An exosome acting path recognition system and method of a smoking inducer fusion

Abstract

The present application relates to the field of healthcare informatics, and discloses an exosome action path recognition system and method fusing smoking inducement factors, comprising: constructing an initial mixed feature model, and calculating a control group residual signal of a non-smoker control group; identifying and excluding contaminated individuals based on the control group residual signal to form a pure control group; constructing a purified mixed feature model based on the pure control group, purifying the signal of a smoker intervention group using the purified mixed feature model, identifying an exosome action path, solving the problem of control group data label contamination through an iterative workflow, obtaining active identification and self-correction ability for label noise, and avoiding analysis failure caused by model contamination.

Description

Technical Field

[0001] This invention relates to an exosome action pathway identification system and method that integrates smoking-inducing factors, belonging to the field of healthcare informatics technology. Background Technology

[0002] One of the core tasks of current computational information processing systems for pathological mechanism identification and attribution analysis of exosome omics data is to extract reliable and interpretable computational models of pathological mechanisms from high-dimensional, high-noise medical and health data, such as multi-omics data and electronic medical records. To this end, various computational analysis tools have been developed in this field, such as differential expression analysis and pathway enrichment analysis. These tools have become the mainstream techniques in current pathological mechanism research and have demonstrated their application value in controlled experimental data. However, with the rise of real-world data and large-scale population cohort studies, the data reality faced by healthcare informatics has changed. Data no longer originates from highly controlled environments but comes from the real world full of confounding variables, such as age, body mass index, individual comorbidities, or chronic inflammation levels. In these real-world data, the target variable used for analysis and these confounding variables are often highly collinear computationally and have a common impact on phenotypic changes in downstream omics data.

[0003] This data reality presents technical challenges to existing mainstream computational tools in the field, such as the enrichment analysis mentioned above. These tools generally rely on the implicit assumption that the target variable used for grouping is the only or at least the dominant signal driver in the dataset. When applied to real-world data full of confounding variables, they cannot computationally distinguish whether phenotypic changes in the data come from the target signal or from confounding signals that are highly correlated with the target, such as aging or chronic inflammation. Existing information processing methods contaminate these signals from different sources, ultimately outputting mixed and unreliable action paths, leading to systematic bias and computational failure in healthcare informatics' core task of etiological attribution.

[0004] This attribution bias caused by confounding signals is prevalent in current pathological mechanism research. It is particularly relevant for screening biomarkers targeting specific causes such as smoking. For example, Chinese invention patent CN118910043A discloses circRNA biomarkers and applications for exosome transport inducing bladder cancer. This involves detecting circRNAs in bladder cancer tissue and adjacent tissues from patients with a history of smoking, and performing differential analysis to screen biomarkers such as hsa_circ_0055033. However, this traditional calculation method, relying on direct comparison between the intervention and control groups, cannot address the aforementioned signal collinearity problem. Calculatedly, it is difficult to distinguish whether the expression differences of the screened circRNAs (such as hsa_circ_0055033) are caused by the target signal of smoking or by confounding signals commonly found in smokers, such as advanced age, chronic inflammation, or high BMI. This limitation in the technical approach leads to inherent uncertainty in the attribution of the screened biomarkers, and the identified smoking-induced pathways are likely contaminated by signals from confounding factors.

[0005] Therefore, the technical problem to be solved by this invention is how to construct a novel computing architecture for healthcare informatics to solve the problem of computational decoupling and active separation between target signals and mixed signals, and to identify highly reliable and specific action paths in real-world data full of mixed variables. Summary of the Invention

[0006] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A system for identifying the exosome action pathway of smoking-inducing factors, comprising a hybrid feature modeling unit, a signal purification calculation unit, and a residual signal path identification unit:

[0007] The confounding feature modeling unit is used to construct an initial confounding feature model based on the confounding factor dataset and exosome omics dataset from a cohort of non-smoker control groups.

[0008] The signal purification unit is used to apply the initial confounding feature model to calculate the residual control signal of each non-smoker control group cohort; based on the residual control signal, it identifies contaminated individuals, excludes contaminated individuals from the non-smoker control group cohort, and forms clean control group data.

[0009] The confounding feature modeling unit is used to construct a purified confounding feature model based on pure control group data;

[0010] The signal purification operation unit is used to purify the confounding feature model, obtain the exosome omics dataset and confounding factor dataset of the smoker intervention group cohort, calculate the predicted confounding exosome spectrum, and generate the residual signal dataset.

[0011] The residual signal path identification unit is used to perform correlation analysis based on the residual signal dataset and the pure control group data, and output the exosome action path.

[0012] Preferably, the construction operation in the confounding feature modeling unit includes: based on the key confounding factors in the confounding factor dataset of the non-smoker control group, computationally dividing the non-smoker control group into multiple data sub-layers; and independently constructing the corresponding confounding feature model within each data sub-layer.

[0013] Preferably, the initial mixed feature model constructed by the mixed feature modeling unit and the purified mixed feature model are multiple linear regression models, ridge regression models, principal component regression models, or partial least squares regression models.

[0014] Preferably, the confounding factor dataset includes at least one or more of age, body mass index (BMI), and C-reactive protein (CRP) levels.

[0015] Preferably, the operation of the signal purification unit to generate the residual signal dataset is implemented by performing the following vector subtraction operation on each individual in the smoker intervention group cohort within the vector space defined by the exosome data: ,in, For any residual signal vector in the residual signal dataset, This represents the actual signal vector in the exosome omics dataset. This represents the predicted contamination vector in the predicted contamination profile of exosomes.

[0016] Preferably, the signal purification operation unit is used to apply the purified confounding feature model to the new non-smoker cohort to calculate the residual signal of the respective new cohort control group; calculate the dynamic bias vector based on the residual signal of the new cohort control group; the operation of the signal purification operation unit to generate the residual signal dataset is limited to subtracting the sum of the predicted confounding exosome spectrum and the dynamic bias vector from the exosome omics dataset.

[0017] Preferably, the system includes a systematic signal-to-noise ratio (SNR) arbitration gating; the systematic SNR arbitration gating is used to perform a global statistical test on the residual signal dataset output by the signal purification operation unit and calculate the systematic signal index; the systematic signal index is compared with a minimum signal threshold; when the systematic signal index is greater than the minimum signal threshold, the execution of the residual signal path identification unit is triggered, and the operation of the signal purification operation unit to identify contaminated individuals is based on the fact that the value of the residual signal of the control group exceeds a preset statistical threshold determined according to the statistical distribution.

[0018] Preferably, the operation of the confounding feature modeling unit to segment the non-smoker control group cohort into multiple data sub-layers is based on a preset clinical threshold or quantile of key confounding factors to segment the data sub-layers.

[0019] Preferably, the operation of the signal purification computing unit for calculating the predicted confounding exosome profile includes: inputting individual data from the confounding factor dataset of the smoker intervention group cohort into the purified confounding feature model, obtaining the output of the purified confounding feature model as the predicted confounding exosome profile.

[0020] A method for identifying the action pathway of exosomes fused with smoking-inducing factors, used to implement a system for identifying the action pathway of exosomes fused with smoking-inducing factors, includes the following steps:

[0021] Step a: Construct an initial confounding feature model based on the confounding factor dataset and exosome omics dataset from the non-smoker control group cohort;

[0022] Step b: Apply the initial confounding feature model to calculate the control group residual signal for each of the non-smoker control group cohorts;

[0023] Step c: Based on the residual signal of the control group, identify contaminated individuals and exclude contaminated individuals from the non-smoker control group cohort to form pure control group data;

[0024] Step d: Based on the data from the pure control group, construct a purified confounding feature model;

[0025] Step e: Using the purified confounding feature model, obtain the exosome omics dataset and confounding factor dataset of the smoker intervention group cohort, calculate the predicted confounding exosome spectrum, and generate the residual signal dataset;

[0026] Step f involves performing a correlation analysis between the residual signal dataset and the pure control group data to output the exosome action pathway.

[0027] Compared with the prior art, the beneficial effects of the present invention are:

[0028] 1. By collaborating with the specific information flow of the confounding feature modeling unit and the signal purification operation unit, a new medical information processing method is constructed. The confounding feature modeling unit learns and constructs a confounding feature model representing the mapping relationship between confounding factors such as age and exosome omics only in the pure control group data environment of non-smokers, where there is no interference from the target signal of smoking. This counterintuitive data flow design enables the constructed model to encapsulate the true computational signature of confounding factors in an unbiased manner in informatics. Subsequently, the signal purification operation unit uses this unbiased model as a high-precision computational tool to apply to the smoker intervention group, predicting and actively stripping the signal components contributed by confounding factors from the total signal, and creating a new and purer data entity in computation, namely the residual signal dataset. This architecture of first isolating the model and then separating it from other groups fundamentally solves the bottleneck of attribution failure caused by the high collinearity between the target signal and the confounding signal in medical informatics.

[0029] 2. By introducing a nonlinear factor hierarchical preprocessing step before the execution of the confounding feature modeling unit, and coordinating it with the execution logic of the signal purification operation unit, the secondary bottleneck of underfitting of the model and incomplete signal stripping caused by the nonlinear relationship between confounding factors and omics data in real-world medical data is solved. Without relying on complex nonlinear algorithms, the confounding feature modeling unit is instructed to build a local simple confounding feature model within each data sub-layer through the information processing architecture of data hierarchical modeling. When processing individual smokers, the signal purification operation unit will determine the data sub-layer to which they belong and call the local model corresponding to that layer to calculate their predicted confounding spectrum. This collaborative mechanism of hierarchical modeling and hierarchical purification reduces the complex global nonlinear fitting problem to multiple parallel local linear fitting problems in terms of computation. While maintaining the high interpretability required by medical informatics, it makes the stripping of informatics subtraction more accurate and fully improves the purity of the final residual signal.

[0030] 3. By constructing an iterative model refinement workflow, we address the core engineering challenge of control group data label contamination in medical informatics when processing real-world data. For example, the non-smoker cohort may contain passive smokers or false reporters. In the first round of execution, the signal purification unit is also applied to the non-smoker control group itself, and the output residual signal is used as a data quality diagnostic indicator. Based on this indicator, contaminated control group individuals are identified and labeled. In the second round of execution, the contamination feature modeling unit is instructed to use the clean control group data excluding contaminated individuals to construct a purified contamination feature model. This iterative information processing closed loop of diagnosis, purification, and remodeling reuses the function of the signal purification unit from an analysis tool to an input data quality control tool, gaining the ability to actively identify and self-correct label noise, and avoiding the failure of the final analysis results due to model contamination. Attached Figure Description

[0031] Figure 1 This is a complete flowchart of the iterative model refinement and signal recognition process of this invention;

[0032] Figure 2 This is a schematic diagram illustrating the effect of the informatics subtraction method for separating mixed signals according to the present invention;

[0033] Figure 3 This is a schematic diagram of the logic unit and hardware deployment architecture of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are only some embodiments of this invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0035] This invention provides a system and method for identifying exosome action pathways incorporating smoking-inducing factors. In terms of information processing architecture, it mainly includes a confounding feature modeling unit, a signal purification and computation unit, and a residual signal pathway identification unit. Through an iterative workflow, the confounding feature modeling unit constructs an initial model using only data from a non-smoker control group. The signal purification and computation unit applies this model to diagnose and remove contaminated individuals from the control group. The confounding feature modeling unit then constructs a purified model based on the pure control group. The signal purification and computation unit uses this purified model to calculate and remove confounding signal components from the signals of the smoker intervention group, generating a residual signal dataset. The residual signal pathway identification unit analyzes this purified data entity to identify high-confidence exosomes. The system's operational path; in specific system deployments, the confounding feature modeling unit, signal purification operation unit, and residual signal path identification unit are all logical modules running on computing devices with specific specifications. The computing devices need to have sufficient memory and processors to process high-dimensional medical datasets, such as floating-point operation units capable of handling vector operations with dimensions up to tens of thousands of dimensions. This addresses the challenges of healthcare informatics, such as the problem of control group data label contamination in real-world cohort studies. Inevitably, non-smoker cohorts contain individuals who smoke or make false reports. Directly using this contaminated data for modeling leads to the confounding feature model incorrectly learning smoking signals as part of the confounding signals. Over-stripping occurs when purifying smoker signals, resulting in the failure of false negative analysis.

[0036] To address this challenge, an iterative information processing closed loop for diagnostic purification and remodeling is implemented, with the following steps: Before acquiring the exosome omics dataset, standardized preprocessing is performed on the raw high-dimensional omics data. This preprocessing applies an inter-group normalization method, such as quantile normalization, to the omics data vector of each organism to correct for non-biological technical biases between different samples. Missing value identification is then performed on the normalized data matrix. For any identified missing data points, a deterministic imputation algorithm, such as the K-nearest neighbor imputation algorithm, is applied to fill the missing data points with the weighted average of the K nearest neighbors in the feature space. This generates a numerically complete and comparable exosome omics dataset, which is then used by the confounding feature modeling unit to construct the initial confounding feature model. The confounding feature modeling unit acquires the confounding factor dataset and exosome omics dataset from the non-smoker control group cohort. The confounding factor dataset is a low-dimensional vector representing an individual's baseline status, including at least key confounding factors recognized in medical informatics, such as age, body mass index (BMI), and C-reactive protein (CRP) levels. The exosome omics dataset is a high-dimensional vector representing the exosome expression profile. Based on this data, the confounding feature modeling unit constructs an initial confounding feature model. In medical informatics applications, a model with high interpretability and computational stability is preferred, such as a multiple linear regression model, ridge regression model, principal component regression model, or partial least squares regression model. Taking the ridge regression model as an example, the unit solves the minimization equation to establish the mapping relationship between confounding factors and exosome profiles. ,in, For exosome spectrum, As a confounding factor, Let be the regression coefficients to be solved. The error term includes an L2 regularization term to prevent overfitting in high-dimensional data; the initial confounding feature model constructed by the signal purification computation unit calculates the predicted confounding exosome profile for each individual in the non-smoker control group cohort. Perform vector subtraction The residual signals of their respective control groups were obtained. The signal purification processing unit identifies contaminated individuals based on residual signals from the control group. The procedure is based on the following technical fact: true non-smokers. It should be close to zero (random noise), while contaminating the individual. This will deviate from zero. To quantify and automate this judgment, this unit calculates the magnitude of the residual signal vector for each individual control group. An optional calculation method is its L2 norm. ,in, The residual signal amplitude, This unit represents the L2 norm. It calculates the statistical distribution of these amplitudes across the entire control group and determines a preset statistical threshold, which can be set to the median of all amplitudes plus three times the median absolute deviation (MAD). , To preset statistical thresholds, The median of the amplitude. This unit represents the absolute deviation of the median. Individuals labeled as contaminated were excluded from the non-smoker control group cohort, resulting in a pure control group with higher data quality. The confounding feature modeling unit was then invoked again, with input limited to the resulting pure control group data, to construct a purified confounding feature model, eliminating the interference of label noise. It can more unbiasedly encapsulate the true calculated signature of confounding factors such as age itself.

[0037] In obtaining the purified confounding feature model Afterwards, the system enters the main signal purification and recognition stage, solving the attribution failure problem caused by the high collinearity between the target signal and the mixed signals in this field. The specific calculation steps are as follows: The signal purification operation unit obtains the exosome data set of the smoker intervention group cohort. Using a purified confounding feature model with a confounding factor dataset. For each individual in the smoker intervention group, a purification operation is performed. This operation first takes the individual data from the confounding factor dataset of the smoker intervention group cohort, such as data of a 65-year-old smoker with a BMI of 28, and inputs it into a purified confounding feature model. The output of the purified confounding feature model is then obtained, which is the predicted confounding exosome profile for this individual. ,this In medical informatics, a vector represents the expected exosome profile of a person assuming they are a pure nonsmoker, given confounding factors such as age 65 and BMI of 28. The signal purification unit, within the vector space defined by the exosome data, performs the following vector subtraction operation on each individual in the smoker intervention cohort: ,in, For any residual signal vector in the residual signal dataset, This represents the actual signal vector in the exosome omics dataset. To predict the confounding vector in the predicted confounding exosome spectrum, this informatics subtraction actively removes signal components contributed by confounding factors. The result... This creates a new, purer data entity, the residual signal dataset, which is computationally largely isolated from interference from confounding signals and primarily enriches signals specifically contributed by smoking-inducing factors. The residual signal path identification unit is then activated. This unit performs correlation analysis between the residual signal dataset and the pure control group data, employing standard bioinformatics analysis procedures from healthcare informatics, such as differential expression analysis between high-dimensional vectors, to identify... Compared to the pure control group, the upregulated or downregulated exosome components, such as specific miRNAs or proteins, are fed into pathway enrichment analysis tools, such as GSEA or KEGG analysis, to output exosome action pathways with high confidence and specificity attributed to smoking-inducing factors. This architecture of first isolating and modeling and then separating different groups solves the attribution problem caused by signal collinearity.

[0038] To improve the system's performance when processing real-world data, the system of this invention may also include a deeper mechanism to address complex data realities such as nonlinearity and queue drift. One such deeper solution is to address secondary bottlenecks caused by nonlinear relationships between confounding factors, such as age and omics data, leading to underfitting of the model and incomplete signal stripping. To this end, the confounding feature modeling unit, during its construction operation, includes data hierarchical preprocessing. Based on at least the key confounding factors in the confounding factor dataset of the non-smoker control group cohort, such as age, the non-smoker control group cohort is computationally divided into multiple data sub-layers. The division is based on a preset clinical threshold or quantile of the key confounding factor. For example, age is divided into three sub-layers: <50 years, 50-70 years, and >70 years. Within each data sub-layer, the confounding feature modeling unit independently constructs a corresponding confounding feature model. When the signal purification operation unit calculates and predicts the confounding exosome spectrum, it determines the confounding factor affiliation of an individual smoker. For example, a 65-year-old smoker belongs to the 50-70 year layer, and the corresponding local confounding feature model for that layer is called. To calculate This collaborative mechanism of hierarchical modeling and hierarchical purification reduces the complex global nonlinear fitting problem to multiple parallel local linear fitting problems in terms of computation. While maintaining the required high interpretability, it makes the stripping of information subtraction more accurate and sufficient.

[0039] Another aspect of this invention's system is addressing the engineering challenges of queue drift or batch effects encountered when applying models across queues and batches. When a model built on queue A is directly applied to queue B, which exhibits data distribution drift, misaligned subtraction and computational failures can occur. To address this, the signal purification unit can be further used to execute a dynamic bias calibration procedure. This procedure is initiated when analyzing a new queue containing both new smokers and new non-smokers. The signal purification unit then purifies the mixed feature model. The new non-smoker cohort within this new cohort is used to calculate the residual signals of their respective new cohort control groups. Theoretically, the mean of these signals should be zero. However, due to cohort drift, the mean will systematically deviate from zero. The unit then calculates a dynamic bias vector based on these residual signals of the new cohort control groups. In informatics, this vector is the computational quantization value of queue drift. A simplified calculation method is to directly calculate the mean vector of the residual signal vectors of all new queue control groups. The operation of the signal purification unit in generating the residual signal dataset involves subtracting the sum of the predicted confounding exosome spectrum and the dynamic bias vector from the exosome omics dataset. ,in, , , The definition is the same as before. As a dynamic bias vector, this modified subtraction cancels out both promiscuous predictions and queue drift.

[0040] Furthermore, to mitigate the risk of false positives in detecting patterns within noise—specifically, in scenarios where the target signal, such as a smoking signal, is extremely weak or absent, causing the downstream analysis module to output incorrect paths from purely random noise—the system of this invention may further include a systematic signal-to-noise ratio (SNR) arbitration gating. This gating is triggered after the signal cleansing unit and before the residual signal path identification unit. The systematic SNR arbitration gating performs a global statistical test on the residual signal dataset output by the signal cleansing unit, regardless of the specific path, to examine whether the dataset as a whole carries a systematic, non-random signal. A feasible universally applicable test method is multivariate analysis of variance (MANOVA) for this purpose. The system determines whether the mean vector deviates from the zero vector or the mean vector of the pure control group, and outputs a systematic signal index, such as a p-value. The gating then compares the systematic signal index with a minimum signal threshold; for example, a p-value less than 0.05 indicates the presence of a systematic signal. When the systematic signal index is greater than the minimum signal threshold (p-value less than 0.05), the residual signal path identification unit is triggered. If it fails (p-value greater than or equal to 0.05), execution is blocked. This arbitration gating mechanism avoids ineffective analysis of random noise, ensuring the decision-making security and scientific rigor of the final system output. The system and method of this invention strictly adhere to data security and privacy protection regulations when processing healthcare informatics data. The exosome omics dataset and confounding factor dataset involved are standardized, de-identified, or anonymized data after obtaining necessary authorization. All calculation steps are applied to these desensitized objective data. The final output exosome action path is a neutral technical indicator and does not constitute any form of individual diagnostic conclusion; it is only used to assist professionals in subsequent scientific research and interpretation.

[0041] Example 1: This example is a specific operational instance in a particular application scenario. It demonstrates an iterative model refinement workflow for addressing the issue of data label contamination in real-world (RWD) data from healthcare informatics. Based on a large-scale population cohort study, the study examines the exosome proteomics data from serum samples of a smoker intervention group and a non-smoker control group. The study focuses on the exosome pathway, where the expression level induced by smoking factors may be relatively weak. As a typical RWD study, the non-smoker control group inevitably suffers from data label contamination. The cohort includes unreported active smokers or those who misreported; these individuals are collectively referred to as contaminated individuals in this paper. (Biological sample data...) In addition to signals contributed by confounding factors such as age and BMI, the data also includes target signals induced by passive smoking that are only present in the intervention group. If a system without the iterative purification mechanism of this invention is used, the confounding feature modeling unit will construct a confounding confounding feature model based on the contaminated control group data. The model will incorrectly learn the passive smoking signals carried by those contaminated individuals, mistakenly identifying these signals as a normal part of the confounding factors. When the signal purification calculation unit uses this confounding model to purify the signal of the smoker intervention group... At that time, the calculated This will be excessive because it contains not only spurious signals but also passively smoked signals from incorrect learning, leading to execution... Over-stripping occurs during vector subtraction, i.e. The smoking-specific signals that should have been preserved are incorrectly weakened or even canceled out, ultimately preventing the residual signal path identification unit from detecting the signal, leading to a systemic false negative failure. In this scenario, the model refinement workflow operates as follows: In the first round of execution (diagnosis round), the confounding feature modeling unit constructs an initial confounding feature model based on the contaminated control group cohort; the signal purification calculation unit applies this initial model to calculate the residual control signals for each of the non-smoker control group cohorts. In this calculation, for uncontaminated non-smokers, their Compared with model predictions Basically the same, The vector magnitude is close to zero; for contaminated individuals, its It contains passive smoking signals that the initial model could not predict, leading to Vector magnitude Deviation from zero, exceeding the preset statistical threshold The signal purification processing unit identifies and marks these contaminated individuals, excluding them from the non-smoker control group cohort to form clean control group data.

[0042] In the second round of execution (purification round), the workflow is reset, and the confounding feature modeling unit restarts. The input is limited to using only the pure control group data generated in the previous step to construct the purified confounding feature model. Since the modeling data source no longer contains interference from (dynamic) smoking signals, this The model construction only characterizes the computational signatures of the effects of confounding factors such as age and BMI on the exosome profile; finally, the signal purification unit uses this purified confounding feature model to perform the intervention on the smoker intervention group cohort. Calculation, due to This represents the signal contributed by confounding factors. The informatics subtraction method is more accurate in computation, especially in the residual signal dataset. The iterative workflow retains the smoking-specific signal, embodying a collaborative mechanism of the present invention. The output of the signal purification operation unit serves as the input data quality control tool for the hybrid feature modeling unit. This closed-loop information flow of output diagnosis input provides the ability to proactively identify and self-correct unreliable labels in real-world data.

[0043] Example 2: This example compares the iterative model refinement workflow of this invention with the standard calculation method that does not use this workflow, and examines the effectiveness of addressing the label contamination problem of the control group data in a typical scenario. This experiment uses a semi-artificial dataset method to obtain the confounding factor dataset and exosome omics dataset of a pure non-smoker cohort (N=180) verified by art standards, specifically through serum cotinine detection, as the basic pure control group. Next, data from N=20 individuals are extracted from an independent smoker cohort as the contaminated individual set. This N=20 contaminated individual set is then computationally merged with the N=180 basic pure control group to form a contaminated control group cohort with a total sample size of N=200. This cohort is used to simulate a 10% label contamination rate in real-world data. Finally, data from an independent smoker intervention group cohort (N=100) are used for signal purification and path recognition testing. The confounding factor datasets of all cohorts include age, body mass index (BMI), and C-reactive protein (CRP) levels, and their exosome omics datasets are all high-dimensional proteomic data.

[0044] Based on the above dataset, two analysis groups were set up: control group A and the present invention group B. Control group A used the standard calculation method, and its confounding feature modeling unit directly used the contaminated control group cohort data (N=200) to construct a confounding feature model. This model was then used to perform informatics subtraction on the smoker intervention group cohort (N=100). The present invention group B followed the complete iterative workflow described in the specific implementation: the confounding feature modeling unit constructed an initial confounding feature model based on the contaminated control group cohort (N=200); the signal purification calculation unit applied this initial model to calculate the residual control group signal for the N=200 cohort itself. ,based on The quantitative procedures, in which The residual signal amplitude, To identify and label contaminated individuals according to the preset statistical threshold determined in the specific implementation method, 18 out of 20 contaminated individuals were labeled in this experimental data, and 2 individuals from the baseline pure control group were incorrectly labeled. The confounding feature modeling unit was called again, with the input limited to using only the pure control group data of N=180 remaining after excluding the above total of 20 labeled individuals, to construct a purified confounding feature model. Its signal purification processing unit uses this purified promiscuous feature model. Informatics subtraction was performed on the intervention group cohort of N=100 smokers to generate their residual signal dataset. The residual signal datasets generated by the control sample group A and the sample group B of the present invention are compared. The data are input into the same residual signal pathway identification unit, and GSEA pathway enrichment analysis is performed to calculate whether each biological pathway is enriched in the residual signal. The results are expressed as the FDR false detection rate q-value, where q-value < 0.05 is the general significance threshold (see Table 1).

[0045] Table 1: Comparison of the impact of different calculation methods on specific path recognition capability

[0046]

[0047] Table 1 shows that the confounding feature model of control group A was affected by interference from 10% of contaminated individuals during its construction. This model learned some smoking-induced signals and underwent excessive stripping during purification, failing to detect two known smoking-induced signaling pathways, the AHR pathway and the CYP1A1 / 1B1 metabolic pathway, with q-values ​​of 0.187 and 0.091 respectively, which did not reach the significance threshold; this was a false negative result. In contrast, the iterative model refinement workflow implemented in sample group B of this invention, where the signal purification operation unit and the confounding feature modeling unit worked synergistically to eliminate most of the contamination. For individuals, a purified confounding feature model was constructed without excessive stripping. The residual signal data of sample group B of this invention centrally retains smoking-specific signals, enabling the residual signal pathway identification unit to detect the AHR pathway with a q-value of 0.003 and the CYP1A1 / 1B1 metabolic pathway with a q-value of 0.011. Neither method detected known confounding factor-related pathways, such as the Wnt pathway and the IL-6 / STAT3 inflammatory pathway, nor did it detect random noise pathways, such as the ABC pathway and the XYZ pathway, and no false positives were introduced.

[0048] Example 3: This example combines Figures 1 to 3This document describes a system and method for identifying the action pathway of an exosome containing a smoking-inducing factor, as follows: Figure 1 As shown, the process begins with a confounding factorial + omics dataset from a non-smoker control group cohort, which includes contaminated individuals. The data is input into a confounding feature modeling unit to construct an initial confounding feature model. A signal purification unit calculates the residual signal from the control group, and based on this signal, it proceeds to identify contaminated individuals. Based on a statistical threshold for the residual signal, individuals identified as contaminated are excluded. Those not identified as contaminated form a clean group, resulting in a clean control group data set that excludes contaminated individuals. This clean control group data is then iteratively fed back into the confounding feature modeling unit through an iterative feedback remodeling path to construct a purified confounding feature model. This step may include a hierarchical modeling mechanism, outputting a purified confounding feature model. On the other hand, the confounding factor + omics dataset of the smoker intervention group cohort, along with this purified confounding feature model, is input into the signal purification operation unit to calculate the predicted confounding spectrum and generate a residual signal dataset. The calculation method is: residual signal = actual signal - predicted confounding signal. The generated residual signal dataset has been stripped of confounding signals. The dataset enters the systematic signal-to-noise ratio arbitration gating to perform a global statistical test on the systematic signal index. If it is determined that the signal is insufficient (p > threshold), the blocking is performed. If the systematic signal does not reach the threshold, it is determined that the signal is p < threshold. This dataset is sent to the residual signal path identification unit to perform association analysis and pathway enrichment, and the output of the high confidence and specificity of the exosome action pathway is attributed to smoking.

[0049] like Figure 2 As shown, this figure illustrates the signal purification process in coordinate form. The vertical axis represents signal intensity, and the horizontal axis represents exosome feature indices. The figure uses actual exosome signals to represent the original high-dimensional omics signals of individual smokers. The predicted confounding exosome spectrum represents the signal components attributable to confounding factors predicted by the purified confounding feature model. The purified residual signal shows the result after subtracting the predicted confounding exosome spectrum from the actual exosome signal. This purified residual signal is the residual signal dataset, which exhibits reduced fluctuations and primarily enriches specific signals. Figure 3 As shown, the core of this architecture is a high-dimensional medical data computing server, which carries the core computing and analysis logic. It includes a processor and memory. The logical units deployed on the server include a heterogeneous feature modeling unit, a signal purification operation unit, a systematic signal-to-noise ratio arbitration gating unit, and a residual signal path identification unit. The data flow between units shows that there is bidirectional interaction between the signal purification operation unit and the systematic signal-to-noise ratio arbitration gating unit, while the systematic signal-to-noise ratio arbitration gating unit points unidirectionally to the residual signal path identification unit. The system's data input comes from clinical and omics databases, which store heterogeneous factors and exosome omics data. Researchers' workstations are used for user access, task submission, and result reception. The workstation sends an analysis request to the server. After performing the calculation, the server generates an analysis result report, i.e., the exosome action path, and returns it to the workstation.

[0050] Example 4: This example illustrates the operational procedure of the confounding feature modeling unit in constructing a model by dividing the non-smoker control group cohort into multiple data sub-layers based on key confounding factors. It addresses the computational problems of underfitting the model and incomplete signal stripping caused by nonlinear effects between key confounding factors and omics data in real-world data. In the application scenario, after executing the iterative model refinement workflow, pure control group data and a purified confounding feature model have been obtained. ,this This is a global regression model constructed based on all data from the pure control group; a model underfitting check is performed once. Applied to the pure control group itself, the residual signal of the control group is calculated. Under the linear assumption, this The data should exhibit a random noise distribution across all individuals; however, in this scenario, the age factor's effect on the exosome spectrum exhibits a non-linear characteristic, with its influence accelerating after age 60, leading to a global linear model. Model underfitting occurs when system analysis When considering the correlation with age, it was found that in individuals >70 years of age, The mean systematically deviates from zero, indicating that the global model It failed to adequately capture and remove nonlinear hybrid signals from elderly individuals.

[0051] To address this underfitting model, a data sub-layer-based modeling mechanism is activated. This involves the coordinated operation of the confounding feature modeling unit construction step and the signal purification operation unit purification step. The operational procedure is as follows: Based on the underfitting verification results, the confounding feature modeling unit identifies age as a key nonlinear confounding factor. Using clinical thresholds or data quantiles, the pure control group cohort is computationally divided into three data sub-layers: Data Sub-layer 1 (age < 50 years), Data Sub-layer 2 (age 50-70 years), and Data Sub-layer 3 (age > 70 years). This unit instructs that instead of building a model globally, a locally purified confounding feature model is independently constructed within each data sub-layer, generating a set of models. , and Since each local model only needs to fit a narrower age range with a stronger linear relationship, the underfitting problem is avoided. When the system executes the signal purification operation unit to purify the smoker intervention group cohort, the calculation logic is corrected: for each individual in the smoker intervention group, the age value in the confounding factor dataset is obtained. Taking an individual as an example, if the age is 68 years old, it is determined that the individual belongs to data sub-layer 2 (50-70 years old); the signal purification operation unit instructs to call the local model corresponding to data sub-layer 2. The confounding factor dataset for this 68-year-old individual includes age, BMI, and CRP. (Input...) To calculate the predicted hybrid exosome profile Finally, perform informatics subtraction. This generates the individual's residual signal; this synergistic mechanism of hierarchical modeling and hierarchical purification, through data sub-layer segmentation at the system architecture level, transforms the global nonlinear fitting problem computationally into multiple parallel local linear fitting problems. This mechanism reuses a universal regression model, ensuring both the interpretability required by healthcare informatics and... The model's fitting accuracy within each data sublayer makes the informatics subtraction stripping performed by the signal cleansing computation unit more accurate, resulting in a better residual signal dataset. The purity of the signal is improved, providing a data foundation for the residual signal path identification unit to identify the smoking induction path.

[0052] Example 5: In engineering applications, to address queue drift, the system of the present invention executes a pre-set dynamic bias vector. The calibration procedure is initiated before analyzing new cohort data. A subset of new non-smoker data is extracted from the new cohort, and the signal purification unit will use a purified confounding feature model built based on the reference cohort. Applied to these new nonsmokers, the residual signal of the new cohort control group was calculated. Then, the system calculates all of these. The mean vector of the vectors is used to determine the dynamic bias vector specific to this new queue. In subsequent cleanup operations, when the system processes new smokers in this new queue, it always starts from... Subtract With that The sum; in addition, to calibrate the minimum signal threshold required for systematic signal-to-noise ratio arbitration gating, an offline permutation test procedure is executed. The input of this procedure is the pure control group data, which undergoes 1000 calculation iterations. In each iteration, the pure control group is randomly divided into a pseudo-intervention group and a pseudo-control group. A global statistical test is performed on the pseudo-intervention group to obtain the pseudo-systematic signal index. The system collects the data generated from these 1000 iterations. The value is used to construct a statistical distribution that characterizes the baseline of pure random noise, extract the value of the 95th percentile of the distribution, and determine the value as the minimum signal threshold for arbitration gating judgment.

[0053] Example 6: This example specifically illustrates an offline standardized procedure for model hyperparameter optimization and threshold calibration performed by the system of the present invention before formal deployment in a cohort study. This eliminates the uncertainty in the setting of key algorithm parameters and logical thresholds. The input to this procedure is the obtained pure control group data, which can come from a cohort that has been rigorously verified by the standards in this field or obtained by executing an iterative workflow as in Example 1. The model used to calibrate the hybrid feature modeling unit is used, taking the ridge regression model as an example. The key L2 regularization term alpha value is determined through a K-Fold Cross-Validation process. The pure control group data is randomly divided into K subsets. Taking K=10 as an example, the system iterates 10 times, trains the model using 9 subsets, and tests on the remaining 1 subset. The preset list of alpha candidate values ​​is traversed, and the alpha value that minimizes the mean squared prediction error (MSE) of 10 iterations is selected as the optimal regularization parameter.

[0054] Using the hierarchical modeling mechanism described in Example 4, the optimal data sub-layer segmentation points for key confounding factors were further determined: multiple candidate segmentation schemes based on clinical common sense or data quantiles were tested. Taking age as an example, Scheme 1 (<50, 50-70, >70 years old) and Scheme 2 (<45, 45-65, >65 years old) were tested. For each segmentation scheme, the aforementioned K-fold cross-validation process was independently executed within each data sub-layer to determine the optimal alpha value of each local model, and the overall validation set MSE under that segmentation scheme was calculated. The optimal MSE was selected based on the overall validation set MSE. The optimal data sub-layer partitioning scheme for the dataset, i.e., Scheme 1 or Scheme 2, is determined by the minimum MSE of the dataset. This procedure is also used to calibrate the preset statistical threshold coefficients required for identifying contaminated individuals in the signal purification processing unit: In the clean control group, the system randomly simulates the injection of known contaminated individual data from a smoker cohort at a preset ratio (e.g., 5%) to form a calibration contaminated dataset; the system tests a series of MAD (Median Absolute Deviation) multiples as candidate threshold coefficients, such as 2.0, 2.5, 3.0, and 3.5; for each candidate coefficient, the identification... The process involves calculating the F1-Score for identifying contaminated individuals. This F1-Score is an evaluation metric balancing precision and recall. Finally, the coefficient that maximizes the F1-Score, such as 3.0, is selected as the preset statistical threshold coefficient for performing identification operations on this dataset. This procedure is also used to calibrate the number of iterations required for permutation testing in systematic signal-to-noise ratio arbitration gating: the system does not use a fixed number of iterations when performing permutation testing, but rather uses p-value convergence judgment; the system starts calculating the maximum number of iterations from an initial number of iterations, such as N=500. For small signal thresholds, the number of iterations is increased to N=1000, and the relative rate of change between the two threshold results is calculated. If the rate of change is greater than the convergence criterion (e.g., 0.1%), the system continues to increase the number of iterations to N=2000 and recalculates the rate of change. This process is repeated until the relative rate of change of the minimum signal threshold calculated in two consecutive iterations (e.g., N=1000 and N=2000) is lower than the convergence criterion. At this point, the system stops iterating, considering that the p-value has reached the statistical stability required for calculation, and the current number of iterations (e.g., N=1000) is determined as the standard number of iterations for this dataset.

[0055] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A system for identifying an exosome action pathway of a smoking inducer fusion factor, characterized by, The system comprises a confounding feature modeling unit, a signal purification operation unit and a residual signal path identification unit: The confounding feature modeling unit is configured to construct an initial confounding feature model based on a confounder dataset of a non-smoker control group cohort and an exosome omics dataset; The signal purification operation unit is configured to apply the initial confounding feature model to calculate a respective control group residual signal of the non-smoker control group cohort, identify a contaminated individual based on the control group residual signal, exclude the contaminated individual from the non-smoker control group cohort to form a pure control group data, and calculate a predicted confounding exosome profile based on the exosome omics dataset and the confounder dataset of the smoker intervention group cohort to generate a residual signal dataset; The confounding feature modeling unit is configured to construct a purified confounding feature model based on the pure control group data; The signal purification operation unit is configured to apply the purified confounding feature model to calculate a predicted confounding exosome profile based on the exosome omics dataset and the confounder dataset of the smoker intervention group cohort to generate a residual signal dataset; The residual signal path identification unit is configured to perform an association analysis based on the residual signal dataset and the pure control group data to output an exosome action path. The confounding feature modeling unit comprises a construction operation, which comprises: based on a key confounder in the confounder dataset of the non-smoker control group cohort, the non-smoker control group cohort is divided into multiple data sub-layers in calculation; In each data sub-layer, a corresponding confounding feature model is independently constructed.

2. The exosome-pathway recognition system of claim 1, wherein the fusion smoking-inducing factor is selected from the group consisting of: and. The initial confounding feature model and the purified confounding feature model constructed by the confounding feature modeling unit are multiple linear regression models, ridge regression models, principal component regression models or partial least squares regression models.

3. The exosome-pathway recognition system of claim 1, wherein the fusion smoking inducer is selected from the group consisting of: and. The confounder dataset at least includes one or more of age, body mass index (BMI) and C-reactive protein (CRP) level.

4. The exosome-pathway recognition system of claim 1, wherein the fusion smoking inducer is selected from the group consisting of: 5-HT, dopamine, and norepinephrine. The signal purification operation unit generates a residual signal dataset by performing the following vector subtraction operation on each individual in the cohort of smokers intervention group in the vector space defined by the exosomic data: wherein, is any residual signal vector in the residual signal dataset, is the actual signal vector in the exosomic dataset, is the predicted confounder vector in the predicted confounder exosome profile.

5. The exosome action pathway identification system for incorporating smoking-inducing factors according to claim 1, characterized in that, The signal purification operation unit is configured to apply the purified confounding feature model to a new non-smoker cohort to calculate a respective new cohort control group residual signal, calculate a dynamic bias vector based on the new cohort control group residual signal, and generate a residual signal dataset by subtracting the sum of the predicted confounding exosome profile and the dynamic bias vector from the actual signal vector in the exosome omics dataset.

6. The exosome-pathway recognition system of claim 1, wherein the fusion smoking inducer is selected from the group consisting of: and. The system comprises a systematic signal-to-noise ratio arbitration gate, which is configured to perform a global statistical test on the residual signal dataset output by the signal purification operation unit to calculate a systematic signal index, compare the systematic signal index with a minimum signal threshold, and trigger the execution of the residual signal path identification unit when the systematic signal index is greater than the minimum signal threshold.

7. The exosome-pathway recognition system of claim 1, wherein the fusion smoking inducer is selected from the group consisting of: and. The signal purification operation unit is configured to identify a contaminated individual based on the value of the control group residual signal exceeding a preset statistical threshold determined according to a statistical distribution.

8. The exosome-pathway recognition system of claim 1, wherein the fusion smoking inducer is selected from the group consisting of: and . The confounding feature modeling unit divides the non-smoker control group cohort into multiple data sub-layers based on a preset clinical threshold or quantile of the key confounder. The signal purification operation unit is configured to calculate a predicted confounding exosome profile by inputting individual data in the confounder dataset of the smoker intervention group cohort into the purified confounding feature model and obtaining the output of the purified confounding feature model as the predicted confounding exosome profile.

9. A method for identifying an exosome acting pathway of a smoking-inducing factor, characterized by, The exosome action path identification system for realizing the fusion smoking induction factor comprises the following steps: Step a, based on the confounder dataset and the exosome dataset of the non-smoker control group queue, an initial confounder feature model is constructed; Step b, the initial confounder feature model is applied to calculate the respective control group residual signal of the non-smoker control group queue; Step c, based on the control group residual signal, the contaminated individuals are identified and excluded from the non-smoker control group queue to form the pure control group data; Step d, based on the pure control group data, a purified confounder feature model is constructed; Step e, using the purified confounder feature model, the exosome dataset and the confounder dataset of the smoking intervention group queue are obtained, the predicted confounder exosome spectrum is calculated, and the residual signal dataset is generated; Step f, based on the residual signal dataset and the pure control group data, correlation analysis is carried out, and the exosome action path is output.

Images