A liver abnormal metabolism data processing method and system based on multi-modal feature fusion

Abstract

This invention provides a method and system for processing abnormal liver metabolic data based on multimodal feature fusion, relating to the field of artificial intelligence data processing technology. The method includes acquiring and concatenating multidimensional parameters of non-invasive ultrasound images and blood biochemical clinical feature vectors of the target object to obtain multimodal heterogeneous feature data; using a preset dimensionality reduction algorithm to filter features from the multimodal heterogeneous feature data and counting the selected frequencies to construct a core multimodal feature combination; inputting the core multimodal feature combination into a machine learning classification and prediction model constructed based on an extreme gradient boosting tree algorithm, and mining cross-modal correlation features between physical acoustic dimensions and biochemical indicator dimensions through the model's gradient boosting mechanism; and obtaining an objective probability score of the abnormal liver metabolic state corresponding to the target object based on the model output. This application overcomes the problems of heterogeneous data dimension barriers and single algorithm feature selection bias, achieving high-precision, objective, and quantitative digital risk stratification output.

Description

Technical Field

[0001] This invention relates to the fields of medical artificial intelligence and multimodal data processing technology, and in particular to a method and system for processing abnormal liver metabolic data based on multimodal feature fusion. Background Technology

[0002] With the development of medical informatization, using non-invasive data to assist in the assessment of liver physiological status has become a trend. However, existing data processing methods have significant technical shortcomings when dealing with high-dimensional, heterogeneous medical feature data. Current conventional models often rely solely on single-modal data, such as ultrasound parameters or blood indicators, or simply perform a linear combination of physical acoustic data and biochemical indicators. Due to the inherent scale gap between the dimensions of physical acoustics and biochemical indicators, this coarse-grained data processing approach cannot effectively align heterogeneous data spaces and is prone to introducing a large amount of redundant noise. Furthermore, existing feature dimensionality reduction methods often rely on a single algorithm, resulting in severe feature selection bias and an inability to comprehensively extract a subset of features with stable representational capabilities. These shortcomings can lead to model failure in complex nonlinear mappings, resulting in serious deviations in the output probability of the target object's state, and even causing disastrous false positives or false negatives, greatly limiting the reliability and safety of data risk assessment in assisting clinical decision-making.

[0003] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] To address the problems in existing technologies, such as the difficulty in aligning the spatial dimensions of heterogeneous medical data and the susceptibility of single dimensionality reduction algorithms to noise interference leading to severe output probability deviations, this application aims to provide a method and system for processing abnormal liver metabolic data based on multimodal feature fusion. By invoking multiple dimensionality reduction algorithms and constructing a core multimodal feature combination based on the frequency of feature selection, the gradient boosting mechanism of extreme gradient boosting trees is utilized to mine cross-modal correlation features between multidimensional parameters of non-invasive ultrasound images and clinical feature vectors of blood biochemistry. This solves the technical pain points in existing technologies, such as poor anti-interference ability of data processing models and lack of objectivity and high accuracy in output risk assessment results due to the barriers of heterogeneous data dimensions and insufficient representation of single features.

[0005] This invention provides a method for processing abnormal liver metabolic data based on multimodal feature fusion, comprising: Obtain multidimensional parameters of non-invasive ultrasound images and clinical feature vectors of blood biochemistry of the target object; concatenate the multidimensional parameters of non-invasive ultrasound images and clinical feature vectors of blood biochemistry to obtain multimodal heterogeneous feature data; Multiple preset dimensionality reduction algorithms are called to perform feature filtering on multimodal heterogeneous feature data to obtain multiple candidate feature subsets; the selection frequency of each feature in multiple candidate feature subsets is counted, and features whose selection frequency meets preset conditions are extracted to construct core multimodal feature combination; The core multimodal feature combination is input into a pre-trained machine learning classification and prediction model. The machine learning classification and prediction model is built based on the extreme gradient boosting tree algorithm. The gradient boosting mechanism of the extreme gradient boosting tree algorithm is used to perform nonlinear weighting and node splitting on the features in the core multimodal feature combination to explore the cross-modal correlation features between non-invasive ultrasound image multidimensional parameters and blood biochemistry clinical feature vectors. Based on the output of the machine learning classification and prediction model, obtain the objective probability score of the abnormal liver metabolic state corresponding to the target object.

[0006] In some optional embodiments, before concatenating the multidimensional parameters of non-invasive ultrasound images and the clinical feature vectors of blood biochemistry to obtain multimodal heterogeneous feature data, the method further includes: Missing values ​​in the multidimensional parameters of non-invasive ultrasound images and the clinical feature vectors of blood biochemistry were imputed. The Z-score normalization method was used to normalize the multidimensional parameters of the interpolated non-invasive ultrasound images and the clinical feature vectors of blood biochemistry, mapping the physical acoustic dimensions of the multidimensional parameters of the non-invasive ultrasound images and the biochemical index dimensions of the clinical feature vectors of blood biochemistry to the same standard normal distribution interval.

[0007] In some optional embodiments, multiple preset dimensionality reduction algorithms are invoked to perform feature filtering on the multimodal heterogeneous feature data, resulting in multiple candidate feature subsets, including: The algorithm employs a forward stepwise regression algorithm, a recursive feature elimination algorithm based on random forest, a recursive feature elimination algorithm based on support vector machine, an elastic network regression algorithm, and a LASSO regression algorithm in parallel to evaluate and filter the feature importance of multimodal heterogeneous feature data, resulting in five candidate feature subsets.

[0008] In some optional embodiments, the frequency of selection of each feature in multiple candidate feature subsets is counted, features whose selection frequency meets preset conditions are extracted, and a core multimodal feature combination is constructed, including: Calculate the total number of occurrences of any feature in the multimodal heterogeneous feature data across 5 candidate feature subsets; extract features that occur 3 or more times in total, and merge the extracted features to construct a core multimodal feature combination.

[0009] In some optional embodiments, the training process of the machine learning classification prediction model includes: Obtain a historical multimodal heterogeneous feature dataset containing known risk classification labels; The historical multimodal heterogeneous feature dataset was divided into a training set and a validation set in a 7:3 ratio; On the training set, the learning rate and maximum tree depth parameters of the extreme gradient boosting tree algorithm were tuned using a grid search method. The training set is fitted using the extreme gradient boosting tree algorithm after parameter tuning, and the area under the receiver operating characteristic curve is used as the evaluation metric on the validation set. The training is iterated until the preset convergence condition is met, and a machine learning classification and prediction model is obtained.

[0010] In some optional embodiments, after obtaining the machine learning classification prediction model, the training process also includes model calibration and decision evaluation steps: Input the feature data of the validation set into the machine learning classification and prediction model to obtain the predicted probability sequence; Based on the predicted probability sequence and the known risk classification labels in the validation set, a calibration curve is plotted, and the output probability of the machine learning classification prediction model is calibrated by calculating the Brier score between the predicted probability sequence and the actual observed distribution. Perform decision curve analysis to calculate the net benefit value of the machine learning classification prediction model under different threshold probabilities, and fix the final weight parameters of the machine learning classification prediction model based on the distribution of the net benefit value.

[0011] In some optional embodiments, acquiring multimodal heterogeneous feature data of the target object includes: The ultrasound-derived fat fraction and shear wave velocity of the target object were obtained as multidimensional parameters of non-invasive ultrasound imaging. Obtain the values ​​of aspartate aminotransferase, alkaline phosphatase, and hemoglobin of the target object as clinical feature vectors of blood biochemistry.

[0012] In some optional embodiments, the method further includes parallel computation and alignment steps based on traditional single-modal features: Based on the target subjects' blood biochemical clinical characteristic vector, the aspartate aminotransferase and platelet ratio indices, as well as four fibrosis indices based on age, aspartate aminotransferase, alanine aminotransferase, and platelet count, were calculated simultaneously. The objective probability scores were aligned with the distribution spaces of aspartate aminotransferase, platelet ratio index, and four fibrosis indices, and their weights were verified.

[0013] In some optional embodiments, based on the output of a machine learning classification prediction model, an objective probability score for the abnormal liver metabolic state corresponding to the target object is obtained, and the method further includes: Based on the maximum value of the Youden index from the receiver operating characteristic curve, a preset stratification threshold is determined; the objective probability score is then compared with the preset stratification threshold. If the objective probability score is greater than or equal to the stratification threshold, then output the first classification label indicating the first risk state. If the objective probability score is less than the stratification threshold, then output a second classification label indicating the second risk state.

[0014] This invention also provides a liver abnormal metabolic data processing system based on multimodal feature fusion, comprising: The data acquisition and stitching module is used to acquire multidimensional parameters of non-invasive ultrasound images and clinical feature vectors of blood biochemistry of the target object; and stitch the multidimensional parameters of non-invasive ultrasound images and clinical feature vectors of blood biochemistry to obtain multimodal heterogeneous feature data. The feature collaborative dimensionality reduction and fusion module is used to call multiple preset dimensionality reduction algorithms to perform feature filtering on multimodal heterogeneous feature data, obtain multiple candidate feature subsets; count the selection frequency of each feature in multiple candidate feature subsets, extract features whose selection frequency meets preset conditions, and construct core multimodal feature combination; The machine learning classification and prediction module is used to input the core multimodal feature combination into the pre-trained machine learning classification and prediction model. The machine learning classification and prediction model is built based on the extreme gradient boosting tree algorithm. The gradient boosting mechanism of the extreme gradient boosting tree algorithm is used to perform nonlinear weighting and node splitting on the features in the core multimodal feature combination to explore the cross-modal correlation features between the multidimensional parameters of non-invasive ultrasound images and the clinical feature vectors of blood biochemistry. The risk outcome output module is used to obtain the objective probability score of the abnormal liver metabolic state corresponding to the target object based on the output of the machine learning classification prediction model.

[0015] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure.

[0016] The liver tissue assessment method, system, storage medium, and electronic device based on multimodal data of the present invention have the following beneficial effects: This application employs multiple preset dimensionality reduction algorithms to perform feature selection on multimodal heterogeneous feature data containing multidimensional parameters of non-invasive ultrasound images and clinical feature vectors of blood biochemistry. It also counts the selection frequency of each feature in multiple candidate feature subsets to construct a core multimodal feature combination. This technology effectively eliminates redundant noise and spurious variables in the heterogeneous medical data space through cross-validation and frequency voting mechanisms of multiple algorithms, avoids feature selection bias that is easily generated by a single dimensionality reduction algorithm, and ensures that the data input to the downstream machine learning model has extremely high stable representation ability. This application inputs the selected core multimodal features into a machine learning classification and prediction model built on an extreme gradient boosting tree. Through the algorithm's nonlinear weighting and node splitting mechanism, it deeply mines the cross-modal correlation features between the multidimensional parameters of non-invasive ultrasound images with physical acoustic dimensions and the clinical feature vectors of blood biochemistry with biochemical indicators. This technology breaks through the technical bottleneck of traditional linear models being unable to handle cross-dimensional heterogeneous data, accurately captures the deep nonlinear mapping law between changes in physical acoustic morphology and abnormalities in biochemical indicators, and effectively solves the technical defects of insufficient representation by single features. Thus, it outputs objective probability scores for high-risk metabolic risk data states that are highly accurate, objectively quantified, and non-invasive. Attached Figure Description

[0017] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.

[0018] Figure 1 A flowchart of a method for processing abnormal liver metabolic data based on multimodal feature fusion provided in an embodiment of this application; Figure 2 A Venn diagram illustrating feature filtering of multimodal heterogeneous feature data provided in an embodiment of this application; Figure 3 A decision curve (DCA) analysis diagram of a machine learning classification prediction model provided in an embodiment of this application; Figure 4 A probability calibration curve of a machine learning classification prediction model provided in an embodiment of this application; Figure 5 This is a structural diagram of a liver abnormal metabolic data processing system based on multimodal feature fusion, provided in an embodiment of this application. Detailed Implementation

[0019] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0020] Furthermore, the accompanying drawings are merely illustrative of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0021] The flowchart shown in the attached diagram is merely an illustrative example and does not necessarily include all steps. For example, some steps may be broken down, while others may be combined or partially combined. Therefore, the actual execution order may change depending on the specific circumstances.

[0022] All target data involved in this invention, including but not limited to ultrasound-derived fat fraction (UDFF), shear wave velocity (Vs), and various blood biochemical test data, have undergone rigorous irreversible anonymization and desensitization processing before data collection, feature extraction, and model training. This process completely removes all sensitive and potentially identifiable information. Informed consent was obtained from the subjects for all data collection and use, and the principles of the Declaration of Helsinki were strictly followed. The data has been reviewed and approved by the relevant medical institution's ethics committee. The data processing procedures of this invention fully comply with national ethical standards and legal regulations regarding data security, personal privacy protection, and human medical research.

[0023] To facilitate understanding of the embodiments of this application, the following definitions and explanations are provided for some of the technical terms and abbreviations involved in this application. Non-invasive ultrasound imaging multidimensional parameters, specifically including ultrasound-derived fat fraction (UDFF) and shear wave velocity (Vs) in this application, are physical acoustic dimensional parameters used to characterize the acoustic attenuation or scattering characteristics and biomechanical stiffness characteristics within the target tissue, respectively. Blood biochemical clinical feature vectors, specifically including aspartate aminotransferase (AST), alkaline phosphatase (ALP), and hemoglobin (Hb) in this application, are biochemical index dimensional parameters used to characterize cell and tissue damage, metabolic stress, and blood background status of the target object. A set of dimensionality reduction algorithms, including forward stepwise regression (SR-FS), recursive feature elimination algorithm based on random forest (RF-RFE), recursive feature elimination algorithm based on support vector machine (SVM-RFE), elastic network regression algorithm (EN), and minimum absolute shrinkage and selection operator regression algorithm (LASSO), are used for feature importance assessment and elimination of heterogeneous high-dimensional data. Extreme Gradient Boosting Tree (XGBoost), a nonlinear machine learning algorithm based on decision tree ensemble, is used in this application to mine deep correlations between cross-modal features. Z-score normalization, a data normalization method, is used to map data of different dimensions to a standard normal distribution interval with a mean of 0 and a standard deviation of 1. The Aspartate Aminotransferase and Platelet Ratio Index (APRI), the FIB-4 index based on age-related aspartate aminotransferase, alanine aminotransferase, and platelet counts, the Nonalcoholic Fatty Liver Fibrosis Score (NFS), and the Fatty Liver Index (FLI) are existing clinical single-modal assessment models constructed based on conventional biochemical indicators. In this application, they are configured as a baseline comparison reference for verifying the synergistic gain effect of cross-modal correlation mining, and as a priori reference covariate for downstream distribution space data alignment and weight verification. The Brier score, a statistical index of mean squared error used to measure the consistency between the predicted probability output of a probabilistic prediction model and the actual observed distribution, is used in this application to characterize and perform output probability calibration of the model. The net benefit value, in decision curve analysis, is a weighted difference between the benefit of a true positive prediction and the cost of a false positive prediction at a specific risk threshold probability. This application uses this numerical distribution to evaluate the overall decision-making effectiveness of the model and fix the weight parameters. The Youden index is a function value calculated by subtracting one from the sum of the sensitivity and specificity of the corresponding point on the receiver operating characteristic curve. This application uses the maximum value of this function to locate the optimal preset stratification threshold for mapping continuous probability scores to discrete classification labels.

[0024] like Figure 1 As shown in the figure, this application provides a method for processing abnormal liver metabolic data based on multimodal feature fusion. This method is executed by a computer device and specifically includes the following steps: Step S100: Obtain the non-invasive ultrasound image multidimensional parameters and blood biochemistry clinical feature vector of the target object; concatenate the non-invasive ultrasound image multidimensional parameters and blood biochemistry clinical feature vector to obtain multimodal heterogeneous feature data. In this embodiment, the non-invasive ultrasound image multidimensional parameters have physical acoustic dimensions, and the blood biochemistry clinical feature vector has biochemical index dimensions. Before feature concatenation, imputation is performed on the missing values ​​in the non-invasive ultrasound image multidimensional parameters and blood biochemistry clinical feature vector. The Z-score normalization method is used to normalize the imputed non-invasive ultrasound image multidimensional parameters and blood biochemistry clinical feature vector, mapping the physical acoustic dimensions of the non-invasive ultrasound image multidimensional parameters and the biochemical index dimensions of the blood biochemistry clinical feature vector to the same standard normal distribution interval, so as to eliminate the feature weight dominance effect caused by the difference in numerical scale. Subsequently, the normalized non-invasive ultrasound image multidimensional parameters and blood biochemistry clinical feature vector are matrix concatenated in the data dimension to generate multimodal heterogeneous feature data.

[0025] Step S200: Multiple preset dimensionality reduction algorithms are invoked to perform feature filtering on the multimodal heterogeneous feature data, resulting in multiple candidate feature subsets. The selection frequency of each feature within these candidate feature subsets is counted, and features whose selection frequency meets preset conditions are extracted to construct a core multimodal feature combination. Combined with... Figure 2 As shown, the Intersections bar chart illustrates the feature aggregation distribution under the intersection of different algorithms; the Numbers bar chart shows the scale of feature selection results for each individual algorithm. By extracting features whose frequency meets preset conditions, they are merged to construct a core multimodal feature combination with principal component dimensionality reduction. Multiple dimensionality reduction algorithms are invoked in parallel: forward stepwise regression, recursive feature elimination algorithm based on random forest, recursive feature elimination algorithm based on support vector machine, elastic network regression algorithm, and LASSO regression algorithm. These algorithms perform feature importance evaluation on the multimodal heterogeneous feature data based on information criteria, Gini impurity, or mixed regularization penalty terms, shrinking the weight coefficients of low-contribution features, and outputting five candidate feature subsets. The total number of times any feature in the multimodal heterogeneous feature data appears in the five candidate feature subsets is calculated. Features with a total occurrence count greater than or equal to 3 are extracted. This frequency extraction mechanism combines... Figure 2 The candidate feature subsets shown intersect, and cross-algorithm ensemble voting is performed to filter out locally optimal features and highly collinear variables generated by a single algorithm, and merge them to construct a core multimodal feature combination with principal component dimension reduction.

[0026] Step S300: The core multimodal feature combination is input into a pre-trained machine learning classification and prediction model. This model is constructed based on the extreme gradient boosting tree algorithm. The gradient boosting mechanism of the extreme gradient boosting tree algorithm performs nonlinear weighting and node splitting on the features in the core multimodal feature combination, uncovering cross-modal correlation features between multidimensional parameters of non-invasive ultrasound images and clinical feature vectors of blood biochemistry. During node splitting in the classification and regression tree, the model traverses the physical acoustic features and biochemical index features in the core multimodal feature combination, calculates the first and second-order gradient statistics corresponding to each split point, and seeks the splitting path that maximizes the information gain of the objective function. The co-residence and nonlinear combination of the multidimensional parameters of non-invasive ultrasound images (physical acoustic dimension) and the clinical feature vectors of blood biochemistry (biochemical index dimension) on the splitting path constitute cross-modal correlation features, realizing the mathematical correlation between physical tissue morphology representation and underlying metabolic environment indicators in the same high-dimensional feature space.

[0027] Step S400: Based on the output of the machine learning classification prediction model, obtain the objective probability score of the abnormal liver metabolic state corresponding to the target object. After the extreme gradient boosting tree algorithm completes node splitting and cross-modal feature mining, the weights of the leaf nodes in each decision tree where the core multimodal feature combination of the target object falls are extracted and summed. The summed weights are input to the Sigmoid activation function for nonlinear mapping, transforming and outputting a continuous probability value with a numerical range between 0 and 1. This continuous probability value is the objective probability score. Mathematically, this objective probability score characterizes the degree to which multimodal heterogeneous feature data is judged to deviate from normal physiological benchmarks in the high-dimensional feature space of the current model. Generating the objective probability score directly through the underlying weight mapping and activation function avoids linear calculation errors caused by differences in the original dimensions of heterogeneous data.

[0028] After obtaining the objective probability score of the abnormal liver metabolic state corresponding to the target object based on the output of the machine learning classification prediction model, the system further includes determining a preset stratification threshold based on the maximum value of the Youden index of the receiver operating characteristic curve; the objective probability score is then compared with the preset stratification threshold. If the objective probability score is greater than or equal to the stratification threshold, a first classification label indicating a first risk state is output; if the objective probability score is less than the stratification threshold, a second classification label indicating a second risk state is output. The first risk state corresponds to the data distribution state where the heterogeneity parameters of the target object deviate from the safety baseline, and the second risk state corresponds to the data distribution state where the baseline is not exceeded. The first and second classification labels are configured as discrete numerical identifiers readable by the computer system. The output classification labels are used to provide a basis for automated data routing and distribution for downstream medical information processing terminals or data management systems. Data packets carrying the first classification label are routed by the system to a high-priority review queue or trigger dynamic data monitoring tasks. The above operations utilize the continuous output values ​​of the machine learning model to perform hard threshold segmentation based on the Youden index, realizing the conversion of high-dimensional heterogeneous data sets into discrete system control instructions.

[0029] In a specific downstream auxiliary application scenario, the aforementioned first risk state, at the data level, characterizes the target object's metabolic abnormal parameters as deviating extremely from the normal range. Correspondingly, in the clinical pathology dimension, the first classification label can be used to indicate that the target object has a very high probability of being in high-risk metabolic-associated steatohepatitis (MASH), i.e., the pathological progression stage of MASH accompanied by significant fibrosis; correspondingly, the second classification label can be used to indicate that the target object is in a normal stage of non-high-risk MASH. It should be stated that the objective probability score and classification label of the high-risk MASH state output by the system in this application are essentially intermediate mathematical results output by the computer using high-dimensional data space mapping. This result serves only as a digital reference feature to trigger warnings in downstream medical information systems or to assist in non-diagnostic risk screening, and is not equivalent to, nor does it replace, the final disease diagnosis conclusion made by a doctor based on the human body.

[0030] In a preferred embodiment of this application, when acquiring multimodal heterogeneous feature data of the target object, the multidimensional parameters of non-invasive ultrasound imaging are obtained as ultrasound-derived fat fraction and shear wave velocity, and the clinical feature vectors of blood biochemistry are obtained as aspartate aminotransferase, alkaline phosphatase, and hemoglobin values. The above feature combination possesses a cross-modal synergistic mechanism, wherein the combination of ultrasound-derived fat fraction and shear wave velocity is used to simultaneously quantify and characterize the acoustic heterogeneity and biomechanical strength within the tissue, and aspartate aminotransferase and alkaline phosphatase jointly characterize cellular-level metabolic damage and membrane pressure. Specifically, hemoglobin is introduced as a physical compensation factor, and the hemoglobin concentration is used to calibrate the ultrasound backscatter signal drift caused by differences in the subject's blood background, thereby improving the underlying signal-to-noise ratio of the ultrasound image parameters.

[0031] In some embodiments, non-invasive ultrasound imaging multidimensional parameters and blood biochemical clinical feature vectors of the target subject are obtained. Specifically, this includes obtaining the ultrasound-derived fat fraction and shear wave velocity of the target subject as non-invasive ultrasound imaging multidimensional parameters; and obtaining the aspartate aminotransferase, alkaline phosphatase, and hemoglobin values ​​of the target subject as blood biochemical clinical feature vectors. In this feature combination, the ultrasound-derived fat fraction and shear wave velocity are used to simultaneously extract physical features reflecting acoustic heterogeneity and biomechanical intensity within the tissue. The hemoglobin value is configured as a physical compensation factor for the acoustic background to calibrate the calculated drift of the ultrasound backscattering coefficient caused by differences in the subject's blood background, thereby suppressing environmental noise during heterogeneous data fusion. Aspartate aminotransferase and alkaline phosphatase are used to extract biochemical features characterizing cellular-level metabolic damage.

[0032] In some embodiments, the training process of the machine learning classification prediction model includes acquiring a historical multimodal heterogeneous feature dataset containing known risk classification labels. The known risk classification labels are based on clinically known pathology gold standards. Using a random hold-out method, the historical multimodal heterogeneous feature dataset is divided into a training set and a validation set in a 7:3 ratio. 70% of the data is used for iterative learning of the model's internal parameters, and 30% is used to evaluate the model's generalization ability and prevent overfitting during training. On the training set, a grid search method is used to tune the learning rate and maximum tree depth parameters of the extreme gradient boosting tree algorithm. The trained extreme gradient boosting tree algorithm is then used to fit the training set. During iterative training, the residual between the true classification label and the previous round of predicted probabilities is calculated. A classification regression tree is generated to fit this residual by minimizing a binary cross-entropy loss function that includes a structural risk regularization term. On the validation set, the area under the receiver operating characteristic (ROC) curve is used as the evaluation metric. Iterative training continues until a preset convergence condition is met, resulting in the machine learning classification prediction model.

[0033] After obtaining the machine learning classification prediction model, the training process also includes model calibration and decision evaluation steps. The feature data of the validation set is input into the machine learning classification prediction model to obtain a predicted probability sequence. Based on the predicted probability sequence and the known risk classification labels in the validation set, combined with... Figure 4 The probability calibration curve of the machine learning classification prediction model shown is used to calibrate the output probability of the model by calculating the Brier score between the predicted probability sequence and the actual observed distribution. Figure 4 In the diagram, the horizontal axis, Predicted Probability, corresponds to the continuous predicted probabilities of the model output, while the vertical axis, Observed Probability, corresponds to the actual distribution proportion of the validation set samples. The degree of overlap between the calibration curve and the diagonal dashed line characterizes the reliability of the model's predictions. Decision curve analysis is then performed, combined with... Figure 3 The decision curve analysis graph of the machine learning classification prediction model is shown, and the net benefit value of the machine learning classification prediction model under different threshold probabilities is calculated. Figure 3 In the diagram, the horizontal axis, Threshold Probability, corresponds to the preset threshold probabilities for various strata, while the vertical axis, Net Benefit, corresponds to the net technical benefit of the model performing risk identification at a specific threshold. The final weight parameters of the machine learning classification prediction model are fixed based on the distribution of the net benefit values.

[0034] Multimodal heterogeneous feature data from 290 target subjects across four independent clinical data collection centers was acquired as a test set. This set included 118 samples at first risk and 172 samples at second risk. Simultaneously, independent multimodal heterogeneous feature data from 80 target subjects was acquired as an external validation cohort. The test set was input into a machine learning classification prediction model to obtain classification prediction results, and parallel comparative tests were performed. Figure 3 and Figure 4 As shown, this application also provides decision curves and calibration curves of Logistic Regression (Logistic), Random Forest (RF), and Support Vector Machine (SVM) as baseline comparison models on the same validation set. The shape and distribution of the objective curves show that the preferred Extreme Gradient Boosting Tree (XGBoost) model in this application has a higher net benefit value within a wider threshold probability range, and its calibration curve fits the diagonal dashed line more closely.

[0035] Table 1. Comparison of XGBoost model with commonly used non-invasive assessment models in clinical practice

[0036] Referring to the performance comparison table in Table 1, the area under the curve (AUC) of the preferred extreme gradient boosting tree machine learning classification prediction model of this application is 0.804. In the comparative data distribution space based on traditional single-modality features, the AUC of the aspartate aminotransferase and platelet ratio index (APRI model) is only 0.666, and the AUC of the four fibrosis indices based on age, aspartate aminotransferase, alanine aminotransferase, and platelet count (FIB-4 model) is only 0.533. The significance of the p-values ​​for multiple prediction indicators demonstrates the nonlinear mapping advantage of the model of this application compared with conventional biochemical feature models.

[0037] Table 2. Performance comparison of XGBoost model with univariate UDFF and Vs models

[0038] Furthermore, referring to the performance comparison table shown in Table 2, in the single-modal data distribution space without cross-modal feature stitching, the area under the curve (AUC) of the single ultrasound-derived fat fraction feature model is 0.688, and the AUC of the single shear wave velocity feature model is 0.604. Through parallel comparison and distribution space mapping of the objective data in Tables 1 and 2, the data distribution patterns and step-like synergistic gain characteristics of physical acoustic dimensional features and biochemical index dimensional features in cross-modal association mining are verified. Inputting the external validation queue into the machine learning classification prediction model, the AUC of the obtained receiver operating characteristic curve is 0.782, verifying the cross-domain stability of the model weight parameters. It should be noted that the "positive predictive value" and "negative predictive value" recorded in Tables 1 and 2 are statistical terms used for evaluating the performance of classification prediction algorithms. In this embodiment, they are used to characterize the mathematical classification conformity of the computer algorithm for the first and second risk states, respectively, and do not directly represent or constitute a clinical disease diagnosis conclusion for humans.

[0039] In some embodiments, the method further includes parallel computation and alignment steps based on traditional single-modality features. Based on the target object's blood biochemical clinical feature vector, aspartate aminotransferase (AST) and platelet ratio indices, as well as a four-item fibrosis index based on age, AST, alanine aminotransferase (ALT), and platelet count, are calculated simultaneously. The objective probability scores are then aligned with the distribution spaces of AST, platelet ratio, and the four-item fibrosis index, and their weights are checked. Specifically, the data alignment operation includes using AST, platelet ratio, and the four-item fibrosis index as prior covariates, and calculating the cumulative distribution function values ​​of these prior covariates in the historical sample distribution space, thereby mapping the original clinical dimensional values ​​of the prior covariates to a baseline distribution probability between zero and one. The weight verification operation includes calculating the deviation between the objective probability score and the baseline probability distribution. If the objective probability score is greater than or equal to a preset stratification threshold and the baseline probability distribution is less than the normal distribution reference value, and the deviation is greater than the preset verification threshold, a weight penalty mechanism is triggered. The objective probability score is multiplied by a preset attenuation penalty coefficient to obtain the verified objective probability score. This operation utilizes a traditional clinical unimodal model as the underlying prior constraint, performing cross-validation between multimodal predicted values ​​and unimodal baseline values ​​within a unified continuous probability space. By penalizing anomalous data with extreme deviations in the distribution space, the algorithm structure suppresses the risk of false positives caused by extreme jumps in a single feature. After obtaining the verified objective probability score based on the output of the machine learning classification prediction model, the process also includes determining a preset stratification threshold based on the maximum value of the Youden exponent from the receiver operating characteristic curve; the verified objective probability score is then compared with the preset stratification threshold. If the verified objective probability score is greater than or equal to the stratification threshold, a first classification label indicating the first risk state is output; if the verified objective probability score is less than the stratification threshold, a second classification label indicating the second risk state is output. The above data alignment and weight verification operations, combined with hierarchical threshold comparison based on the Youden index, map continuous objective probability scores into discrete risk classification labels, providing a digital risk stratification output for the physiological state of the target object.

[0040] like Figure 5As shown, this application also provides a liver abnormal metabolic data processing system based on multimodal feature fusion. This system, as a virtual device or set of functional modules running on a computer device, is used to implement the liver abnormal metabolic data processing method based on multimodal feature fusion of any of the above embodiments. Specifically, it includes a data acquisition and stitching module M100, a feature collaborative dimensionality reduction and fusion module M200, a machine learning classification and prediction module M300, and a risk result output module M400. The data acquisition and stitching module M100 is used to acquire multidimensional parameters of non-invasive ultrasound images and blood biochemical clinical feature vectors of the target object; wherein, the multidimensional parameters of non-invasive ultrasound images have physical acoustic dimensions, and the blood biochemical clinical feature vectors have biochemical index dimensions; the multidimensional parameters of non-invasive ultrasound images and the blood biochemical clinical feature vectors are stitched together to obtain multimodal heterogeneous feature data. The Feature Collaborative Dimensionality Reduction and Fusion Module M200 calls multiple preset dimensionality reduction algorithms to filter features from multimodal heterogeneous feature data, obtaining multiple candidate feature subsets. It then counts the selection frequency of each feature within these subsets, extracts features whose selection frequency meets preset conditions, and constructs a core multimodal feature combination. The Machine Learning Classification and Prediction Module M300 inputs the core multimodal feature combination into a pre-trained machine learning classification and prediction model. This model is built based on the extreme gradient boosting tree algorithm, which uses gradient boosting to nonlinearly weight and split nodes in the core multimodal feature combination, uncovering cross-modal correlation features between multidimensional parameters of non-invasive ultrasound images and clinical feature vectors of blood biochemistry. The Risk Result Output Module M400 obtains the objective probability score of the abnormal liver metabolic state corresponding to the target object based on the output of the machine learning classification and prediction model.

[0041] After obtaining the objective probability score of the abnormal liver metabolic state corresponding to the target object based on the output of the machine learning classification prediction model, the process also includes determining a preset stratification threshold based on the maximum value of the Youden index of the receiver operating characteristic curve; the objective probability score is then compared with the preset stratification threshold. If the objective probability score is greater than or equal to the stratification threshold, a first classification label indicating the first risk state is output; if the objective probability score is less than the stratification threshold, a second classification label indicating the second risk state is output.

[0042] It should be noted that, at the data representation level, the first risk state corresponds to the deviation of the target object's tissue acoustic heterogeneity and metabolic abnormality parameter quantification values ​​from exceeding a specific baseline; the second risk state corresponds to the normal distribution state that does not exceed this baseline. The first and second classification labels are discrete numerical identifiers readable by a computer system.

[0043] The aforementioned data alignment and weight verification operations, combined with stratified threshold comparison based on the Youden index, map continuous objective probability scores to discrete risk classification labels. The engineering purpose of outputting these classification labels is to provide downstream medical information processing terminals or data management systems with automated data routing and distribution criteria. For example, data carrying the first classification label will be automatically assigned to a high-priority review queue, or trigger automatic early warning interaction commands and high-frequency dynamic data monitoring tasks for that target object. This mechanism significantly optimizes the resource scheduling efficiency and objective stratified early warning capabilities of medical data processing terminals without directly issuing clinical diagnostic conclusions.

[0044] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of this application and should not be construed as limiting the specific implementation of this application to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of this application, and all such modifications or substitutions should be considered within the scope of protection of this application.

Claims

1. A method for processing abnormal liver metabolic data based on multimodal feature fusion, characterized in that, include: Obtain multidimensional parameters of non-invasive ultrasound images and clinical feature vectors of blood biochemistry of the target object; concatenate the multidimensional parameters of non-invasive ultrasound images and the clinical feature vectors of blood biochemistry to obtain multimodal heterogeneous feature data; Multiple preset dimensionality reduction algorithms are invoked to perform feature filtering on the multimodal heterogeneous feature data, resulting in multiple candidate feature subsets; The frequency of each feature being selected in the multiple candidate feature subsets is counted, and features whose selection frequency meets the preset conditions are extracted to construct a core multimodal feature combination. The core multimodal feature combination is input into a pre-trained machine learning classification and prediction model; the machine learning classification and prediction model is constructed based on the extreme gradient boosting tree algorithm, and the gradient boosting mechanism of the extreme gradient boosting tree algorithm is used to perform nonlinear weighting and node splitting on the features in the core multimodal feature combination to mine the cross-modal correlation features between the multidimensional parameters of the non-invasive ultrasound image and the clinical feature vector of blood biochemistry. Based on the output of the machine learning classification and prediction model, an objective probability score of the abnormal liver metabolic state corresponding to the target object is obtained.

2. The method for processing abnormal liver metabolic data based on multimodal feature fusion according to claim 1, characterized in that, Before concatenating the multidimensional parameters of the non-invasive ultrasound image and the clinical feature vector of blood biochemistry to obtain multimodal heterogeneous feature data, the method further includes: Missing values ​​in the non-invasive ultrasound image multidimensional parameters and the blood biochemistry clinical feature vector are imputed. The Z-score normalization method was used to normalize the multidimensional parameters of the interpolated non-invasive ultrasound images and the clinical feature vectors of blood biochemistry, mapping the physical acoustic dimensions of the multidimensional parameters of the non-invasive ultrasound images and the biochemical index dimensions of the clinical feature vectors of blood biochemistry to the same standard normal distribution interval.

3. The method for processing abnormal liver metabolic data based on multimodal feature fusion according to claim 1, characterized in that, The process involves calling multiple preset dimensionality reduction algorithms to perform feature filtering on the multimodal heterogeneous feature data, resulting in multiple candidate feature subsets, including: The forward stepwise regression algorithm, the recursive feature elimination algorithm based on random forest, the recursive feature elimination algorithm based on support vector machine, the elastic network regression algorithm, and the LASSO regression algorithm are called in parallel to evaluate and filter the feature importance of the multimodal heterogeneous feature data, resulting in 5 candidate feature subsets.

4. The method for processing abnormal liver metabolic data based on multimodal feature fusion according to claim 3, characterized in that, The method involves statistically analyzing the selection frequency of each feature within the multiple candidate feature subsets, extracting features whose selection frequency meets preset conditions, and constructing a core multimodal feature combination, including: Calculate the total number of occurrences of any feature in the multimodal heterogeneous feature data in the five candidate feature subsets; extract the features that occur more than or equal to three times, and merge the extracted features to construct the core multimodal feature combination.

5. The method for processing abnormal liver metabolic data based on multimodal feature fusion according to claim 1, characterized in that, The training process of the machine learning classification prediction model includes: Obtain a historical multimodal heterogeneous feature dataset containing known risk classification labels; The historical multimodal heterogeneous feature dataset is divided into a training set and a validation set in a 7:3 ratio; On the training set, the learning rate and maximum tree depth parameters of the extreme gradient boosting tree algorithm are tuned using a grid search method; The training set is fitted using the hyperparameter-tuned extreme gradient boosting tree algorithm, and the area under the receiver operating characteristic curve (ROC) is used as the evaluation metric on the validation set. The training is iterated until the preset convergence condition is met, and the machine learning classification prediction model is obtained.

6. The method for processing abnormal liver metabolic data based on multimodal feature fusion according to claim 5, characterized in that, After obtaining the machine learning classification prediction model, the training process also includes model calibration and decision evaluation steps: The feature data of the validation set is input into the machine learning classification prediction model to obtain a prediction probability sequence; Based on the predicted probability sequence and the known risk classification labels in the validation set, a calibration curve is plotted, and the output probability of the machine learning classification prediction model is calibrated by calculating the Brier score between the predicted probability sequence and the actual observed distribution. Perform decision curve analysis to calculate the net benefit value of the machine learning classification prediction model under different threshold probabilities, and fix the final weight parameters of the machine learning classification prediction model according to the distribution of the net benefit value.

7. The method for processing abnormal liver metabolic data based on multimodal feature fusion according to claim 1, characterized in that, The acquisition of multimodal heterogeneous feature data of the target object includes: The ultrasound-derived fat fraction and shear wave velocity of the target object are obtained as multidimensional parameters of the non-invasive ultrasound image. The values ​​of aspartate aminotransferase, alkaline phosphatase, and hemoglobin of the target object are obtained as the blood biochemical clinical feature vector.

8. The method for processing abnormal liver metabolic data based on multimodal feature fusion according to claim 7, characterized in that, The method also includes parallel computation and alignment steps based on traditional single-modal features: Based on the target subject's blood biochemical clinical feature vector, aspartate aminotransferase and platelet ratio indices, as well as four fibrosis indices based on age, aspartate aminotransferase, alanine aminotransferase, and platelet count are calculated simultaneously. The objective probability scores were aligned and weighted with the distribution spaces of the aspartate aminotransferase, platelet ratio index, and the four fibrosis indices.

9. The method for processing abnormal liver metabolic data based on multimodal feature fusion according to claim 1, characterized in that, The step of obtaining an objective probability score of the abnormal liver metabolic state corresponding to the target object based on the output of the machine learning classification prediction model, and then further includes: A preset stratification threshold is determined based on the maximum value of the Youden index from the receiver operating characteristic curve; the objective probability score is then compared with the preset stratification threshold. If the objective probability score is greater than or equal to the stratification threshold, then a first classification label indicating the first risk state is output. If the objective probability score is less than the hierarchical threshold, a second classification label indicating the second risk state is output.

10. A liver abnormal metabolic data processing system based on multimodal feature fusion, characterized in that, include: The data acquisition and stitching module is used to acquire multidimensional parameters of non-invasive ultrasound images and clinical feature vectors of blood biochemistry of the target object; and stitch the multidimensional parameters of non-invasive ultrasound images and the clinical feature vectors of blood biochemistry to obtain multimodal heterogeneous feature data. The feature collaborative dimensionality reduction and fusion module is used to call multiple preset dimensionality reduction algorithms to perform feature filtering on the multimodal heterogeneous feature data respectively, and obtain multiple candidate feature subsets; The frequency of each feature being selected in the multiple candidate feature subsets is counted, and features whose selection frequency meets the preset conditions are extracted to construct a core multimodal feature combination. The machine learning classification and prediction module is used to input the core multimodal feature combination into a pre-trained machine learning classification and prediction model. The machine learning classification and prediction model is constructed based on the extreme gradient boosting tree algorithm. The gradient boosting mechanism of the extreme gradient boosting tree algorithm is used to perform nonlinear weighting and node splitting on the features in the core multimodal feature combination to mine the cross-modal correlation features between the multidimensional parameters of the non-invasive ultrasound image and the clinical feature vector of blood biochemistry. The risk result output module is used to obtain an objective probability score of the abnormal liver metabolic state corresponding to the target object based on the output of the machine learning classification prediction model.

Images