A method and system for constructing an environment sensitive profile of a patient with chronic obstructive pulmonary disease

Abstract

This invention relates to the interdisciplinary application of digital health, smart healthcare, and artificial intelligence, and discloses a method and system for constructing an environmental sensitivity profile for COPD patients. The method includes: constructing a multimodal patient historical data profile dataset; reconstructing biological effects based on respiratory pathology mechanisms; constructing and training a causal feature learning network; calculating individualized environmental sensitivity; and generating clinical decision support information. Its core lies in mining multimodal longitudinal historical data of patients to construct a computational framework that deeply integrates respiratory pathology mechanisms and causal inference algorithms (a biological effect reconstruction module based on a multi-mechanism model, a causal feature learning module, and a dual machine learning module based on hypernetwork generation). This framework outputs an interpretable, individualized digital profile of environmental sensitivity, capable of quantifying a patient's pathological susceptibility to specific environmental factors (such as particulate matter), and can be combined with environmental forecasting to provide precise quantitative support for clinical decision-making.

Description

Technical Field

[0001] This invention relates to the interdisciplinary application of digital health, smart healthcare and artificial intelligence, and in particular to a method and system for constructing environmentally sensitive profiles of patients with COPD. Background Technology

[0002] Chronic obstructive pulmonary disease (COPD) is a common chronic respiratory disease, and along with cardiovascular disease, diabetes, and cancer, it is recognized by the WHO as one of the four major chronic diseases. The prevention and treatment of COPD face two core challenges: firstly, how to effectively slow the long-term, irreversible progression of the disease, characterized by a continuous decline in lung function (such as forced expiratory volume in one second, FEV1); and secondly, how to prevent short-term, life-threatening acute exacerbations of COPD (AECOPD). AECOPD not only accelerates permanent damage to lung function but is also a major cause of hospitalization and death.

[0003] Numerous authoritative epidemiological and clinical studies have confirmed that outdoor air pollution, particularly fine particulate matter (PM2.5) and ozone (O3), is a key environmental risk factor in the development of COPD. For patients already diagnosed with COPD, a rapid increase in environmental pollutant concentrations within a short period is usually the most common environmental trigger for adverse events following COPD (AECOPD). Therefore, accurate quantification of environmental exposure and proactive avoidance are indispensable key aspects of the entire COPD management process.

[0004] However, the field of personalized environmental risk management for COPD still faces significant technological bottlenecks. Its core deficiency lies in the lack of scientific and accurate attribution capabilities for individualized environmental risks, specifically manifested in the following two aspects:

[0005] First, the "one-size-fits-all" approach to risk assessment and the neglect of individual heterogeneity:

[0006] Existing environmental health risk assessment tools, such as the Air Quality Index (AQI), are based on the fundamental assumption that "the same environment poses the same harm to everyone," which is seriously inconsistent with clinical observations. The development of COPD is a multi-factorial process, resulting from the combined effects of external environmental factors and individual internal factors. Current technologies cannot quantify this individual heterogeneity—the differences in how different individuals respond to the same environmental stimuli—thus failing to achieve truly precise protection.

[0007] Second, the confounding bias of attribution logic:

[0008] In epidemiological terms, intrinsic personal factors are called confounding factors. These factors influence both an individual's baseline health risk and may be related to environmental exposure levels. Existing simplified health index models typically employ only simple linear weighting calculations, failing to accurately isolate the independent pathogenic effects of environmental factors on COPD from complex confounding factors. This confounding bias in attribution logic severely limits the accuracy and reliability of clinical guidance, making it difficult for physicians to develop targeted prevention strategies.

[0009] Therefore, there is an urgent need to develop a novel method for constructing environmental sensitivity profiles of COPD patients to solve the medical challenge of personalized prevention in existing technologies, where individual heterogeneity of COPD patients leads to significant differences in their responses to environmental factors (such as air pollution). Summary of the Invention

[0010] This invention provides a method for constructing an environmental sensitivity profile for COPD patients. The invention aims to address the medical challenge of significantly different responses to environmental factors (such as air pollution) among COPD patients due to individual heterogeneity, where existing technologies cannot accurately quantify the modification of this effect, thus hindering personalized prevention. This method delves into the causal level, quantifying the modifying effect of individual characteristics on environmental pathogenicities by mining patients' longitudinal historical data, thereby constructing a unique digital profile of COPD environmental sensitivity for each patient. The specific technical solution is as follows:

[0011] A method for constructing environmental sensitivity profiles of COPD patients includes the following steps:

[0012] Step S1: Construct a multimodal patient historical data profile dataset. Specifically, this involves acquiring and organizing multi-source heterogeneous patient data, and constructing a structured dataset containing environmental exposures, effect modifiers, and health outcomes through spatiotemporal alignment and feature extraction.

[0013] Step S2, biological effect reconstruction based on respiratory pathology mechanism, specifically: receiving the environmental exposure data from step S1, classifying it through a multi-mechanism model, and converting it into a unified biological effect vector;

[0014] Step S3: Construct and train a causal feature learning network. Specifically, receive the biological effect vector output from step S2 and the effect modification factor obtained from step S1, extract high-order causal features through a deep network containing temporal convolution and feature fusion, and train an auxiliary prediction function for subsequent causal inference.

[0015] Step S4: Calculate individualized environmental sensitivity. Specifically, based on the biological effect vector obtained in step S2, the health outcome data obtained in step S1, and the auxiliary prediction function obtained in step S3, the individualized environmental sensitivity vector of the patient is obtained using a dual machine learning architecture and a hypernetwork model.

[0016] Step S5: Generate clinical decision support information, specifically by fusing the patient's individualized environmental sensitivity vector obtained in step S4 with external environmental prediction data to obtain a visualized clinical analysis report.

[0017] Preferably, the specific steps in step S1 for constructing the multimodal patient historical data profile dataset include:

[0018] Step S1.1: Acquire and organize multi-source heterogeneous data, specifically: acquire environmental exposure data, effect modifier data, and health outcome data;

[0019] Step S1.2: Perform multi-scale data alignment, specifically: downsample the environmental exposure data to daily frequency to obtain aligned environmental exposure data; upsample and expand the effect modification factor data to daily time series to obtain aligned effect modification factor data.

[0020] Step S1.3: Separate the aligned environmental exposure sequences into pollutant groups and meteorological element groups based on their physicochemical properties; integrate the aligned effect modification factor data and health outcome data to construct a structured dataset containing environmental exposure, effect modification factors, and health outcomes.

[0021] Preferably, obtaining environmental exposure data specifically includes: connecting to an environmental data interface and generating a daily high spatiotemporal resolution environmental exposure sequence for each patient, including particulate matter, gaseous pollutants, and meteorological elements, through spatial interpolation and temporal aggregation;

[0022] Obtaining effect modifier data specifically includes: extracting and quantifying genetic risk characteristics and clinical characteristics from patients' electronic medical records and gene testing reports; clinical characteristics include lifestyle characteristics and basic clinical indicator characteristics;

[0023] Obtaining health outcome data specifically includes: collecting and encoding patients' daily symptom load, acute exacerbations, and physiological fluctuation indicators.

[0024] Preferably, constructing effector modifier data includes extracting a genetic risk feature based on pathological functional clusters, specifically including:

[0025] Step ①: Multiple significant genetic susceptibility loci from genome-wide association analysis were mapped to three independent pathological functional clusters based on their biological functions: epithelial barrier defense, oxidative stress metabolism, and protease-antiprotease homeostasis. Specifically, epithelial barrier defense was mapped using... =1 indicates that oxidative stress metabolism is employed. =2 indicates that protease-antiprotease steady state is adopted. =3 indicates;

[0026] Step 2: For each pathological functional cluster, independently calculate the weighted sum of effect values ​​for its internal risk sites to obtain the components. ;

[0027] Step ③: Take the components from step ② Combining these elements creates a multidimensional genetic risk vector. ,in: .

[0028] Preferably, constructing effect modifier data also includes the quantification and combination of clinical features, specifically including:

[0029] Step ①: Calculate the smoking exposure index, physical activity level based on metabolic equivalents, and a dietary inflammation index based on a weighted sum of multiple nutritional parameters to obtain behavioral characteristics;

[0030] Step 2: Standardize the three basic clinical indicators, which include continuous physiological parameters, heterogeneous inflammatory markers, and discrete biological parameters, to obtain the basic clinical indicator characteristics; the continuous physiological parameters include age, BMI, and FEV1 / FVC; the heterogeneous inflammatory marker is selected as blood eosinophil count; the discrete biological parameter is selected as sex.

[0031] Step ③: Concatenate the behavioral characteristics from Step ① with the basic clinical indicator characteristics from Step ② to form a clinical feature vector. .

[0032] Preferably, the biological effect reconstruction in step S2 includes the following steps:

[0033] Step 2.1, Calculation of bioeffective dose of inhalable pollutants, specifically: For particulate matter and gaseous pollutants, perform outdoor-indoor exposure correction, respiratory deposition dose calculation and inflammatory hysteresis and cumulative effect calculation in series;

[0034] The quantification of the biostimulatory effects of meteorological elements specifically involves: calculating the wind-cold coupled stimulation index (WCSI) for temperature, humidity, and wind speed; calculating the WCSI used to measure the intensity of characteristic cold air stimuli; and calculating the humidity deviation index. calculate;

[0035] Step 2.2: Perform a characteristic cascade of the bioeffective dose of inhalable pollutants and the biostimulation effects of meteorological elements to form a unified bioeffect vector. .

[0036] Preferably, the outdoor-indoor exposure correction specifically involves: employing a steady-state mass balance equation and introducing a factor determined by building characteristics and patient... Dynamic hourly air exchange rate parameter determined by real-time window opening behavior This converts outdoor pollutant concentrations into individual microenvironment exposure concentrations.

[0037] The calculation of respiratory deposition dose specifically involves: using the ICRP lung deposition model and introducing a dose based on the patient's... Dynamic hourly ventilation parameters obtained from real-time physical activity intensity The concentration of pollutants exposed to the microenvironment was converted into the mass of pollutants deposited in the lungs on day T. ;

[0038] The calculation of inflammatory hysteresis and cumulative effects specifically involves: employing an exponentially decaying convolutional model based on biological half-life to analyze past... The mass of pollutants deposited in the lungs within a day is weighted and summed to output a bioeffective dose that quantifies the hysteresis effect of inflammation. The formula is as follows:

[0039] ;

[0040] in: The number of days in lag; The specific biological half-life of the p-th environmental factor for the rate of inflammatory response; Total number of days in the past; The time scale is a day, representing the cumulative value of 24 hours.

[0041] Preferably, the wind-cold coupling stimulation index is calculated as follows: a nonlinear model based on physiological thresholds is used to couple actual temperature with wind speed to calculate the wind-cold coupling stimulation index for the intensity of characteristic cold air stimulation, as expressed below:

[0042] ;

[0043] in: The mean wind-cold coupling stimulation index on day T; The average actual temperature on day T; The average perceived effective temperature on day T;

[0044] Average perceived effective temperature on day T The following empirical formula is used for construction:

[0045] ;

[0046] in: The average wind speed on day T; This is the wind chill correction factor;

[0047] The humidity deviation index is calculated by taking the difference between the daily relative humidity and the median humidity suitable for respiratory function, in order to quantify the intensity of stimulation from dry or humid environments. The expression is as follows:

[0048] ;

[0049] in: The average humidity deviation index for day T; The average relative humidity on day T; The humidity level is the optimal level for respiratory function.

[0050] Preferably, the causal feature learning network in step S3 includes an environmental temporal pattern extraction module and a dual-head prediction architecture with input feature isolation design. The environmental temporal pattern extraction module is configured with multiple one-dimensional convolutional kernels of different sizes in parallel to capture the dynamic patterns of environmental exposure at three time scales: acute, short-term, and chronic. The dual-head prediction architecture with input feature isolation design trains auxiliary prediction functions for subsequent causal inference, which include functions for estimating the underlying risk. and the function used to estimate exposure propensity .

[0051] Preferably, the step S4 of calculating individualized environmental sensitivity includes the following steps:

[0052] Step S4.1: Use the auxiliary prediction function obtained in step S3 for subsequent causal inference to perform residual purification, and obtain the health outcome residual and environmental exposure residual;

[0053] Step S4.2: Construct a dual machine learning architecture and supernetwork model consisting of a parameter generator subnetwork HyperNet and a causal inference main network PrimaryNet. The parameter generator subnetwork HyperNet is the first... Using the patient's state feature vector as input, a patient's individualized environmental sensitivity vector is dynamically generated. The causal inference main network, PrimaryNet, utilizes the aforementioned... As regression coefficients, they fit the residuals of healthy outcomes within the residual space;

[0054] The patient's individualized environmental sensitivity vector Represented as , For the first Environmental factors on the first The weight of each patient.

[0055] Preferably, the mathematical model of the causal inference main network PrimaryNet is:

[0056] ;

[0057] in: This refers to the overall dimension of environmental factors; It reflects the patient's genes and condition. Modified environmental exposure residuals of health outcomes The intensity of the impact; To predict the outcome residuals; For the first Environmental exposure to various environmental factors.

[0058] Preferably, the loss function of the hypernetwork model includes a mean squared error loss function containing an L2 regularization term, the mathematical form of which is as follows:

[0059] ;

[0060] in: This represents the total loss function value. Total patient sample size; The length of the observation time window, in days; Regularization coefficient These are the global parameters of the hypernetwork itself that need to be trained; It is an L2 norm.

[0061] Preferably, generating clinical decision support information in step S5 includes: constructing a predictive risk index calculation model based on parameter coupling, which couples and weights the future environmental forecast values ​​with the individualized environmental sensitivity vector.

[0062] Preferably, the formula for calculating the predicted risk index PRI is as follows:

[0063] ;

[0064] in: For the first Heavenly The predicted environmental risk index for each patient represents the expected pathological burden after the excess environmental exposure is amplified by individual sensitivity. For the first Heavenly Meteorological forecast values ​​for various environmental factors; For the first Safety baseline values ​​for various environmental factors.

[0065] The effect of applying the technical solution of this invention is:

[0066] The method of this invention includes: constructing a multimodal patient historical data profile dataset; reconstructing biological effects based on respiratory pathological mechanisms; constructing and training a causal feature learning network; calculating individualized environmental sensitivity; and generating clinical decision support information. The mechanism is as follows: through a bioeffective dose reconstruction method, particularly by introducing dynamically changing air exchange rate and minute ventilation parameters, and employing an inflammation convolution model based on biological half-life, coarse environmental physical exposure data is transformed into precise, pathologically significant biological effect vectors, laying a solid data foundation for causal inference of COPD environmental pathogenic factors; through pathological pathway mapping and hypernetwork architecture design, explicit modeling of effect modification mechanisms is achieved; by deconstructing genetic risk into independent pathological functional clusters, input features are given clear biological meaning; the hypernetwork mathematically simulates the biological process of "genotype determining response parameters," resulting in highly clinically interpretable results; and by combining dual machine learning with the hypernetwork, a model capable of dynamically generating causal parameters is constructed. For each patient, this invention can calculate their unique, multi-dimensional environmental sensitivity vector, realizing the precise construction of a parametric profile that is "personalized" rather than a "one-size-fits-all" statistical pattern. Through the predictive risk index calculation model, the static individual sensitivity profile is coupled with the dynamic future environmental forecast, forming a complete technical closed loop from diagnosis to prediction to intervention, providing a tool with high clinical practical value for the precise prevention and self-management of COPD patients.

[0067] This invention also discloses a system for constructing environmental sensitivity profiles of COPD patients, implementing the aforementioned method for constructing environmental sensitivity profiles of COPD patients; the device includes:

[0068] The data collection unit is used to construct a multimodal patient historical data profile dataset;

[0069] Data processing unit: Equipped with a processor and memory, it is used for reconstructing biological effects based on respiratory pathology mechanisms, building and training causal feature learning networks, and calculating individualized environmental sensitivities;

[0070] User interface and output unit: Used to generate clinical decision support information. Attached Figure Description

[0071] 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.

[0072] Figure 1 This is a flowchart illustrating the method for constructing environmentally sensitive profiles of COPD patients in an embodiment of the present invention.

[0073] Figure 2 yes Figure 1 Detailed flowchart.

[0074] 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

[0075] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0076] Example:

[0077] This embodiment discloses a method for constructing an environmental sensitivity profile for patients with chronic obstructive pulmonary disease (COPD). Specifically, it discloses a method based on multimodal longitudinal health data, which quantifies the modifying effects of genetic susceptibility and lifestyle on environmental pathogenicity, to construct an individualized digital profile of environmental sensitivity for COPD patients, and provides a precise dynamic risk assessment accordingly. Figure 1 and Figure 2 As shown. This method mines multimodal longitudinal data of patients and utilizes a deep learning and dual machine learning architecture to quantify the modifying effects of genetic susceptibility and lifestyle on environmental pathogenicity. The specific scheme is as follows:

[0078] Step S1: Construct a multimodal patient historical data profile dataset. Specifically, this involves acquiring and organizing multi-source heterogeneous patient data, and constructing a structured dataset containing environmental exposures, effect modifiers, and health outcomes through spatiotemporal alignment and feature extraction.

[0079] Step S1 aims to provide high-quality, multi-dimensional input data for subsequent causal inference. The system interfaces with a biobank via a non-invasive data interface to collect multi-source heterogeneous data from target patients within a historical time window (the past 12 to 24 months), constructing a structured tensor dataset containing environmental exposures, effector modifiers, and health outcomes. Specific steps include:

[0080] Step S1.1: Acquire and organize multi-source heterogeneous data, specifically: acquire environmental exposure data, effect modifier data, and health outcome data. In this embodiment, preferably, acquiring environmental exposure data includes: connecting to an environmental data interface and generating a daily high spatiotemporal resolution environmental exposure sequence for each patient, containing particulate matter, gaseous pollutants, and meteorological elements, through spatial interpolation and temporal aggregation; acquiring effect modifier data specifically includes: extracting and quantifying genetic risk characteristics and clinical characteristics from the patient's electronic medical record and gene testing report; clinical characteristics include lifestyle characteristics and basic clinical indicator characteristics; acquiring health outcome data specifically includes: collecting and encoding the patient's daily symptom load, acute exacerbations, and physiological fluctuation indicators. Details are as follows:

[0081] (1) Obtain high spatiotemporal resolution environmental exposure data streams, i.e., multivariate environmental data, as follows:

[0082] Connect to the four-dimensional gridded environmental dataset provided by the meteorological bureau or the national environmental monitoring station. ,in: :longitude, :latitude, : Time (hour), p: Environmental factors. The spatial resolution of this dataset is 1km×1km, and the temporal resolution is on the order of hours.

[0083] Environmental factors were collected, covering core indicators closely related to its respiratory pathophysiology, specifically including:

[0084] ① Particulate matter index: Fine particulate matter (PM2.5) 2.5 ) and inhalable particulate matter (PM2.5) 10 As a major physical factor inducing airway oxidative stress;

[0085] ② Gaseous pollutants: Nitrogen dioxide (NO2), ozone (O3), and sulfur dioxide (SO2) are the main factors that induce chemical inflammation of the airway mucosa;

[0086] ③ Meteorological elements: daily average temperature, relative humidity and wind speed, used to assess the impact of meteorological stimuli on airway spasm.

[0087] For each patient in the queue (patient number is...) First, obtain its geographic coordinates at time t within the historical time window. Subsequently, bilinear interpolation was used to extract the patient's environmental exposure value at time t from the four-dimensional gridded environmental dataset E. To match the daily frequency of clinical outcome data, the system performs time aggregation on the aforementioned hourly exposure sequences and calculates the number of exposures using the following formula. Daily average :

[0088] ;

[0089] in: The observation window length is measured in days. Environmental factors.

[0090] Arrange all aggregated daily average value sequences in chronological order, which is the [number]th [sequence]. A two-dimensional time-series matrix is ​​generated for each patient – ​​a high spatiotemporal resolution environmental exposure data stream. .

[0091] (2) Obtain individualized effect modifier data These are effect modifiers (genetic / lifestyle / clinical data), specifically: based on the "gene-environment-behavior" interaction mechanism, two key features are extracted from patients' electronic medical records and gene testing reports to quantify individual differences in susceptibility to the environment.

[0092] In this embodiment, the preferred method for constructing effect modifier data includes extracting genetic risk features based on pathological functional clusters, specifically constructing a genetic risk vector. ,as follows:

[0093] Unlike traditional methods that aggregate all genetic variations into a single risk value, this embodiment employs a structured feature engineering approach based on biological pathways to deconstruct patients' specific defense deficiencies against different environmental stresses. The specific implementation is as follows:

[0094] Step 1: Pre-prepared COPD genetic susceptibility loci database. A functional annotation filtering algorithm is used to select significant loci (genome-wide significance level) from genome-wide association analysis (GWAS) statistics from authoritative databases (UK Biobank). The following seven core marker genes were selected and mapped to three independent pathological functional clusters based on their gene regions and biological pathways:

[0095] Cluster 1 (Epithelial Barrier Defense Cluster): FAM13A (sequence similarity family 13 member A gene) and HHIP (Hedgehog interacting protein gene). This group of genes determines the regenerative and repair capacity of airway epithelial cells after physical injury and the maintenance capacity of alveolar structure. Risk single nucleotide polymorphisms (SNPs) are extracted from the gene regions, and their effect-weighted sums are calculated to generate the first component of the vector. ;

[0096] Cluster 2 (Oxidative Stress Metabolism Cluster): IREB2 (iron-response element-binding protein 2 gene), EPHX1 (microsomal epoxide hydrolase 1 gene), and GSTM1 (glutathione S-transferase Mu-1 gene). This group of genes regulates intracellular iron homeostasis, hydrolysis of exogenous toxins, and the activity of the glutathione antioxidant system, determining the lung's biochemical defense capacity to clear environmental oxidative pollutants and carcinogens from tobacco smoke. Risk SNPs located in the IREB2, EPHX1, and GSTM1 gene regions were extracted, and their effect-weighted sums were calculated to generate the second component of the vector. ;

[0097] Cluster 3 (Protein-Antiprotein Homeostasis Cluster): SERPINA1 (a member of the serine protease inhibitor family A, gene 1) and MMP12 (a matrix metalloproteinase 12 gene). These genes encode... Antitrypsin and macrophage metalloelastase are responsible for regulating the balance between the degradation and protection of elastic fibers in the alveolar wall, determining the structural stability of lung tissue against emphysema formation under chronic inflammation. Risk SNPs are extracted from gene regions, their effect-weighted sums are calculated, and the third component of the generated vector is determined. .

[0098] Step 2, Vectorization Output: The three components mentioned above are Z-score standardized and combined to form the patient's three-dimensional genetic risk vector. :

[0099] .

[0100] In this embodiment, the construction of effect modifier data also includes the quantification and combination of clinical features, specifically the construction of a lifestyle and basic clinical indicator vector (i.e., a clinical feature vector). This includes: extracting patients' historical COPD-related records and basic clinical indicators, and constructing feature vectors for the acquired modification of characteristic environmental effects through mathematical transformation and standardization. Specific processing steps include:

[0101] Cumulative calculation and normalization of smoking exposure index, and statistical analysis of patients' historical pack-years of smoking. :

[0102] .

[0103] For patients who have quit smoking, their cumulative dose before quitting is retained. Considering the long-tailed distribution of smoking data, a logarithmic transformation is performed on the original pack-year count to smooth out extreme values, followed by normalization. Mapping to the [0,1] interval yields This represents the basic degree of damage to the airway mucosa.

[0104] Weighted quantification of physical activity levels: Calculation logic: Based on the physical activity questionnaire, patients' activity types are divided into three categories: low intensity, moderate intensity, and high intensity. Different intensities of activity are assigned corresponding metabolic equivalent (MET) weights (walking = 3.3, moderate = 4.0, high intensity = 8.0), and the weekly average metabolic equivalent is calculated. :

[0105] .

[0106] The calculation results were Z-score standardized to eliminate dimensional differences. This represents the patient's cardiopulmonary tolerance reserve.

[0107] The weighted summation of the dietary anti-inflammatory index is as follows: Based on the dietary frequency questionnaire, the intake of 45 nutritional parameters by patients is extracted: the pro-inflammatory group (positive weight parameters) includes total energy, total fat, saturated fatty acids, trans fatty acids, cholesterol, and carbohydrates; the anti-inflammatory group (negative weight parameters) includes dietary fiber, polyunsaturated fatty acids, and carbohydrates. Micronutrients (magnesium, zinc, selenium), vitamins (vitamins A, B6, C, D, E), and phytochemicals (such as isoflavones and allicin).

[0108] The total dietary inflammation index (DCI) was obtained using a standardized weighted summation algorithm. ):

[0109] ;

[0110] in: Representing the patient on the first The standardized score of an individual's intake of a nutrient relative to a global reference baseline. For the first The weight of each nutrient.

[0111] This index is used to characterize the baseline level of systemic inflammation in patients due to dietary habits (positive values ​​represent a pro-inflammatory tendency, and negative values ​​represent an anti-inflammatory tendency).

[0112] Three basic clinical indicators, including continuous physiological parameters, heterogeneous inflammatory markers, and discrete biological parameters, were standardized to obtain the basic clinical indicator characteristics. Continuous physiological parameters included age, BMI, and FEV1 / FVC; the heterogeneous inflammatory marker was selected as serum eosinophil count; and the discrete biological parameter was selected as sex. (See below:)

[0113] Three data types that determine the pathological features and severity grading of COPD were selected for processing, as follows:

[0114] Continuous physiological parameters: Three indicators were selected: age, body mass index (BMI), and pulmonary function obstruction index (FEV1 / FVC). Z-Score standardization was used to eliminate dimensional differences in nutrition and metabolic status.

[0115] Heterogeneous inflammatory markers (EOS): Serum eosinophil count (EOS) was selected as a key indicator of characteristic airway inflammation phenotypes. Given its significantly skewed distribution and large numerical range in the population, a log-normal transformation was performed before standardization.

[0116] Discrete biological parameter processing: The sex index was selected and binary one-hot encoding was used.

[0117] The above components are concatenated to output the clinical feature subvector. ,as follows:

[0118] ;

[0119] in: These are standardized age-related physiological parameters; This is the standardized body mass index; Standardized pulmonary obstruction index; These are standardized heterogeneous inflammatory markers; As a gender indicator, Represents male, It represents women.

[0120] After quantifying smoking, physical activity, diet, and basic clinical indicators, the acquired modified features were combined into a unified, fixed-dimensional clinical feature vector through feature cascading. ,as follows:

[0121] ;

[0122] in: This is a concatenation operation for vectors; The behavioral feature sub-vector is calculated using the following formula:

[0123] .

[0124] After constructing the genetic risk vector and the clinical characteristic vector, this embodiment integrates these two representative vectors to form the first... One complete individualized effector modifier data for each patient. :

[0125] .

[0126] (3) Constructing a clinical health outcome vector That is, health outcome data, specifically:

[0127] Collect the patient's first data within the historical time window Daily disease records are quantified and encoded to construct target response variables for training causal inference models, capturing short-term fluctuations and key exacerbation events of COPD disease from multiple dimensions:

[0128] Daily symptom burden The internationally recognized COPD assessment test score (CAT) was used. This questionnaire contains 8 items (cough, sputum production, chest tightness, dyspnea, etc.), each scored from 0 to 5 points based on severity, for a total score of 0 to 40 points. Following clinical guidelines, the CAT score was discretized into high symptom burden states for the classification task.

[0129] .

[0130] Acute exacerbation of COPD (AECOPD) Based on patients' electronic medical records, medication records, or self-reported logs, events are tagged as binary variables according to the Global Initiative for Disability and Health (GOLD) standards.

[0131] .

[0132] Daily physiological fluctuations When patients are equipped with a home portable spirometer, peak expiratory flow rates are collected twice daily, morning and evening. and ,calculate Intraday variability to capture airway instability:

[0133] .

[0134] The above quantified indicators are combined into the first... Clinical health outcome vector of a patient :

[0135] .

[0136] Step S1.2: Perform multi-scale data alignment, specifically: downsample the environmental exposure data to daily frequency to obtain aligned environmental exposure data; upsample and expand the effect modification factor data to daily time series to obtain aligned effect modification factor data.

[0137] This embodiment specifically involves data spatiotemporal alignment, fusion, and multidimensional tensor construction, including:

[0138] Unified benchmark: The system sets the calendar day (UTC 00:00 - 23:59) as the unified analysis time unit and alignment benchmark for all data streams.

[0139] Time scale alignment: To address the differences in temporal resolution between data sources, the following classification and processing strategies are employed: High-frequency data downsampling: For hourly environmental data streams... The daily mean aggregation method described in S1.1 was used to downsample the data to daily frequency, aligning it with daily clinical outcome data. Low-frequency / static data upsampling: For the collected effect modifier data X, a time-filling strategy was used to expand it into a daily time series: for genetic risk vectors... For lifelong unchanging features such as gender, their values ​​are copied and filled into each day of the entire historical time window. Quasi-static features: For slowly changing features measured at a specific time point (such as FEV1 / FVC, smoking history, dietary habits, BMI), a forward filling strategy is used: it is assumed that the feature value remains unchanged between two measurements until the next measurement update.

[0140] Spatial interpolation: To address spatial misalignment caused by the patient's location not being at the center of the grid, a bilinear interpolation algorithm is used, specifically temporal interpolation: high spatiotemporal resolution environmental exposure data stream. For short-term (< 24 hours) missing data, linear interpolation was used; for long-term (≥ 24 hours) missing data, Kriging spatial interpolation based on historical data from other monitoring stations in the same area was used for completion. Clinical health outcome vector. For occasional missing daily symptom scores (CAT), a forward / backward imputation method was used to ensure the continuity of the time series.

[0141] Step S1.3: Separate the aligned environmental exposure sequences into pollutant groups and meteorological element groups based on physicochemical properties; integrate the aligned effect modification factor data and health outcome data to construct a structured dataset containing environmental exposure, effect modification factors, and health outcomes. Specifically, this embodiment involves aligning and interpolating high spatiotemporal resolution environmental exposure data streams. Two sets of features were separated:

[0142] Group 1: Pollutant Group Including particulate matter (PM) 2.5、 PM 10 ), gaseous pollutants (NO) 2、 O 3、 SO2); Group 2: Meteorological Elements This includes temperature, relative humidity, and wind speed.

[0143] Step S2, Bioeffective Dose Reconstruction Based on Respiratory Pathology Mechanism, specifically involves receiving the environmental exposure data from Step S1, classifying it using a multi-mechanism model, and transforming it into a unified biological effect vector.

[0144] This step, based on mathematical models of respiratory physiology, aerosol dynamics, and inflammation dynamics, transforms the high spatiotemporal resolution environmental exposure data stream generated in step S1. Based on the pathophysiological mechanisms of different environmental factors, a classification strategy is adopted to transform them into a unified biological effect vector. The reconstruction process comprises the following three core computational modules:

[0145] Step 2.1, Bioeffective dose of inhalable pollutants Reconstruction, specifically the environmental sequence of pollutants, refers to the reconstruction of particulate matter and gaseous pollutants. The calculation involves performing outdoor-indoor exposure correction, respiratory deposition dose calculation, and inflammatory hysteresis and cumulative effect calculation in series, i.e., performing the following three cascaded calculation sub-steps:

[0146] First, outdoor-indoor concentration exposure correction (i.e., indoor-outdoor environmental exposure correction):

[0147] Considering the significant differences in indoor living environments among COPD patients, in order to convert the macroscopic outdoor environmental concentration into a more accurate individual exposure dose that reflects the patient's actual situation, the present invention designs the following exposure correction module in a preferred embodiment:

[0148] ;

[0149] in: This refers to environmental factors; for Time of the first Indoor concentrations of various environmental factors; for Time of the first Outdoor concentrations of various environmental factors (directly taken from) ); Permeability coefficient; for The air exchange rate at any given time is a dynamic variable determined by the building age (baseline exchange rate) and real-time patient behavior (such as opening windows for ventilation); For the first Indoor settlement rate due to various environmental factors.

[0150] In this embodiment, the specific parameters are set according to the ASHRAE standard:

[0151] : It is 1.0; It is 0.5; It is 0.7; It is 0.8; It is 0.7.

[0152] : It is 0.2 ACH. For 3ACH, It is 0.5 ACH. It is 0.8 ACH. It is 1.5 ACH.

[0153] The air exchange rate is a baseline value obtained from a standard building code database based on the age and type of the building where the patient resides. The duration of time the patient opens their window each day is determined by... (4 hours corresponds to 0.5 ACH), simulated infiltration ventilation rate caused by wind pressure and thermal pressure. The value obtained by summing up.

[0154] Second, the calculation of respiratory deposition dose is as follows:

[0155] The actual mass of contaminants deposited in the lungs was calculated using the ICRP lung deposition model. :

[0156] ;

[0157] in: If the patient is Always outdoors, value If the patient is Always indoors, value ; The total hourly ventilation is based on activity intensity, expressed as minute ventilation. The time unit conversion yields the result. Values: 5-8 L / min for resting state, 10-15 L / min for light activity, and 20-40 L / min for moderate activity; This is a preset constant specific to the pollutant, representing the total deposition fraction of the pollutant in the trachea-bronchus and alveolar regions, based on the official model of the ICRP (International Commission on Radiological Protection).

[0158] Obtain the mass of pollutants deposited in the lungs on day T. Then we have:

[0159] ;

[0160] Where: h is the hourly time index.

[0161] Third, the calculation of the hysteresis and cumulative effects of inflammation (i.e., the post-inflammatory phase and the cumulative portion) is as follows:

[0162] Based on a first-order dynamical convolution model, and through biological half-life... A parameterized exponential decay function for the past The mass of pollutants deposited in the lungs within a day is weighted and summed to output a bioeffective dose that quantifies the hysteresis effect of inflammation. The formula is as follows:

[0163] ;

[0164] in: The bioeffective dose for the delayed inflammatory effect on day T; The number of days in lag; The specific biological half-life of the p-th environmental factor for the rate of inflammatory response; Total number of days in the past; The time scale is a day, representing the cumulative value of 24 hours.

[0165] Quantification module of biostimulation effects of meteorological elements (i.e., the meteorological environment sequence part), specifically: for temperature, humidity, and wind speed, perform wind-cold coupled stimulation index calculation, calculate the wind-cold coupled stimulation index (WCSI) for characteristic cold air stimulation intensity, and calculate the humidity deviation index. The calculation is as follows:

[0166] Specialized meteorological element group Including temperature Relative humidity (H) and wind speed (V).

[0167] Wind-cold coupling stimulation index Considering the amplifying effect of wind speed on cold stimulation, this module does not directly use the original temperature, but instead calculates the effective stimulation index under wind-cold coupling, as shown in the following expression:

[0168] ;

[0169] in: The mean wind-cold coupling stimulation index on day T; The average actual temperature on day T; The average perceived effective temperature on day T;

[0170] Average perceived effective temperature on day T The following empirical formula is used for construction:

[0171] ;

[0172] in: The average wind speed on day T; This is the correction factor for wind chill.

[0173] The humidity deviation index is calculated as follows:

[0174] ;

[0175] Among them: Among them: The average humidity deviation index for day T; The average relative humidity on day T; The humidity level is the optimal level for respiratory function.

[0176] The module for quantifying the biostimulation effects of all meteorological elements is as follows:

[0177] .

[0178] Step 2.2: Perform a characteristic cascade of the bioeffective dose of inhalable pollutants and the biostimulation effects of meteorological elements to form a unified bioeffect vector. (i.e., the part that obtains the biological effect vector) Specifically, the outputs of the two modules mentioned above are concatenated to form a unified biological effect vector that includes multiple biological mechanisms. .

[0179] In addition, this embodiment also includes constructing a total dataset for causal inference, specifically, constructing a total dataset for the causal inference model. ,as follows:

[0180] ;

[0181] in: To count the total number of days; For the first Individualized effector data for each patient. ; For the first A vector of clinical health outcomes for each patient.

[0182] Step S3: Construct and train a causal feature learning network. Specifically, receive the biological effect vector output from step S2 and the effect modification factor obtained from step S1, extract high-order causal features through a deep network containing temporal convolution and feature fusion, and train an auxiliary prediction function for subsequent causal inference.

[0183] In this embodiment, a causal feature learning network based on temporal convolution and feature fusion is preferably constructed. The aim is to build an efficient feature extraction and auxiliary prediction network responsible for two core tasks: First, extracting dynamic patterns from environmental sequences using multi-scale convolution. Specifically, the environmental temporal pattern extraction module is configured with multiple one-dimensional convolutional kernels of different sizes in parallel to capture dynamic patterns at acute, short-term, and chronic time scales of environmental exposure. Second, integrating the discrete modifying factors (genes, lifestyle) constructed in step S1 into a unified patient state feature vector using a deep feature fusion layer. This provides accurate information describing individual characteristics for the design of step S4. Specifically, a dual-head prediction architecture with input feature isolation is used to train auxiliary prediction functions for subsequent causal inference. These auxiliary prediction functions include functions for estimating baseline risk. and the function used to estimate exposure propensity .

[0184] More preferably, the network receives the total dataset constructed in step S2. As input, the design includes the following four key functional modules:

[0185] First, the environmental timing pattern extraction module, as follows:

[0186] To address the limitations of traditional methods that treat genetic traits as independent numerical values, this paper explicitly models the topological structure and synergistic effects of gene loci in biological pathways.

[0187] This method extracts pathologically significant dynamic exposure patterns from daily biological effect vector sequences, addressing the problem of information loss caused by traditional methods that only use daily values.

[0188] Input: A sequence of biological effect vectors over the past L days (preferably L=14) (i.e., historical biological effect vectors) The patient's genetic risk vector .

[0189] Specific operations:

[0190] Parallel processing using multi-scale one-dimensional causal convolution (i.e., the multi-scale causal convolution part):

[0191] ;

[0192] in: Here, BatchNum is the activation function; BatchNum is the batch normalization function. This represents a one-dimensional causal convolution operation with a kernel size of k; Let be the dilation rate of the k-th convolutional branch; k represents the number of parallel convolutional kernels with kernel sizes of k and k respectively. Specifically: k Capture the instantaneous stimulus characteristics of the day; k To capture short-term, sustained fluctuation characteristics; k 7. Capture weekly chronic cumulative features; perform global max pooling on the outputs of the three convolutional layers and concatenate them to form an environmental temporal pattern vector. .

[0193] Second, the deep fusion module for individualized modification features, as follows:

[0194] The genetic risk vector of the patient preprocessed in step S1 (i.e., genetic risk vector) and lifestyle and clinical trait vector (i.e., clinical feature vectors) undergo nonlinear mapping and dimensionality reduction to generate high-density patient state features:

[0195] Feature concatenation is performed using a fully connected embedding layer: Implicit associations between features are learned through a multilayer perceptron (MLP): Ultimately, a fixed-dimensional patient state feature vector is generated. (i.e., the part involving deep integration of individual characteristics);

[0196] Third, the dual-head auxiliary prediction module, as follows:

[0197] Two auxiliary models are trained using the extracted features to remove confounding factors in step S4.

[0198] Input: Joint features combining environmental patterns and patient status .

[0199] Double-head structure:

[0200] 1. Basic Risk Prediction Head The basic risk component is as follows:

[0201] Predicting baseline disease without environmental exposure bias:

[0202] ;

[0203] in: This represents the predicted probability of symptoms occurring on that day. It is a learnable weight matrix, where each weight element represents the marginal contribution rate of a single feature to the risk of underlying disease. This is a scalar parameter representing the intercept of the average baseline incidence rate in the population.

[0204] 2. Exposure propensity prediction head The part about exposure tendency is as follows:

[0205] Predicting expected environmental exposure levels based on individual socioeconomic status and other characteristics:

[0206] ;

[0207] in: This is the predicted effective biological dose vector for the day; This is a learnable parameter matrix, representing the influence of environmental factors on patient characteristics; This is the bias vector, representing the average level of environmental exposure.

[0208] Fourth, optimize the objective and loss function as follows:

[0209] Training strategy: Employ multi-task supervised learning. Utilize real-world clinical health outcomes at historical target moments. The actual bioeffective dose (i.e., bioeffect vector) at the historical target time. As a supervisory label, it is used to jointly update the gradient of the feature extraction layer and the prediction head parameters in the network.

[0210] Loss function design: A composite loss function is constructed, employing the binary cross-entropy loss function (BCE) for the basic risk prediction head and the mean squared error loss function (MSE) for the exposure propensity prediction head. By minimizing the weighted sum of the two, the network can extract the environmental temporal patterns and patient state features with the highest causal discrimination.

[0211] Step S4: Calculate the individualized environmental sensitivity. Specifically, based on the biological effect vector obtained in step S2, the health outcome data obtained in step S1, and the auxiliary prediction function obtained in step S3, the individualized environmental sensitivity vector of the patient is obtained using a dual machine learning architecture and a hypernetwork model.

[0212] This step is based on a dual machine learning architecture, designing a parameter generator that, based on the patient state features provided in step S3, The corresponding causal effect parameter matrix is ​​dynamically generated, which mathematically simulates the biological mechanism that "individual constitution determines its response coefficient to the environment".

[0213] In this embodiment, the preferred steps for calculating individualized environmental sensitivity include:

[0214] Step S4.1: Use the auxiliary prediction function obtained in step S3 for subsequent causal inference to perform residual purification, and obtain the health outcome residual and environmental exposure residual (i.e., the residual purification part).

[0215] Step S4.2: Construct a dual machine learning architecture and supernetwork model consisting of a parameter generator subnetwork HyperNet and a causal inference main network PrimaryNet. The parameter generator subnetwork HyperNet is the first... The patient and the first Using the state feature vectors of various environmental factors as input, a patient's individualized environmental sensitivity vector is dynamically generated. The causal inference main network, PrimaryNet, utilizes the aforementioned... As regression coefficients, they fit the residuals of healthy outcomes within the residual space.

[0216] Details are as follows:

[0217] First, residual purification, as follows:

[0218] This step eliminates confounding biases introduced by the patient's baseline physical condition and exposure preferences, constructing a pure causal inference space. Using the auxiliary head trained in step S3, the residuals are calculated:

[0219] ① Residual of the ending: This represents symptom fluctuations that cannot be explained by the underlying medical condition. The actual clinical health outcome at the target historical moment (COPD incidence: a binary variable of 0 or 1, where 1 represents the onset of the disease). The predicted probability of symptom occurrence on the current day (a continuous variable between 0 and 1); The residual is essentially a mathematical centering of the binary outcome, responsible for stripping away the baseline probability components explained by confounding factors and extracting the pure residual signal containing only environmental causal effects.

[0220] ② Exposed residuals: This represents a sudden environmental stimulus. The actual bioeffective dose (bioeffect vector) at the historical target time. This is the predicted effective biological dose vector for the day; The residual performs a debiasing operation on the treatment variable, which is responsible for removing the endogeneous exposure bias caused by confounding factors such as the patient's socioeconomic status or lifestyle habits, and extracting the exogenous stimulus signals caused only by external environmental fluctuations.

[0221] Second, the Med-Hyper-PrimaryNet hypernetwork, which modifies environmental effects, is constructed. This step aims to build a generative causal inference model and designs a hypernetwork architecture that incorporates the patient's state feature vectors. The mapping is the weight parameter space of the main network. Given that the individualized environmental sensitivity vector of patients is not directly observable in reality, this invention designs a coupled architecture for parameter generation and effect verification. By fitting the observable health outcome residuals, the unobservable causal coefficients are solved in reverse. Specifically, it includes the following modules:

[0222] The parameter generator subnetwork HyperNet fits the latent vector; this step uses a nonlinear mapping operator. The patient state feature vector in the feature space Project the data onto the parameter space to generate a causal effect coefficient vector specific to that patient.

[0223] Input layer: Receives patient state feature vectors from the output of S3. .

[0224] Hidden layers: A multi-layer fully connected neural network is used, with batch standardization between layers to prevent covariate shift, and the LeakyReLU activation function is used.

[0225] Output layer: A linear transformation layer is used to directly output the patient's individualized environmental sensitivity vector. , represented as , For the first Environmental factors on the first The weights for each patient are calculated. This layer does not use a non-linear activation function, as follows:

[0226] ;

[0227] in: These are the global parameters of the hypernetwork itself that need to be trained; These are the local model parameters for the i-th patient.

[0228] The causal inference main network, PrimaryNet, involves constructing a supervision loop. This step designs a dynamic linear regressor whose weight parameters are derived from the parameter generator. Its core function is to establish a mapping relationship from potential causal coefficients to observable health outcomes, providing a pathway for gradient backpropagation for the supernetwork.

[0229] Input data: Received environmental exposure residuals ;

[0230] Forward propagation: Performs a dot product operation between the input residuals and dynamic weights to fit the healthy outcome residuals, obtaining the predicted outcome residuals. ,as follows:

[0231] ;

[0232] in: This refers to the overall dimension of environmental factors; Representing the The patient on the first The conditionally averaged treatment effect of various environmental factors reflects the patient's genes and condition. Modified environmental exposure residuals of health outcomes The intensity of the impact; To predict the outcome residuals; For the first Environmental exposure to various environmental factors.

[0233] Third, the joint optimization objective and loss function are as follows:

[0234] Training strategy: Supervised learning is adopted because Unlabeled, the design will predict the outcome residuals. The residual between the actual ending and the final outcome The error between them is used as a supervision label for the hypernetwork parameters. Perform gradient updates.

[0235] Gradient backpropagation path: The error signal is first calculated relative to... The gradient is then backpropagated to the linear operation. Finally, the global parameters of (HyperNet) are further updated. .

[0236] Loss function design: Construct a mean squared error loss function and fit the predicted outcome residuals. Residual from the actual ending Minimize the difference between the model-predicted causal effect and the actual disease fluctuations after baseline removal:

[0237] ;

[0238] in: This represents the total loss function value. Total patient sample size; The length of the observation time window, in days; Regularization coefficient These are the global parameters of the hypernetwork itself that need to be trained; It is an L2 norm.

[0239] This design forces the hypernetwork to learn the intrinsic interpretation of its parameters by minimizing the loss function: the reasons for the fluctuations in the patient's condition under the current environmental exposure conditions.

[0240] The generation of individualized disease-environment sensitivity profiles involves, specifically: after training, the state vector of any patient... Inputting Med-enHyperNet directly yields a personalized environment sensitivity vector. .

[0241] Step S5: Generate clinical decision support information. Specifically, this involves fusing the individualized environmental sensitivity vector of the patient obtained in Step S4 with external environmental prediction data to obtain a visualized clinical analysis report. This includes environmental sensitivity digital profiling and dynamic closed-loop intervention, as follows:

[0242] This step combines the individualized environmental sensitivity vector output from step S4 with real-time environmental forecast data to transform it into executable clinical decision instructions. This step includes the following three processing modules:

[0243] First, sensitivity phenotype mapping based on vector space: specifically, mapping high-dimensional continuous individualized environmental sensitivity vectors to discrete clinical phenotype spaces. The system pre-configures percentile threshold vectors based on large-scale population distributions. For each environmental factor Execution threshold determination:

[0244] ;

[0245] in: To determine the environmental sensitivity type of the i-th patient based on the threshold results, This is a sensitivity phenotype.

[0246] according to The numerical characteristics of different components are classified as follows:

[0247] Sensitivity phenotype Q-segmentation includes:

[0248] ① Particulate matter susceptible type: , This indicates that the airway is highly sensitive to oxidative stress caused by the deposition of physical particles.

[0249] ② Chemically susceptible type: , , This indicates that the airway mucosa is highly sensitive to chemical damage from oxidizing or acidic gases.

[0250] ③ Meteorological and physical susceptibility type: , This indicates that the airway smooth muscle is highly sensitive to physical stimuli such as cold air (wind chill effect) or abnormal humidity.

[0251] ④. Complex susceptible type: simultaneously meets the criteria for two or more of types ①-③.

[0252] ⑤ Environmentally Tolerant Type: None of the above components exceed the preset high sensitivity threshold.

[0253] Final output: A set of discretized digital profile labels for the patient.

[0254] Second, the calculation and intervention recommendations for the predicted environmental risk index based on parameter coupling (i.e., environmental risk index prediction) specifically involves: constructing a coupling model between individual sensitivity parameters and future environmental fluctuations to quantify the expected risk of disease fluctuations within a specific future time window.

[0255] ;

[0256] in: For the first Heavenly The predicted environmental risk index for each patient represents the expected pathological burden after the excess environmental exposure is amplified by individual sensitivity. For the first Heavenly Meteorological forecast values ​​for various environmental factors; For the first Safety baseline values ​​for various environmental factors.

[0257] Based on these quantitative indicators, feasible intervention recommendations can be formulated.

[0258] Third, the generation of a visualization report for the multi-dimensional sensitivity profile involves rendering the above analysis results into a graphical user interface, including:

[0259] Sensitivity radar chart: The various components are mapped onto the radar axis, visually displaying the patient's risk weaknesses.

[0260] Population positioning map: Shows the percentile ranking of the patient's sensitivity among people of the same age / disease stage.

[0261] Clinical recommendations: Based on classification, generate reference prescription suggestions.

[0262] The core of this invention lies in constructing a computational framework that deeply integrates respiratory pathology mechanisms and causal inference algorithms by mining multimodal longitudinal historical data of patients, as detailed below:

[0263] 1. The biological effect reconstruction module based on a multi-mechanism model specifically classifies and processes multi-dimensional environmental factors. For inhalable pollutants, this module converts physical concentration into a bioeffective dose by introducing dynamic physiological parameters and a first-order inflammatory dynamics convolution model. For meteorological elements, it quantifies their biostimulation intensity through a nonlinear model based on physiological thresholds, ultimately outputting a unified biological effect vector.

[0264] 2. Causal Feature Learning Module: Specifically, it constructs a feature extraction network and uses multi-scale one-dimensional causal convolution to extract the dynamic exposure mode of biological effect sequences; at the same time, it uses a deep fusion layer to nonlinearly map the patient's genetic and clinical characteristics into a high-dimensional patient state feature vector.

[0265] 3. The dual machine learning module based on the hypernetwork is as follows: a parameter-generating causal inference architecture is constructed, the features extracted by the aforementioned module are used for orthogonalized residual purification, and the patient's state feature vector is mapped to the weight parameter space of the main network through the hypernetwork. Finally, the patient's multidimensional individualized environmental sensitivity vector is dynamically generated and calculated.

[0266] The solution of this invention ultimately outputs an interpretable, individualized digital profile of environmental sensitivity. This profile can quantify a patient's pathological susceptibility to specific environmental factors (such as particulate matter, cold air, etc.) and, combined with environmental forecasts, provide precise quantitative support for clinical decision-making. The solution of this invention has important clinical application value for realizing individualized risk stratification and prospective intervention for COPD.

[0267] Example 2:

[0268] This embodiment discloses a system for constructing environmental sensitivity profiles of COPD patients, implementing the method for constructing environmental sensitivity profiles of COPD patients described in Embodiment 1; the system includes:

[0269] The data collection unit is used to construct a multimodal patient historical data profile dataset;

[0270] Data processing unit: Equipped with a processor and memory, it is used for reconstructing biological effects based on respiratory pathology mechanisms, building and training causal feature learning networks, and calculating individualized environmental sensitivities;

[0271] User interface and output unit: Used to generate clinical decision support information.

[0272] The above description only discloses specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or modifications that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing an environmental sensitivity profile of a patient with COPD, characterized in that, Includes the following steps: Step S1: Construct a multimodal patient historical data profile dataset. Specifically, this involves: acquiring and organizing multi-source heterogeneous patient data, and constructing a structured dataset containing environmental exposure, effect modifiers, and health outcomes through spatiotemporal alignment and feature extraction. Acquiring and organizing multi-source heterogeneous patient data includes acquiring environmental exposure data, effect modifier data, and health outcome data. Acquiring environmental exposure data specifically includes: connecting to environmental data interfaces and generating daily high spatiotemporal resolution environmental exposure sequences for each patient, including particulate matter, gaseous pollutants, and meteorological elements, through spatial interpolation and temporal aggregation; acquiring effect modification factor data specifically includes: extracting and quantifying genetic risk characteristics and clinical characteristics from patients' electronic medical records and genetic testing reports; clinical characteristics include lifestyle characteristics and basic clinical indicators; acquiring health outcome data specifically includes: collecting and encoding patients' daily symptom load, acute exacerbations, and physiological fluctuation indicators; Step S2, biological effect reconstruction based on respiratory pathology mechanisms, specifically involves receiving the environmental exposure data from step S1, classifying it using a multi-mechanism model, and transforming it into a unified biological effect vector. This includes the following steps: Step 2.1, Calculation of bioeffective dose of inhalable pollutants, specifically: For particulate matter and gaseous pollutants, perform outdoor-indoor exposure correction, respiratory deposition dose calculation and inflammatory hysteresis and cumulative effect calculation in series; The quantification of the biostimulatory effects of meteorological elements specifically involves: calculating the wind-cold coupled stimulation index (WCSI) for temperature, humidity, and wind speed; calculating the WCSI used to measure the intensity of characteristic cold air stimuli; and calculating the humidity deviation index. calculate; Step 2.2: Perform a characteristic cascade between the bioeffective dose of inhalable pollutants and the biostimulation effects of meteorological elements to form a unified bioeffect vector. ; Step S3: Construct and train a causal feature learning network. Specifically, the network receives the biological effect vector output from step S2 and the effect modification factor obtained from step S1. It then extracts high-order causal features through a deep network containing temporal convolution and feature fusion, and trains an auxiliary prediction function for subsequent causal inference. The causal feature learning network includes an environmental temporal pattern extraction module and a dual-head prediction architecture with input feature isolation. The environmental temporal pattern extraction module is configured with multiple one-dimensional convolutional kernels of different sizes in parallel to capture the dynamic patterns of acute, short-term, and chronic environmental exposures at different time scales. The dual-head prediction architecture with input feature isolation trains auxiliary prediction functions for subsequent causal inference, including a function for estimating the underlying risk. and the function used to estimate exposure propensity ; Step S4: Calculate individualized environmental sensitivity. Specifically, based on the biological effect vector obtained in step S2, the health outcome data obtained in step S1, and the auxiliary prediction function obtained in step S3, the individualized environmental sensitivity vector of the patient is obtained using a dual machine learning architecture and a hypernetwork model. This includes the following steps: Step S4.1: Use the auxiliary prediction function obtained in step S3 for subsequent causal inference to perform residual purification, and obtain the health outcome residual and environmental exposure residual; Step S4.2: Construct a dual machine learning architecture and supernetwork model consisting of a parameter generator subnetwork HyperNet and a causal inference main network PrimaryNet. The parameter generator subnetwork HyperNet is used as the first... Using the patient's state feature vector as input, a patient's individualized environmental sensitivity vector is dynamically generated. The causal inference main network, PrimaryNet, utilizes the aforementioned... As regression coefficients, they fit the residuals of healthy outcomes within the residual space; Step S5: Generate clinical decision support information, specifically by fusing the patient's individualized environmental sensitivity vector obtained in step S4 with external environmental prediction data to obtain a visualized clinical analysis report.

2. The method for constructing an environmental sensitivity profile of COPD patients according to claim 1, characterized in that, The specific steps in step S1 for constructing the multimodal patient historical data profile dataset include: Step S1.1: Acquire and organize multi-source heterogeneous data, specifically: acquire environmental exposure data, effect modifier data, and health outcome data; Step S1.2: Perform multi-scale data alignment, specifically: downsample the environmental exposure data to daily frequency to obtain aligned environmental exposure data; upsample and expand the effect modification factor data to daily time series to obtain aligned effect modification factor data. Step S1.3: Separate the aligned environmental exposure sequences into pollutant groups and meteorological element groups based on their physicochemical properties; integrate the aligned effect modification factor data and health outcome data to construct a structured dataset containing environmental exposure, effect modification factors, and health outcomes.

3. The method for constructing an environmental sensitivity profile of COPD patients according to claim 2, characterized in that, The construction of effector modifier data includes the extraction of genetic risk features based on pathological functional clusters, specifically including: Step ①: Multiple significant genetic susceptibility loci from genome-wide association analysis were mapped to three independent pathological functional clusters based on their biological functions: epithelial barrier defense, oxidative stress metabolism, and protease-antiprotease homeostasis. Specifically, epithelial barrier defense was mapped using... =1 indicates that oxidative stress metabolism is employed. =2 indicates that protease-antiprotease steady state is adopted. =3 indicates; Step 2: For each pathological functional cluster, independently calculate the weighted sum of effect values ​​for its internal risk sites to obtain the components. ; Step ③: Take the components described in step ② Combining these elements creates a multidimensional genetic risk vector. ,in: .

4. The method for constructing an environmental sensitivity profile for COPD patients according to claim 3, characterized in that, Constructing effect modifier data also includes the quantification and combination of clinical features, specifically including: Step ①: Calculate the smoking exposure index, physical activity level based on metabolic equivalents, and a dietary inflammation index based on a weighted sum of multiple nutritional parameters to obtain behavioral characteristics; Step 2: Standardize the three basic clinical indicators, which include continuous physiological parameters, heterogeneous inflammatory markers, and discrete biological parameters, to obtain the basic clinical indicator characteristics; the continuous physiological parameters include age, BMI, and FEV1 / FVC; the heterogeneous inflammatory marker is selected as blood eosinophil count; the discrete biological parameter is selected as sex. Step ③: Concatenate the behavioral characteristics from Step ① with the basic clinical indicator characteristics from Step ② to form a clinical feature vector. .

5. The method for constructing an environmental sensitivity profile for COPD patients according to claim 1, characterized in that, The outdoor-indoor exposure correction specifically involves: employing a steady-state mass balance equation and introducing a factor determined by building characteristics and patient factors. Dynamic hourly air exchange rate parameter determined by real-time window opening behavior This converts outdoor pollutant concentrations into individual microenvironment exposure concentrations. The calculation of respiratory deposition dose specifically involves: using the ICRP lung deposition model and introducing a dose based on the patient's... Dynamic hourly ventilation parameters obtained from real-time physical activity intensity Obtain the mass of pollutants deposited in the lungs on day T. ; The calculation of inflammatory hysteresis and cumulative effects specifically involves: employing an exponentially decaying convolutional model based on biological half-life to analyze past... The mass of pollutants deposited in the lungs within a day is weighted and summed to output a bioeffective dose that quantifies the hysteresis effect of inflammation, as shown in the following formula: ； in: The bioeffective dose for the delayed inflammatory effect on day T; The number of days in lag; The specific biological half-life of the p-th environmental factor for the rate of inflammatory response; This represents the total number of days in the past.

6. The method for constructing an environmental sensitivity profile of COPD patients according to claim 1, characterized in that, The wind-cold coupling stimulation index is calculated as follows: a nonlinear model based on physiological thresholds is used to couple actual temperature with wind speed to calculate the wind-cold coupling stimulation index for the intensity of characteristic cold air stimulation. The expression is as follows: ； in: The mean wind-cold coupling stimulation index on day T; The average actual temperature on day T; The average perceived effective temperature on day T; Average perceived effective temperature on day T The following empirical formula is used for construction: ； in: Let T be the average wind speed on day T. This is the wind chill correction factor; The humidity deviation index is calculated by taking the difference between the daily relative humidity and the median humidity suitable for respiratory function, in order to quantify the intensity of stimulation from dry or humid environments. The expression is as follows: ； in: The average humidity deviation index for day T; The average relative humidity on day T; The humidity level is the optimal level for respiratory function.

7. The method for constructing an environmental sensitivity profile of COPD patients according to claim 1, characterized in that, The patient's individualized environmental sensitivity vector Represented as , For the first Environmental factors on the first The weight of each patient.

8. The method for constructing an environmental sensitivity profile of COPD patients according to claim 7, characterized in that, The mathematical model of the causal inference main network, PrimaryNet, is as follows: ； in: This refers to the overall dimension of environmental factors; It reflects the patient's genes and condition. Modified environmental exposure residuals of health outcomes The intensity of the impact; To predict the outcome residuals; For the first Environmental exposure to various environmental factors.

9. The method for constructing an environmentally sensitive profile of a COPD patient according to claim 8, characterized in that, The loss function of the hypernetwork model includes a mean squared error loss function with an L2 regularization term, and its mathematical form is as follows: ； in: This represents the total loss function value. Total patient sample size; The length of the observation time window, in days; The regularization coefficient is used. These are the global parameters of the hypernetwork itself that need to be trained; It is an L2 norm.

10. The method for constructing an environmentally sensitive profile of a COPD patient according to claim 1, characterized in that, The step S5 of generating clinical decision support information includes: constructing a predictive risk index calculation model based on parameter coupling, which couples and weights the future environmental forecast values ​​with the individualized environmental sensitivity vector.

11. The method for constructing an environmentally sensitive profile of a COPD patient according to claim 10, characterized in that, The formula for calculating the predicted risk index PRI is as follows: ； in: For the first Heavenly The predicted environmental risk index for each patient represents the expected pathological burden after the excess environmental exposure is amplified by individual sensitivity. For the first Heavenly Meteorological forecast values ​​for various environmental factors; For the first Safety baseline values ​​for various environmental factors.

12. A system for constructing environmentally sensitive profiles of COPD patients, characterized in that, The method for constructing environmentally sensitive profiles of COPD patients as described in any one of claims 1-11; the system includes: The data collection unit is used to construct a multimodal patient historical data profile dataset; Data processing unit: Equipped with a processor and memory, it is used for reconstructing biological effects based on respiratory pathology mechanisms, building and training causal feature learning networks, and calculating individualized environmental sensitivities; User interface and output unit: Used to generate clinical decision support information.

Images