Chronic kidney disease mineral and bone disorder marker prediction method and system

Abstract

The application discloses a chronic kidney disease mineral and bone abnormality marker prediction method and system. The method comprises the following steps: obtaining historical data, calculating derived electrical characteristics based on the historical data, and constructing a multi-dimensional input feature system; automatically preprocessing and reducing the dimension of the multi-dimensional input feature system data; training an integrated learning regression model based on an extreme random tree by using the reduced features and corresponding serum biochemical index true values, and obtaining an optimal prediction model through automatic hyperparameter optimization; inputting real-time data of a to-be-tested person into the model to obtain prediction values of key biomarkers, and quantifying the contribution degrees of each input feature based on SHAP values. The application realizes non-invasive and high-precision prediction of chronic kidney disease mineral and bone abnormality markers, effectively replaces traditional invasive blood detection, has excellent model interpretability, and provides a reliable basis for clinical diagnosis and treatment.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical prediction technology, and in particular to a method and system for predicting mineral and bone abnormality markers in chronic kidney disease. Background Technology

[0002] Chronic kidney disease (CKD) has become a major public health challenge worldwide, often accompanied by the systemic complication of CKD-mineral and bone disorder (CKB). This condition arises from decreased kidney function leading to impaired phosphate excretion, which in turn triggers compensatory increases in fibroblast growth factor 23 (GF-23) and parathyroid hormone (PGH). Long-term progression can lead to severe clinical outcomes such as renal osteodystrophy and vascular calcification, making CKB a key factor influencing patient prognosis. Currently, the core of CKD-MBD management in clinical practice relies on regular monitoring of key biochemical indicators such as serum calcium, phosphorus, and whole-fractional parathyroid hormone (WF). However, current standard testing methods require venous blood sampling combined with laboratory biochemical analysis. This invasive procedure causes continuous discomfort for patients requiring lifelong monitoring, especially hemodialysis patients. Frequent arteriovenous fistula punctures not only increase patient discomfort but also significantly increase the risk of local infection and thrombosis, resulting in generally poor patient compliance.

[0003] Limited by operational complexity, medical costs, and patient affordability, these biochemical tests are typically conducted monthly or quarterly. The low-frequency, static data makes it difficult to capture the dynamic fluctuations of indicators and reflect the body's immediate response to treatment plans, thus affecting clinicians' ability to make precise adjustments to treatment. The DOPPS study shows that the achievement rate of CKD-MBD management in my country, especially in primary healthcare institutions, remains unsatisfactory, and the lack of convenient and efficient monitoring methods is a significant reason for this situation. From an informational perspective, traditional biochemical tests only provide isolated concentration values ​​of serum ions. These values ​​are essentially the final manifestation of the complex regulatory network within the body and cannot provide deeper pathophysiological information about the body's fluid status, cell membrane function, nutritional status, etc.—information that is closely related to the progression of CKD-MBD.

[0004] Given this context, bioimpedance analysis (BIA), as a non-invasive and rapid detection method, can accurately quantify the distribution of body fluids and cellular structural characteristics by measuring parameters such as extracellular resistance, intracellular resistance, and cell membrane capacitance, providing a potential pathway for assessing mineral metabolism. However, directly using BIA parameters to predict CKD-MBD biochemical indicators faces significant technical bottlenecks: the relationship between bioelectrical properties and serum mineral concentration is highly nonlinear and influenced by multiple physiological factors such as individual body size differences and tissue complexity, making it difficult for traditional linear statistical analysis methods to establish an effective mapping relationship. Although existing studies have attempted to use BIA parameters at single or limited frequencies for prediction, they have failed to fully utilize the rich tissue information contained in the broadband bioimpedance spectrum, resulting in limited prediction accuracy and generalization ability. Furthermore, existing methods lack interpretability analysis of model prediction results, making it difficult for clinicians to understand the intrinsic correlation between electrophysiological characteristics and metabolic indicators, hindering the widespread application of this technology in clinical practice. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the problems of existing technologies that rely on invasive detection, have insufficient prediction accuracy, and lack interpretability.

[0006] To address the aforementioned technical problems, this invention provides a method for predicting mineral and bone abnormality markers in chronic kidney disease, comprising the following steps:

[0007] S1: Acquire historical data, which includes the patient's basic physiological information, basic bioelectrical impedance parameters, and bioelectrical impedance spectrum data; calculate derived electrical characteristics based on the basic bioelectrical impedance parameters and basic physiological information, and integrate the basic physiological information, basic bioelectrical impedance parameters, derived electrical characteristics, and bioelectrical impedance spectrum data to construct a multi-dimensional input feature system;

[0008] S2: Perform automated preprocessing on the data in the multidimensional input feature system, and perform dimensionality reduction and compression on the bioelectrical impedance spectrum data in the multidimensional input feature system to obtain the feature data after preprocessing and dimensionality reduction;

[0009] S3: Using the preprocessed and dimensionality-reduced feature data and their corresponding true values ​​of serum biochemical indicators, train the learning regression model based on extreme random trees, and optimize the learning regression model to obtain the optimal prediction model.

[0010] S4: Input the real-time data of the subject into the optimal prediction model to obtain the predicted values ​​of key biomarkers of mineral and bone abnormalities in chronic kidney disease, and quantify the contribution of each input feature to the individual prediction result.

[0011] In one embodiment of the present invention, in S2, the bioelectrical impedance spectrum data in the multidimensional input feature system is subjected to dimensionality reduction and compression to obtain feature data after preprocessing and dimensionality reduction. The method is as follows:

[0012] The numerical matrix formed by the bioelectrical impedance spectrum data is zero-mean normalized, and the covariance matrix of the zero-mean normalized data is calculated. The covariance matrix is ​​then subjected to eigenvalue decomposition to obtain the eigenvalues ​​and corresponding eigenvectors of the covariance matrix. Subsequently, the eigenvectors corresponding to the m largest eigenvalues ​​whose cumulative variance contribution rate exceeds a set threshold are selected according to the eigenvalues ​​to form a projection matrix. The original bioelectrical impedance spectrum data matrix is ​​multiplied by the projection matrix to obtain the dimensionality-reduced low-dimensional feature matrix.

[0013] In one embodiment of the present invention, the method for training and optimizing a learning regression model based on extreme random trees to obtain the optimal prediction model is as follows: The preprocessed and dimensionality-reduced feature data and their corresponding true values ​​of serum biochemical indicators are used as the dataset; the dataset is divided into a training set and a test set according to a preset ratio; K-fold cross-validation is used on the training set, and an automated hyperparameter optimization framework is combined to train and optimize the parameters of the extreme random tree model; wherein, during the construction of each decision tree, the extreme random tree model simultaneously performs random feature sampling and random selection of the splitting threshold when splitting nodes.

[0014] In one embodiment of the present invention, when using K-fold cross-validation on the training set, the first K-1 fold training data in the training set are processed as follows: uniformly distributed noise of a preset amplitude is applied to the weight, body mass index, extracellular resistance, and intracellular resistance in the training data, and Gaussian distributed noise of a preset parameter is applied to the cell membrane capacitance to form perturbed basic physical quantities; the derived electrical features are recalculated based on the perturbed basic physical quantities, and the recalculated derived electrical features are mixed with the original data to form an enhanced training dataset.

[0015] In one embodiment of the present invention, in S4, the method for quantifying the contribution of each input feature to a single prediction result is as follows: for a single prediction result, based on the SHAP value calculation method, the model prediction output is decomposed into the sum of the marginal contributions of each input feature for calculation.

[0016] In one embodiment of the present invention, in S1, the derived electrical characteristics include at least a body-size-normalized extracellular electrical impedance index, a body-size-normalized intracellular electrical impedance index, and a cell membrane capacitance per unit mass of tissue.

[0017] The body-size-standardized extracellular electrical impedance index is calculated by dividing the extracellular resistance by the body mass index; the body-size-standardized intracellular electrical impedance index is calculated by dividing the intracellular resistance by the body mass index; and the cell membrane capacitance per unit mass of tissue is calculated by dividing the cell membrane capacitance by the body weight.

[0018] In one embodiment of the present invention, in S1, the basic bioelectrical impedance parameters include extracellular resistance, intracellular resistance, and cell membrane capacitance obtained under single-frequency measurement; the bioelectrical impedance spectrum data are impedance magnitude and phase angle data measured at multiple discrete frequency points within a preset frequency range.

[0019] This invention also provides a predictive system for mineral and bone abnormality markers in chronic kidney disease, comprising the following modules:

[0020] The feature construction module is used to acquire historical data, which includes the patient's basic physiological information, basic bioelectrical impedance parameters, and bioelectrical impedance spectrum data; based on the basic bioelectrical impedance parameters and basic physiological information, derived electrical features are calculated, and the basic physiological information, basic bioelectrical impedance parameters, derived electrical features, and bioelectrical impedance spectrum data are integrated to construct a multi-dimensional input feature system;

[0021] The data processing module is used to automatically preprocess the data in the multidimensional input feature system and to reduce and compress the bioelectrical impedance spectrum data in the multidimensional input feature system to obtain the feature data after preprocessing and dimensionality reduction.

[0022] The model training module is used to train the ensemble learning regression model based on extreme random trees using the preprocessed and dimensionality-reduced feature data and the corresponding true values ​​of serum biochemical indicators, and to optimize the learning regression model to obtain the optimal prediction model.

[0023] The predictive analysis module is used to input the real-time data of the test subject into the optimal prediction model to obtain the predicted values ​​of key biomarkers of mineral and bone abnormalities in chronic kidney disease, and to quantify the contribution of each input feature to a single prediction result.

[0024] The present invention also provides an electronic device comprising a processor, a memory, and a bus system, wherein the processor and the memory are connected via the bus system, the memory is used to store instructions, and the processor is used to execute the instructions stored in the memory to implement the method for predicting mineral and bone abnormality markers in chronic kidney disease.

[0025] The present invention also provides a computer storage medium, characterized in that the computer storage medium stores a computer software product, the computer software product including a plurality of instructions for causing a computer device to execute the method for predicting mineral and bone abnormality markers in chronic kidney disease.

[0026] The technical solution of the present invention has the following advantages compared with the prior art:

[0027] This invention achieves non-invasive and accurate prediction of mineral bone abnormality markers in chronic kidney disease by integrating bioelectrical impedance spectroscopy and machine learning techniques. This method requires only a single non-invasive measurement for assessment, effectively avoiding the risks of traditional blood tests and significantly improving patient adherence to long-term monitoring. A high-precision prediction model was established by constructing a multi-dimensional feature system and combining it with ensemble learning algorithms, demonstrating significantly superior performance compared to traditional methods. Data augmentation and feature dimensionality reduction strategies enhance the model's robustness and generalization ability. Furthermore, interpretability analysis techniques are introduced to make the prediction results clinically interpretable, providing a reliable basis for treatment and strongly promoting the clinical application of this non-invasive assessment technology. Attached Figure Description

[0028] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0029] Figure 1 This is a flowchart illustrating the method for predicting mineral and bone abnormality markers in chronic kidney disease provided in this embodiment of the invention.

[0030] Figure 2 This is a schematic diagram of the structure of the extreme random tree ensemble learning regression model in an embodiment of the present invention;

[0031] Figure 3 This is a schematic diagram of the results of interpretability analysis of the serum calcium prediction model using the SHAP method in Experiment 1;

[0032] Figure 4 This is a schematic diagram of the results of interpretability analysis of the parathyroid hormone prediction model using the SHAP method in Experiment 1.

[0033] Figure 5 This is a schematic diagram of the results of interpretability analysis of the serum phosphorus prediction model using the SHAP method in Experiment 1;

[0034] Figure 6 This is a scatter plot of the regression fit between the predicted and actual values ​​of Ca in Experiment 1.

[0035] Figure 7 This is a scatter plot of the regression fit between the predicted and actual values ​​of P in Experiment 1.

[0036] Figure 8 This is a scatter plot of the regression fit between the predicted and actual values ​​of PTH in Experiment 1.

[0037] Figure 9 This is the residual distribution diagram of Ca in Experiment 1;

[0038] Figure 10 This is the residual distribution diagram of P in Experiment 1;

[0039] Figure 11 This is the residual distribution diagram of PTH in Experiment 1;

[0040] Figure 12 This is a schematic diagram of the structure of the chronic kidney disease mineral and bone abnormality marker prediction system provided in this embodiment of the invention. Detailed Implementation

[0041] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0042] Example 1:

[0043] like Figure 1 As shown, this invention provides a method for predicting mineral and bone abnormality markers in chronic kidney disease, comprising the following steps:

[0044] S1: Acquire historical data, which includes the patient's basic physiological information, basic bioelectrical impedance parameters, and bioelectrical impedance spectrum data; calculate derived electrical characteristics based on the basic bioelectrical impedance parameters and basic physiological information, and integrate the basic physiological information, basic bioelectrical impedance parameters, derived electrical characteristics, and bioelectrical impedance spectrum data to construct a multi-dimensional input feature system;

[0045] S2: Preprocess the data in the multidimensional input feature system and perform dimensionality reduction and compression on the bioelectrical impedance spectrum data in the multidimensional input feature system to obtain the feature data after preprocessing and dimensionality reduction;

[0046] S3: Using the preprocessed and dimensionality-reduced feature data and their corresponding true values ​​of serum biochemical indicators, train the learning regression model based on extreme random trees, and optimize the learning regression model to obtain the optimal prediction model.

[0047] S4: Input the real-time data of the subject into the optimal prediction model to obtain the predicted values ​​of key biomarkers of mineral and bone abnormalities in chronic kidney disease, and quantify the contribution of each input feature to the individual prediction result.

[0048] The technical solution of this invention is based on the deep fusion of multi-source bioelectrical impedance data and machine learning algorithms. By constructing a multi-dimensional feature system including basic physiological parameters, single-frequency impedance characteristics, and broadband spectrum information, and combining it with an optimized ensemble learning model, high-precision prediction of key biochemical indicators is achieved. Compared with traditional blood tests, this solution only requires non-invasive measurement to obtain reliable test results, which not only reduces patient suffering but also creates conditions for long-term dynamic monitoring. At the same time, the ensemble learning model has good interpretability, clearly demonstrating the contribution of each feature to the prediction results, providing strong support for clinical diagnostic decisions, and has excellent clinical application prospects.

[0049] Specifically, in S1, historical data collection and the construction of a high-quality dataset were systematically carried out. For each historical patient, three types of key information were collected simultaneously: First, basic physiological information, including basic indicators such as age, gender, height, and weight, and the body mass index (BMI) was calculated; second, bioelectrical impedance information, which contains two levels of data. The first level is basic bioelectrical impedance parameters measured at a specific frequency of 50kHz, including extracellular resistance (…). ), intracellular resistance ( ) and cell membrane capacitance ( The three core electrical parameters are: first, bioelectrical impedance spectrum data measured at 50 discrete frequency points uniformly selected within a wide frequency range of 5 kHz to 1 MHz, with impedance modulus and phase angle recorded at each frequency point, collectively forming rich frequency domain features; and second, target label information, which consists of the true values ​​of serum biochemical indicators collected synchronously with the bioelectrical impedance data, including the concentrations of serum calcium (mmol / L), phosphorus (mmol / L), and full-range parathyroid hormone (pg / mL). After rigorous data cleaning, outlier removal, and integrity verification, three specialized prediction datasets were constructed: a calcium (Ca) prediction dataset, a phosphorus (P) prediction dataset, and a parathyroid hormone (PTH) prediction dataset. Each dataset ensures the representativeness of the samples and the reliability of the data.

[0050] Based on high-quality datasets, this invention innovatively constructs a comprehensive, multi-level, multi-dimensional input feature system. The design of this feature system fully considers the pathophysiological characteristics of mineral and bone abnormalities in chronic kidney disease, achieving a comprehensive representation of disease features. The basic feature set directly utilizes collected basic physiological information (age, sex, height, weight, BMI) and basic bioelectrical impedance parameters (…). , , These features provide the model with the most basic individual information and electrical properties.

[0051] To further improve the physiological relevance and predictive stability of the model, a derived feature set with clear physiological significance is constructed. The basic parameters are then standardized by mathematical transformations, specifically including:

[0052] The body-standardized extracellular electrical impedance index is calculated using the following formula:

[0053] ,

[0054] Body size-standardized extracellular electrical impedance index eliminates the interference of body size factors on extracellular resistance and more accurately reflects the distribution of extracellular fluid.

[0055] The body size-standardized intracellular electrical impedance index is calculated using the following formula:

[0056] ,

[0057] The body size-standardized intracellular electrical impedance index eliminates the influence of body size differences and better characterizes the distribution characteristics of intracellular fluid.

[0058] The cell membrane capacitance per unit mass of tissue is calculated using the following formula:

[0059] ,

[0060] Cell membrane capacitance per unit mass of tissue reflects the functional state of the cell membrane per unit body weight.

[0061] The introduction of these derived features significantly improves the model's ability to adapt to individual differences and enhances the stability and reliability of the prediction results.

[0062] The construction of the frequency domain feature set fully leverages the rich information in bioelectrical impedance spectroscopy data. By integrating impedance modulus and phase angle data at 50 frequency points, a frequency domain feature set containing 100 features was formed. These features comprehensively record the electrical response characteristics of tissues over a wide frequency range. Low-frequency features primarily reflect extracellular fluid information, high-frequency features characterize intracellular fluid states more closely, while mid-frequency features contain detailed information on cell membrane function. This wide-band feature acquisition method enables the model to deeply mine electrophysiological features related to abnormal mineral metabolism at multiple scales.

[0063] The multidimensional feature system not only comprehensively characterizes the pathophysiological states related to mineral and bone abnormalities in chronic kidney disease from multiple dimensions, but also provides a solid data foundation for building high-precision, highly generalizable predictive models through feature complementarity and enhancement. Each feature level has a clear physiological meaning and computational basis, together forming a complete feature system that is both theoretically supported and practically valuable.

[0064] Furthermore, in S2, the constructed multidimensional input feature system undergoes systematic data cleaning and preprocessing, including one-hot encoding of categorical features to convert discrete categorical variables into numerical forms, which facilitates processing by machine learning models; at the same time, missing values ​​in all numerical features are filled with the median of the feature. This method can effectively maintain the distribution characteristics of the data and avoid introducing bias.

[0065] To further eliminate the impact of dimensional differences between different features on model training, this embodiment applies Z-score standardization to all numerical features, ensuring they follow a standard normal distribution with a mean of 0 and a variance of 1. This not only solves the problem of model training instability caused by inconsistent feature scales but also significantly improves the convergence speed and final performance of the model. The standardized feature data allows for a relatively balanced weight distribution among different features in the model, thereby improving the model's prediction accuracy and generalization ability.

[0066] The bioelectrical impedance spectrum data in the multidimensional input feature system is subjected to dimensionality reduction and compression to obtain preprocessed and dimensionality-reduced feature data. The method is as follows: the numerical matrix composed of the bioelectrical impedance spectrum data is zero-mean normalized, the covariance matrix of the zero-mean normalized data is calculated, the covariance matrix is ​​eigenvalued to obtain the eigenvalues ​​and corresponding eigenvectors of the covariance matrix, and then the eigenvectors corresponding to the m largest eigenvalues ​​whose cumulative variance contribution rate exceeds a set threshold are selected according to the eigenvalues ​​to form a projection matrix. The original bioelectrical impedance spectrum data matrix is ​​multiplied by the projection matrix to obtain the dimensionality-reduced low-dimensional feature matrix.

[0067] Specifically, for a high-dimensional frequency domain feature set consisting of 100 features, principal component analysis (PCA) is used for dimensionality reduction, a process implemented through rigorous mathematical calculations. A preset cumulative variance contribution rate threshold η is set, preferably 95%, to effectively remove data redundancy and noise while preserving core information, thereby reducing the risk of model overfitting.

[0068] Let the original feature matrix be:

[0069] ,

[0070] Where n is the number of samples, and p=111 is the number of original features. For the first The first sample Each feature value.

[0071] By performing zero-mean normalization on the data to eliminate the influence of dimensions and bias, we can obtain the zero-mean characteristic matrix. :

[0072] ,

[0073] in, Represents an n×1 column vector of all 1s. The feature mean vector, .

[0074] Then calculate the covariance matrix. The formula is:

[0075] .

[0076] By performing eigenvalue decomposition on the covariance matrix, we can obtain:

[0077] ,

[0078] in, Let be the eigenvalue of the k-th principal component. Let it be its corresponding feature vector, and have .

[0079] Furthermore, the cumulative variance contribution rate constraint is calculated, and the variance contribution rate of a single principal component is defined:

[0080] ,

[0081] in, For the first The variance contribution rate of each principal component;

[0082] Define the cumulative variance contribution rate of the first m principal components. :

[0083] .

[0084] In this embodiment, the cumulative variance contribution rate is set to be greater than or equal to a threshold. Then the number of principal components m satisfies:

[0085] ;

[0086] The projection matrix is ​​formed by taking the eigenvectors corresponding to the first m largest eigenvalues. :

[0087] ;

[0088] The new feature matrix obtained after dimensionality reduction of the original features by PCA for:

[0089] ,

[0090] in Features are used as inputs to the regression model.

[0091] The 111 original features (including 11 related features and 100 original impedance spectrum frequency domain features) are compressed into m linearly independent principal components. While retaining more than 95% of the core information, this effectively reduces the model complexity and the risk of overfitting, providing high-quality feature inputs for the subsequent establishment of a high-precision prediction model.

[0092] Furthermore, such as Figure 2 As shown in S3, the ensemble learning regression model based on the Extra Trees model is trained using the preprocessed and dimensionality-reduced feature data and its corresponding true values ​​of serum biochemical indicators. The method is as follows: the preprocessed and dimensionality-reduced feature data and its corresponding true values ​​of serum biochemical indicators are used as the dataset; the dataset is divided into a training set and a test set according to a preset ratio; K-fold cross-validation is used on the training set, and the Extra Trees model is trained and its parameters are optimized using an automated hyperparameter optimization framework; wherein, in the process of constructing each decision tree, the Extra Trees model simultaneously performs random feature sampling and random selection of the splitting threshold when splitting nodes.

[0093] When using K-fold cross-validation on the training set, for each round of cross-validation, the training set is divided into K mutually exclusive subsets. The first K-1 folds of training data are used as the training data each time, and the remaining K-th subset is used as the validation data. A data augmentation strategy is implemented on the training data of the current round (i.e., the first K-1 folds of training data). Specifically, uniformly distributed noise of a preset amplitude (weight ±3%, others ±2%) is applied to the weight, body mass index, extracellular resistance, and intracellular resistance in the training data. Gaussian distributed noise with a mean of 1.0 and a standard deviation of 0.005 is applied to the cell membrane capacitance to form perturbed basic physical quantities. Based on the perturbed basic physical quantities, the derived electrical features (including body-size-standardized extracellular impedance index, intracellular impedance index, and cell membrane capacitance per unit mass of tissue) are recalculated. The recalculated derived electrical features are mixed with the original training data to form the augmented training dataset. The validation data (i.e., the remaining subset of the current round) is not subjected to data augmentation and is directly used as validation data for model validation.

[0094] In this embodiment, preferably, the complete dataset is randomly divided into a training set and an independent test set according to a preset ratio of 80%:20% to ensure the objectivity of model evaluation. During the training phase, a 5-fold cross-validation strategy is adopted, further dividing the training set into 5 mutually exclusive subsets, and performing training and validation 5 times in a loop. Each time, 4 subsets are used as training data, and the remaining subset is used as validation data, so as to fully explore the potential of the data and reliably evaluate the model performance.

[0095] During cross-validation, an innovative physical consistency data augmentation strategy was implemented on the first four folds of training data. Specifically, random noise consistent with the actual measurement error distribution was applied to the basic physical quantities in the training data: ±3% uniformly distributed noise was applied to body weight, ±2% uniformly distributed noise was applied to body mass index, extracellular resistance, and intracellular resistance, and Gaussian distributed noise with a mean of 1.0 and a standard deviation of 0.005 was applied to cell membrane capacitance. This noise parameter setting, based on statistical analysis of actual measurement errors, effectively simulates measurement variability in a real clinical environment. Based on the perturbed basic physical quantities, derived electrical characteristics, including the body size-standardized extracellular impedance index, were recalculated according to preset physical formulas. Body size-standardized intracellular electrical impedance index and cell membrane capacitance per unit mass of tissue This process creates enhanced data samples that maintain consistent physical relationships. The recalculated derived electrical features are then mixed with the original data to construct an enhanced training dataset, significantly improving the model's robustness to measurement noise and its generalization ability.

[0096] For model training, Extremely Random Trees (ERTs) are chosen as the core regression algorithm. The ERT model is a further development of the Bagging concept in ensemble learning, consisting of multiple independently generated random decision trees. For a regression problem, let there be a total of T trees, and the regression function corresponding to the t-th tree is... The overall output of the model is the arithmetic mean of the predictions from each tree. :

[0097] ,

[0098] in, The eigenvectors after dimensionality reduction using PCA.

[0099] The construction process of each decision tree employs a dual randomization mechanism: at each split node, K candidate features are randomly selected. Candidate splitting thresholds are generated completely randomly within the value range of each feature. Choose the partition that minimizes the mean square error of the split region:

[0100] ,

[0101] in, The optimal splitting feature is the one selected from K randomly chosen candidate features that is chosen as the best feature for splitting at the current node. The optimal splitting threshold is for the optimal features. An optimal numerical threshold was found; L and R represent the left and right sub-regions after partitioning, respectively. The mean of the target values ​​for the corresponding region samples; if the characteristics of the samples The value is less than or equal to If the result is positive, the sample is assigned to the left subregion (L); otherwise, it is assigned to the right subregion (R). Indicates finding a specific pair The parameters make the following summation expression... The value is the smallest.

[0102] This dual random mechanism of random feature sampling and completely random selection of the partition threshold effectively enhances the diversity and generalization ability of the model and avoids the risk of overfitting.

[0103] To fully leverage the performance potential of extreme random tree ensemble learning regression models and ensure optimal accuracy in real-world prediction tasks, this invention introduces the advanced automated hyperparameter optimization framework Optuna to construct a systematic, repeatable, and scalable hyperparameter optimization process.

[0104] First, this invention defines a complete hyperparameter search space based on the model structure and task characteristics, covering:

[0105] Number of decision trees (n_estimators): 100–1000;

[0106] Maximum tree depth (max_depth): 5–50;

[0107] Minimum number of leaf node samples (min_samples_leaf): 1–20;

[0108] Number of random samples for a feature (max_features): Auto, sqrt, log2, or a numerical range;

[0109] Key parameters include the minimum number of samples required for node splitting (min_samples_split).

[0110] Based on this, the average coefficient of determination is validated using K-fold cross-validation. Maximizing the objective function, an automatic optimization system is constructed. Optuna employs an adaptive search algorithm based on a tree-structured Parzen estimator, without directly modeling... Instead, it uses a reverse modeling approach to apply historical evaluation results to performance thresholds. The system is divided into high-performance and low-performance regions, and the conditional densities of two classes are estimated separately:

[0111] ,

[0112] ;

[0113] And its probabilistic model is constructed using kernel density estimation:

[0114] ,

[0115] in For kernel function, For bandwidth, This is a record of historical experiments.

[0116] The sampling criterion for the tree-structured Parzen estimator is to maximize the following ratio:

[0117] ,

[0118] That is, we prioritize the selection of hyperparameter combinations that have a higher probability in good sample distributions and a lower probability in bad sample distributions, thereby accelerating the exploration of regions with high performance potential and achieving the maximum expected improvement in an approximate sense.

[0119] To further improve efficiency, Optuna introduces an experimental pruning mechanism to terminate experiments prematurely for hyperparameter combinations that perform poorly midway: when the intermediate results of an experiment... If the value falls below the baseline value at the same time in historical trials (the median is used in this preferred example), pruning is triggered:

[0120] ,

[0121] in For historical experiments to time The intermediate result set, denoted as , It is the first The hyperparameter test at time step The verification performance, It has already run to step 1 The number of trials; the threshold is determined by Optuna based on Historical performance is calculated automatically.

[0122] Pruning avoids unnecessary computation, while introducing asynchronous parallel computing to support simultaneous evaluation of different experiments in a multi-threaded / multi-process environment; and introducing database persistence to ensure the traceability and recoverability of large-scale experiments.

[0123] After 100–500 rounds of hyperparameter sampling and evaluation, Optuna automatically selects the parameter combination with the best cross-validation performance and retrains it on the complete training set to obtain the final model. Results show that the extreme random tree model optimized by Optuna significantly improves performance on the independent test set. The system effectively balances model complexity and generalization ability by using key indicators such as MAE and RMSE, enabling the Chronic Kidney Disease-Mineral and Bone Abnormality Marker Prediction System of this invention to meet the accuracy requirements for clinical application and providing reliable technical support for non-invasive detection.

[0124] In the model evaluation phase, the coefficient of determination (COP) is used. The three metrics—mean absolute error (MAE) and root mean square error (RMSE)—are used to comprehensively evaluate performance on independent test sets.

[0125] RMSE is a sample standard deviation used to measure the difference between predicted and actual values. It is more sensitive to samples with large errors and can effectively reflect the dispersion of model predictions. The calculation formula is as follows:

[0126] ;

[0127] MAE is used to calculate the average absolute error between predicted and actual values. It intuitively reflects the actual magnitude of the prediction error and has good interpretability. The calculation formula is as follows:

[0128] ;

[0129] Used to characterize the model's ability to explain the variance of the dependent variable, it is a key indicator for measuring goodness of fit. The closer its value is to 1, the stronger the model's ability to explain data variability. The calculation formula is as follows:

[0130] ;

[0131] in, The total number of samples in the test set; Indicates the first The true value of each sample indicator; The model represents the first Predicted values ​​for each sample; This is the arithmetic mean of the true values ​​of all samples in the test set.

[0132] Furthermore, once the optimal prediction model is obtained, it can be put into practical application. For subjects who need to be assessed for mineral and bone abnormalities in chronic kidney disease, their basic physiological information and bioelectrical impedance data are collected. After undergoing the same feature construction, preprocessing, and dimensionality reduction process as in the training phase, the processed real-time data is input into the trained optimal prediction model, which can quickly output the predicted values ​​of key biomarkers (including serum calcium, phosphorus, and whole-segment parathyroid hormone), achieving non-invasive and rapid clinical assessment.

[0133] To enhance the clinical interpretability of the model's predictions and build physicians' trust in non-invasive assessment methods, this implementation method employs the SHAP (Shapley Additive Explanations) value calculation method to quantify the feature contribution of individual prediction results. Based on the Shapley value theory in cooperative game theory, the SHAP method assigns a specific numerical value, the SHAP value, to each feature. This value accurately characterizes the marginal contribution of that feature in driving the model's prediction from the baseline value to the final output value.

[0134] Specifically, for the feature vector of a single sample and its model prediction output The SHAP method decomposes the model output into the sum of the contributions of each feature using the following formula. :

[0135] ,

[0136] in, This represents the average of all predicted values ​​for all samples, i.e., the baseline output of the model; The number of features; It is the first The SHAP value of a feature represents the specific contribution of that feature to the prediction result. A positive value indicates that the feature improves the prediction result, while a negative value indicates that the feature reduces the prediction result. The absolute value reflects the strength of the influence.

[0137] Each feature SHAP value Defined as:

[0138] ,

[0139] in, Indicates that it does not contain features Feature subset, For subset Size, weighting factor This ensures that all possible feature orders are considered fairly. Indicates the inclusion of features The model's predicted value, while Indicates that it does not contain features The difference between the model predictions and the predicted values ​​represents the feature. Join a subset Marginal contribution over time.

[0140] In this embodiment, six key electrical characteristics with clear physiological significance were specifically selected. , , , , , Independent SHAP analysis was performed to quantify the specific contribution direction and degree of each feature to the prediction of the target indicator. The analysis results showed that in the prediction of serum calcium, extracellular resistance... It is the most important positive driving factor, and the standardized extracellular electrical impedance index Standardized intracellular electrical impedance index and intracellular resistance It shows a negative correlation; in parathyroid hormone prediction, the standardized intracellular electrical impedance index It is the most critical negative impact characteristic, followed by intracellular resistance. and standardized extracellular electrical impedance index In serum phosphorus prediction, cell membrane capacitance per unit mass and cell membrane capacitance These are the two most influential factors, showing a strong negative correlation and a significant positive correlation, respectively.

[0141] This interpretability analysis based on SHAP values ​​not only verified the reliability and physiological consistency of the model predictions and revealed the intrinsic relationship between electrophysiological characteristics and mineral metabolism indicators, but also provided a new quantitative research approach for exploring the potential pathological mechanisms of mineral and bone abnormalities in chronic kidney disease from an electrophysiological perspective, significantly enhancing the acceptance and application value of this non-invasive assessment method in clinical practice.

[0142] Experiment 1:

[0143] This experiment aims to validate the non-invasive predictive ability of a bioelectrical impedance spectroscopy-based ensemble learning model with extreme randomized trees for key biomarkers of mineral and bone abnormalities in chronic kidney disease (CKD)—serum calcium, phosphorus, and whole-segment parathyroid hormone (PSH). It also aims to evaluate the model's accuracy, robustness, and clinical interpretability. Experimental data were collected from CKD patients who visited the Department of Nephrology at Wuxi People's Hospital between 2017 and June 2025. All data were collected synchronously before hemodialysis. As mentioned earlier, the datasets were divided into three categories based on gender and target indicators: the calcium prediction dataset contained 3462 samples (2162 males and 1300 females), the phosphorus prediction dataset contained 2246 samples (1435 males and 811 females), and the PSH prediction dataset contained 1185 samples (749 males and 436 females).

[0144] The feature set is constructed as described in the previous embodiments, including a basic feature set (age, gender, height, weight, BMI, extracellular resistance, intracellular resistance, and cell membrane capacitance), derived electrical features (body size-standardized extracellular / intracellular impedance index and cell membrane capacitance per unit mass of tissue), and a frequency domain feature set (100-dimensional features consisting of impedance moduli and phase angles at 50 frequency points). In the data preprocessing stage, one-hot encoding, median padding, and Z-score normalization are employed. An innovative physical consistency data augmentation strategy is introduced, applying random noise consistent with the actual measurement error distribution to weight, BMI, resistance parameters, and cell membrane capacitance during training, recalculating the derived features, and effectively improving the model's robustness.

[0145] In the feature reduction and model training stages, principal component analysis was used to reduce the dimensionality of the 100-dimensional frequency domain features, retaining the core components with a 95% cumulative variance contribution rate. As mentioned earlier, the ensemble learning regression model based on extreme random trees employed a dual randomization mechanism during training, combined with 5-fold cross-validation and the Optuna automated hyperparameter optimization framework, to establish a high-precision prediction model. Model evaluation results show that on the independent test set, the determination coefficient for serum calcium prediction reached 0.899, with a mean absolute error of 0.036; the determination coefficient for serum phosphorus prediction was as high as 0.942, with a mean absolute error of 0.061; and the determination coefficient for whole-segment parathyroid hormone prediction was 0.891, with a mean absolute error of 25.77. All indicators demonstrated excellent predictive performance. The model performance evaluation results are shown in Table 1.

[0146] Table 1:

[0147]

[0148] Table 2 shows the best-performing data extracted from the performance evaluation results of the prediction models for each marker. Random Forest, XGBoost, and CatBoost showed stable performance, but ExtraTrees achieved the best performance and high training efficiency in all three metrics.

[0149] Table 2:

[0150]

[0151] To further enhance the clinical acceptability of the model, the SHAP value analysis method was used to conduct an interpretability study of the prediction results. The SHAP method, based on cooperative game theory, decomposes the model's prediction output into the sum of the marginal contributions of each feature. In the SHAP plot, each point represents a sample, the Y-axis is sorted by feature importance, and the X-axis represents the SHAP value, indicating the direction and degree of the feature's influence on the prediction result: a SHAP value greater than 0 indicates that the feature improves the prediction result, while a value less than 0 indicates that it decreases the prediction result; the color of the point represents the magnitude of the feature value, with red for high values ​​and blue for low values.

[0152] To clarify the physical relationship between bioelectrical impedance parameters and CKD-MBD biomarkers, we independently trained a model based on six core electrical features. , , , , , The sub-model ensures that interpretability corresponds one-to-one with actual physical characteristics. SHAP analysis results show:

[0153] like Figure 3 As shown, in the prediction of serum calcium, The most important positive feature is , , It is a negative feature;

[0154] like Figure 4 As shown, in the parathyroid hormone prediction, This is a key negative feature. , Next;

[0155] like Figure 5 As shown, in serum phosphorus prediction, and These are the two most influential factors, showing a strong negative correlation and a significant positive correlation, respectively. and Similarly, a positive correlation exists, and and The impact is relatively small, with the points concentrated near the center line, exhibiting both negative and slight positive effects. It is the least important feature, and its points are distributed very concentratedly.

[0156] This experiment successfully constructed a non-invasive prediction system for CKD-MBD biomarkers based on bioelectrical impedance spectroscopy and ensemble learning. Through multidimensional feature architecture, physical consistency data augmentation, PCA dimensionality reduction, and ExtraTrees model optimization, high-precision prediction of serum calcium, phosphorus, and whole-segment parathyroid hormone was achieved. In the model evaluation phase, the coefficient of determination (R²), mean absolute error (MAE), and root mean square error (RMSE) were used to comprehensively evaluate the performance on independent test sets. Figure 6 , Figure 7 , Figure 8 As shown, the scatter plot of the regression fit between predicted and actual values ​​for Ca, P, and PTH visually demonstrates the correlation and goodness of fit between the model's predicted values ​​and the actual values; as... Figure 9 , Figure 10 , Figure 11 As shown, the residual distribution plots, including Ca, P, and PTH, reflect the distribution characteristics of the prediction error and their consistency with the model.

[0157] According to experimental results, the highest R² of this invention can reach 0.942. Combined with SHAP interpretability analysis, the contribution of each electrophysiological feature and its physiological correlation with the target indicator were clarified, providing a transparent and reliable non-invasive monitoring tool for clinical use, with good prospects for clinical application and translation.

[0158] Example 2:

[0159] Based on the same inventive concept as Embodiment 1, the present invention also provides a predictive system for mineral and bone abnormality markers in chronic kidney disease, used to implement the steps of the predictive method for mineral and bone abnormality markers in chronic kidney disease described in Embodiment 1. Figure 12 As shown, the chronic kidney disease mineral and bone abnormality marker prediction system includes the following modules:

[0160] The feature construction module is used to acquire the basic physiological information of the subject, the basic bioelectrical impedance parameters measured at a single frequency, and the bioelectrical impedance spectrum data measured at multiple discrete frequencies; based on the basic bioelectrical impedance parameters and basic physiological information, the module calculates derived electrical features, and integrates the basic physiological information, basic bioelectrical impedance parameters, derived electrical features, and the bioelectrical impedance spectrum data to construct a multi-dimensional input feature system;

[0161] The data processing module is used to automatically preprocess the data in the multidimensional input feature system and to reduce and compress the bioelectrical impedance spectrum data in the multidimensional input feature system to obtain the feature data after preprocessing and dimensionality reduction.

[0162] The model training module is used to train the ensemble learning regression model based on extreme random trees using the preprocessed and dimensionality-reduced feature data and the corresponding true values ​​of serum biochemical indicators, and to optimize the model parameters using an automated hyperparameter optimization framework to obtain the optimal prediction model.

[0163] The predictive analysis module is used to input the real-time data of the test subject into the optimal prediction model to obtain the predicted values ​​of key biomarkers of mineral and bone abnormalities in chronic kidney disease, and to quantify the contribution of each input feature to a single prediction result.

[0164] This embodiment proposes a predictive system for mineral and bone abnormality markers in chronic kidney disease, which is used to implement the predictive method for mineral and bone abnormality markers in chronic kidney disease in Embodiment 1. Therefore, the specific implementation of the predictive system for mineral and bone abnormality markers in chronic kidney disease can be found in a part of the embodiment, and will not be repeated here to avoid redundancy.

[0165] Example 3:

[0166] The present invention also provides an electronic device, which includes a processor, a memory, and a bus system. The processor and the memory are connected through the bus system. The memory is used to store instructions, and the processor is used to execute the instructions stored in the memory to implement the method for predicting mineral and bone abnormality markers in chronic kidney disease as described in Embodiment 1.

[0167] Example 4:

[0168] The present invention also provides a computer storage medium storing a computer software product, the computer software product including a plurality of instructions for causing a computer device to execute the method for predicting mineral and bone abnormality markers in chronic kidney disease as described in Embodiment 1.

[0169] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0170] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0171] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0172] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0173] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A predictive system for mineral and bone abnormality markers in chronic kidney disease, characterized in that, Includes the following modules: The feature construction module is used to acquire historical data, which includes the patient's basic physiological information, basic bioelectrical impedance parameters, and bioelectrical impedance spectrum data; based on the basic bioelectrical impedance parameters and basic physiological information, derived electrical features are calculated, and the basic physiological information, basic bioelectrical impedance parameters, derived electrical features, and bioelectrical impedance spectrum data are integrated to construct a multi-dimensional input feature system; The data processing module is used to automatically preprocess the data in the multidimensional input feature system and to reduce and compress the bioelectrical impedance spectrum data in the multidimensional input feature system to obtain the feature data after preprocessing and dimensionality reduction. The model training module is used to train the ensemble learning regression model based on extreme random trees using the preprocessed and dimensionality-reduced feature data and the corresponding true values ​​of serum biochemical indicators, and to optimize the learning regression model to obtain the optimal prediction model. The predictive analysis module is used to input the real-time data of the test subject into the optimal prediction model to obtain the predicted values ​​of key biomarkers of mineral and bone abnormalities in chronic kidney disease, and to quantify the contribution of each input feature to a single prediction result. The basic bioelectrical impedance parameters include extracellular resistance, intracellular resistance, and cell membrane capacitance obtained under single-frequency measurement; the bioelectrical impedance spectrum data are impedance magnitude and phase angle data measured at multiple discrete frequency points within a preset frequency range. The derived electrical characteristics include at least a body-size-standardized extracellular electrical impedance index, a body-size-standardized intracellular electrical impedance index, and a cell membrane capacitance per unit mass of tissue; wherein the body-size-standardized extracellular electrical impedance index is calculated by dividing the extracellular resistance by the body mass index, the body-size-standardized intracellular electrical impedance index is calculated by dividing the intracellular resistance by the body mass index, and the cell membrane capacitance per unit mass of tissue is calculated by dividing the cell membrane capacitance by the body weight. The data processing module performs dimensionality reduction and compression on the bioelectrical impedance spectrum data as follows: the numerical matrix formed by the bioelectrical impedance spectrum data is zero-mean normalized, the covariance matrix of the zero-mean normalized data is calculated, the covariance matrix is ​​eigenvalue decomposition is performed on the covariance matrix to obtain the eigenvalues ​​and corresponding eigenvectors of the covariance matrix, and then the eigenvectors corresponding to the m largest eigenvalues ​​whose cumulative variance contribution rate exceeds a set threshold are selected according to the eigenvalues ​​to form a projection matrix. The original bioelectrical impedance spectrum data matrix is ​​multiplied by the projection matrix to obtain the dimensionality-reduced low-dimensional feature matrix.

2. The predictive system for mineral and bone abnormality markers in chronic kidney disease according to claim 1, characterized in that: In the model training module, a learning regression model based on extreme random trees is trained and optimized to obtain the optimal prediction model. This includes: using preprocessed and dimensionality-reduced feature data and their corresponding true values ​​of serum biochemical indicators as a dataset; dividing the dataset into a training set and a test set according to a preset ratio; using K-fold cross-validation on the training set and combining it with an automated hyperparameter optimization framework to train and optimize the parameters of the extreme random tree model; wherein, during the construction of each decision tree, the extreme random tree model simultaneously performs random feature sampling and random selection of the splitting threshold when splitting nodes.

3. The predictive system for mineral and bone abnormality markers in chronic kidney disease according to claim 2, characterized in that: When using K-fold cross-validation on the training set, the first K-1 fold training data in the training set are processed as follows: uniformly distributed noise of a preset amplitude is applied to the weight, body mass index, extracellular resistance, and intracellular resistance in the training data, and Gaussian distributed noise of a preset parameter is applied to the cell membrane capacitance to form the perturbed basic physical quantities. The derived electrical features are recalculated based on the perturbated basic physical quantities, and the recalculated derived electrical features are mixed with the original data to form an enhanced training dataset.

4. The predictive system for mineral and bone abnormality markers in chronic kidney disease according to claim 1, characterized in that: In the predictive analysis module, the contribution of each input feature to a single prediction result is quantified, including: for a single prediction result, the model prediction output is decomposed into the sum of the marginal contributions of each input feature based on the SHAP value calculation method.

5. An electronic device, characterized in that, The electronic device includes a processor, a memory, and a bus system, the processor and the memory being connected via the bus system, the memory being used to store instructions, and the processor being used to execute the instructions stored in the memory to run the chronic kidney disease mineral and bone abnormality marker prediction system according to any one of claims 1 to 4.

6. A computer storage medium, characterized in that, The computer storage medium stores a computer software product, the computer software product including a plurality of instructions for causing a computer device to run the chronic kidney disease mineral and bone abnormality marker prediction system according to any one of claims 1 to 4.

Images