AI screening method and device for AD marker of urine exosome

Abstract

The invention discloses an AI screening method and device for a urine exosome AD marker. The method comprises the following steps: acquiring a urine exosome, extracting proteome, miRNA group and metabolite group data, acquiring time sequence data of historical sampling of a detection target, and extracting dynamic change characteristics of the AD marker; processing data of the proteome, the miRNA group and the metabolite group to obtain a protein-miRNA embedded vector and a cross-modal feature vector, processing the dynamic change feature to obtain a time sequence dynamic feature vector, and fusing the time sequence dynamic feature vector into a multi-modal feature vector; performing KL divergence detection on the multi-modal feature vector; and after the KL divergence detection is passed, analyzing the multi-modal feature vector to output key marker data, analyzing the contribution degree of a key marker, and outputting a screening result. According to the scheme, the AD marker can be accurately screened, the model can be dynamically updated, and the method has the advantages of high sensitivity, high specificity detection and low detection cost.

Description

Technical Field

[0001] This invention relates to the field of urine exosome AD detection technology, and in particular to an AI screening method, device, medical testing terminal, and readable storage medium for urine exosome AD markers. Background Technology

[0002] Alzheimer's disease (AD) is a common neurodegenerative disease, and early diagnosis is crucial for slowing disease progression. Traditional diagnostic methods rely on clinical symptoms and neuroimaging examinations, but these methods have limitations. In recent years, urinary exosomes have attracted attention as a potential source of biomarkers. Urinary exosomes are nanoscale vesicles (30-150 nm) secreted by cells in the urinary system, carrying biomolecules such as proteins, nucleic acids (e.g., miRNA, lncRNA), lipids, and metabolites. Studies have shown that brain-derived exosomes from Alzheimer's disease (AD) patients can cross the blood-brain barrier into the bloodstream and are eventually filtered into the urine by the kidneys. Their contents, such as phosphorylated Tau protein and Aβ oligomers, can directly reflect pathological features in the brain, making urinary exosomes an ideal target for non-invasive diagnosis of AD. However, with the popularization of technologies such as proteomics and transcriptomics, exosome multi-omics data has experienced explosive growth. Traditional single-marker screening methods that rely on experimental validation (such as ELISA targeted quantification) are inefficient and struggle to uncover multi-molecule synergistic signals. Meanwhile, conventional bioinformatics tools such as differential expression analysis and enrichment analysis, limited to single-modality data processing and suffering from high false-positive rates, cannot meet the demand for high-precision AD marker discovery. Against this backdrop, artificial intelligence technology, by integrating multi-omics data, simulating molecular interaction networks, and automatically extracting nonlinear correlation features, can significantly improve the sensitivity and specificity of marker screening, becoming a key technological means to advance the early warning window for AD in my country.

[0003] The current mainstream technology in the field of non-invasive Alzheimer's disease screening is the ELISA-based diagnostic method for Alzheimer's disease by detecting p-tau181 in urinary exosomes. This method uses ultracentrifugation to separate exosomes from urine samples, resuspends them in PBS buffer to obtain purified exosomes, and then uses enzyme-linked immunosorbent assay (ELISA) to quantitatively detect phosphorylated tau protein (p-tau181) in the exosome lysate. Finally, the individual's AD risk level is determined based on clinically validated thresholds. This technology uses urine as the non-invasive biological sample source and establishes a complete experimental pathway from exosome isolation to target protein detection, providing an operational methodological framework for early screening of neurodegenerative diseases based on body fluid biomarkers.

[0004] The aforementioned method suffers from three core drawbacks: First, relying on a single biomarker, p-tau181, leads to cross-reactions between antibodies and non-target proteins due to the complex composition of urine matrix (containing high concentrations of urea, salts, and metabolic debris) and the extremely low abundance of p-tau181 in exosomes (<0.01% of total protein). Clinical validation shows a false positive rate as high as 23.2%. Second, the entire process is complex, and differences in operator skill introduce significant batch-to-batch variation. We need to establish a new urine self-test targeting method. Third, the biomarker is singular; relying solely on p-tau181 concentration thresholds to determine disease status fails to capture multi-pathological features of Alzheimer's disease (such as abnormal miRNA regulation and metabolic disorders), resulting in a detection rate of less than 40% for patients with early-stage mild cognitive impairment. These drawbacks collectively limit the potential application of this method in precision medicine and dynamic monitoring of Alzheimer's disease.

[0005] In view of this, it is necessary to propose improvements to the diagnostic method for Alzheimer's disease using urinary exosomes p-tau181. Summary of the Invention

[0006] To address at least one of the aforementioned technical problems, the main objective of this invention is to provide an AI screening method and apparatus for urinary exosome AD markers.

[0007] To achieve the above objectives, one technical solution adopted by the present invention is: providing an AI screening method for urinary exosome AD markers, comprising:

[0008] Urine exosomes from urine samples of the target to be detected are obtained, and proteomic, miRNA, and metabolomic data are extracted from the urine exosomes to be detected. Time-series data of the target to be detected in historical sampling are also obtained, and dynamic change characteristics of AD markers are extracted from the time-series data. The proteomic, miRNA, and metabolomic data are related to AD markers.

[0009] The proteome, miRNAome, and metabolome data were processed to obtain protein-miRNA embedding vectors and cross-modal feature vectors, and the dynamic change features were processed to obtain time-series dynamic feature vectors. The protein-miRNA embedding vectors, cross-modal feature vectors, and time-series dynamic feature vectors were fused to obtain multimodal feature vectors.

[0010] Perform KL divergence detection on multimodal feature vectors;

[0011] After the multimodal feature vectors are detected by KL divergence, the multimodal feature vectors are analyzed to output key marker data. The contribution of key markers is analyzed using SHAP values, and the screening results of AD markers are output.

[0012] The processing of proteomic, miRNA, and metabolomic data to obtain protein-miRNA embedding vectors and cross-modal feature vectors includes:

[0013] By fusing proteomics and miRNAiome data, protein-miRNA interaction network data was obtained. A graph neural network model was then used to calculate the protein-miRNA embedding vector from the protein-miRNA interaction network data.

[0014] The protein-miRNA embedding vector is mapped to metabolomics data to obtain the mapped feature vector, and the cross-modal attention model is used to calculate the cross-modal feature vector.

[0015] Wherein, the dynamic change feature is a temporal fusion feature vector, and the process of processing the dynamic change feature to obtain the temporal dynamic feature vector includes:

[0016] Projecting and dimensionality reduction onto the temporal fusion feature vector yields a dynamic feature vector of a preset dimension;

[0017] Spatiotemporal attention is calculated on dynamic feature vectors in a preset dimension to output temporal dynamic feature vectors.

[0018] The process of extracting proteomic, miRNA, and metabolomic data from the exosomes in the urine to be tested further includes:

[0019] The concentrations of the proteome and metabolome were corrected using a urinary creatinine correction method, yielding corrected concentration data for both.

[0020] The same amount of standard reference RNA was added to the urine sample to correct the miRNA group data.

[0021] The KL divergence detection of the multimodal feature vectors includes:

[0022] Detecting the difference in probability distribution between current data and historical data in multimodal feature vectors based on the KL divergence formula;

[0023] If the KL divergence value of the multimodal feature vector is greater than the set value, then update the parameters of the top-level classifier and use the composite loss function to balance the current data with the historical data.

[0024] The set value is 0.3, and the top-level classifier includes a top-level classifier for a cross-modal attention model and a graph neural network model.

[0025] Specifically, the process of parsing multimodal feature vectors outputs key biomarker data, analyzes the contribution of key biomarkers using SHAP values, and outputs the screening results for AD biomarkers, including:

[0026] The attention scores of proteins, miRNAs, and metabolites are analyzed using the attention mechanism, and key biomarker combinations are selected based on the scores.

[0027] The contribution of each key biomarker data is quantified based on the SHAP algorithm, and the screening results of AD biomarkers are output.

[0028] The extraction of proteomic, miRNA, and metabolomic data from the exosomes in the urine sample includes:

[0029] Proteomic data were extracted using liquid chromatography-tandem mass spectrometry.

[0030] Using sequencing technology corrected for unique molecular identifiers to screen differentially expressed miRNA genome data related to blood-brain barrier function; and

[0031] Metabolomic data closely related to blood-brain barrier permeability were extracted using an ultra-high performance liquid chromatography-quadrupole time-of-flight mass spectrometry platform.

[0032] To achieve the above objectives, one technical solution adopted by the present invention is: providing an AI screening device for urinary exosome AD markers, comprising:

[0033] The extraction module is used to acquire urinary exosomes from urine samples of the target to be detected, and extract proteomic, miRNA, and metabolomic data from the urinary exosomes to be detected, as well as acquire time-series data of the target in historical sampling, and extract dynamic change characteristics of AD markers from the time-series data, wherein the proteomic, miRNA, and metabolomic data are related to AD markers;

[0034] The fusion module is used to process proteomic, miRNA, and metabolomic data to obtain protein-miRNA embedding vectors and cross-modal feature vectors, as well as to process dynamic change features to obtain time-series dynamic feature vectors. The protein-miRNA embedding vectors, cross-modal feature vectors, and time-series dynamic feature vectors are fused to obtain multimodal feature vectors.

[0035] The detection module is used to perform KL divergence detection on multimodal feature vectors;

[0036] The parsing module is used to parse the multimodal feature vectors after they have passed the KL divergence test, output key marker data, analyze the contribution of key markers using SHAP values, and output the screening results of AD markers.

[0037] To achieve the above objectives, another technical solution adopted by the present invention is to provide a medical testing terminal, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the steps in the above method.

[0038] To achieve the above objectives, another technical solution adopted by the present invention is to provide a readable storage medium on which a computer program is stored, wherein when the computer program is executed by a processor, the steps in the above method are implemented.

[0039] The technical solution of this invention mainly employs linear acquisition of urine exosomes from urine samples of the target to be detected, and extracts proteome, miRNA, and metabolome data from the exosomes to be detected. It also acquires time-series data of the target in historical sampling, and extracts dynamic change features of AD biomarkers from the time-series data. The proteome, miRNA, and metabolome data are then processed to obtain protein-miRNA embedding vectors and cross-modal feature vectors. The dynamic change features are further processed to obtain time-series dynamic feature vectors. These vectors are then fused to obtain multimodal feature vectors. KL divergence detection is performed on the multimodal feature vectors. Finally, after the multimodal feature vectors pass KL divergence detection, the key biomarker data is output from the multimodal feature vectors. The contribution of the key biomarkers is analyzed using SHAP values, and the screening results of AD biomarkers are output. This approach integrates multimodal data from the proteome, miRNA, and metabolome of urinary exosomes, along with historical clinical sampling data. Based on a dynamic model using the Transformer architecture and incremental learning technology, and combined with SHAP value analysis to quantify the contribution of biomarkers, it outputs screening results for AD biomarkers. This approach achieves highly sensitive and specific accurate detection. In addition, this approach also has the advantages of low detection cost, high detection efficiency, and dynamically updated detection model. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0041] Figure 1 This is a flowchart of an AI screening method for urinary exosome AD markers according to an embodiment of the present invention;

[0042] Figure 2This is a flowchart illustrating the AI ​​screening method for urinary exosome AD markers according to an embodiment of the present invention;

[0043] Figure 3 This is a block diagram of an AI screening device for urinary exosome AD markers according to an embodiment of the present invention;

[0044] Figure 4 This is a block diagram of a medical testing terminal according to an embodiment of the present invention.

[0045] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0047] It should be noted that the descriptions involving "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature. Furthermore, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. When the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed by this invention.

[0048] Unlike existing ELISA-based methods for detecting p-tau181 in urinary exosomes for Alzheimer's disease, which rely on single protein biomarkers, suffer from static threshold models that cannot adapt to individual differences, lack the ability to integrate multi-omics data, and have insufficient clinical interpretability, this approach proposes an AI-based screening method for urinary exosome AD biomarkers. This method aims to provide a low-cost, dynamically updated approach with a clear pathological mechanism correlation for accurate AD screening. The specific steps of this AI-based screening method for urinary exosome AD biomarkers are detailed in the following embodiments.

[0049] Please refer to Figure 1 and Figure 2 , Figure 1 This is a flowchart of an AI screening method for urinary exosome AD markers according to an embodiment of the present invention; Figure 2This is a flowchart illustrating the AI ​​screening method for urinary exosome AD biomarkers according to an embodiment of the present invention. In this embodiment, the AI ​​screening method for urinary exosome AD biomarkers is applied in precision medicine. The AI ​​screening of urinary exosome AD biomarkers includes the following steps:

[0050] S110. Obtain urine exosomes from urine samples of the target to be detected, and extract proteome, miRNA and metabolome data from the urine exosomes to be detected, and obtain time-series data of the target in historical sampling, and extract dynamic change characteristics of AD markers from the time-series data, wherein the proteome, miRNA and metabolome data are related to AD markers.

[0051] In this embodiment, the target of detection is the human body, and the urine sample of the target is obtained by ultracentrifugation. The urine sample is placed at 4°C and centrifuged at 1000g for 10 min to discard cell debris precipitate; the supernatant is collected and placed at 4°C and centrifuged at 17,000g for 15 min to discard large vesicle precipitate; the supernatant is collected and transferred to a clean dedicated ultracentrifuge tube and centrifuged at 4°C and 200,000g for 1 h; 90% of the supernatant is discarded, and about 1 mL of supernatant is retained for resuspending the precipitate. The resuspended solution is transferred to a new ultracentrifuge tube and centrifuged again at 4°C and 200,000g for 1 h; the supernatant is discarded, and 100 μL of sterile 1xPBS is added to each ultracentrifuge tube and vortexed to resuspend the precipitate to obtain exosomes. Exosomes were isolated and lysed, and a multi-omics screening system combining proteomics, miRNAiomics, and metabolomics was used to extract proteomic, miRNAiomics, and metabolomics data to comprehensively capture AD biomarker characteristics.

[0052] S120. Process the proteome, miRNA, and metabolome data to obtain protein-miRNA embedding vectors and cross-modal feature vectors, and process the dynamic change features to obtain time-series dynamic feature vectors. Fuse the protein-miRNA embedding vectors, cross-modal feature vectors, and time-series dynamic feature vectors to obtain multimodal feature vectors.

[0053] In this step, three types of data—protein-miRNA embedding vectors, cross-modal feature vectors, and temporal dynamic feature vectors—are integrated through a multimodal AI model to construct multimodal feature vectors. This enables the discovery of cross-omics synergistic biomarker combinations, overcoming the limitations of single known targets. miRNAs (microRNAs) are a class of non-coding single-stranded RNA molecules, approximately 22 nucleotides in length, encoded by endogenous genes. They participate in post-transcriptional gene expression regulation in plants and animals.

[0054] S130. Perform KL divergence detection on the multimodal feature vectors. KL divergence detection can measure the difference between the current data and historical data, which helps to reduce detection errors.

[0055] S140. After the multimodal feature vectors pass the KL divergence test, the key marker data is output by parsing the multimodal feature vectors. The contribution of the key markers is analyzed using SHAP values, and the selection results of AD markers are output. In this step, if the KL divergence test fails, the parameters of the top-level classifier of the model need to be updated, and the KL divergence test needs to be performed again until it passes before parsing the key marker data. After the key marker data is quantified by SHAP values, the AD marker detection results can be output. The AD marker detection results can be output as an electronic document or a paper report. SHAP values ​​(SHapley Additive ex Planations) are a tool used to interpret the predictions of machine learning models. Based on the Shapley value theory in game theory, it aims to assign an "importance" score to each feature, which reflects the magnitude of the feature's contribution to the model's prediction.

[0056] In one specific embodiment, the processing of proteomic, miRNA, and metabolomic data to obtain protein-miRNA embedding vectors and cross-modal feature vectors includes:

[0057] By fusing proteomics and miRNAiome data, protein-miRNA interaction network data was obtained. A graph neural network model was then used to calculate the protein-miRNA embedding vector from the protein-miRNA interaction network data.

[0058] The protein-miRNA embedding vector is mapped to metabolomics data to obtain the mapped feature vector, and the cross-modal attention model is used to calculate the cross-modal feature vector.

[0059] In this step, an interaction graph between the proteome and miRNA group is constructed based on the STRING database, i.e., interaction network data. This addresses the problem that traditional machine learning methods such as logistic regression, which rely on manual feature engineering, cannot capture the functional synergy between proteins and miRNAs. The interaction network data is then used to create a graph neural network (GNN) that updates the feature representation through neighbor node aggregation, as shown in the following formula:

[0060]

[0061] in, Let W be the feature vector of node v at layer l, N(v) be the set of neighbors of node v (defined by the STRING database), and W be the feature vector of node v at layer l. (l) Let be the trainable weight matrix of the l-th layer, and σ be the nonlinear activation function that enhances feature representation. Aggregate features for neighboring nodes. (This involves) X... * ij Inputting the corrected protein concentration, site status, and miRNA expression level into the above formula will output a protein-miRNA functional co-embedding vector (e.g., miR-132-3p downregulates Tau phosphorylase PPP2R5B, indirectly increasing p-Tau181 concentration).

[0062] Because the protein-miRNA embedding vector (256-dimensional) and metabolite features (300-dimensional) are in different dimensions and distribution spaces, their correlation cannot be directly calculated. Therefore, a learnable weight matrix W is used instead. Q W K W V By mapping these three elements to a unified semantic space, data from different omics systems are "translated" into a common, conversational language. The specific mapping formula is as follows:

[0063] Q = X 嵌入 ·W Q (W Q ∈R 256×128 )

[0064] K = X 代谢物 ·W K (W K ∈R 300×128 V = X 代谢物 ·W V (W V ∈R 300×128 )

[0065] Here, X represents the input feature, and W is the projection matrix, which is essentially a "feature importance filter" that automatically learns which features are crucial for AD prediction during training. The above formula independently and linearly maps protein, miRNA, and metabolite data to generate Q, K, and V. The protein-Q question vector carries "pathological feature questions" (e.g., "[Is p-Tau81 abnormality related to metabolism?]"), the miRNA-K clue vector provides "regulatory signal answers" (e.g., "[Sphingolipid pathway activity = 0.85, glycolysis activity = 0.2]"), and the metabolite-V evidence vector contributes "microenvironment information" (e.g., C16-sphingosine = 0.61, lactate = 0.03]).

[0066] In specific testing, key indicators will be dynamically focused on (e.g., p-Tau181 first, then combined with sphingolipid metabolism), rather than treating all data equally. This scheme simulates this process using Q / K / V triples and uses the Transformer attention mechanism to calculate joint weights, achieving cross-modal attention fusion. The formula is as follows:

[0067]

[0068] Among them, QK T To measure similarity, the matching degree between the question vector Q and the clue vector K is calculated. A higher matching degree indicates that the question vector Q and the clue vector K are highly correlated.

[0069] For example:

[0070]

[0071] Scaling To prevent gradient explosion, Softmax normalization transforms similarity into a probability distribution, d k This represents the dimension of the clue vector. An example is shown below:

[0072]

[0073] The model allocates 82% of its attention to sphingolipid metabolism and 18% to glycolysis. Finally, a weighted V-value is used to generate fusion features, revealing the regulatory strength of metabolites on the molecular synergistic network with quantifiable weights. This overcomes the bottleneck of traditional methods in analyzing the "metabolism-protein-RNA" cascade effect, providing a target basis for precise intervention. The output data is as follows:

[0074] Output = 0.85 × V 鞘氨醇 +0.18×V 乳酸 = Fusion Vector

[0075] Among them, [p-Tau181-sphingolipid synergistic strength = 0.82, other metabolic contributions = 0.18].

[0076] In a specific embodiment, the dynamic change feature is a temporal fusion feature vector, and the process of processing the dynamic change feature to obtain the temporal dynamic feature vector includes:

[0077] Projecting and dimensionality reduction onto the temporal fusion feature vector yields a dynamic feature vector of a preset dimension;

[0078] Spatiotemporal attention is calculated on dynamic feature vectors in a preset dimension to output temporal dynamic feature vectors.

[0079] In this step, time-series data from multiple samplings of the same detection target are input into TimeSformer to extract dynamic change features such as the slope and variance of biomarker concentrations. This further quantifies the rate of progression from mild cognitive impairment to Alzheimer's disease (AD) (e.g., sphingolipid metabolite increase → miR-132-3p decrease → p-Tau181 accumulation → accelerated AD conversion). Dynamic weights reveal the "sphingolipid metabolism → miRNA → protein" transmission chain (e.g., sphingosine weight → miR-132-3p weight → p-Tau181 weight = 0.38 → 0.42 → 0.45), providing dynamic evidence for early intervention. Specifically, the input data is a temporal fusion feature vector (256-dimensional) sampled multiple times from the same detection target. The model first reduces the dimensionality to 128-dimensional using linear projection, and then performs spatiotemporal attention calculations: temporal attention observes the changing trend of each biomarker (e.g., the monthly growth rate of p-Tau181), and spatial attention analyzes the dynamic correlation between different biomarkers (e.g., the synchronicity between the increase of sphingosine and the decrease of miR-132-3p). Finally, it outputs 128-dimensional dynamic features, which include quantitative indicators such as slope and variance, and also encode the co-evolutionary pattern across biomarkers. The final multimodal feature vector (640-dimensional) is composed of three parts: protein-miRNA interaction features extracted by GNN (256-dimensional), cross-modal fusion features of Transformer (256-dimensional), and dynamic features of TimeSformer (128-dimensional).

[0080] In one specific embodiment, after extracting proteomic, miRNA, and metabolomic data from the exosomes in the urine to be tested, the method further includes:

[0081] The concentrations of the proteome and metabolome were corrected using a urinary creatinine correction method, yielding corrected concentration data for both.

[0082] The same amount of standard reference RNA was added to the urine sample to correct the miRNA group data.

[0083] In this embodiment, urine sample concentrations may fluctuate due to individual differences in water intake. A urine creatinine correction method is used to correct for protein and metabolite concentrations, eliminating the influence of urine concentration differences on quantitative results and ensuring the biological significance of the data is accurate and reliable. The calculation formula is:

[0084]

[0085] In addition, in miRNA detection, by adding the same amount of standard reference RNA (cel-miR-39) to each urine sample, the technical error caused by the different sequencing data volume is corrected, so that the miRNA detection results of different samples can be directly compared.

[0086] Furthermore, technical biases introduced by different experimental batches (e.g., different dates, testing centers, reagent kits) can mask true biological differences. The ComBat algorithm is used to adjust the data distribution across different experimental batches, providing standardized input to the model and preventing misjudgments due to systematic bias. The ComBat algorithm formula is:

[0087]

[0088] Among them, X ij μ is the original detection value of the i-th sample in the j-th batch. batch Let σ be the mean of all samples in the j-th batch. batch Let μ be the standard deviation of all samples in the j-th batch. global σ is the mean of all batches of samples globally. global Let X be the standard deviation of all batches of samples globally. * ij This is the corrected detection value.

[0089] In a specific embodiment, the KL divergence detection of the multimodal feature vectors includes:

[0090] Detecting the difference in probability distribution between current data and historical data in multimodal feature vectors based on the KL divergence formula;

[0091] If the KL divergence value of the multimodal feature vector is greater than the set value, then update the parameters of the top-level classifier and use the composite loss function to balance the current data with the historical data.

[0092] The set value is 0.3, and the top-level classifier includes a top-level classifier for a cross-modal attention model and a graph neural network model.

[0093] In this step, considering the data distribution shift caused by the high baseline biomarker concentration of APOEε4 exosomes and the difference in the progression rate of different AD biomarker subtypes, the KL divergence is used to measure the difference in probability distribution between the new data and historical data. The specific formula is as follows:

[0094]

[0095] KL divergence is based on the probability distribution difference of complete feature vectors, rather than the inter-modal vector difference, where P is the distribution of the new data (e.g., a newly added cohort of young AD patients with an average age of 50 years), and Q is the distribution of the old model training data (e.g., a cohort of elderly AD patients). If the KL divergence > 0.3 (e.g., in a newly added cohort of young AD patients, the KL divergence detects a difference in the dynamic slope of the markers), a model update is triggered to allow the model to continuously adapt to the new data. At this time, the underlying weights of GNN and Transformer are fixed to retain the general feature extraction capability (retaining knowledge of elderly AD), and only the parameters of the top-level classifier are updated to adapt to the progression rate of the younger population, while avoiding the loss of historical knowledge. A composite loss function is used to balance the new and old knowledge, as shown in the following formula:

[0096]

[0097] in, Cross-entropy loss is used to ensure that new data is classified correctly; Constrain the consistency of the output distribution between the old and new models to prevent abrupt changes in prediction; Parameter smoothing regularization limits the magnitude of parameter changes and prevents excessive deviation from historical knowledge. A composite loss function ensures model stability during updates and reduces the false positive rate.

[0098] In one specific embodiment, the step of parsing the multimodal feature vector outputs key marker data, using SHAP values ​​to analyze the contribution of key markers, and outputting the screening results of AD markers includes:

[0099] The attention scores of proteins, miRNAs, and metabolites are analyzed using the attention mechanism, and key biomarker combinations are selected based on the scores.

[0100] The contribution of each key biomarker data is quantified based on the SHAP algorithm, and the screening results of AD biomarkers are output.

[0101] In this step, attention scores of proteins, miRNAs, and metabolites are analyzed using an attention mechanism to screen for the top three contributing key biomarker combinations, including phosphorylated Tau protein, miR-132-3p, and sphingolipid metabolites. The contribution of each feature to the prediction results is quantified based on the SHAP algorithm, using the following formula:

[0102]

[0103] Where F represents the set of all features, S is a subset excluding feature i, and f(S) is the model's prediction on subset S. A SHAP value > 0 indicates that the feature improves the AD risk score, while a SHAP value < 0 indicates that the feature reduces the risk. The algebraic sum of the SHAP values ​​of multiple features directly determines the final risk score. Based on the SHAP value, the contribution of key biomarkers is quantified to generate an AD risk score of 0-100. A score ≥ 75 triggers a high-risk warning. The trend of biomarker concentration over time is displayed simultaneously, and its correlation with the AD pathological mechanism is indicated.

[0104] In addition, the output of AD biomarker screening results can be an AD biomarker risk report, which can be used to further determine the risk assessment and trend prediction of AD for the detection target.

[0105] Specifically, the extraction of proteomic, miRNA, and metabolomic data from the exosomes of the urine to be tested includes:

[0106] Proteomic data were extracted using liquid chromatography-tandem mass spectrometry.

[0107] Using sequencing technology corrected for unique molecular identifiers to screen differentially expressed miRNA genome data related to blood-brain barrier function; and

[0108] Metabolomic data closely related to blood-brain barrier permeability were extracted using an ultra-high performance liquid chromatography-quadrupole time-of-flight mass spectrometry platform.

[0109] In this step, liquid chromatography-tandem mass spectrometry (LC-MS / MS) was used to quantitatively analyze over 2000 proteins at a detection sensitivity of 1 pg / mL, covering a broad range of targets including Aβ42 and p-Tau181, yielding proteomic data. For miRNA profiling, next-generation sequencing (NGS) technology with unique molecular identifier (UMI) correction was employed to effectively eliminate polymerase chain reaction (PCR) amplification bias and accurately screen differentially expressed miRNAs related to blood-brain barrier function. Polymerase chain reaction (PCR) is a molecular biology technique used to amplify specific DNA fragments; it can be viewed as a special form of DNA replication outside the organism. The most significant characteristic of PCR is its ability to significantly increase minute amounts of DNA. Metabolomics analysis was performed using an ultra-high performance liquid chromatography-quadrupole time-of-flight mass spectrometry (UPLC-Q-TOF) platform to detect over 500 metabolites, focusing on changes in metabolic pathways closely related to blood-brain barrier permeability, such as sphingolipids, yielding metabolomic data.

[0110] In summary, this approach innovatively integrates multi-omics data (proteins, miRNAs, metabolites) from urinary exosomes with clinical information to construct an AI model based on the Transformer architecture and dynamic incremental learning technology for screening AD biomarkers. Specifically, it comprehensively captures AD biomarker characteristics by isolating exosomes from urine and employing a multi-omics joint screening system; it uses urinary creatinine correction and the ComBat algorithm to process data, eliminating individual differences and batch effects; it achieves cross-modal data fusion through graph neural networks (GNNs) and the Transformer attention mechanism, and extracts dynamic change features of biomarkers by combining time-series Transformers; and it introduces a dynamic incremental learning mechanism, enabling the model to adapt to new data in real time without retraining, significantly improving the model's adaptability and efficiency while avoiding false positives; finally, it analyzes the contribution of key biomarkers using SHAP values ​​to generate AD risk scores and dynamic monitoring reports. Furthermore, this approach uses a non-invasive urine sample testing method, significantly reducing testing costs and time while improving diagnostic accuracy and reliability.

[0111] Compared with existing technologies, this invention significantly improves the sensitivity and specificity of detection, and its accuracy is far higher than traditional methods. Simultaneously, this solution employs a non-invasive urine sample detection method, which significantly reduces detection costs and shortens detection time compared to traditional PET imaging methods. Furthermore, the AI ​​screening method of this solution adapts to new data in real time through a dynamic incremental learning mechanism, eliminating the need for retraining and significantly reducing computational resource consumption. It also analyzes the contribution of key biomarkers through SHAP values, providing interpretable evidence for clinical diagnosis. In large-scale screening experiments, this invention can complete the detection and analysis of the target in a short time, and its simple operation makes it easy to promote in medical institutions at all levels.

[0112] Please refer to Figure 3 , Figure 3 This is a block diagram of an AI screening device for urinary exosome AD markers according to an embodiment of the present invention. In this embodiment, the AI ​​screening device for urinary exosome AD markers includes:

[0113] Extraction module 110 is used to acquire urine exosomes from urine samples of the target to be detected, and extract proteome, miRNA and metabolome data from the urine exosomes to be detected, as well as acquire time series data of the target in historical sampling, and extract dynamic change features of AD markers from the time series data, wherein the proteome, miRNA and metabolome data are related to AD markers;

[0114] The fusion module 120 is used to process proteome, miRNA, and metabolome data to obtain protein-miRNA embedding vectors and cross-modal feature vectors, as well as to process dynamic change features to obtain time-series dynamic feature vectors. The protein-miRNA embedding vectors, cross-modal feature vectors, and time-series dynamic feature vectors are fused to obtain multimodal feature vectors.

[0115] Detection module 130 is used to perform KL divergence detection on multimodal feature vectors;

[0116] The parsing module 140 is used to parse the multimodal feature vectors after they have passed the KL divergence test, output key marker data, analyze the contribution of key markers using SHAP values, and output the screening results of AD markers.

[0117] Please see Figure 4 , Figure 4 This is a block diagram of a medical testing terminal according to an embodiment of the present invention. This medical testing terminal can be used to implement the AI ​​screening method for urinary exosome AD markers in the aforementioned embodiments. Figure 4 As shown, the medical testing terminal includes a memory 301, a processor 302, a bus 303, and a computer program stored in the memory 301 and executable on the processor 302. The memory 301 and the processor 302 are connected via the bus 303. When the processor 302 executes the computer program, it implements the functions described in the previous embodiment. The number of processors can be one or more.

[0118] The memory 301 can be a high-speed random access memory (RAM) or a non-volatile memory, such as a disk storage device. The memory 301 is used to store executable program code, and the processor 302 is coupled to the memory 301.

[0119] Furthermore, embodiments of this application also provide a computer-readable storage medium, which may be disposed in the medical testing terminal of the above embodiments, and the computer-readable storage medium may be the aforementioned... Figure 4 The memory in the illustrated embodiment.

[0120] The computer-readable storage medium stores a computer program that, when executed by a processor, implements the AI ​​screening method for urinary exosome AD markers described in the foregoing embodiments. Furthermore, the computer-readable storage medium can also be a USB flash drive, external hard drive, read-only memory (ROM), RAM, magnetic disk, or optical disk, or any other medium capable of storing program code.

[0121] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0122] The modules described as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0123] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0124] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned readable storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0125] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0126] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0127] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. An AI screening method for urinary exosome AD markers, characterized in that, include: Urine exosomes from urine samples of the target to be detected are obtained, and proteomic, miRNA, and metabolomic data are extracted from the urine exosomes to be detected. Time-series data of the target to be detected in historical sampling are also obtained, and dynamic change characteristics of AD markers are extracted from the time-series data. The proteomic, miRNA, and metabolomic data are related to AD markers. The proteome, miRNAome, and metabolome data were processed to obtain protein-miRNA embedding vectors and cross-modal feature vectors, and the dynamic change features were processed to obtain time-series dynamic feature vectors. The protein-miRNA embedding vectors, cross-modal feature vectors, and time-series dynamic feature vectors were fused to obtain multimodal feature vectors. Perform KL divergence detection on multimodal feature vectors; After the multimodal feature vectors are detected by KL divergence, the multimodal feature vectors are analyzed to output key marker data. The contribution of key markers is analyzed using SHAP values, and the screening results of AD markers are output.

2. The AI ​​screening method for urinary exosome AD markers as described in claim 1, characterized in that, The processing of proteomic, miRNA, and metabolomic data yields protein-miRNA embedding vectors and cross-modal feature vectors, including: By fusing proteomics and miRNAiome data, protein-miRNA interaction network data was obtained. A graph neural network model was then used to calculate the protein-miRNA embedding vector from the protein-miRNA interaction network data. The protein-miRNA embedding vector is mapped to metabolomics data to obtain the mapped feature vector, and the cross-modal attention model is used to calculate the cross-modal feature vector.

3. The AI ​​screening method for urinary exosome AD markers as described in claim 2, characterized in that, The dynamic change feature is a temporal fusion feature vector. The process of processing the dynamic change feature to obtain the temporal dynamic feature vector includes: Projecting and dimensionality reduction onto the temporal fusion feature vector yields a dynamic feature vector of a preset dimension; Spatiotemporal attention is calculated on dynamic feature vectors in a preset dimension to output temporal dynamic feature vectors.

4. The AI ​​screening method for urinary exosome AD markers as described in claim 1, characterized in that, After extracting proteomic, miRNA, and metabolomic data from the exosomes in the urine to be tested, the process also includes: The concentrations of the proteome and metabolome were corrected using a urinary creatinine correction method, yielding corrected concentration data for both. The same amount of standard reference RNA was added to the urine sample to correct the miRNA group data.

5. The AI ​​screening method for urinary exosome AD markers as described in claim 2, characterized in that, The KL divergence detection of the multimodal feature vectors includes: Detecting the difference in probability distribution between current data and historical data in multimodal feature vectors based on the KL divergence formula; If the KL divergence value of the multimodal feature vector is greater than the set value, then update the parameters of the top-level classifier and use the composite loss function to balance the current data with the historical data. The set value is 0.3, and the top-level classifier includes a top-level classifier for a cross-modal attention model and a graph neural network model.

6. The AI ​​screening method for urinary exosome AD markers as described in claim 2, characterized in that, The process involves analyzing multimodal feature vectors to output key biomarker data, using SHAP values ​​to analyze the contribution of key biomarkers, and outputting the selection results for AD biomarkers, including: The attention scores of proteins, miRNAs, and metabolites are analyzed using the attention mechanism, and key biomarker combinations are selected based on the scores. The contribution of each key biomarker data is quantified based on the SHAP algorithm, and the screening results of AD biomarkers are output.

7. The AI ​​screening method for urinary exosome AD markers as described in claim 1, characterized in that, The data on proteomics, miRNAs, and metabolites extracted from the exosomes of the urine sample to be tested include: Proteomic data were extracted using liquid chromatography-tandem mass spectrometry. Using sequencing technology corrected for unique molecular identifiers to screen differentially expressed miRNA genome data related to blood-brain barrier function; and Metabolomic data closely related to blood-brain barrier permeability were extracted using an ultra-high performance liquid chromatography-quadrupole time-of-flight mass spectrometry platform.

8. An AI screening device for urinary exosome AD markers, characterized in that, include: The extraction module is used to acquire urinary exosomes from urine samples of the target to be detected, and extract proteomic, miRNA, and metabolomic data from the urinary exosomes to be detected, as well as acquire time-series data of the target in historical sampling, and extract dynamic change characteristics of AD markers from the time-series data, wherein the proteomic, miRNA, and metabolomic data are related to AD markers; The fusion module is used to process proteomic, miRNA, and metabolomic data to obtain protein-miRNA embedding vectors and cross-modal feature vectors, as well as to process dynamic change features to obtain time-series dynamic feature vectors. The protein-miRNA embedding vectors, cross-modal feature vectors, and time-series dynamic feature vectors are fused to obtain multimodal feature vectors. The detection module is used to perform KL divergence detection on multimodal feature vectors; The parsing module is used to parse the multimodal feature vectors after they have passed the KL divergence test, output key marker data, analyze the contribution of key markers using SHAP values, and output the screening results of AD markers.

9. A medical testing terminal, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements the steps of the method according to any one of claims 1-7.

10. A readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-7.

Images