Artificial intelligence driven medical test multi-index joint analysis method
By employing an AI-driven multi-indicator joint analysis method, the problem of capturing the temporal evolution characteristics and correcting drift in multi-batch data in medical testing has been solved, thereby improving the accuracy of disease diagnosis and clinical adaptability, and possessing strong clinical application value.
Patent Information
- Application Number
- CN202611434827.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-16
- Publication Date
- 2026-11-10
AI Technical Summary
Current medical testing technologies neglect the temporal evolution characteristics of multiple batches of test data, and cannot effectively capture short-term fluctuations, medium-term trends and long-term steady-state abnormalities of indicators. Furthermore, they lack targeted anomaly filtering and cross-batch drift correction mechanisms, leading to a decrease in the accuracy of disease diagnosis.
An AI-driven multi-indicator joint analysis method is adopted. By acquiring multiple batches of medical test data, cross-batch time-varying drift correction and outlier filtering are performed to construct an indicator correlation topological feature matrix. Attention-weighted temporal evolution network is used to extract dynamic evolution trajectory features, and cross-scale feature aggregation and disease risk level classification are performed.
It achieves standardized purification of multiple batches of test data, accurately captures the synergistic correlation characteristics of indicators over time, improves the accuracy of disease diagnosis, and enhances clinical suitability and diagnostic accuracy through reverse tracing and personalized test recommendations.
Smart Images

Figure CN123072597A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical testing, specifically to an artificial intelligence-driven method for joint analysis of multiple indicators in medical testing. Background Technology
[0002] Multi-indicator data from medical tests are the core foundation for early screening of clinical diseases, dynamic assessment of disease status, and development of individualized treatment plans. Accurately identifying the temporal changes and correlations of multiple test indicators is of great significance for improving the accuracy of diagnosis of difficult diseases and achieving precision medicine.
[0003] Currently, routine clinical laboratory analysis methods are mostly based on static threshold comparisons of a single time point and a single indicator to assess disease risk. They generally ignore the temporal evolution characteristics of patients' long-term, multi-batch test data, and cannot effectively capture short-term fluctuations, medium-term trends, and long-term steady-state abnormal patterns of indicators. The analysis dimensions are singular and one-sided.
[0004] Traditional analytical methods lack targeted anomaly filtering and cross-batch drift correction mechanisms, resulting in insufficient consistency and reliability of raw test data. This directly leads to a significant decrease in the accuracy of subsequent feature analysis and disease discrimination. Meanwhile, most existing technologies analyze and judge various test indicators independently, ignoring the inherent dynamic correlations and mutually restraining topological relationships between human physiological indicators. They often use fixed static correlation coefficients to analyze indicator correlations, failing to adapt to the dynamic fluctuation characteristics of physiological indicators and making it difficult to uncover implicit synergistic anomalies among indicators. Summary of the Invention
[0005] To address the aforementioned technical problems, this technical solution provides an AI-driven method for joint analysis of multiple indicators in medical testing, resolving the issues raised in the background section.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0007] Artificial intelligence-driven multi-indicator joint analysis methods for medical testing include:
[0008] Obtain multiple batches of raw medical test data for the individual to be analyzed, and perform cross-batch time-varying drift correction and outlier filtering on the multiple batches of raw medical test data to obtain a standardized time-series test dataset.
[0009] A dynamic correlation topology among multiple indicators is constructed on the standardized time-series test dataset to generate an indicator correlation topology feature matrix.
[0010] Based on the topological feature matrix of the index association, the dynamic evolution trajectory features of each test index are extracted using a time-series evolution network weighted by the attention mechanism, and the time-series evolution feature vector is obtained.
[0011] The temporal evolution feature vector is input into a multi-granularity attention fusion network for cross-scale feature aggregation to obtain a joint discriminative feature representation;
[0012] Disease risk levels are classified based on the joint discriminant feature representation, and the results of multi-indicator joint analysis of individualized medical tests are output.
[0013] Preferably, the step of acquiring multiple batches of raw medical test data for the individual to be analyzed, and performing cross-batch time-varying drift correction and outlier filtering on the multiple batches of raw medical test data to obtain a standardized time-series test dataset specifically includes:
[0014] Raw values of multiple test items, including blood biochemistry, immunology, hematology and urine analysis, are extracted from the medical records of the same individual to be analyzed, forming a raw test matrix covering N types of test indicators and spanning T time points;
[0015] For each type of test indicator in the original test matrix, the local rate of change of that indicator between two adjacent time points is calculated. The local rate of change is compared with the preset physiological fluctuation threshold range for that type of indicator. If the local rate of change exceeds the physiological fluctuation threshold range, the indicator value at that time point is marked as a candidate outlier. The formula for calculating the local rate of change is as follows:
[0016] ;
[0017] in, Indicates the first Class indicators in the first The time point relative to the first Local rate of change at each time point Indicates the first Class indicators in the first The original values at each time point ;
[0018] The criteria for determining candidate outliers are:
[0019] ;
[0020] ;
[0021] in, and The first The lower and upper bounds of the preset physiological fluctuation threshold range for the class of indicators;
[0022] For each type of test indicator in the original test matrix, a locally weighted regression curve is established with time point as the independent variable and the indicator value as the dependent variable. The fitting residual value of the locally weighted regression curve at each time point is extracted. Data points whose fitting residual value exceeds the upper limit of the residual quantile of that type of indicator are marked as statistical outliers. The formula for calculating the fitting residual of the locally weighted regression curve is as follows:
[0023] ;
[0024] in, For the first Class indicators in the first The fitting residuals at each time point This is the original value of that point. This represents the fitted value of the locally weighted regression curve at that point. For the first Centered on the first time point, for the first The kernel weights assigned to each data point at each time point;
[0025] The criteria for identifying statistical outliers are:
[0026] ;
[0027] in, For the first The 75th percentile of the fitted residuals of the class index. The interquartile range of the residual values of this indicator;
[0028] By merging candidate outliers and statistical outliers, a complete set of outlier data points is obtained. The corresponding indicator values for this complete set of outlier data points are then replaced with the weighted interpolation results of that indicator at its effective time points before and after it. The weighted interpolation replacement formula is as follows:
[0029] ;
[0030] in, For the first Similar indicators at abnormal time points Replacement value, and These are the most recent valid time points before and after the anomaly. and The corresponding valid indicator value;
[0031] To address the systematic shift caused by instrument model changes or reagent batch differences between different batches of tests, this study uses two batches of data from the same individual within an overlapping time window as a benchmark. The median deviation and interquartile range scaling factor for each indicator between the two batches are calculated. The median deviation is used to perform translation correction on the subsequent batch of data, and the interquartile range scaling factor is used to perform scaling correction on the subsequent batch of data. This process is iteratively repeated batch by batch to complete the drift correction for all batches, ultimately yielding a standardized time-series test dataset. The formula for calculating the median deviation is:
[0032] ;
[0033] in, For the first Batch to number Batch No. The median deviation of the class of indicators and The first Batch and No. The batch within the overlapping time window Numerical sequences of class indicators;
[0034] The formula for calculating the interquartile range scaling factor is:
[0035] ;
[0036] in, For the first Batch to number Batch No. Interquartile range scaling factor for class indicators For the first The interquartile range of the value of this indicator for the batch.
[0037] Preferably, the step of constructing a dynamic correlation topology among multiple indicators in the standardized time-series test dataset to generate an indicator correlation topology feature matrix specifically includes:
[0038] Using N types of test indicators in the standardized time-series test dataset as nodes, calculate the time-by-time Pearson dynamic correlation coefficient sequence for any two types of indicators at all T time points. The Pearson dynamic correlation coefficient sequence consists of the local correlation coefficient values at T time points. The formula for calculating the Pearson dynamic correlation coefficient is:
[0039] ;
[0040] in, For the first Class indicators and the first Class indicators in the first Local Pearson correlation coefficients at each time point and The first Class and First Class indicators in the first Standardized values at each time point and Each of the following is the first Centered on a specific time point, window length Local mean within;
[0041] The Pearson dynamic correlation coefficient sequence is segmented into sliding windows, with each segment consisting of L consecutive time points. The mean and variance of the correlation coefficient values within each segment are calculated. If the variance of a segment is lower than a preset dynamic correlation stability threshold, the mean of that segment is taken as the stable correlation strength value between the two indices within that time period; otherwise, the segment is marked as a correlation mutation segment. The formulas for calculating the mean and variance of the correlation coefficient values within each segment are as follows:
[0042] ;
[0043] ;
[0044] in, For the first Section No. Class and the The mean of the correlation coefficients of the class indicators, The variance of this segment. The segment length of the sliding window, For segment index;
[0045] The distribution of stable correlation strength values between the two types of indicators is statistically analyzed over the entire time period. The median of stable correlation strength values is used as the basic edge weight, and the proportion of correlation mutation segments is used as the instability penalty factor for the edge weight. The comprehensive topological edge weight between the two types of indicators is calculated. The formula for determining and calculating the stable correlation strength value is as follows:
[0046] ;
[0047] in, For the first Class and the Class indicators in the first The stable correlation strength value of the segment, This is a preset dynamic correlation stability threshold;
[0048] Using the combined topological edge weights of all pairwise index pairs as elements, construct an N×N symmetric weight matrix. Apply a threshold filter to the symmetric weight matrix, retaining edges with weights higher than a preset topological sparsity threshold, to obtain the sparsified index-related topological adjacency matrix. The formula for calculating the combined topological edge weights is as follows:
[0049] ;
[0050] in, For the first Class and the The comprehensive topological edge weights between the class indicators The median of all stable association strength values. The number of associated mutation segments, Total number of segments The coefficient of the instability penalty factor;
[0051] The spectral decomposition of the sparse index association topology adjacency matrix is performed to extract the eigenvectors corresponding to the top K largest eigenvalues. The eigenvectors are then concatenated column by column to form the index association topology feature matrix.
[0052] Preferably, the step of extracting the dynamic evolution trajectory features of each test indicator based on the indicator association topological feature matrix and using an attention-weighted temporal evolution network to obtain a temporal evolution feature vector specifically includes:
[0053] The numerical sequence of each test indicator at T time points in the standardized time series test dataset is used as the time series input of that indicator. The row vector corresponding to that indicator in the indicator association topological feature matrix is used as the topological context embedding of that indicator. The time series input and the topological context embedding are concatenated to form a topology-aware time series representation of the indicator. The concatenation formula for the topology-aware time series representation of the indicator is as follows:
[0054] ;
[0055] in, For the first Class indicators in the first Topology-aware representation of each time point This is the time-series input vector for this indicator. The first in the topological feature matrix associated with the index The topological context embedding vector of the row. , This represents a vector concatenation operation;
[0056] Arrange the topology-aware temporal representations of all N types of test indicators along the time point dimension to form a three-dimensional tensor of shape T×N×D, where D is the feature dimension after splicing.
[0057] Multi-head self-attention is calculated for the 3D tensor along the time dimension. For each time point, the feature vectors of all N types of indicators at that time point are used as the query vector, and the feature vectors of all indicators at all time points are used as the key vector and value vector. The attention weights reflect the degree of mutual influence between different indicators at different time points. The formula for calculating multi-head self-attention is as follows:
[0058] ;
[0059] ;
[0060] in, For the first All time points A query matrix composed of index features. and All Key matrix and value matrix at each time point, For the first Learnable projection matrix for each attention head For the number of attention heads, To output the projection matrix;
[0061] The attention fusion representation at each time point is obtained by weighted summation of the value vector using attention weights. The attention fusion representations at all T time points are then concatenated in chronological order to obtain a preliminary temporal evolution feature sequence. The formula for constructing the preliminary temporal evolution feature sequence is as follows:
[0062] ;
[0063] in, For the first Attention fusion representation at each time point;
[0064] The initial temporal evolution feature sequence is bidirectionally recursively encoded, extracting dependencies from both the forward and reverse temporal directions. The forward and reverse encoding results are concatenated at each time point to obtain the final temporal evolution feature vector. The concatenation formula for the bidirectional recursive encoding is as follows:
[0065] ;
[0066] in, For the first The final temporal evolution feature vector at each time point This is the result of forward recursive encoding. This is the result of reverse recursive encoding. To hide the state dimension.
[0067] Preferably, the step of inputting the temporal evolution feature vector into a multi-granularity attention fusion network for cross-scale feature aggregation to obtain a joint discriminative feature representation specifically includes:
[0068] The temporal evolution feature vector is segmented into short-range segments, medium-range segments, and long-range segments according to different time spans. The short-range segments cover consecutive w1 time points, the medium-range segments cover consecutive w2 time points, and the long-range segments cover consecutive w3 time points, with w1 less than w2 less than w3.
[0069] For each granularity of segment, perform intra-segment mean pooling and intra-segment max pooling to obtain mean pooling vectors and max pooling vectors for each granularity. Concatenate the mean pooling vectors and max pooling vectors for the same granularity to form a two-channel segment representation for that granularity. The formulas for intra-segment mean pooling and max pooling are:
[0070] ;
[0071] ;
[0072] in, and The first Particle size, number The mean pooling vector and the max pooling vector of each segment. Corresponding to short-distance, medium-distance, and long-distance, The length of the fragment at this particle size. For the first A set of time points covered by a segment For the first Temporal evolution feature vectors at each time point;
[0073] Learnable granular attention weights are applied to the two-channel segment representations of all three granularities. These granular attention weights are obtained by inputting the two-channel segment representations of each granularity into a lightweight fully connected network and then normalizing them using softmax. The lightweight fully connected network uses the norm of each granularity segment representation as an auxiliary input. The concatenation formula for the two-channel segment representations is as follows:
[0074] ;
[0075] in, For the first Seed particle size Dual-channel segment representation of a segment;
[0076] The formula for calculating granular attention weights is:
[0077] ;
[0078] in, For the first Attention weights at various granularities This represents the average two-channel representation of all segments at this granularity. Its L2 norm, and For the weight matrix and bias vector of a lightweight fully connected network;
[0079] By using granular attention weights to perform a weighted summation of the three granularity dual-channel segment representations, a granular fusion representation is obtained. The weighted summation formula for the granular fusion representation is as follows:
[0080] ;
[0081] in, This represents granularity fusion.
[0082] The granular fusion representation and the temporal evolution feature vector are concatenated using residual concatenation. Layer normalization and nonlinear activation transformation are then applied to the concatenated vector to obtain the joint discriminative feature representation. The calculation formula for the joint discriminative feature representation is as follows:
[0083] ;
[0084] in, For joint discriminative feature representation, This represents the global mean of the temporal evolution feature vector. This indicates a splicing operation. For layer normalization, It is a non-linear activation function.
[0085] Preferably, the step of classifying disease risk levels based on joint discriminant feature representation and outputting personalized medical test multi-indicator joint analysis results specifically includes:
[0086] The joint discriminant feature representation is input into the hierarchical classifier, which is composed of a first-level classifier head and a second-level classifier head connected in series. The first-level classifier head maps the joint discriminant feature representation to the probability distribution of the major disease category. The second-level classifier head maps the joint discriminant feature representation to the probability distribution of the specific disease subclass under the major disease category determined by the first-level classifier head.
[0087] The first-level classifier calculates the cosine similarity between the joint discriminative feature representation and the preset M disease category prototype vectors. The cosine similarity is then temperature-scaled and input into the softmax function to obtain the disease category probability distribution. The disease category with the highest probability value is selected as the target category. The formulas for calculating the cosine similarity and probability distribution of the first-level classifier are as follows:
[0088] ;
[0089] ;
[0090] in, For joint discriminative feature representation With the Prototype vectors of major disease categories cosine similarity, For the temperature scaling parameters of the first level, For prediction of the first The probability of each major disease category;
[0091] The formula for selecting the target category is:
[0092] ;
[0093] in, Index for the target category;
[0094] The second-level classification head conditionally modulates the joint discriminative feature representation using the prototype vector corresponding to the target major category. It then calculates the cosine similarity between the modulated features and the prototype vectors of the S pre-defined disease subclasses under that major category. After temperature scaling and softmax processing, it obtains the disease subclass probability distribution. The disease subclass with the highest probability value is selected as the target subclass. The probability distribution calculation formula for the second-level classification head is as follows:
[0095] ;
[0096] ;
[0097] in Modulated features With the Major Category prototype vectors of disease subclasses cosine similarity, For the second level of temperature scaling parameters, In the target category Predicted as the first under the condition The probability of each subclass;
[0098] The target major category and target subclass are combined, and their respective probability values and joint discriminant features are attached to the contribution distribution of attention weights at each granularity, which is then output as the result of multi-indicator joint analysis of personalized medicine test.
[0099] Preferably, after outputting the results of the multi-indicator joint analysis of personalized medical tests, the method further includes interpretable reverse tracing analysis of the joint discriminant feature representation, specifically including:
[0100] Based on the classification decision boundary corresponding to the target major class and target subclass output by the hierarchical classifier, the projection component of the joint discriminative feature representation in the direction of the decision boundary normal vector is calculated. Feature dimensions whose absolute value of the projection component is greater than a preset significance threshold are marked as key decision dimensions. The formula for calculating the projection component in the direction of the decision boundary normal vector is as follows:
[0101] ;
[0102] in, For joint discriminative feature representation In the Normal vectors along the decision boundary on each feature dimension The absolute value of the projected component, Target Category With target subclass The corresponding classification decision boundary normal vector;
[0103] The criteria for determining key decision dimensions are as follows:
[0104] ;
[0105] in, A set of indexes for all key decision dimensions. The preset significance threshold;
[0106] For each key decision dimension, we trace back to the corresponding granular segment representation in the multi-granularity attention fusion network to determine which temporal granularity segment primarily contributes to that key decision dimension. The formula for determining the contribution granularity is as follows:
[0107] ;
[0108] in, The index with the highest granularity that contributes most to key decision-making dimensions. For the first In the representation of granular fragments, the first The weight contribution of each dimension;
[0109] For a given contribution granularity segment, we further trace back to the corresponding time point range in the temporal evolution feature vector, extract the attention weight distribution within that time point range, and select the top p indicators with the highest attention weight values as the core driving indicators for this key decision dimension. The selection of core driving indicators is based on the attention weight ranking.
[0110] ;
[0111] in, For the first The set of core driving indicators corresponding to each key decision-making dimension For the first Class indicators in the first Attention weight values at each time point;
[0112] All key decision dimensions and their corresponding core driving indicators, contribution time ranges and contribution granularity types are summarized to generate a key indicator contribution list indexed by indicator name and sorted by contribution intensity.
[0113] Extract the core driving indicators with the highest contribution intensity (q) from the list of key indicators, and check whether the standardized time series test dataset already contains historical test data of the core driving indicators.
[0114] For core driving indicators that already contain historical monitoring data, calculate the deviation between the indicator's value range distribution at its historical monitoring time points and its current value. If the deviation exceeds the individual historical fluctuation range of the indicator, then the indicator is marked as a fluctuation indicator that needs to be closely monitored. The formula for calculating the deviation is:
[0115] ;
[0116] in, For the first The degree of deviation of the core driving indicators. For this indicator at the current time point Standardized values, and These are the mean and standard deviation of the indicator at historical detection time points, respectively;
[0117] The criteria for determining volatility indicators are:
[0118]
[0119] in, A set of volatility indicators, For the first Threshold for the individual historical fluctuation range of the index.
[0120] For core driver metrics that do not include historical testing data, mark the metric as a candidate metric for suggesting new testing.
[0121] The volatility index and candidate indices are sorted in descending order of their contribution intensity in the key index contribution list to generate an individualized supplementary recommendation list. Each entry in the individualized supplementary recommendation list includes the index name, the type of reason for recommendation, and a priority level based on the degree of deviation or contribution intensity. The formula for calculating the priority level is as follows:
[0122] ;
[0123] in, For the first Priority levels of category indicators This represents the contribution intensity value of this indicator in the list of key indicator contributions. This is a set of candidate new indicators.
[0124] Preferably, after generating the individualized test supplementary recommendation list, the method further includes screening for potential collaborative anomalies in the remaining indicators in the standardized time-series test dataset that were not included in the key indicator contribution list, specifically including:
[0125] From the indicator association topology feature matrix, extract the associated indicators that are directly connected to any core driving indicator in the key indicator contribution list and whose edge weight is higher than a preset collaboration threshold, to form an associated indicator candidate set. The extraction conditions for the associated indicator candidate set are as follows:
[0126] ;
[0127] in, For the candidate set of related indicators, The set of core driver metrics in the key performance indicator (KPI) contribution list. The first in the topological feature matrix associated with the index Class and the Side weights of class indicators, The preset collaboration threshold;
[0128] For each indicator in the candidate set of related indicators, extract its standardized numerical sequence at all T time points, and calculate the mutual information value between this sequence and the standardized numerical sequence of its corresponding core driving indicator. The mutual information value is obtained by calculating the ratio of the joint probability distribution to the marginal probability distribution after discretizing the two sequences by equal frequency. The formula for calculating the mutual information value is as follows:
[0129] ;
[0130] in, For the first Core-driven indicator sequence With the Class-related indicator sequence Mutual information value between them This is the joint probability distribution after isofrequency discretization. and This represents the corresponding marginal probability distribution;
[0131] The mutual information value is compared with a preset threshold for determining collaborative anomalies. If the mutual information value is higher than the threshold and the direction of change of the value of the indicator in the candidate set of the related indicators is opposite to that of its core driving indicator at the nearest time point, then the indicator is marked as a potential collaborative anomaly indicator. The determination condition for the opposite direction of change is as follows:
[0132] ;
[0133] in, and These are related indicators With core driving indicators In the The change at each point in time;
[0134] All potential collaboration anomaly indicators are sorted in descending order of mutual information value to generate a sorted list of potential collaboration anomaly indicators. The criteria for determining potential collaboration anomaly indicators are as follows:
[0135] ;
[0136] in, A set of potential collaborative anomaly indicators, This is the threshold for determining collaborative anomalies.
[0137] Preferably, after generating the potential collaborative anomaly index ranking list, the method further includes consistency verification and analysis result correction of the output results of the hierarchical classifier and the potential collaborative anomaly index ranking list, specifically including:
[0138] Read the target major category and target subclass output by the hierarchical classifier, and retrieve the set of known biomarkers associated with the target subclass from the preset disease knowledge graph;
[0139] The intersection of the indicators in the potential synergistic anomaly ranking list with the known biomarker set is calculated. If the intersection is not empty, the consistency score of the intersection indicator is obtained by multiplying its ranking weight in the ranking list with its prior importance score in the known biomarker set. The intersection calculation formula is as follows:
[0140] ;
[0141] in, For the set of intersecting indicators, A sorted list of potential collaboration anomalies. It is the set of known biomarkers retrieved from a disease knowledge graph;
[0142] The formula for calculating the consistency score is:
[0143] ;
[0144] in, Intersection Indicators Consistency score, This refers to the ranking weight of this indicator in the potential collaborative anomaly indicator ranking list. This is the prior importance score of the indicator within the known set of biomarkers;
[0145] If the consistency score is higher than the preset high consistency threshold, the original output of the hierarchical classifier will remain unchanged.
[0146] If the consistency score is lower than the preset low consistency threshold, the highest-ranking indicator in the potential collaboration anomaly indicator ranking list that belongs to the intersection will be used as the correction trigger indicator. The selection formula for the correction trigger indicator is as follows:
[0147] ;
[0148] in, To correct the trigger indicator index;
[0149] Feature shift compensation for the joint discriminant feature representation is performed using the temporal patterns of the modified triggering index in the standardized time-series test dataset. Specifically, the directional difference between the temporal pattern vector of the modified triggering index and the joint discriminant feature representation is calculated, and the directional difference is proportionally superimposed onto the joint discriminant feature representation to obtain the modified joint discriminant feature representation. The temporal pattern vector of the modified triggering index is:
[0150] ;
[0151] in, For the first Time-series pattern vector of class indicators, For this indicator in the 1st Topology-aware representation of each time point;
[0152] The formulas for compensating for orientation difference and feature offset are as follows:
[0153] ;
[0154] ;
[0155] in, To correct the orthogonal direction difference components between the trigger indicator time-series pattern vector and the joint discriminative feature representation, For compensation ratio coefficient, This is the modified joint discriminant feature representation;
[0156] The corrected joint discriminant feature representation is re-inputted into the hierarchical classifier to obtain the corrected disease risk hierarchical classification result. The corrected disease risk hierarchical classification result replaces the original output result as the final result of the multi-indicator joint analysis of individualized medical tests.
[0157] Compared with existing technologies, this invention provides an artificial intelligence-driven method for joint analysis of multiple indicators in medical testing, which has the following beneficial effects:
[0158] This invention achieves standardized purification of multi-batch time-series inspection data through dual-layer anomaly screening and batch drift correction, significantly improving the reliability and consistency of the original data;
[0159] By constructing a dynamic indicator association topology, the collaborative association characteristics of multiple indicators changing over time can be accurately captured, breaking through the limitations of traditional static correlation analysis.
[0160] By combining topology-aware temporal networks with multi-granularity attention fusion mechanisms, we can fully explore the deep features of short-term mutations and long-term evolution of indicators.
[0161] Simultaneously, a hierarchical disease classification system is constructed to achieve precise disease identification at both coarse and fine granular levels. Through reverse tracing, individualized test recommendations, collaborative anomaly screening, and knowledge graph consistency correction, the results are interpretable, traceable, and self-optimizable.
[0162] This invention takes into account both data patterns and prior medical knowledge, significantly improving the accuracy and clinical applicability of disease risk assessment, and has strong clinical application value. Attached Figure Description
[0163] Figure 1 This is a schematic diagram of the method flow for S101-S105 in this invention;
[0164] Figure 2 This is a schematic diagram of the method flow for S201-S205 in this invention;
[0165] Figure 3 This is a schematic diagram of the method flow for S301-S305 in this invention;
[0166] Figure 4 This is a schematic diagram of the method flow for S401-S405 in this invention;
[0167] Figure 5 This is a schematic diagram of the method flow for S501-S505 in this invention. Detailed Implementation
[0168] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.
[0169] Example 1
[0170] Please refer to Figure 1 As shown, an AI-driven multi-indicator joint analysis method for medical testing includes:
[0171] S101. Obtain multiple batches of raw medical test data for the individual to be analyzed, and perform cross-batch time-varying drift correction and outlier filtering on the multiple batches of raw medical test data to obtain a standardized time-series test dataset.
[0172] S102. Construct a dynamic correlation topology among multiple indicators for the standardized time series test dataset and generate an indicator correlation topology feature matrix.
[0173] S103. Based on the topological feature matrix of the index association, the dynamic evolution trajectory features of each test index are extracted using a time-series evolution network weighted by the attention mechanism to obtain the time-series evolution feature vector.
[0174] S104. Input the temporal evolution feature vector into a multi-granularity attention fusion network to perform cross-scale feature aggregation and obtain a joint discriminative feature representation;
[0175] S105. Classify disease risk levels based on the joint discriminant feature representation and output the results of multi-indicator joint analysis of individualized medical tests.
[0176] Example 2
[0177] Please refer to Figure 2 As shown, multiple batches of raw medical test data for the individual to be analyzed are obtained. Cross-batch time-varying drift correction and outlier filtering are performed on the multiple batches of raw medical test data to obtain a standardized time-series test dataset, specifically including:
[0178] S201. Extract raw values of multiple test items, including blood biochemistry, immunology, hematology and urine analysis, from the medical records of the same individual to be analyzed, and form a raw test matrix covering N types of test indicators and spanning T time points.
[0179] S202. For each type of test indicator in the original test matrix, calculate the local rate of change of that indicator between two adjacent time points. Compare the local rate of change with the preset physiological fluctuation threshold range for that type of indicator. If the local rate of change exceeds the physiological fluctuation threshold range, mark the indicator value at that time point as a candidate outlier. The formula for calculating the local rate of change is:
[0180] ;
[0181] in, Indicates the first Class indicators in the first The time point relative to the first Local rate of change at each time point Indicates the first Class indicators in the first The original values at each time point ;
[0182] The criteria for determining candidate outliers are:
[0183] ;
[0184] ;
[0185] in, and The first The lower and upper bounds of the preset physiological fluctuation threshold range for the class of indicators;
[0186] S203. For each type of test indicator in the original test matrix, establish a locally weighted regression curve with time point as the independent variable and the indicator value as the dependent variable. Extract the fitting residual value of the locally weighted regression curve at each time point. Mark data points whose fitting residual value exceeds the upper limit of the residual quantile of that type of indicator as statistical outliers. The formula for calculating the fitting residual of the locally weighted regression curve is as follows:
[0187] ;
[0188] in, For the first Class indicators in the first The fitting residuals at each time point This is the original value of that point. This represents the fitted value of the locally weighted regression curve at that point. For the first Centered on the first time point, for the first The kernel weights assigned to each data point at each time point;
[0189] The criteria for identifying statistical outliers are:
[0190] ;
[0191] in, For the first The 75th percentile of the fitted residuals of the class index. The interquartile range of the residual values of this indicator;
[0192] S204. Merge candidate outliers and statistical outliers to obtain a complete set of outlier data points. Replace the indicator values corresponding to the complete set of outlier data points with the weighted interpolation results of the indicator at its effective time points before and after it. The weighted interpolation replacement formula is:
[0193] ;
[0194] in, For the first Similar indicators at abnormal time points Replacement value, and These are the most recent valid time points before and after the anomaly. and The corresponding valid indicator value;
[0195] S205. To address the systematic shift caused by instrument model changes or reagent batch differences between different batches of tests, using two batches of data for the same individual within an overlapping time window as a benchmark, calculate the median deviation and interquartile range scaling factor for each indicator between the two batches. Use the median deviation to perform translation correction on the subsequent batch of data, and use the interquartile range scaling factor to perform scaling correction on the subsequent batch of data. Iterate batch by batch to complete the drift correction for all batches, ultimately obtaining a standardized time-series test dataset. The formula for calculating the median deviation is:
[0196] ;
[0197] in, For the first Batch to number Batch No. The median deviation of the class of indicators and The first Batch and No. The batch within the overlapping time window Numerical sequences of class indicators;
[0198] The formula for calculating the interquartile range scaling factor is:
[0199] ;
[0200] in, For the first Batch to number Batch No. Interquartile range scaling factor for class indicators For the first The interquartile range of the value of this indicator for the batch.
[0201] Specifically, this embodiment completes the cleaning and batch normalization of multiple batches of medical test data. The core principle is a dual-layer anomaly detection + batch drift joint correction mechanism. First, anomalies are screened by combining medical prior knowledge and statistical characteristics. Coarse anomalies that violate the normal physiological changes of the human body are screened using physiological fluctuation thresholds. Then, local weighted regression is used to fit the time-series trend, and fine anomalies that deviate from the time-series evolution pattern are screened based on the residual distribution, avoiding the problems of missed detection and false detection caused by single anomaly detection. For the detected abnormal data, weighted interpolation of the effective time points before and after is used to repair the data, preserving the individual time-series change trend to the greatest extent. At the same time, in response to the systematic bias of instruments, reagents and batches that are common in medical testing, a batch alignment strategy of median shift + interquartile range scaling is adopted to iteratively eliminate time-varying system drift batch by batch, realizing the uniform distribution of test data collected at different times and with different equipment, and providing a highly consistent and reliable standardized dataset for subsequent correlation modeling and time-series feature extraction.
[0202] Example 3
[0203] Please refer to Figure 3 As shown, a dynamic correlation topology is constructed among multiple indicators in the standardized time-series test dataset to generate an indicator correlation topology feature matrix, specifically including:
[0204] S301. Using the N types of test indicators in the standardized time-series test dataset as nodes, calculate the time-by-time Pearson dynamic correlation coefficient sequence for any two types of indicators at all T time points. The Pearson dynamic correlation coefficient sequence consists of the local correlation coefficient values at T time points. The formula for calculating the Pearson dynamic correlation coefficient is:
[0205] ;
[0206] in, For the first Class indicators and the first Class indicators in the first Local Pearson correlation coefficients at each time point and The first Class and First Class indicators in the first Standardized values at each time point and Each of the following is the first Centered on a specific time point, window length Local mean within;
[0207] S302. Divide the Pearson dynamic correlation coefficient sequence into sliding window segments, each segment consisting of L consecutive time points. Calculate the mean and variance of the correlation coefficient values within each segment. If the variance of a segment is lower than a preset dynamic correlation stability threshold, then the mean of that segment is taken as the stable correlation strength value between the two indices within that time period; otherwise, the segment is marked as a correlation mutation segment. The formulas for calculating the mean and variance of the correlation coefficient values within each segment are as follows:
[0208] ;
[0209] ;
[0210] in, For the first Section No. Class and the The mean of the correlation coefficients of the class indicators, The variance of this segment. The segment length of the sliding window, For segment index;
[0211] S303. Statistically analyze the distribution of stable correlation strength values between the two types of indicators over the entire time period. Use the median of stable correlation strength values as the basic edge weight, and the proportion of correlation mutation segments as the instability penalty factor for the edge weight. Calculate the comprehensive topological edge weight between the two types of indicators. The formula for determining and calculating the stable correlation strength value is as follows:
[0212] ;
[0213] in, For the first Class and the Class indicators in the first The stable correlation strength value of the segment, This is a preset dynamic correlation stability threshold;
[0214] S304. Using the comprehensive topological edge weights of all pairwise index pairs as elements, construct an N×N symmetric weight matrix. Perform threshold filtering on the symmetric weight matrix, retaining edges with weight values higher than a preset topological sparsity threshold, to obtain the sparsified index-related topological adjacency matrix. The formula for calculating the comprehensive topological edge weights is:
[0215] ;
[0216] in, For the first Class and the The comprehensive topological edge weights between the class indicators The median of all stable association strength values. The number of associated mutation segments, Total number of segments The coefficient of the instability penalty factor;
[0217] S305. Perform spectral decomposition on the sparse index association topology adjacency matrix, extract the eigenvectors corresponding to the top K largest eigenvalues, and concatenate the eigenvectors column by column to form the index association topology feature matrix.
[0218] Specifically, the core principle of this embodiment is dynamic temporal correlation topology modeling, which overcomes the shortcomings of traditional static global correlation analysis and achieves accurate characterization of the dynamic correlation relationships of multiple test indicators over time. By calculating local Pearson correlations at each time point, it captures the instantaneous correlation characteristics between indicators, replacing the fixed global correlation coefficient and adapting to the dynamic fluctuation characteristics of human physiological indicators. It utilizes a sliding window to segmentally statistically analyze correlation stability, distinguishing between stable correlation periods and abrupt correlation periods. It combines the median of stable correlations with the proportion of abrupt changes to calculate the comprehensive edge weights, preserving long-term stable physiological correlation patterns while suppressing topological interference caused by instantaneous noise fluctuations. Threshold sparsity removes weakly correlated redundant edges, simplifying the topological structure. Finally, it uses graph spectral decomposition to embed low-dimensional features of the topological structure, transforming the abstract indicator correlation network into a computable and input-model topological feature matrix.
[0219] Example 4
[0220] Please refer to Figure 4 As shown, based on the indicator association topological feature matrix, a time-series evolution network weighted by an attention mechanism is used to extract the dynamic evolution trajectory features of each test indicator, resulting in a time-series evolution feature vector, which specifically includes:
[0221] S401. Take the numerical sequence of each type of test indicator in the standardized time series test dataset at T time points as the time series input of the indicator. Take the row vector corresponding to the type of indicator in the indicator association topological feature matrix as the topological context embedding of the indicator. Concatenate the time series input and the topological context embedding to form a topology-aware time series representation of the indicator. The concatenation formula for the topology-aware time series representation of the indicator is as follows:
[0222] ;
[0223] in, For the first Class indicators in the first Topology-aware representation of each time point This is the time-series input vector for this indicator. The first in the topological feature matrix associated with the index The topological context embedding vector of the row. , This represents a vector concatenation operation;
[0224] S402. Arrange the topology-aware temporal representations of all N types of test indicators according to the time point dimension to form a three-dimensional tensor of shape T×N×D, where D is the feature dimension after splicing.
[0225] S403. Perform multi-head self-attention calculation on the three-dimensional tensor along the time dimension. For each time point, the multi-head self-attention calculation uses the feature vectors of all N-class indicators at that time point as the query vector, and the feature vectors of all indicators at all time points as the key vector and value vector, to calculate the attention weight. The attention weight reflects the degree of mutual influence between different indicators at different time points. The formula for calculating multi-head self-attention is:
[0226] ;
[0227] ;
[0228] in, For the first All time points A query matrix composed of index features. and All Key matrix and value matrix at each time point, For the first Learnable projection matrix for each attention head For the number of attention heads, To output the projection matrix;
[0229] S404. The value vectors are weighted and summed using attention weights to obtain the attention fusion representation at each time point. All T attention fusion representations at each time point are concatenated in chronological order to obtain a preliminary temporal evolution feature sequence. The formula for constructing the preliminary temporal evolution feature sequence is as follows:
[0230] ;
[0231] in, For the first Attention fusion representation at each time point;
[0232] S405. Perform bidirectional recursive encoding on the preliminary temporal evolution feature sequence, extracting dependencies from both the forward and reverse temporal directions. Concatenate the forward and reverse encoding results at each time point to obtain the final temporal evolution feature vector. The concatenation formula for bidirectional recursive encoding is:
[0233] ;
[0234] in, For the first The final temporal evolution feature vector at each time point This is the result of forward recursive encoding. This is the result of reverse recursive encoding. To hide the state dimension.
[0235] Specifically, the core of this embodiment is topology-aware temporal deep feature extraction, which solves the problems of traditional temporal modeling ignoring the physiological correlation between indicators and the single feature representation. By concatenating temporal numerical features with topological correlation features, the temporal changes of each indicator carry the global indicator correlation context, achieving a dual representation of "single indicator temporal evolution + multi-indicator correlation constraints". A multi-head self-attention mechanism is introduced to adaptively capture the long-range dependencies between different time points and different test indicators, automatically strengthening the feature weights of key indicators and key time nodes, and suppressing invalid noise features. Combined with bidirectional recursive encoding, it simultaneously learns the bidirectional temporal dependencies of the past on the future and the future on the past, accurately fitting the continuous evolution law of human test indicators, and outputting a high-dimensional, high-discrimination temporal evolution feature vector.
[0236] Example 5
[0237] Please refer to Figure 5 As shown, the temporal evolution feature vector is input into a multi-granularity attention fusion network for cross-scale feature aggregation to obtain a joint discriminative feature representation, specifically including:
[0238] S501. The time series evolution feature vector is divided into short-range segments, medium-range segments and long-range segments according to different time spans. The short-range segments cover consecutive w1 time points, the medium-range segments cover consecutive w2 time points, and the long-range segments cover consecutive w3 time points, and w1 is less than w2 and less than w3.
[0239] S502. For each granularity of the segment, perform intra-segment mean pooling and intra-segment max pooling to obtain the mean pooling vector and max pooling vector for each granularity. Concatenate the mean pooling vector and max pooling vector for the same granularity to form the two-channel segment representation for that granularity. The formulas for intra-segment mean pooling and max pooling are:
[0240] ;
[0241] ;
[0242] in, and The first Particle size, number The mean pooling vector and the max pooling vector of each segment. Corresponding to short-distance, medium-distance, and long-distance, The length of the fragment at this particle size. For the first A set of time points covered by a segment For the first Temporal evolution feature vectors at each time point;
[0243] S503. Learnable granular attention weights are applied to the two-channel segment representations of all three granularities. These granular attention weights are obtained by inputting the two-channel segment representations of each granularity into a lightweight fully connected network and then normalizing them using softmax. The lightweight fully connected network uses the norm of each granular segment representation as an auxiliary input. The concatenation formula for the two-channel segment representations is as follows:
[0244] ;
[0245] in, For the first Seed particle size Dual-channel segment representation of a segment;
[0246] The formula for calculating granular attention weights is:
[0247] ;
[0248] in, For the first Attention weights at various granularities This represents the average two-channel representation of all segments at this granularity. Its L2 norm, and For the weight matrix and bias vector of a lightweight fully connected network;
[0249] S504. The granularity attention weights are used to perform a weighted summation of the three granularity dual-channel segment representations to obtain the granularity fusion representation. The weighted summation formula for the granularity fusion representation is as follows:
[0250] ;
[0251] in, This represents granularity fusion.
[0252] S505. The granularity fusion representation and the temporal evolution feature vector are concatenated using residual concatenation. The concatenated vector is then subjected to layer normalization and nonlinear activation transformation to obtain the joint discriminative feature representation. The calculation formula for the joint discriminative feature representation is as follows:
[0253] ;
[0254] in, For joint discriminative feature representation, This represents the global mean of the temporal evolution feature vector. This indicates a splicing operation. For layer normalization, It is a non-linear activation function.
[0255] Specifically, the core principle of this embodiment is multi-scale adaptive attention feature fusion, which addresses the technical limitation of single-scale temporal features in simultaneously capturing short-term abrupt fluctuations and long-term chronic evolution. By segmenting time segments into short, medium, and long granularities, it captures the instantaneous abnormal fluctuations, medium-term trends, and long-term steady-state evolution characteristics of the test indicators, respectively. Mean pooling captures the overall horizontal shift, and max pooling captures extreme value abrupt changes, forming a dual-channel complementary representation. A learnable granular attention mechanism is introduced to adaptively allocate weights to features at different time scales, automatically highlighting scale features more critical for judging human physiological states and weakening redundant scale information. Finally, by residual concatenation, normalization, and nonlinear transformation, fine-grained temporal features and multi-scale aggregated features are fused, preserving both shallow details and deep global semantics, outputting highly robust and discriminative joint features, providing accurate feature support for subsequent disease risk assessment and physiological state evaluation.
[0256] Example 6
[0257] Disease risk levels are classified based on the joint discriminant feature representation, and the results of multi-indicator joint analysis of personalized medical tests are output, specifically including:
[0258] The joint discriminant feature representation is input into the hierarchical classifier, which is composed of a first-level classifier head and a second-level classifier head connected in series. The first-level classifier head maps the joint discriminant feature representation to the probability distribution of the major disease category. The second-level classifier head maps the joint discriminant feature representation to the probability distribution of the specific disease subclass under the major disease category determined by the first-level classifier head.
[0259] The first-level classifier calculates the cosine similarity between the joint discriminative feature representation and the preset M disease category prototype vectors. The cosine similarity is then temperature-scaled and input into the softmax function to obtain the disease category probability distribution. The disease category with the highest probability value is selected as the target category. The formulas for calculating the cosine similarity and probability distribution of the first-level classifier are as follows:
[0260] ;
[0261] ;
[0262] in, For joint discriminative feature representation With the Prototype vectors of major disease categories cosine similarity, For the temperature scaling parameters of the first level, For prediction of the first The probability of each major disease category;
[0263] The formula for selecting the target category is:
[0264] ;
[0265] in, Index for the target category;
[0266] The second-level classification head conditionally modulates the joint discriminative feature representation using the prototype vector corresponding to the target major category. It then calculates the cosine similarity between the modulated features and the prototype vectors of the S pre-defined disease subclasses under that major category. After temperature scaling and softmax processing, it obtains the disease subclass probability distribution. The disease subclass with the highest probability value is selected as the target subclass. The probability distribution calculation formula for the second-level classification head is as follows:
[0267] ;
[0268] ;
[0269] in Modulated features With the Major Category prototype vectors of disease subclasses cosine similarity, For the second level of temperature scaling parameters, In the target category Predicted as the first under the condition The probability of each subclass;
[0270] The target major category and target subclass are combined, and their respective probability values and joint discriminant features are attached to the contribution distribution of attention weights at each granularity, which is then output as the result of multi-indicator joint analysis of personalized medicine test.
[0271] Specifically, this embodiment employs a two-level nested hierarchical classification mechanism to achieve progressive and accurate discrimination from coarse disease category classification to fine disease subcategory classification, addressing the shortcomings of traditional single-level classification models, such as coarse disease differentiation and easy confusion of similar diseases. The first level relies on the prototype vector of the disease category to measure feature similarity and optimizes the smoothness of the softmax probability distribution through temperature scaling, avoiding probability polarization or insufficient discrimination, and quickly identifying the disease's category. The second level modulates conditional features based on the category results, eliminating interference from irrelevant category features, and completing refined subcategory discrimination within the defined disease semantic space, conforming to the clinical diagnostic logic of "category classification - subtype subdivision." Simultaneously, it outputs classification probabilities and multi-granularity feature contribution weights, making the results quantifiable and traceable.
[0272] Example 7
[0273] After outputting the results of the multi-indicator joint analysis of personalized medicine testing, the method also includes interpretable reverse tracing analysis of the joint discriminant feature representation, specifically including:
[0274] Based on the classification decision boundary corresponding to the target major class and target subclass output by the hierarchical classifier, the projection component of the joint discriminative feature representation in the direction of the decision boundary normal vector is calculated. Feature dimensions whose absolute value of the projection component is greater than a preset significance threshold are marked as key decision dimensions. The formula for calculating the projection component in the direction of the decision boundary normal vector is as follows:
[0275] ;
[0276] in, For joint discriminative feature representation In the Normal vectors along the decision boundary on each feature dimension The absolute value of the projected component, Target Category With target subclass The corresponding classification decision boundary normal vector;
[0277] The criteria for determining key decision dimensions are as follows:
[0278] ;
[0279] in, A set of indexes for all key decision dimensions. The preset significance threshold;
[0280] For each key decision dimension, we trace back to the corresponding granular segment representation in the multi-granularity attention fusion network to determine which temporal granularity segment primarily contributes to that key decision dimension. The formula for determining the contribution granularity is as follows:
[0281] ;
[0282] in, The index with the highest granularity that contributes most to key decision-making dimensions. For the first In the representation of granular fragments, the first The weight contribution of each dimension;
[0283] For a given contribution granularity segment, we further trace back to the corresponding time point range in the temporal evolution feature vector, extract the attention weight distribution within that time point range, and select the top p indicators with the highest attention weight values as the core driving indicators for this key decision dimension. The selection of core driving indicators is based on the attention weight ranking.
[0284] ;
[0285] in, For the first The set of core driving indicators corresponding to each key decision-making dimension For the first Class indicators in the first Attention weight values at each time point;
[0286] All key decision dimensions and their corresponding core driving indicators, contribution time ranges and contribution granularity types are summarized to generate a key indicator contribution list indexed by indicator name and sorted by contribution intensity.
[0287] Specifically, this embodiment constructs a multi-dimensional chain-based reverse tracing and interpretability mechanism to address the industry pain point of deep learning models' "black box discrimination and inability to explain the etiological basis." Based on the classification decision boundary normal vector projection, it filters key feature dimensions that play a decisive role in disease discrimination, accurately removing invalid and redundant features. Through hierarchical reverse backtracking logic, it sequentially traces the temporal granularity, temporal interval, and core test indicators of feature contributions, establishing a full-link attribution relationship of "classification result - feature dimension - temporal scale - original test indicators." Finally, it quantifies the contribution intensity of each test indicator to disease risk discrimination, forming a structured contribution list that aligns with clinicians' diagnostic needs for interpretable test results.
[0288] Example 8
[0289] Extract the core driving indicators with the highest contribution intensity (q) from the list of key indicators, and check whether the standardized time series test dataset already contains historical test data of the core driving indicators.
[0290] For core driving indicators that already contain historical monitoring data, calculate the deviation between the indicator's value range distribution at its historical monitoring time points and its current value. If the deviation exceeds the individual historical fluctuation range of the indicator, then the indicator is marked as a fluctuation indicator that needs to be closely monitored. The formula for calculating the deviation is:
[0291] ;
[0292] in, For the first The degree of deviation of the core driving indicators. For this indicator at the current time point Standardized values, and These are the mean and standard deviation of the indicator at historical detection time points, respectively;
[0293] The criteria for determining volatility indicators are:
[0294]
[0295] in, A set of volatility indicators, For the first Threshold for the individual historical fluctuation range of the index.
[0296] For core driver metrics that do not include historical testing data, mark the metric as a candidate metric for suggesting new testing.
[0297] The volatility index and candidate indices are sorted in descending order of their contribution intensity in the key index contribution list to generate an individualized supplementary recommendation list. Each entry in the individualized supplementary recommendation list includes the index name, the type of reason for recommendation, and a priority level based on the degree of deviation or contribution intensity. The formula for calculating the priority level is as follows:
[0298] ;
[0299] in, For the first Priority levels of category indicators This represents the contribution intensity value of this indicator in the list of key indicator contributions. This is a set of candidate new indicators.
[0300] Specifically, this embodiment translates model-based discrimination results into clinical testing guidance strategies. The core principle is individualized longitudinal comparison combined with contribution quantification and grading. Unlike traditional general reference standards, this approach constructs an individual baseline distribution based on the patient's historical testing data. It then uses deviations from the normal distribution to quantify abnormal fluctuations in indicators, adapting to the individual differences in patient constitutions. Combining the contribution intensity of indicators obtained from model attribution, it performs dual screening for abnormally fluctuating indicators and missing key indicators, and quantifies testing priorities. This achieves an individualized testing recommendation logic of "high contribution, high fluctuation prioritizes retesting; high contribution, no historical data prioritizes new testing," addressing the problems of homogeneity and lack of specificity in traditional testing protocols.
[0301] Example 9
[0302] After generating the individualized test supplementary recommendation list, the process also includes screening for potential collaborative anomalies in the remaining indicators in the standardized time-series test dataset that were not included in the key indicator contribution list. Specifically, this includes:
[0303] From the indicator association topology feature matrix, extract the associated indicators that are directly connected to any core driving indicator in the key indicator contribution list and whose edge weight is higher than a preset collaboration threshold, to form an associated indicator candidate set. The extraction conditions for the associated indicator candidate set are as follows:
[0304] ;
[0305] in, For the candidate set of related indicators, The set of core driver metrics in the key performance indicator (KPI) contribution list. The first in the topological feature matrix associated with the index Class and the Side weights of class indicators, The preset collaboration threshold;
[0306] For each indicator in the candidate set of related indicators, extract its standardized numerical sequence at all T time points, and calculate the mutual information value between this sequence and the standardized numerical sequence of its corresponding core driving indicator. The mutual information value is obtained by calculating the ratio of the joint probability distribution to the marginal probability distribution after discretizing the two sequences by equal frequency. The formula for calculating the mutual information value is as follows:
[0307] ;
[0308] in, For the first Core-driven indicator sequence With the Class-related indicator sequence Mutual information value between them This is the joint probability distribution after isofrequency discretization. and This represents the corresponding marginal probability distribution;
[0309] The mutual information value is compared with a preset threshold for determining collaborative anomalies. If the mutual information value is higher than the threshold and the direction of change of the value of the indicator in the candidate set of the related indicators is opposite to that of its core driving indicator at the nearest time point, then the indicator is marked as a potential collaborative anomaly indicator. The determination condition for the opposite direction of change is as follows:
[0310] ;
[0311] in, and These are related indicators With core driving indicators In the The change at each point in time;
[0312] All potential collaboration anomaly indicators are sorted in descending order of mutual information value to generate a sorted list of potential collaboration anomaly indicators. The criteria for determining potential collaboration anomaly indicators are as follows:
[0313] ;
[0314] in, A set of potential collaborative anomaly indicators, This is the threshold for determining collaborative anomalies.
[0315] Specifically, this embodiment, based on the topological relationships of physiological indicators, mines implicit synergistic anomalies beyond the primary key indicators, compensating for information omissions caused by relying solely on attribution indicators in a single model. It utilizes the previously constructed indicator topological weights to screen strongly correlated indicators, and quantifies the nonlinear correlation between two sets of time series through mutual information, overcoming the limitations of traditional linear correlation. Combined with the inverse constraint of the time series change direction, it screens out potential abnormal indicators with "high correlation and inverse fluctuations," corresponding to pairs of physiological indicators that mutually restrain and antagonize each other in clinical practice. This achieves comprehensive risk screening from "explicit key indicator discrimination" to "implicit synergistic anomaly mining," significantly improving the comprehensiveness of laboratory analysis.
[0316] Example 10
[0317] After generating the potential collaborative anomaly index ranking list, the process also includes consistency verification and analysis of the output results of the hierarchical classifier and the potential collaborative anomaly index ranking list, specifically including:
[0318] Read the target major category and target subclass output by the hierarchical classifier, and retrieve the set of known biomarkers associated with the target subclass from the preset disease knowledge graph;
[0319] The intersection of the indicators in the potential synergistic anomaly ranking list with the known biomarker set is calculated. If the intersection is not empty, the consistency score of the intersection indicator is obtained by multiplying its ranking weight in the ranking list with its prior importance score in the known biomarker set. The intersection calculation formula is as follows:
[0320] ;
[0321] in, For the set of intersecting indicators, A sorted list of potential collaboration anomalies. It is the set of known biomarkers retrieved from a disease knowledge graph;
[0322] The formula for calculating the consistency score is:
[0323] ;
[0324] in, Intersection Indicators Consistency score, This refers to the ranking weight of this indicator in the potential collaborative anomaly indicator ranking list. This is the prior importance score of the indicator within the known set of biomarkers;
[0325] If the consistency score is higher than the preset high consistency threshold, the original output of the hierarchical classifier will remain unchanged.
[0326] If the consistency score is lower than the preset low consistency threshold, the highest-ranking indicator in the potential collaboration anomaly indicator ranking list that belongs to the intersection will be used as the correction trigger indicator. The selection formula for the correction trigger indicator is as follows:
[0327] ;
[0328] in, To correct the trigger indicator index;
[0329] Feature shift compensation for the joint discriminant feature representation is performed using the temporal patterns of the modified triggering index in the standardized time-series test dataset. Specifically, the directional difference between the temporal pattern vector of the modified triggering index and the joint discriminant feature representation is calculated, and the directional difference is proportionally superimposed onto the joint discriminant feature representation to obtain the modified joint discriminant feature representation. The temporal pattern vector of the modified triggering index is:
[0330] ;
[0331] in, For the first Time-series pattern vector of class indicators, For this indicator in the 1st Topology-aware representation of each time point;
[0332] The formulas for compensating for orientation difference and feature offset are as follows:
[0333] ;
[0334] ;
[0335] in, To correct the orthogonal direction difference components between the trigger indicator time-series pattern vector and the joint discriminative feature representation, For compensation ratio coefficient, This is the modified joint discriminant feature representation;
[0336] The corrected joint discriminant feature representation is re-inputted into the hierarchical classifier to obtain the corrected disease risk hierarchical classification result. The corrected disease risk hierarchical classification result replaces the original output result as the final result of the multi-indicator joint analysis of individualized medical tests.
[0337] Specifically, the implementation plan constructs a closed-loop result correction mechanism driven by both "model prediction" and "medical prior knowledge" to address the issues of pure data-driven models being susceptible to noise interference and having discrimination bias. By introducing clinical prior biomarkers through a disease knowledge graph, potential abnormal indicators mined by the model are medically consistent with the data, quantifying the degree of matching between data features and clinical common sense. For biased results with low consistency, the global discrimination features are offset compensated based on the temporal distribution characteristics of highly reliable collaborative abnormal indicators, correcting model feature bias errors and iterating again to output accurate results. This achieves a deep integration of data patterns and medical knowledge, greatly improving the reliability and clinical consistency of disease discrimination results.
[0338] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.
Claims
1. An AI-driven method for joint analysis of multiple indicators in medical laboratory tests, characterized in that: include: Obtain multiple batches of raw medical test data for the individual to be analyzed, and perform cross-batch time-varying drift correction and outlier filtering on the multiple batches of raw medical test data to obtain a standardized time-series test dataset. A dynamic correlation topology among multiple indicators is constructed on the standardized time-series test dataset to generate an indicator correlation topology feature matrix. Based on the topological feature matrix of the index association, the dynamic evolution trajectory features of each test index are extracted using a time-series evolution network weighted by the attention mechanism, and the time-series evolution feature vector is obtained. The temporal evolution feature vector is input into a multi-granularity attention fusion network for cross-scale feature aggregation to obtain a joint discriminative feature representation; Disease risk levels are classified based on the joint discriminant feature representation, and the results of multi-indicator joint analysis of individualized medical tests are output.
2. The AI-driven multi-indicator joint analysis method for medical testing according to claim 1, characterized in that, The process of acquiring multiple batches of raw medical test data for the individual to be analyzed, and performing cross-batch time-varying drift correction and outlier filtering on the multiple batches of raw medical test data to obtain a standardized time-series test dataset, specifically includes: Raw values of multiple test items, including blood biochemistry, immunology, hematology and urine analysis, are extracted from the medical records of the same individual to be analyzed, forming a raw test matrix covering N types of test indicators and spanning T time points; For each type of test index in the original test matrix, calculate the local rate of change of the index between two adjacent time points, compare the local rate of change with the preset physiological fluctuation threshold range of the index, and if the local rate of change exceeds the physiological fluctuation threshold range, mark the index value at that time point as a candidate abnormal point. For each type of test index in the original test matrix, a locally weighted regression curve is established with time point as independent variable and the value of the index as dependent variable. The fitting residual value of the locally weighted regression curve at each time point is extracted, and data points whose fitting residual value exceeds the upper limit of the residual quantile of the index are marked as statistical outliers. Merge candidate outliers and statistical outliers to obtain a set of all outlier data points, and replace the indicator values corresponding to the set of all outlier data points with the weighted interpolation results of the indicator at the effective time points before and after it. To address the systematic shifts between different batches of tests caused by instrument model changes or reagent batch differences, this study uses two batches of data from the same individual within an overlapping time window as a benchmark. The median deviation and interquartile range scaling factor for each indicator between the two batches are calculated. The median deviation is used to perform translation correction on the data of the latter batch, and the interquartile range scaling factor is used to perform scaling correction on the data of the latter batch. The drift correction is completed iteratively batch by batch to complete the drift correction for all batches, and finally a standardized time-series test dataset is obtained.
3. The AI-driven multi-indicator joint analysis method for medical testing according to claim 2, characterized in that, The step of constructing a dynamic correlation topology among multiple indicators in the standardized time-series test dataset to generate an indicator correlation topology feature matrix specifically includes: Using N types of test indicators in the standardized time series test dataset as nodes, calculate the time-by-time Pearson dynamic correlation coefficient sequence of any two types of indicators at all T time points. The Pearson dynamic correlation coefficient sequence is composed of the local correlation coefficient values at T time points. The Pearson dynamic correlation coefficient sequence is segmented into sliding windows, with each segment consisting of L consecutive time points. The mean and variance of the correlation coefficient values within each segment are calculated. If the variance of a segment is lower than the preset dynamic correlation stability threshold, the mean of that segment is taken as the stable correlation strength value of the two types of indicators within that time period; otherwise, the segment is marked as a correlation mutation segment. The distribution of stable correlation strength values of the two types of indicators is statistically analyzed over the entire time period. The median of stable correlation strength values is used as the basic edge weight, and the proportion of correlation mutation segments is used as the instability penalty factor for the edge weight. The comprehensive topological edge weight between the two types of indicators is then calculated. Using the comprehensive topological edge weights of all pairwise index pairs as elements, construct an N×N symmetric weight matrix. Perform threshold filtering on the symmetric weight matrix to retain edges with weight values higher than the preset topological sparsity threshold, and obtain the sparsified index-related topological adjacency matrix. The spectral decomposition of the sparse index association topology adjacency matrix is performed to extract the eigenvectors corresponding to the top K largest eigenvalues. The eigenvectors are then concatenated column by column to form the index association topology feature matrix.
4. The AI-driven multi-indicator joint analysis method for medical testing according to claim 3, characterized in that, The method, based on the indicator-related topological feature matrix, utilizes an attention-weighted temporal evolution network to extract the dynamic evolution trajectory features of each test indicator, resulting in a temporal evolution feature vector. Specifically, this includes: The numerical sequence of each test index at T time points in the standardized time series test dataset is used as the time series input of the index. The row vector corresponding to the index in the index association topology feature matrix is used as the topology context embedding of the index. The time series input and the topology context embedding are concatenated to form a topology-aware time series representation of the index. Arrange the topology-aware temporal representations of all N types of test indicators along the time point dimension to form a three-dimensional tensor of shape T×N×D, where D is the feature dimension after splicing. Multi-head self-attention computation is performed on the three-dimensional tensor along the time point dimension. For each time point, the multi-head self-attention computation uses the feature vectors of all N-type indicators at that time point as the query vector, and the feature vectors of indicators at all time points as the key vector and value vector to calculate the attention weight. The attention fusion representation at each time point is obtained by weighted summation of the value vector using attention weights. The attention fusion representations at all T time points are concatenated in chronological order to obtain a preliminary temporal evolution feature sequence. The preliminary temporal evolution feature sequence is bidirectionally recursively encoded, and the dependencies are extracted from both the forward and reverse temporal directions. The forward encoding results and the reverse encoding results are concatenated at each time point to obtain the final temporal evolution feature vector.
5. The AI-driven multi-indicator joint analysis method for medical testing according to claim 4, characterized in that, The step of inputting the temporal evolution feature vector into a multi-granularity attention fusion network for cross-scale feature aggregation to obtain a joint discriminative feature representation specifically includes: The temporal evolution feature vector is segmented into short-range segments, medium-range segments, and long-range segments according to different time spans. The short-range segments cover consecutive w1 time points, the medium-range segments cover consecutive w2 time points, and the long-range segments cover consecutive w3 time points, with w1 less than w2 less than w3. For each granularity of the segment, mean pooling and max pooling within the segment are performed separately to obtain mean pooling vector and max pooling vector for each granularity. The mean pooling vector and max pooling vector for the same granularity are concatenated to form a two-channel segment representation for that granularity. Learnable granular attention weights are applied to the two-channel segment representations of all three granularities. The granular attention weights are obtained by inputting the two-channel segment representations of each granularity into the lightweight fully connected network and normalizing them with softmax. The lightweight fully connected network uses the norm of each granular segment representation as an auxiliary input. By using granular attention weights to perform a weighted summation of the two-channel segment representations at three different granularities, a granular fusion representation is obtained. The granular fusion representation and the temporal evolution feature vector are concatenated by residual connection. The concatenated vector is then subjected to layer normalization and nonlinear activation transformation to obtain the joint discriminative feature representation.
6. The AI-driven multi-indicator joint analysis method for medical testing according to claim 5, characterized in that, The process of classifying disease risk levels based on joint discriminant feature representation and outputting personalized medical test results with multi-indicator joint analysis specifically includes: The joint discriminant feature representation is input into the hierarchical classifier, which is composed of a first-level classifier head and a second-level classifier head connected in series. The first-level classifier head maps the joint discriminant feature representation to the probability distribution of the major disease category. The second-level classifier head maps the joint discriminant feature representation to the probability distribution of the specific disease subclass under the major disease category determined by the first-level classifier head. The first-level classifier calculates the cosine similarity between the joint discriminative feature representation and the preset M disease category prototype vectors. The cosine similarity is then temperature-scaled and input into the softmax function to obtain the disease category probability distribution. The disease category with the highest probability value is selected as the target category. The formulas for calculating the cosine similarity and probability distribution of the first-level classifier are as follows: ; ; in, For joint discriminative feature representation With the Prototype vectors of major disease categories cosine similarity, For the temperature scaling parameters of the first level, For prediction of the first The probability of each major disease category; The formula for selecting the target category is: ; in, Index for the target category; The second-level classification head conditionally modulates the joint discriminative feature representation using the prototype vector corresponding to the target major category. It then calculates the cosine similarity between the modulated features and the prototype vectors of the S pre-defined disease subclasses under that major category. After temperature scaling and softmax processing, it obtains the disease subclass probability distribution. The disease subclass with the highest probability value is selected as the target subclass. The probability distribution calculation formula for the second-level classification head is as follows: ; ; in Modulated features With the Major category prototype vectors of disease subclasses cosine similarity, For the second level of temperature scaling parameters, In the target category Predicted as the first under the condition The probability of each subclass; The target major category and target subclass are combined, and their respective probability values and joint discriminant features are attached to the contribution distribution of attention weights at each granularity, which is then output as the result of multi-indicator joint analysis of personalized medicine test.
7. The AI-driven multi-indicator joint analysis method for medical testing according to claim 6, characterized in that, Following the output of the multi-indicator joint analysis results for personalized medicine testing, the method also includes interpretable reverse tracing analysis of the joint discriminant feature representation, specifically including: Based on the classification decision boundary corresponding to the target major class and target subclass output by the hierarchical classifier, the projection component of the joint discriminative feature representation in the direction of the decision boundary normal vector is calculated. Feature dimensions whose absolute value of the projection component is greater than a preset significance threshold are marked as key decision dimensions. The formula for calculating the projection component in the direction of the decision boundary normal vector is as follows: ; in, For joint discriminative feature representation In the Normal vectors along the decision boundary on each feature dimension The absolute value of the projected component, Target Category With target subclass The corresponding classification decision boundary normal vector; The criteria for determining key decision dimensions are as follows: ; in, A set of indexes for all key decision dimensions. The preset significance threshold; For each key decision dimension, we trace back to the corresponding granular segment representation in the multi-granularity attention fusion network to determine which temporal granularity segment primarily contributes to that key decision dimension. The formula for determining the contribution granularity is as follows: ; in, The index with the highest granularity that contributes most to key decision-making dimensions. For the first In the representation of granular fragments, the first The weight contribution of each dimension; For a given contribution granularity segment, we further trace back to the corresponding time point range in the temporal evolution feature vector, extract the attention weight distribution within that time point range, and select the top p indicators with the highest attention weight values as the core driving indicators for this key decision dimension. The selection of core driving indicators is based on the attention weight ranking. ; in, For the first The set of core driving indicators corresponding to each key decision-making dimension For the first Class indicators in the first Attention weight values at each time point; All key decision dimensions and their corresponding core driving indicators, contribution time ranges and contribution granularity types are summarized to generate a key indicator contribution list indexed by indicator name and sorted by contribution intensity.
8. The AI-driven multi-indicator joint analysis method for medical testing according to claim 7, characterized in that, Following the generation of the key indicator contribution list, the process also includes individualized verification and supplementary recommendations based on the key indicator contribution list, specifically including: Extract the core driving indicators with the highest contribution intensity (q) from the list of key indicators, and check whether the standardized time series test dataset already contains historical test data of the core driving indicators. For core driving indicators that already contain historical monitoring data, calculate the deviation between the indicator's value range distribution at its historical monitoring time points and its current value. If the deviation exceeds the individual historical fluctuation range of the indicator, then the indicator is marked as a fluctuation indicator that needs to be closely monitored. The formula for calculating the deviation is: ; in, For the first The degree of deviation of the core driving indicators. For this indicator at the current time point Standardized values, and These are the mean and standard deviation of the indicator at historical testing points, respectively. The criteria for determining volatility indicators are: in, A set of volatility indicators, For the first Threshold for the individual historical fluctuation range of a category indicator; For core driver metrics that do not include historical testing data, mark the metric as a candidate metric for suggesting new testing. The volatility index and candidate indices are sorted in descending order of their contribution intensity in the key index contribution list to generate an individualized supplementary recommendation list. Each entry in the individualized supplementary recommendation list includes the index name, the type of reason for recommendation, and a priority level based on the degree of deviation or contribution intensity. The formula for calculating the priority level is as follows: ; in, For the first Priority levels of category indicators This represents the contribution intensity value of this indicator in the list of key indicator contributions. This is a set of candidate new indicators.
9. The AI-driven multi-indicator joint analysis method for medical testing according to claim 8, characterized in that, After generating the individualized test supplementary recommendation list, the process also includes screening for potential collaborative anomalies in the remaining indicators in the standardized time-series test dataset that were not included in the key indicator contribution list. Specifically, this includes: From the indicator association topology feature matrix, extract the associated indicators that are directly connected to any core driving indicator in the key indicator contribution list and whose edge weight is higher than a preset collaboration threshold, to form an associated indicator candidate set. The extraction conditions for the associated indicator candidate set are as follows: ; in, For the candidate set of related indicators, The set of core driver metrics in the key performance indicator (KPI) contribution list. The first in the topological feature matrix associated with the index Class and the Side weights of class indicators, The preset collaboration threshold; For each indicator in the candidate set of related indicators, extract its standardized numerical sequence at all T time points, and calculate the mutual information value between this sequence and the standardized numerical sequence of its corresponding core driving indicator. The mutual information value is obtained by calculating the ratio of the joint probability distribution to the marginal probability distribution after discretizing the two sequences by equal frequency. The formula for calculating the mutual information value is as follows: ; in, For the first Core Driver Indicator Sequence With the Class-related indicator sequence Mutual information value between them This is the joint probability distribution after isofrequency discretization. and This represents the corresponding marginal probability distribution; The mutual information value is compared with a preset threshold for determining collaborative anomalies. If the mutual information value is higher than the threshold and the direction of change of the value of the indicator in the candidate set of the related indicators is opposite to that of its core driving indicator at the nearest time point, then the indicator is marked as a potential collaborative anomaly indicator. The determination condition for the opposite direction of change is as follows: ; in, and These are related indicators With core driving indicators In the The change at each point in time; All potential collaboration anomaly indicators are sorted in descending order of mutual information value to generate a sorted list of potential collaboration anomaly indicators. The criteria for determining potential collaboration anomaly indicators are as follows: ; in, A set of potential collaborative anomaly indicators, This is the threshold for determining collaborative anomalies.
10. The AI-driven multi-indicator joint analysis method for medical testing according to claim 9, characterized in that, After generating the potential collaborative anomaly index ranking list, the process also includes consistency verification and analysis result correction between the output of the hierarchical classifier and the potential collaborative anomaly index ranking list, specifically including: Read the target major category and target subclass output by the hierarchical classifier, and retrieve the set of known biomarkers associated with the target subclass from the preset disease knowledge graph; The intersection of the indicators in the potential synergistic anomaly ranking list with the known biomarker set is calculated. If the intersection is not empty, the consistency score of the intersection indicator is obtained by multiplying its ranking weight in the ranking list with its prior importance score in the known biomarker set. The intersection calculation formula is as follows: ; in, For the set of intersecting indicators, A sorted list of potential collaboration anomalies. It is the set of known biomarkers retrieved from a disease knowledge graph; The formula for calculating the consistency score is: ; in, Intersection Indicators Consistency score, This refers to the ranking weight of this indicator in the potential collaborative anomaly indicator ranking list. This is the prior importance score of the indicator within the known set of biomarkers; If the consistency score is higher than the preset high consistency threshold, the original output of the hierarchical classifier will remain unchanged. If the consistency score is lower than the preset low consistency threshold, the highest-ranking indicator in the potential collaboration anomaly indicator ranking list that belongs to the intersection will be used as the correction trigger indicator. The selection formula for the correction trigger indicator is as follows: ; in, To correct the trigger indicator index; Feature shift compensation for the joint discriminant feature representation is performed using the temporal patterns of the modified triggering index in the standardized time-series test dataset. Specifically, the directional difference between the temporal pattern vector of the modified triggering index and the joint discriminant feature representation is calculated, and the directional difference is proportionally superimposed onto the joint discriminant feature representation to obtain the modified joint discriminant feature representation. The temporal pattern vector of the modified triggering index is: ; in, For the first Time-series pattern vector of class indicators, For this indicator in the 1st Topology-aware representation of each time point; The formulas for compensating for orientation difference and feature offset are as follows: ; ; in, To correct the orthogonal direction difference components between the trigger indicator time-series pattern vector and the joint discriminative feature representation, For compensation ratio coefficient, This is the modified joint discriminant feature representation; The corrected joint discriminant feature representation is re-inputted into the hierarchical classifier to obtain the corrected disease risk hierarchical classification result. The corrected disease risk hierarchical classification result replaces the original output result as the final result of the multi-indicator joint analysis of individualized medical tests.
