obtain a comprehensive feature vector; The comprehensive feature vector is subjected to vector standardization, bi-clustering algorithm parameter optimization, cluster generation and fusion in sequence to obtain the feature clustering result; Based on the feature clustering results, the neural network is iteratively trained using a preset comprehensive objective optimization function to obtain the original prediction classification model; Labeled image data, indicator data, and text data of patients with kidney stones were extracted from historical databases to obtain fine-tuned small sample data. The original prediction classification model is fine-tuned through meta-learning, cluster correction, and cluster labeling using the fine-tuned small sample data to obtain the stone risk prediction model.

[0006] Preferably, the imaging data includes: CT images, ultrasound images, X-ray images, and microscopic images of urinary sediment crystals; the indicator data includes: urinary biochemical indicators, blood biochemical indicators, and basic clinical indicators; the urinary biochemical indicators include: pH value, specific gravity, calcium, oxalate, uric acid, and citric acid; the blood biochemical indicators include: creatinine, blood urea nitrogen, electrolytes, and parathyroid hormone; the basic clinical indicators include: age, gender, BMI, and relative abundance of intestinal flora; and the text data includes: electronic medical record text and examination report text.

[0007] Preferably, the neural network comprises: an input layer, a cross-modal attention fusion layer, a feature enhancement layer, a hidden layer, and an output layer connected in sequence; the expression of the cross-modal attention fusion layer includes: ;in, Let be the weight of the m-th mode; It is a multilayer perceptron; For feature vectors; This is the query vector.

[0008] Preferably, the comprehensive feature vector is subjected to vector standardization, bi-clustering algorithm parameter optimization, cluster generation and fusion in sequence to obtain the feature clustering result, including: The comprehensive feature vector is normalized. The normalized comprehensive feature vector is clustered using DBSCAN and DPC respectively to obtain the first clustering result and the second clustering result; The similarity between the first clustering result and the second clustering result is calculated using the Hungarian algorithm to obtain a similarity matrix. Then, the first clustering result and the second clustering result are fused using a weighted voting method based on the similarity matrix to obtain the feature clustering result.

[0009] Preferably, the expression for the comprehensive objective optimization function is: ;in, This is the total loss function; , , These are the first weight, the second weight, and the third weight, respectively. This represents the cluster dispersion loss; This represents the cluster separation loss. This is a regularization term.

[0010] Preferably, the original prediction classification model is fine-tuned through meta-learning, clustering correction, and cluster labeling using the fine-tuned small sample data to obtain the kidney stone risk prediction model, including: The parameters of the original prediction classification model are divided based on the MAML framework to obtain the updated parameter range; Based on the fine-tuned small sample data, the original prediction classification model is iteratively adjusted using a fine-tuning loss function that mitigates class imbalance loss and cluster pseudo-label loss. After the iteration is completed, each cluster is labeled according to the label distribution state to obtain the stone risk prediction model.

[0011] Preferably, the image data is subjected to multimodal image standardization, morphological and texture feature extraction, feature fusion and dimensionality reduction processing in sequence to obtain an image feature vector, including: The microscopic image of the urine sediment crystallization was subjected to Gaussian denoising and binarization segmentation to obtain a preprocessed microscopic image; The CT images, ultrasound images, and X-ray images were subjected to feature extraction using an image processing model based on 3DU-Net and ViT to obtain local-global features; The Inception-ResNet-v2 model was used to extract features from the preprocessed microscopic image to obtain texture features; The image feature vector is obtained by using principal component analysis to perform feature fusion and dimensionality reduction on the local-global features and the texture features.

[0012] Preferably, the indicator data is sequentially cleaned, standardized and encoded, and feature-filtered to obtain an indicator feature vector, including: The outlier data of the indicator data is removed using the median absolute deviation, and the missing values ​​of the indicator data are filled using the KNN algorithm. The numerical indicators in the indicator data are subjected to Z-Score standardization, and the categorical indicators in the indicator data are subjected to one-hot encoding. The indicator data is filtered using PLS-DA to obtain the indicator feature vector.

[0013] Preferably, the text data is subjected to vector transformation and semantic enhancement to obtain a text semantic vector, including: The text data is sequentially segmented, stop words are removed, and entity recognition is performed to obtain the original text embedding; The original text embedding is used to extract semantic features using the ClinicalBERT model to obtain the text semantic vector.

[0014] Preferably, an electronic device includes: at least one processor and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the aforementioned machine learning-based method for predicting kidney stone risk.

[0015] The present invention discloses the following technical effects: This invention provides a machine learning-based method and electronic device for predicting kidney stone risk. By employing a semi-supervised learning strategy of unlabeled data clustering and small-sample adjustment, it addresses the problem of strong dependence on labeled data in existing technologies, significantly reducing the need for labeled data. Through the fusion of DBSCAN and DPC algorithms for clustering and unsupervised iteration, it solves the problem of low reliability under unlabeled training, achieving reliable pre-training on unlabeled data. Furthermore, by combining multimodal data fusion, clustering iteration, and meta-learning fine-tuning, it addresses the problems of insufficient standardization and poor robustness in multimodal fusion in existing technologies, improving model robustness. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A schematic diagram of a machine learning-based stone risk prediction process provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the neural network architecture provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the feature clustering process provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the small-sample fine-tuning process provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the image feature extraction process provided in an embodiment of the present invention. Detailed Implementation

[0018] 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.

[0019] The purpose of this invention is to provide a method and electronic device for predicting the risk of kidney stones based on machine learning, which solves the problems of strong dependence on labeled data, insufficient standardization of multimodal fusion, and poor robustness in the existing technology.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Figure 1 This is a schematic diagram of a machine learning-based stone risk prediction process provided in an embodiment of the present invention, such as... Figure 1 As shown, this invention provides a machine learning-based method for predicting the risk of kidney stones, comprising: Step 100: Collect the imaging data and indicator data of the patient to be tested to obtain the raw data to be tested; Step 200: Input the raw data to be detected into the pre-trained stone risk prediction model for identification, and obtain the stone risk prediction result; the stone risk prediction result includes: stone presence determination status and stone risk level; the construction process of the stone risk prediction model includes: Step 201: Extract unlabeled image data, indicator data, and text data of historical patients from the historical database; Step 202: Perform multimodal image standardization, morphological and texture feature extraction, feature fusion and dimensionality reduction on the image data in sequence to obtain image feature vectors; Step 203: Perform data cleaning, standardization and encoding, and feature filtering on the indicator data in sequence to obtain the indicator feature vector; Step 204: Perform vector transformation and semantic enhancement on the text data to obtain a text semantic vector; Step 205: Use a cross-modal residual attention network to fuse the image feature vector, the index feature vector, and the text semantic vector to obtain a comprehensive feature vector; Step 206: Perform vector standardization, bi-clustering algorithm parameter optimization, cluster generation and fusion on the comprehensive feature vector in sequence to obtain the feature clustering result; Step 207: Based on the feature clustering results, iteratively train the neural network using a preset comprehensive objective optimization function to obtain the original prediction classification model; Step 208: Extract labeled image data, indicator data, and text data of patients with kidney stones from the historical database to obtain fine-tuned small sample data; Step 209: Use the fine-tuned small sample data to perform meta-learning fine-tuning, cluster correction, and cluster labeling on the original prediction classification model to obtain the stone risk prediction model.

[0022] Specifically, the imaging data includes: CT images, ultrasound images, X-ray images, and microscopic images of urinary sediment crystals; the indicator data includes: urinary biochemical indicators, blood biochemical indicators, and basic clinical indicators; the urinary biochemical indicators include: pH value, specific gravity, calcium, oxalate, uric acid, and citric acid; the blood biochemical indicators include: creatinine, blood urea nitrogen, electrolytes, and parathyroid hormone; the basic clinical indicators include: age, gender, BMI, and relative abundance of intestinal flora; and the text data includes: electronic medical record text and examination report text.

[0023] refer to Figure 2 The neural network comprises: an input layer, a cross-modal attention fusion layer, a feature enhancement layer, a hidden layer, and an output layer connected in sequence; the expression for the cross-modal attention fusion layer includes: ;in, Let be the weight of the m-th mode; It is a multilayer perceptron; For feature vectors; This is the query vector.

[0024] refer to Figure 3 The comprehensive feature vector is subjected to vector standardization, bi-clustering algorithm parameter optimization, cluster generation and fusion in sequence to obtain the feature clustering result, including: Step 20601: Normalize the comprehensive feature vector; Step 20602: Use DBSCAN and DPC to cluster the normalized comprehensive feature vector to obtain the first clustering result and the second clustering result; Step 20603: Calculate the similarity between the first clustering result and the second clustering result using the Hungarian algorithm to obtain a similarity matrix. Then, based on the similarity matrix, fuse the first clustering result and the second clustering result using a weighted voting method to obtain the feature clustering result.

[0025] Specifically, the expression for the comprehensive objective optimization function is: ;in, This is the total loss function; , , These are the first weight, the second weight, and the third weight, respectively. This represents the cluster dispersion loss; This represents the cluster separation loss. This is a regularization term.

[0026] refer to Figure 4 The original prediction classification model is fine-tuned through meta-learning, clustering correction, and cluster labeling using the fine-tuned small sample data to obtain the kidney stone risk prediction model, including: Step 20901: Based on the MAML framework, divide the parameters of the original prediction classification model to obtain the updated parameter range; Step 20902: Based on the fine-tuned small sample data, the original prediction classification model is iteratively adjusted using a fine-tuning loss function that mitigates class imbalance loss and cluster pseudo-label loss. After the iteration is completed, each cluster is labeled according to the label distribution state to obtain the stone risk prediction model.

[0027] refer to Figure 5 The image data is subjected to multimodal image standardization, morphological and texture feature extraction, feature fusion and dimensionality reduction processing in sequence to obtain image feature vectors, including: Step 20201: Perform Gaussian denoising and binarization segmentation on the microscopic image of the urine sediment crystallization to obtain a preprocessed microscopic image; Step 20202: Use an image processing model based on 3DU-Net and ViT to extract features from the CT image, the ultrasound image, and the X-ray image to obtain local-global features; Step 20203: Use the Inception-ResNet-v2 model to extract features from the preprocessed microscopic image to obtain texture features; Step 20204: Use principal component analysis to perform feature fusion and dimensionality reduction on the local-global features and the texture features to obtain the image feature vector.

[0028] Preferably, the indicator data is sequentially cleaned, standardized and encoded, and feature-filtered to obtain an indicator feature vector, including: The outlier data of the indicator data is removed using the median absolute deviation, and the missing values ​​of the indicator data are filled using the KNN algorithm. The numerical indicators in the indicator data are subjected to Z-Score standardization, and the categorical indicators in the indicator data are subjected to one-hot encoding. The indicator data is filtered using PLS-DA to obtain the indicator feature vector.

[0029] Furthermore, the text data undergoes vector transformation and semantic enhancement to obtain a text semantic vector, including: The text data is sequentially segmented, stop words are removed, and entity recognition is performed to obtain the original text embedding; The original text embedding is used to extract semantic features using the ClinicalBERT model to obtain the text semantic vector.

[0030] Specifically, unlabeled data is collected and expanded through multiple channels. To ensure comprehensive data coverage, this embodiment gathers unlabeled data from various sources, primarily including image data, laboratory indicator data, and text information. Simultaneously, existing technologies are used to optimize and expand the dimensions of this data, ensuring both data quality and diversity.

[0031] 1) Imaging Data: In addition to common CT, ultrasound, and X-ray images, urinary sediment crystallization microscopic images were also used. These images provide more detailed information, helping to better understand the patient's condition. Computed tomography (CT) scans, used to obtain images of internal body structures, clearly show the morphology and pathological conditions of various organs. Images formed by ultrasound reflection are commonly used to examine soft tissues and blood vessels, offering the advantages of being non-invasive and real-time. X-rays are used to penetrate human tissue and form images. Observing crystals in urine under a microscope can help determine the presence of kidney stones or other metabolic diseases.

[0032] 2) Laboratory Indicator Data: Integrating the following three categories of key indicators: Urine Biochemical Indicators: These detect various chemical components in urine, such as pH, specific gravity, and key metabolites like calcium, oxalate, uric acid, and citric acid. These indicators reflect kidney function and metabolic status. pH: Measures the acidity or alkalinity of urine, helping to determine if there is an acid-base imbalance. Specific Gravity: Reflects the degree of urine concentration, aiding in assessing the kidney's concentrating function. Calcium, Oxalate, Uric Acid, and Citric Acid: These are important metabolites in urine; abnormal levels may be related to kidney stone formation. Blood Biochemical Indicators: These detect various chemical components in blood, such as creatinine, blood urea nitrogen, electrolytes, and parathyroid hormone. These indicators reflect kidney function and endocrine status. Creatinine and Blood Urea Nitrogen: These are important indicators of kidney function; elevated levels usually indicate impaired kidney function. Electrolytes: Including sodium, potassium, and chloride, these maintain fluid balance and neuromuscular function. Parathyroid Hormone: Regulates calcium and phosphorus metabolism and is closely related to kidney disease.

[0033] We collect basic information about patients, such as age, gender, and BMI, as well as the relative abundance values ​​of various gut microbiota genera. This information helps us comprehensively assess the patient's overall health status. Age, gender, and BMI are commonly used clinical indicators that reflect the patient's basic physiological characteristics. Relative abundance values ​​of gut microbiota genera are closely related to the occurrence and development of various diseases by detecting the composition of the gut microbiota.

[0034] 3) Text Data: Collect electronic medical record texts, examination report texts, and medical order records. This text data contains rich clinical information, providing important supplementary information for model analysis. Electronic medical record texts: Contain detailed information such as the patient's medical history, symptoms, diagnosis, and treatment process. Examination report texts: Record the results of various examinations. Medical order records: Record the doctor's medication and treatment recommendations.

[0035] By collecting and integrating this multi-source, unlabeled data, we can gain a more comprehensive understanding of patients' conditions and provide a more reliable foundation for subsequent research.

[0036] Further, data preprocessing and feature extraction: 1) Image data preprocessing and feature extraction: Convert CT, ultrasound, and X-ray images to DICOM format. Images of crystallized urine sediment require Gaussian denoising and binarization segmentation.

[0037] Improved Feature Extraction Structure: This embodiment employs a hybrid model of 3DU-Net+ViT and Inception-ResNet-v2. 3DU-Net+ViT is responsible for extracting local and global features from CT, ultrasound, and X-ray images. Inception-ResNet-v2 is used to extract morphological and texture features from crystalline images.

[0038] Principal component analysis (PCA) is used to merge image features from different sources into a single image feature vector.

[0039] 2) Indicator data preprocessing and vector mapping: First, outliers are removed using the median absolute deviation (MAD) method, and then the KNN algorithm is used to complete the missing data.

[0040] Numerical metrics were standardized using Z-scores, and categorical metrics were encoded using one-heat encoding. For genus data, logarithms were taken first, followed by normalization.

[0041] The PLS-DA method is used to select key indicators, and finally, indicator feature vectors are generated.

[0042] 3) Text data preprocessing and vector mapping: First, words are segmented based on a medical dictionary. Then, stop words are removed, entities are identified, and key information such as disease history, medication history, and symptoms is extracted.

[0043] The ClinicalBERT model is used, and word vectors are clustered together using a self-attention mechanism to generate text semantic vectors.

[0044] Optionally, the neural network architecture can be optimized as follows: 1) The input layer receives image feature vectors, indicator feature vectors, and text feature vectors; 2) A cross-modal attention fusion layer introduces an inter-modal attention mechanism to calculate the weight coefficients of each modal feature and achieve dynamic fusion. The core formula is: in, Let be the weight of the m-th mode; It is a multilayer perceptron; For feature vectors; This is the query vector.

[0045] 3) The feature enhancement layer reduces the dimensionality of the fused features through a fully connected layer. This module adds residual connections to avoid gradient vanishing. 4) The hidden layer adopts a three-layer structure, with ReLU as the activation function. Each layer is followed by Dropout and batch normalization operations. 5) The output layer is used to output the comprehensive feature vector (for subsequent clustering analysis).

[0046] Preferably, a dual-algorithm fusion strategy combining DBSCAN and Density Peak Clustering (DPC) is adopted to improve adaptability to complex clinical data distributions: After standardizing the features using the Z-Score standardization method, the above algorithms are used for dual clustering analysis: DBSCAN: determines the neighborhood radius and minimum number of points through Bayesian optimization, with the optimization objective being to maximize the silhouette coefficient; DPC: core parameters include local density (reflecting the degree of sample clustering) and relative distance (reflecting the proximity relationship between samples and high-density samples); clustering results are generated by DBSCAN and DPC respectively, and the similarity matrix of the two clusters is calculated based on the Hungarian algorithm. The results are then fused using a weighted voting method.

[0047] Optionally, the comprehensive objective optimization function designed in this embodiment for iterating the above-mentioned neural network is: in, This is the total loss function; , , These are the first weight, the second weight, and the third weight, respectively. This represents the cluster dispersion loss; This represents the cluster separation loss. This is a regularization term.

[0048] Furthermore, several clinically confirmed marker data were collected, including the presence or absence of stones, risk level, and stone composition labels. The small sample data was then augmented using the SMOTE-NC algorithm. The MAML framework was employed, dividing the neural network parameters into shared parameters and task parameters. The shared parameters remained fixed, while only the task parameters were fine-tuned. The loss function obtained from the fine-tuning is as follows: in, To predict losses; The learnable coefficient; , These are the mitigation class imbalance loss and the clustering pseudo-label loss, respectively.

[0049] Based on the limited amount of fine-tuning data mentioned above, the initial prediction classification model was repeatedly adjusted by fine-tuning the loss function until it was completed. Then, according to the label distribution, each cluster was labeled, resulting in the kidney stone risk prediction model.

[0050] Furthermore, in subsequent inspection steps, the data to be predicted only needs to be input into the pre-trained neural network. The system will then provide a judgment on the presence of stones and a risk level assessment, while simultaneously generating a risk assessment report.

[0051] As an optional implementation, this embodiment also provides an electronic device, including: at least one processor, and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the aforementioned machine learning-based method for predicting kidney stone risk.

[0052] The beneficial effects of this invention are as follows: This invention significantly reduces the need for labeled data through a semi-supervised learning strategy of unlabeled data clustering and few-sample adjustment; it achieves pre-training on unlabeled data by fusing clustering and unsupervised iteration using DBSCAN and DPC dual algorithms, thus improving the reliability of unlabeled training; and it enhances the robustness of the model by combining multimodal data fusion, clustering iteration, and meta-learning fine-tuning.

[0053] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0054] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A machine learning based stone risk prediction method, characterized by, include: Collect imaging data and indicator data of the patients to be tested to obtain the raw data to be tested; The raw data to be detected is input into a pre-trained stone risk prediction model for identification, and the stone risk prediction result is obtained. The stone risk prediction results include: the status of stone presence and the stone risk level; the construction process of the stone risk prediction model includes: Extract unlabeled image data, indicator data, and text data from historical patients in the historical database; The image data is subjected to multimodal image standardization, morphological and texture feature extraction, feature fusion and dimensionality reduction processing in sequence to obtain image feature vectors; The indicator data is sequentially cleaned, standardized and encoded, and feature filtered to obtain the indicator feature vector; The text data is subjected to vector transformation and semantic enhancement to obtain a text semantic vector; A cross-modal residual attention network is used to fuse the image feature vector, the index feature vector, and the text semantic vector to obtain a comprehensive feature vector; The comprehensive feature vector is subjected to vector standardization, bi-clustering algorithm parameter optimization, cluster generation and fusion in sequence to obtain the feature clustering result; Based on the feature clustering results, the neural network is iteratively trained using a preset comprehensive objective optimization function to obtain the original prediction classification model; Labeled image data, indicator data, and text data of patients with kidney stones were extracted from historical databases to obtain fine-tuned small sample data. The original prediction classification model is fine-tuned through meta-learning, cluster correction, and cluster labeling using the fine-tuned small sample data to obtain the stone risk prediction model.

2. The machine learning based stone risk prediction method of claim 1, wherein, The imaging data includes: CT images, ultrasound images, X-ray images, and microscopic images of urinary sediment crystals; the indicator data includes: urinary biochemical indicators, blood biochemical indicators, and basic clinical indicators; the urinary biochemical indicators include: pH value, specific gravity, calcium, oxalate, uric acid, and citric acid; the blood biochemical indicators include: creatinine, blood urea nitrogen, electrolytes, and parathyroid hormone; the basic clinical indicators include: age, gender, BMI, and relative abundance of intestinal flora; the text data includes: electronic medical record text and examination report text. 3.The machine learning based stone risk prediction method of claim 1, wherein, The neural network comprises: an input layer, a cross-modal attention fusion layer, a feature enhancement layer, a hidden layer, and an output layer connected in sequence; the expression for the cross-modal attention fusion layer includes: ; wherein, is a weight for the mth modality; is a multi-layer perceptron; is a feature vector; is a query vector.

4. The method for predicting kidney stone risk based on machine learning according to claim 1, characterized in that, The comprehensive feature vector is subjected to vector standardization, bi-clustering algorithm parameter optimization, cluster generation and fusion in sequence to obtain the feature clustering result, including: The comprehensive feature vector is normalized. The normalized comprehensive feature vector is clustered using DBSCAN and DPC respectively to obtain the first clustering result and the second clustering result; The similarity between the first clustering result and the second clustering result is calculated using the Hungarian algorithm to obtain a similarity matrix. Then, the first clustering result and the second clustering result are fused using a weighted voting method based on the similarity matrix to obtain the feature clustering result.

5. The method for predicting kidney stone risk based on machine learning according to claim 1, characterized in that, The expression for the comprehensive objective optimization function is: ;in, This is the total loss function; , , These are the first weight, the second weight, and the third weight, respectively. This represents the cluster dispersion loss; This represents the cluster separation loss; This is a regularization term.

6. The machine learning based stone risk prediction method of claim 1, wherein, The original prediction classification model is fine-tuned through meta-learning, clustering correction, and cluster labeling using the fine-tuned small sample data to obtain the kidney stone risk prediction model, including: The parameters of the original prediction classification model are divided based on the MAML framework to obtain the updated parameter range; Based on the fine-tuned small sample data, the original prediction classification model is iteratively adjusted using a fine-tuning loss function that mitigates class imbalance loss and cluster pseudo-label loss. After the iteration is completed, each cluster is labeled according to the label distribution state to obtain the stone risk prediction model.

7. The machine learning based stone risk prediction method of claim 2, wherein, The image data is subjected to multimodal image standardization, morphological and texture feature extraction, feature fusion and dimensionality reduction processing in sequence to obtain image feature vectors, including: The microscopic image of the urine sediment crystallization was subjected to Gaussian denoising and binarization segmentation to obtain a preprocessed microscopic image; The CT images, ultrasound images, and X-ray images were subjected to feature extraction using an image processing model based on 3DU-Net and ViT to obtain local-global features; The Inception-ResNet-v2 model was used to extract features from the preprocessed microscopic image to obtain texture features; The image feature vector is obtained by using principal component analysis to perform feature fusion and dimensionality reduction on the local-global features and the texture features.

8. The machine learning based stone risk prediction method of claim 2, wherein, The indicator data is sequentially cleaned, standardized and encoded, and feature-filtered to obtain the indicator feature vector, including: The outlier data of the indicator data is removed using the median absolute deviation, and the missing values ​​of the indicator data are filled using the KNN algorithm. The numerical indicators in the indicator data are subjected to Z-Score standardization, and the categorical indicators in the indicator data are subjected to one-hot encoding. The indicator data is filtered using PLS-DA to obtain the indicator feature vector. 9.The machine learning based stone risk prediction method of claim 2, wherein, The text data is subjected to vector transformation and semantic enhancement to obtain a text semantic vector, including: The text data is sequentially segmented, stop words are removed, and entity recognition is performed to obtain the original text embedding; The original text embedding is used to extract semantic features using the ClinicalBERT model to obtain the text semantic vector.

10. An electronic device, comprising: include: At least one processor, and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform a machine learning-based method for predicting the risk of kidney stones according to any one of claims 1 to 9.

Images