Tear total IGE detection method and system based on multi-modal data fusion

Abstract

The invention discloses a tear total IGE detection method and system based on multi-modal data fusion. The method comprises the following steps: acquiring ocular surface images, physiological parameters and environmental allergen data, and acquiring historical medical data from a hospital information system; preprocessing and feature extraction are performed, patient symptom description is analyzed in combination with a natural language processing technology, and related image block features are recognized; adjusting a space-time attention weight based on frequency domain analysis and transfer learning, inputting a lightweight Transform model, and predicting a tear total IgE concentration and an allergy risk score; whether a trace tear sampling program is started or not is judged, and the actually measured concentration of total tear IgE is detected through a disposable micro-fluidic chip; and integrating all information by using a Bayesian adaptive filtering algorithm to determine the total tear IgE level. By implementing the method provided by the invention, the sampling amount is extremely small, the detection time is short, the environmental interference can be dynamically corrected, and the absolute concentration value is output, so that the dual requirements of clinical precise diagnosis and treatment and home monitoring are met.

Description

Technical Field

[0001] This invention relates to big data, and more specifically to a method and system for detecting total IGE in tears based on multimodal data fusion. Background Technology

[0002] Total immunoglobulin E in tears is an important biomarker for ocular surface diseases such as allergic conjunctivitis, vernal keratoconjunctivitis, and some types of dry eye. Currently, commonly used clinical methods for detecting total IgE in tears mainly include enzyme-linked immunosorbent assay (ELISA), immunoturbidimetry, and fluorescence immunochromatography.

[0003] These methods share the following drawbacks: they require the collection of 5–10 μL of tear fluid, while the normal basal tear volume in adults is only about 7 μL, making them extremely unfriendly to infants, dry eye patients, or postoperative individuals, resulting in poor compliance; the centralized laboratory testing model leads to reporting cycles of 2–6 hours, failing to capture minute-level fluctuations in IgE levels during pollen exposure and acute allergic reactions; and existing methods neglect pollen and PM2.5. 2.5 The real-time impact of environmental allergens such as temperature and humidity on IgE levels leads to a high false negative / false positive rate; it only outputs a single concentration value and does not integrate ocular surface morphology (conjunctival hyperemia, papillary hyperplasia), physiological parameters (tear film breakup time, tear secretion rate) and individual history (baseline IgE, medication history), making it difficult to achieve personalized threshold setting; it relies on large-scale immunoassay analyzers, making it difficult to popularize in primary healthcare institutions or home settings.

[0004] In recent years, although some studies have attempted to achieve micro-sampling using tear test strips or microfluidic chips, invasive sampling is still required, and the dynamic modeling of environmental variables and individual characteristics remains unresolved. On the other hand, the application of artificial intelligence in medical image analysis is becoming increasingly mature, but current technologies are limited to static image classification, have not achieved cross-modal fusion of multimodal data, and have not established a closed-loop system for non-invasive prediction and micro-invasive calibration.

[0005] Therefore, it is necessary to design a new method that achieves extremely small sampling volume, short detection time, dynamic correction of environmental interference, and output of absolute concentration values ​​to meet the dual needs of precise clinical diagnosis and home monitoring. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for detecting total IGE in tears based on multimodal data fusion.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a method for detecting total IGE in tears based on multimodal data fusion, comprising:

[0008] Multimodal data is obtained by simultaneously acquiring ocular surface images, physiological parameters, and environmental allergen data using cameras, tear film sensors, and environmental sensors, as well as retrieving users' historical medical data from hospital information systems.

[0009] The multimodal data is preprocessed and aligned to obtain the preprocessing result;

[0010] We obtain patient symptom descriptions, extract keywords from these descriptions using natural language processing (NLP) technology, construct symptom keyword chains, and then perform semantic expansion using a medical knowledge graph to obtain the symptom text.

[0011] Calculate the cosine similarity between ocular surface image patches and symptom text feature vectors within the preprocessed results to identify image patch features related to symptoms;

[0012] By combining spatial correlation maps and causal relationship matrices, and through frequency domain analysis and progressive transfer learning, the spatiotemporal attention weights are dynamically adjusted to enhance the understanding of environmental allergen data.

[0013] The image patch features, spatiotemporal attention weights, physiological parameters, environmental allergen data, and historical medical data are fused and input into a lightweight Transformer model to generate the predicted concentration of total tear IgE and allergy risk score, and to classify the risk level.

[0014] Whether to trigger a micro-tear sampling command is determined based on the predicted concentration of total IgE in the tear fluid and the allergy risk score;

[0015] When the micro-tear sampling command is triggered, the disposable microfluidic chip is activated to collect a 0.5 μL tear sample through a capillary glass tube or tear test strip adsorption device;

[0016] The collected tear samples are imported into a microfluidic chip and detected using a fluorescence immunoassay or surface plasmon resonance technique. The measured concentration of total IgE in the tear is read within a few minutes.

[0017] Based on the Bayesian adaptive filtering algorithm, the spatiotemporal attention weights are dynamically adjusted by combining the predicted concentration of total tear IgE, the measured concentration of total tear IgE, and the baseline IgE value in the historical medical data, and the final reported value is calculated to obtain the total tear IgE value.

[0018] The further technical solution is as follows: Calculating the cosine similarity between the ocular surface image patch and the symptom text feature vector within the preprocessed result to identify image patch features related to the symptoms includes:

[0019] The preprocessed ocular surface image is segmented and encoded using a visual Transformer to obtain encoded image blocks;

[0020] The symptom text is encoded using the BERT model to obtain text feature vectors;

[0021] The cosine similarity between the encoded image patch and the text feature vector is calculated, and image patches that meet the matching requirements related to symptoms are selected to obtain image patch features related to symptoms.

[0022] Its further technical solution is as follows: Combining spatial correlation maps and causal relationship matrices, and dynamically adjusting spatiotemporal attention weights through frequency domain analysis and progressive transfer learning to enhance the understanding of environmental allergen data, including:

[0023] A patient spatial association map and a cross-scale causal relationship matrix are constructed. A fast Fourier transform is performed on the pollen concentration sequence in the environmental allergen data to generate frequency domain attention gain. A progressive transfer weight function is constructed for users whose data volume does not meet the requirements, and the spatiotemporal attention weight is dynamically calculated.

[0024] The further technical solution is as follows: The process involves constructing a patient spatial correlation map and a cross-scale causal relationship matrix; performing a fast Fourier transform on the pollen concentration sequence within the environmental allergen data to generate a frequency domain attention gain; and constructing a progressive transfer weight function for users whose data volume does not meet requirements, dynamically calculating spatiotemporal attention weights, including:

[0025] Obtain patient geographic location information, calculate the geographic distance and pollen exposure similarity for each pair of patients, and construct a patient spatial association map;

[0026] Define time scale hierarchy, extract feature data of each time scale, use Granger causality test to identify causal relationships between features of different time scales, and construct cross-scale causal relationship matrix;

[0027] Pollen concentration time series was extracted from environmental allergen data, preprocessed, and then converted to the frequency domain using Fast Fourier Transform. Attention gain corresponding to different frequency components was calculated to obtain frequency domain attention gain.

[0028] Set a data volume threshold, calculate the user similarity matrix, determine the data sufficiency index of users whose data volume does not meet the requirements, and design a progressive migration weight function.

[0029] By integrating information from patient spatial correlation maps and cross-scale causal relationship matrices, frequency domain attention gain is combined, and a progressive transfer weighting function is applied to adjust the spatiotemporal attention weights.

[0030] The further technical solution is as follows: the lightweight Transformer model includes an input embedding layer, multiple lightweight Transformer encoders, and an output layer; the input embedding layer converts the comprehensive feature vector into a fixed-dimensional embedding vector; multiple lightweight Transformer encoders process the embedding vectors sequentially, each encoder containing a multi-head self-attention mechanism and a feedforward neural network, wherein the multi-head self-attention mechanism is used to capture the correlation between features, and the feedforward neural network is used for nonlinear transformation; the output layer is based on the results output by the encoder.

[0031] The further technical solution is as follows: The image patch features, spatiotemporal attention weights, physiological parameters, environmental allergen data, and historical medical data are fused and input into a lightweight Transformer model to generate a predicted concentration of total tear IgE and an allergy risk score, and to classify risk levels, including:

[0032] Image patch features, spatiotemporal attention weights, physiological parameters, environmental allergen data, and historical medical data are each converted into fixed-length vectors and then concatenated and fused in sequence to form a comprehensive feature vector.

[0033] Convert the composite feature vector into a fixed-dimensional embedding vector;

[0034] The correlation between features in the embedding vector is calculated, self-attention weights are generated, and nonlinear transformation is performed on the features output by the self-attention mechanism to obtain the encoder output.

[0035] Based on the encoder output, the predicted concentration of total tear IgE and the allergy risk score are generated, and the risk level is classified.

[0036] The further technical solution is as follows: Based on the Bayesian adaptive filtering algorithm, combined with the predicted concentration of total tear IgE, the measured concentration of total tear IgE, and the baseline IgE value in the historical medical data, the spatiotemporal attention weight is dynamically adjusted to calculate the final reported value, thereby obtaining the total tear IgE value, including:

[0037] Initialize weight parameters to balance the effects of non-invasive predicted concentrations, measured concentrations, and historical baseline values;

[0038] The difference between the predicted concentration of total tear IgE and the measured concentration of total tear IgE is calculated, and the weight parameters are updated according to the Bayesian adaptive filtering algorithm.

[0039] The predicted concentration of total tear IgE and the measured concentration of total tear IgE are weighted and summed to obtain a fusion concentration value. This fusion concentration value is then combined with the baseline IgE value and weighted to obtain a final reported value, which is the total tear IgE value.

[0040] The further technical solution is as follows: after dynamically adjusting the spatiotemporal attention weights and calculating the final reported value based on the Bayesian adaptive filtering algorithm, combined with the predicted concentration of total tear IgE, the measured concentration of total tear IgE, and the baseline IgE value in the historical medical data, to obtain the total tear IgE value, it also includes:

[0041] The training dataset was expanded using the total IGE value of the tear fluid and the corresponding multimodal data.

[0042] We used the expanded training dataset to fine-tune the spatiotemporal attention weights, image-text alignment parameters, and other parameters online.

[0043] This invention also provides a tear total IGE detection system based on multimodal data fusion, comprising:

[0044] The multimodal data acquisition unit is used to simultaneously acquire ocular surface images, physiological parameters, environmental allergen data, and retrieve users' historical medical data from the hospital information system to obtain multimodal data using a camera, tear sensor, and environmental sensor.

[0045] A preprocessing unit is used to preprocess and align the multimodal data to obtain preprocessing results;

[0046] The text processing unit is used to obtain patient symptom descriptions, extract keywords from the patient symptom descriptions using natural language processing technology and construct symptom keyword chains, and simultaneously perform semantic expansion by combining medical knowledge graphs to obtain symptom text.

[0047] The image patch feature calculation unit is used to calculate the cosine similarity between the ocular surface image patch and the symptom text feature vector in the preprocessing result, so as to identify image patch features related to symptoms;

[0048] The adjustment unit is used to combine spatial correlation maps and causal relationship matrices to dynamically adjust spatiotemporal attention weights through frequency domain analysis and progressive transfer learning, thereby enhancing the understanding of environmental allergen data.

[0049] The prediction unit is used to fuse the image patch features, spatiotemporal attention weights, physiological parameters, environmental allergen data, and historical medical data, input them into a lightweight Transformer model, generate the predicted concentration of total tear IgE and allergy risk score, and classify the risk level.

[0050] The judgment unit is used to determine whether to trigger a micro-tear sampling command based on the predicted concentration of total tear IgE and the allergy risk score.

[0051] The micro-sampling unit is used to activate a disposable microfluidic chip when a micro-tear sampling command is triggered, and collect a 0.5μL tear sample through a capillary glass tube or tear test paper adsorption device.

[0052] The detection unit is used to import the collected tear samples into the microfluidic chip and perform detection based on fluorescence immunoassay or surface plasmon resonance technology, and read the measured concentration of total IgE in the tear within a few minutes;

[0053] The final value calculation unit is used to dynamically adjust the spatiotemporal attention weights based on the Bayesian adaptive filtering algorithm, combined with the predicted concentration of total tear IgE, the measured concentration of total tear IgE, and the baseline IgE value in the historical medical data, to calculate the final reported value and obtain the total tear IgE value.

[0054] The present invention also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0055] The advantages of this invention compared to existing technologies are as follows: This invention simultaneously collects ocular surface images, physiological parameters, and environmental allergen data, and combines this with historical medical data from the hospital information system. It utilizes preprocessing and feature extraction techniques to identify image features related to symptoms; employs frequency domain analysis and progressive transfer learning to dynamically adjust spatiotemporal attention weights to enhance the understanding of environmental allergen data; inputs multiple data sources into a lightweight Transformer model to predict total tear IgE concentration and allergy risk score; determines whether to initiate a micro-tear sampling procedure based on the prediction results, requiring only 0.5 μL of sample; uses a disposable microfluidic chip for actual measurement; and obtains detection results within minutes based on fluorescence immunoassay or surface plasmon resonance technology; finally, applies a Bayesian adaptive filtering algorithm to integrate all information and calculate the final report value. This method achieves extremely small sampling volume, rapid detection time, dynamic correction of environmental interference, and output of absolute concentration values, meeting the dual needs of precise clinical diagnosis and home monitoring.

[0056] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0057] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a flowchart illustrating the tear total IGE detection method based on multimodal data fusion provided in an embodiment of the present invention.

[0059] Figure 2 This is a schematic block diagram of a tear total IGE detection system based on multimodal data fusion provided in an embodiment of the present invention;

[0060] Figure 3 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0062] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0063] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0064] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0065] Please see Figure 1 , Figure 1This is a schematic flowchart illustrating the tear total IgE detection method based on multimodal data fusion provided in this invention. The method is applied to a server that interacts with a high-resolution camera, tear sensor, and environmental sensor. It simultaneously acquires ocular surface images, physiological parameters, and environmental allergen data, and combines this with historical medical data from a hospital information system. Preprocessing and feature extraction techniques are used to identify image patch features related to symptoms. Frequency domain analysis and progressive transfer learning are employed to dynamically adjust spatiotemporal attention weights to enhance the understanding of environmental allergen data. Multiple data points are input into a lightweight Transformer model to predict tear total IgE concentration and allergy risk score. Based on the prediction results, a micro-tear sampling procedure is initiated, and actual measurement is performed using a disposable microfluidic chip. A Bayesian adaptive filtering algorithm is applied to integrate all information and calculate the final report value. This method achieves extremely small sampling volume (0.5 μL), rapid detection time, dynamic correction for environmental interference, and output of absolute concentration values, meeting the needs of precise clinical diagnosis and home monitoring.

[0066] Figure 1 This is a schematic flowchart of the tear total IGE detection method based on multimodal data fusion provided in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps S110 to S200.

[0067] S110 uses a camera, tear sensor, and environmental sensor to simultaneously acquire ocular surface images, physiological parameters, environmental allergen data, and retrieves the user's historical medical data from the hospital information system to obtain multimodal data.

[0068] In this embodiment, multimodal data refers to ocular surface images, physiological parameters, environmental allergen data, and the user's historical medical data collected synchronously through multiple sensors and data sources. These data together constitute the multi-dimensional feature input required by this invention, as detailed below:

[0069] Ocular surface image data:

[0070] Acquisition method: RGB and near-infrared (NIR) images of the ocular surface are acquired using a high-resolution camera (no less than 12 million pixels, 30fps) to observe ocular surface features such as conjunctival congestion and meibomian gland morphology.

[0071] Data characteristics: RGB images can intuitively reflect color changes such as redness and congestion on the ocular surface; NIR images can clearly show the morphological and texture changes of the meibomian glands, helping to identify potential problems such as meibomian gland dysfunction.

[0072] Data significance: Ocular surface images provide a visual foundation for subsequent cross-modal alignment of text and images. By segmenting and encoding the images using a visual Transformer (ViT), high-matching image patches related to symptoms are selected, enabling precise localization and correlation analysis between symptoms and ocular surface lesion areas.

[0073] Physiological parameter data:

[0074] Acquisition method: Physiological parameters such as tear secretion rate, tear film breakup time (BUT), and corneal sensitivity are monitored in real time using a tear sensor, with a sampling frequency of 100Hz.

[0075] Data characteristics: Tear secretion rate reflects tear production; BUT indicates the time tears remain on the ocular surface, reflecting tear film stability; corneal sensitivity is related to the function of ocular surface nerves.

[0076] Significance of the data: Physiological parameters provide important physiological basis for subsequent non-invasive prediction. These parameters are closely related to the physiological function of the ocular surface and can reflect the health status of the ocular surface from different perspectives. After being integrated with other data such as environmental allergen data and ocular surface images, the total IgE level of tears can be assessed more comprehensively.

[0077] Environmental allergen data:

[0078] Acquisition method: using a pollen laser counter, temperature and humidity sensor, and PM2.5. 2.5 The sensor collects environmental allergen data at a sampling frequency of 1 Hz.

[0079] Data characteristics: including pollen concentration (unit: grains / m³) 3 PM 2.5 Concentration (unit: μg / m³) 3 Information such as temperature and humidity (units: %RH and ℃) reflects the user's exposure to allergens in their environment.

[0080] Data significance: Environmental allergen data is used to dynamically correct non-invasive prediction results. Combined with spatiotemporal attention mechanisms, it provides the model with key weight information in the environmental and time dimensions, helping to quantify the impact of environmental factors on total tear IgE and improve the accuracy and personalization of predictions.

[0081] Historical medical data:

[0082] Acquisition method: Retrieve the user's allergy history, previous tear total IgE baseline value and medication records from the hospital information system (HIS) or electronic medical record (EMR) interface, and synchronize them offline in batches.

[0083] Data characteristics: Includes the user's past diagnosis and treatment of allergic diseases, types and dosages of medications, as well as historical measurements and trends of total IgE in tears.

[0084] Significance of the data: Historical medical data provides users with long-term health background information for the model, which helps to build personalized models. By combining baseline IgE values ​​and past responses, the current total IgE level of tears can be predicted more accurately, while also providing a reference for triggering decisions and dynamic calibration.

[0085] S120. The multimodal data is preprocessed and aligned to obtain the preprocessing result.

[0086] In this embodiment, the preprocessing result refers to the multimodal data after standardization, normalization, encoding transformation, and time alignment, enabling effective feature extraction and fusion in subsequent steps, as detailed below:

[0087] Numerical data normalization:

[0088] Target parameters: physiological parameters such as tear secretion rate, tear film breakup time (BUT), and corneal sensitivity, as well as pollen concentration and PM2.5 levels. 2.5 Data on environmental allergens, such as concentration, temperature, and humidity.

[0089] Processing method: The Z-score normalization method is used to process the numerical data to eliminate the influence of different units and ranges of units, so that the data has the characteristics of a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0090] One-hot encoding for categorical data:

[0091] Data processed includes categorized data such as pollen type and type of medication (e.g., antihistamines, glucocorticoids).

[0092] Processing method: Convert categorical data into one-hot encoded vectors. For example, encode pollen types (Artemisia, ragweed, etc.) into binary vector form so that the model can recognize and process categorical information.

[0093] Image data augmentation:

[0094] Processed objects: RGB images and near-infrared (NIR) images of the ocular surface.

[0095] Processing method: Adaptive histogram equalization and gamma correction are applied to the image data to improve image contrast and clarity, enhance the visualization of regions of interest in the image, and facilitate subsequent feature extraction and analysis.

[0096] Timestamp alignment:

[0097] Data processed: All collected multimodal data, including ocular surface images, physiological parameters, environmental allergen data, and historical medical data.

[0098] Processing method: Based on GPS-synchronized real-time clock (RTC), all data are uniformly timestamped to the millisecond level to ensure the consistency and synchronization of data from different modalities in the time dimension, providing a time basis for subsequent feature fusion and dynamic analysis.

[0099] After the above preprocessing steps, the preprocessed results include normalized physiological parameters and environmental allergen data, one-heat encoded classification data, and enhanced image data. All data have a unified timestamp, which can accurately reflect the temporal correspondence of the data. This provides high-quality, synchronous multimodal feature input for subsequent steps such as cross-modal alignment of text and image, spatiotemporal attention fusion, and non-invasive prediction.

[0100] S130. Obtain the patient's symptom description, extract keywords from the patient's symptom description using natural language processing technology and construct a symptom keyword chain, and simultaneously perform semantic expansion by combining medical knowledge graph to obtain the symptom text.

[0101] In this embodiment, symptom text refers to structured text information that accurately reflects the patient's current symptoms, obtained by performing natural language processing and semantic expansion on the patient's symptom description.

[0102] Obtaining Patient Symptom Descriptions: Patient symptom descriptions can be verbal descriptions provided to healthcare professionals during their visit, or textual information entered by the patient through mobile medical applications or smart medical devices. These typically include the patient's subjective feelings of eye discomfort, such as "itchy eyes," "increased eye discharge," "red eyes," "photophobia," and "teary eyes," as well as information on the time, frequency, and duration of symptom occurrence, such as "symptoms have lasted for 3 days, worsening in the morning."

[0103] The text describing the patient's symptoms is segmented into individual words using a word segmentation algorithm. For example, "Itchy eyes, the more you rub them, the redder they become" is broken down into words such as "eyes", "itchy", "the more you rub them", and "the redder they become".

[0104] Based on a predefined medical dictionary or keyword database, keywords related to eye symptoms, such as "itching" and "redness," are selected, while irrelevant stop words, such as "both eyes," "more," and "more," are removed.

[0105] The selected keywords are tagged with parts of speech to identify nouns, verbs, adjectives, etc., so as to better understand the role and meaning of the keywords in the sentence. For example, "itching" is tagged as a noun and "red" is tagged as an adjective.

[0106] Based on the logical relationship of the order in which the patient's symptoms appear or the severity of the symptoms, the extracted keywords are arranged into an ordered sequence to form a symptom keyword chain, such as "itching → congestion → increased tear secretion".

[0107] Based on the importance or frequency of keywords in symptom descriptions, each keyword is assigned a corresponding weight value. These weight values ​​can be used in subsequent steps such as cross-modal alignment of text and images to measure the relevance of keywords to image patches.

[0108] By matching and associating symptom keywords with entities in a medical knowledge graph, the medical concepts corresponding to the keywords can be identified. For example, "itching" is related to "allergic reaction," and "congestion" is related to "conjunctival inflammation."

[0109] By leveraging semantic relationships within medical knowledge graphs, we can uncover potential associations and hierarchical relationships between keywords. For example, "itching" may be a symptom of "allergic conjunctivitis," and "congestion" may be a sign of "conjunctival inflammation."

[0110] Based on semantic relationships, keywords are semantically expanded to supplement related medical terms or potential pathological mechanism-related vocabulary, such as expanding "itching" to "eosinophilic infiltration" and "mast cell degranulation," in order to enrich the semantic information of the symptom text and enhance its expressive power and diagnostic value in subsequent analysis.

[0111] The symptom text obtained after the above steps not only retains the key information of the patient's original symptom description, but also enriches the medical connotation of the symptoms through semantic expansion. This provides high-quality, structured text input for subsequent steps such as cross-modal alignment of text and images and non-invasive prediction, which helps improve the model's ability to understand and analyze patient symptoms, thereby more accurately predicting the total IgE level of tears and the level of allergy risk.

[0112] S140. Calculate the cosine similarity between the ocular surface image patch and the symptom text feature vector in the preprocessing result to identify image patch features related to symptoms.

[0113] In this embodiment, image patch features refer to the feature vectors obtained by segmenting the preprocessed ocular surface image into multiple image patches using a Visual Transformer (ViT) and encoding each image patch. Each image patch feature vector can characterize the visual information of the image patch, such as color, texture, and shape, and is used for subsequent matching with symptom text feature vectors.

[0114] In one embodiment, step S140 described above may include steps S141 to S143.

[0115] S141. Use a visual Transformer to segment and encode the preprocessed ocular surface image to obtain encoded image blocks.

[0116] In this embodiment, the preprocessed ocular surface image is input into the Visual Transformer (ViT) model.

[0117] The ViT model divides an image into multiple fixed-size image blocks, such as non-overlapping blocks of 64×64 pixels.

[0118] Each image patch is flattened and mapped to a high-dimensional space, generating a fixed-length feature vector, such as a 512-dimensional vector.

[0119] These feature vectors together form a set of encoded image patches, with each vector corresponding to a local region feature of the image.

[0120] S142. Encode the symptom text using the BERT model to obtain the text feature vector.

[0121] In this embodiment, the symptom keyword chain extracted and constructed using natural language processing technology is input into the BERT model.

[0122] The BERT model encodes symptom text to generate a fixed-length text feature vector, such as a 768-dimensional vector.

[0123] This feature vector can characterize the semantic information of the symptom keyword chain and reflect the symptom features described in the text.

[0124] S143. Calculate the cosine similarity between the encoded image patch and the text feature vector, and filter out image patches that meet the matching requirements related to symptoms to obtain image patch features related to symptoms.

[0125] In this embodiment, for each encoded image patch feature vector, its cosine similarity with the text feature vector is calculated.

[0126] The formula for calculating cosine similarity is: cosθ=(A·B) / (|A|·|B|), where A represents the image patch feature vector and B represents the text feature vector.

[0127] Based on a set similarity threshold (e.g., 0.85), image patches with a cosine similarity greater than or equal to the threshold are selected.

[0128] The selected image patches are image patch features related to symptoms. These image patches are visually highly correlated with the content of the symptom text description and may be allergy-related feature areas such as conjunctival congestion and meibomian gland morphology abnormalities.

[0129] S150 combines spatial correlation maps and causal relationship matrices, and dynamically adjusts spatiotemporal attention weights through frequency domain analysis and progressive transfer learning to enhance the understanding of environmental allergen data.

[0130] In this embodiment, a patient spatial association map and a cross-scale causal relationship matrix are constructed. A fast Fourier transform is performed on the pollen concentration sequence in the environmental allergen data to generate frequency domain attention gain. A progressive transfer weight function is constructed for users whose data volume does not meet the requirements, and the spatiotemporal attention weight is dynamically calculated.

[0131] In one embodiment, step S150 described above may include steps S151 to S155.

[0132] S151. Obtain patient geographic location information, calculate the geographic distance and pollen exposure similarity for each pair of patients, and construct a patient spatial association map.

[0133] In this embodiment, the patient spatial association graph is a graph structure where nodes represent patients and edges represent the spatial association strength between patients. The specific construction steps are as follows:

[0134] The geographical coordinates of each patient are obtained from the hospital information system or the GPS module of the user's device.

[0135] For each pair of patients, the geographical distance between them can be calculated using the Euclidean distance formula.

[0136] Collect pollen concentration data at the patient's location and calculate the similarity of pollen exposure patterns between each pair of patients. Similarity measurement methods such as Pearson correlation coefficient can be used.

[0137] By combining geographical distance and pollen exposure similarity, a correlation weight is assigned to each patient pair to construct a spatial correlation map of patients. The weight calculation formula can be expressed as:

[0138] w ij =α·exp(-β·d ij )+(1-α)·s ij Among them, w ij It is the association weight between patient i and patient j, d ij It is the geographical distance between them, s ij It represents pollen exposure similarity, with α and β being the modulating parameters.

[0139] S152. Define time scale hierarchy, extract feature data of each time scale, use Granger causality test to identify causal relationships between features of different time scales, and construct cross-scale causal relationship matrix.

[0140] In this embodiment, the cross-scale causal relationship matrix is ​​a matrix used to represent the strength of causal relationships between features at different time scales. The specific construction steps are as follows:

[0141] Based on characteristics such as changes in tear secretion and fluctuations in environmental allergens, time scales such as second-level, minute-level, hour-level, and day-level are defined.

[0142] Feature data for each time scale is extracted and represented, such as the second-level changes in tear secretion rate and the hour-level fluctuations in pollen concentration.

[0143] The Granger causality test is used to identify causal relationships between features at different time scales. For features X and Y, if the past value of X has a statistically significant predictive effect on the current value of Y, then X is considered to have a Granger causal relationship with Y.

[0144] Causal relationships are quantified into matrix form, forming a cross-scale causal relationship matrix. The elements in the matrix represent the strength of the causal relationship between features, with values ​​ranging from [0,1]. The larger the value, the stronger the causal relationship.

[0145] S153. Extract pollen concentration time series from environmental allergen data, preprocess it, apply Fast Fourier Transform to convert it to the frequency domain, calculate the attention gain corresponding to different frequency components, and obtain the frequency domain attention gain.

[0146] In this embodiment, frequency domain attention gain refers to the attention weight gain corresponding to different frequency components in the frequency domain, used to emphasize the pollen concentration fluctuation frequency that has a significant impact on total tear IgE. The specific calculation steps are as follows:

[0147] Pollen concentration time series were extracted from collected environmental allergen data.

[0148] The pollen concentration sequence was preprocessed, such as by noise removal and normalization.

[0149] The pollen concentration sequence was converted to the frequency domain using Fast Fourier Transform (FFT) to obtain the frequency-amplitude spectrum.

[0150] Based on the frequency-amplitude spectrum, calculate the attention gain corresponding to different frequency components. An amplitude-weighted method can be used, assigning higher attention gain to frequency components with larger amplitudes.

[0151] The calculated attention gain is normalized to fit within the range [0,1], thus obtaining the frequency domain attention gain.

[0152] S154. Set a data volume threshold, calculate the user similarity matrix, determine the data sufficiency index for users whose data volume does not meet the requirements, and design a progressive migration weight function.

[0153] In this embodiment, the user similarity matrix is ​​a single matrix representing the similarity between users, calculated based on multi-dimensional data such as user symptoms, physiological characteristics, and environmental exposure. Calculation methods can include cosine similarity, Euclidean distance, etc.

[0154] The progressive migration weight function is a function used to determine the migration weight from similar users based on user data sufficiency and similarity. The design steps are as follows:

[0155] Set a data volume threshold to determine whether a user has sufficient data.

[0156] Calculate the similarity matrix between users.

[0157] For users whose data volume does not meet the requirements, calculate their data sufficiency index. The data sufficiency index can be defined as the ratio of the user's valid data volume to the data volume threshold.

[0158] To design an incremental transfer weight function, an exponential decay function can be used:

[0159] TW i =exp(-γ·D i ); among them, TW i D is the migration weight of user i. i γ is the data sufficiency metric for user i, and γ is the decay coefficient. This function indicates that the lower the data sufficiency, the greater the migration weight.

[0160] S155. By integrating information from the patient spatial correlation map and the cross-scale causal relationship matrix, frequency domain attention gain is combined, and a progressive transfer weight function is applied to adjust the spatiotemporal attention weights.

[0161] In this embodiment, the spatial association information in the patient spatial association map and the causal relationship information in the cross-scale causal relationship matrix are fused using a weighted summation method to obtain a comprehensive spatial-causal relationship matrix. Integrating the frequency domain attention gain into the spatiotemporal attention weight calculation allows the frequency domain attention gain to be used as a weighting factor in the time dimension, combined with the weighting factor in the spatial dimension.

[0162] For users with insufficient data, a progressive transfer weighting function is applied to transfer weights from similar users and adjust the spatiotemporal attention weights. Specifically, the spatiotemporal attention weights of similar users are weighted and averaged, with the weights determined by the progressive transfer weighting function.

[0163] Finally, the adjusted spatiotemporal attention weights are fed back into the multimodal fusion model for subsequent tear total IgE prediction and analysis.

[0164] S160. The image patch features, spatiotemporal attention weights, physiological parameters, environmental allergen data, and historical medical data are fused and input into a lightweight Transformer model to generate the predicted concentration of total tear IgE and allergy risk score, and the risk level is classified.

[0165] In this embodiment, the predicted concentration of total tear IgE refers to the concentration of IgE in tears predicted by the model, in units of IU / mL.

[0166] The allergy risk score refers to the level of allergy risk predicted by the model, and its value ranges from 0 to 1.

[0167] Risk level refers to the level divided according to the allergy risk score, specifically divided into: low risk (0-0.25), medium risk (0.25-0.5), high risk (0.5-0.75) and very high risk (0.75-1).

[0168] The lightweight Transformer model includes an input embedding layer, multiple lightweight Transformer encoders, and an output layer. The input embedding layer converts the comprehensive feature vector into a fixed-dimensional embedding vector. The multiple lightweight Transformer encoders process the embedding vectors sequentially. Each encoder contains a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism is used to capture the correlation between features, and the feedforward neural network is used for nonlinear transformation. The output layer is based on the results output by the encoders.

[0169] Image patch features: Feature vectors extracted from ocular surface images, representing visual information of the ocular surface, such as the degree of conjunctival congestion and the morphology of meibomian glands.

[0170] Spatiotemporal attention weights: Integrating spatial correlation maps, causal relationship matrices, frequency domain attention gains, and weights adjusted by transfer learning to highlight key spatiotemporal features.

[0171] Physiological parameters, including tear secretion rate, BUT (bulk under time) and corneal sensitivity, reflect the physiological state of the ocular surface.

[0172] Environmental allergen data: such as pollen concentration, PM2.5 concentration, etc. 2.5 Concentration and temperature / humidity reflect the influence of the external environment on total tear IgE.

[0173] Historical medical data: Includes past total tear IgE values ​​and medication records, providing users with a long-term health background.

[0174] The aforementioned feature vectors are concatenated sequentially to form a comprehensive feature vector, which is then input into a lightweight Transformer model. This model comprises an input embedding layer, multiple lightweight Transformer encoders, and an output layer. The input embedding layer transforms the comprehensive feature vector into fixed-dimensional embedding vectors. Multiple lightweight Transformer encoders process the embedding vectors sequentially; each encoder includes a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism captures the correlations between features, while the feedforward neural network performs a non-linear transformation. Based on the encoder outputs, the output layer generates the predicted concentration of total tear IgE and an allergy risk score, and classifies the risk level.

[0175] In one embodiment, step S160 described above may include steps S161 to S164.

[0176] S161. Convert image patch features, spatiotemporal attention weights, physiological parameters, environmental allergen data, and historical medical data into fixed-length vectors, and then concatenate and fuse them in order to form a comprehensive feature vector.

[0177] In this embodiment, image patch feature transformation involves arranging the image patch feature vectors extracted by the visual Transformer in sequence to form a one-dimensional vector. Assuming that each image patch feature vector has a dimension of 512, and there are a total of 64 image patches, the length of the arranged vector is 64×512.

[0178] Spatiotemporal attention weight transformation: The spatiotemporal attention weight vector is directly treated as a one-dimensional vector, assuming its dimension is 128.

[0179] Physiological parameter transformation: Physiological parameters (such as tear secretion rate, BUT, corneal sensitivity, etc.) are normalized to form a one-dimensional vector. Assume the normalized physiological parameter vector has a dimension of 10.

[0180] Environmental allergen data conversion: Converting environmental allergen data (such as pollen concentration, PM2.5 concentration, PM2.5 concentration) into environmental allergen data. 2.5 After normalization of concentration, temperature, humidity, etc., a one-dimensional vector is formed. Assume that the normalized environmental allergen data vector has a dimension of 20.

[0181] Historical medical data transformation: Historical medical data (such as past total tear IgE values, medication records, etc.) is encoded into a one-dimensional vector. Assume the encoded historical medical data vector has a dimension of 30.

[0182] Concatenation and Fusion: Following the order of image patch features, spatiotemporal attention weights, physiological parameters, environmental allergen data, and historical medical data, the above vectors are concatenated to form a comprehensive feature vector. The dimensions of the concatenated comprehensive feature vector are 64×512+128+10+20+30=32768+128+60=32956.

[0183] S162. Convert the comprehensive feature vector into a fixed-dimensional embedding vector.

[0184] In this embodiment, the synthesized feature vector is input to the input embedding layer of the lightweight Transformer model. The input embedding layer transforms the synthesized feature vector into a fixed-dimensional embedding vector through a fully connected layer. Assume the dimension of the embedding vector is 512. The weight matrix of the fully connected layer is 512 × 32,956, and the bias vector is 512 × 1. The embedding vector is calculated using matrix multiplication and an activation function (such as ReLU): Embedding vector = ReLU(synthetic feature vector × weight matrix + bias vector).

[0185] S163. Calculate the correlation between features in the embedding vector, generate self-attention weights, and perform nonlinear transformation on the features output by the self-attention mechanism to obtain the encoder output.

[0186] In this embodiment, the multi-head self-attention mechanism is as follows:

[0187] Calculate the Query, Key, and Value matrices: Using the embedding vector as input, calculate the Query matrix (Q), Key matrix (K), and Value matrix (V) through three different fully connected layers. Assuming the embedding vector dimension is 512 and the multi-head self-attention mechanism has 8 heads, then the dimension of each head is 512 / 8 = 64. The weight matrix of the fully connected layer is 512×512, and the bias vector is 512×1.

[0188] Scaling Dot Product Attention: For each head, compute the dot product of the transposes of Q and K to obtain the attention score matrix. Then divide the attention score matrix by a scaling factor. Then, the self-attention weight matrix is ​​calculated using the Softmax function:

[0189] Multiplying the self-attention weight matrix by the V matrix yields the output of the multi-head self-attention mechanism. Then, the outputs from multiple heads are concatenated and linearly transformed through a fully connected layer to obtain the final multi-head self-attention mechanism output: Multi-head self-attention output = concatenated (head1, head2, ..., head8) × weight matrix + bias vector.

[0190] Feedforward Neural Network: The feature vector output by the multi-head self-attention mechanism is input into the feedforward neural network. The feedforward neural network contains two fully connected layers. The first fully connected layer expands the dimension of the feature vector to 2048, using GELU as the activation function; the second fully connected layer reduces the dimension of the feature vector back to 512. The calculation process is as follows:

[0191] Feedforward output = GELU(multi-head self-attention output × weight matrix 1 + bias vector 1) × weight matrix 2 + bias vector 2;

[0192] Residual connections and LayerNorm: The output of the feedforward neural network is added to the output of the multi-head self-attention mechanism to obtain the encoder output. Then, the output is normalized by the LayerNorm layer: Encoder output = LayerNorm(feedforward output + multi-head self-attention output).

[0193] S164. Based on the encoder output, generate the predicted concentration of total tear IgE and the allergy risk score, and classify the risk level.

[0194] In this embodiment, the feature vector output by the encoder is input to the output layer. The output layer contains two branches:

[0195] Predicted concentration branch: The dimension of the feature vector is reduced to 1 through a fully connected layer, and a linear activation function is used to obtain the predicted concentration value of total tear IgE.

[0196] Predicted concentration = linear activation (encoder output × weight vector 1 + bias 1);

[0197] Allergy risk score branch: The dimension of the feature vector is reduced to 1 through a fully connected layer, and the Sigmoid function is used as the activation function to compress the output value to the interval of 0 and 1, thus obtaining the allergy risk score.

[0198] Allergy risk score = Sigmoid(encoder output × weight vector 2 + bias 2);

[0199] Risk Level Classification: A threshold for risk level classification is established based on the allergy risk score. In this embodiment, the risk level is divided into four levels:

[0200] Low risk: Allergy risk score in the range of 0 to 0.25.

[0201] Medium risk: Allergy risk score in the range of 0.25 to 0.5.

[0202] High risk: Allergy risk score in the range of 0.5 to 0.75.

[0203] Extremely high risk: Allergy risk score in the range of 0.75 to 1.

[0204] By comparing allergy risk scores with these thresholds, the corresponding risk level is determined.

[0205] Through the above steps, the lightweight Transformer model can integrate the characteristics and advantages of multimodal data to generate the predicted concentration of total tear IgE and the allergy risk score, and classify the risk level accordingly, providing an important reference for clinical diagnosis and treatment.

[0206] S170. Determine whether to trigger a micro-tear sampling command based on the predicted concentration of total IgE in the tear fluid and the allergy risk score.

[0207] In this embodiment, the trigger threshold for the allergy risk score is set to 0.75. This means that when the allergy risk score exceeds 0.75, the patient is considered to have a high allergy risk, and micro-tear sampling is required to obtain a more accurate measured concentration. Simultaneously, the rate of change in the predicted total tear IgE concentration is also considered; when the rate of change exceeds 30%, a sampling command is also triggered.

[0208] Continuously monitor the allergy risk score and predicted total tear IgE concentration output by the lightweight Transformer model. Calculate whether the allergy risk score exceeds a set threshold of 0.75, and whether the rate of change of the predicted concentration compared to previous measurements exceeds 30%.

[0209] If the allergy risk score exceeds 0.75 or the predicted concentration change rate exceeds 30%, the system will automatically trigger a micro-tear sampling command. Alternatively, users can manually trigger the micro-tear sampling process based on their own needs or their doctor's advice.

[0210] S180. When the micro-tear sampling command is triggered, the disposable microfluidic chip is activated to collect a 0.5μL tear sample through a capillary glass tube or tear test paper adsorption device.

[0211] In this embodiment, upon receiving a trigger command, the system activates a disposable microfluidic chip. This chip is designed for the collection and detection of minute tear fluid samples and features high efficiency, speed, and accuracy.

[0212] The capillary glass tubes or tear test strip adsorption devices in the microfluidic chip are in place, ready to collect tear samples. These adsorption devices can precisely control the volume of tears collected, ensuring that a sufficient but not excessive sample is collected.

[0213] Following the system prompts, users perform slight eye movements and blinking to stimulate natural tear production. The capillary glass tube or tear test strip absorbs approximately 0.5 μL of tear sample via capillary action or adsorption. This volume of tear is sufficient for testing while minimizing patient discomfort.

[0214] S190. The collected tear sample is imported into a microfluidic chip and detected using a fluorescence immunoassay or surface plasmon resonance technique. The measured concentration of total IgE in the tear is read within a few minutes.

[0215] In this embodiment, the collected tear fluid sample is introduced into the detection channel inside the microfluidic chip through the sample inlet. The design of the microfluidic chip ensures precise control and rapid transfer of the sample within the chip.

[0216] Detection was performed using either a competitive fluorescence immunoassay or surface plasmon resonance (SPR) technique. The competitive fluorescence immunoassay utilizes the specific binding of fluorescently labeled antigens or antibodies to IgE in tears, determining the IgE concentration by measuring the intensity of the fluorescence signal. SPR, on the other hand, monitors the binding of IgE to specific antibodies in real time by detecting changes in refractive index caused by interactions between biomolecules, thereby quantitatively analyzing the IgE concentration.

[0217] Within 3 minutes of the sample being introduced into the microfluidic chip, the instrument completes the analysis of the tear sample and reads the measured concentration of total IgE in the tears. This rapid testing process allows patients to obtain results quickly, improving testing efficiency and patient satisfaction.

[0218] S200. Based on the Bayesian adaptive filtering algorithm, combined with the predicted concentration of total tear IgE, the measured concentration of total tear IgE, and the baseline IgE value in the historical medical data, the spatiotemporal attention weight is dynamically adjusted to calculate the final reported value, thereby obtaining the total tear IgE value.

[0219] In this embodiment, the total IGE value of tears refers to...

[0220] In one embodiment, step S200 described above may include steps S201 to S203.

[0221] S201. Initialize weight parameters to balance the influence of non-invasive predicted concentration, measured concentration and historical baseline value.

[0222] In this embodiment, at the beginning, weight parameters α1 and α2 are initialized, where α1 is used to balance the influence of non-invasive predicted concentration and measured concentration, and α2 is used to balance the influence of fused concentration value and historical baseline value. The initial values ​​can be set to α1 = 0.5 and α2 = 0.5, indicating that the predicted concentration and measured concentration are given equal weights initially, and the fused concentration value and historical baseline value are also given equal weights.

[0223] Set weight adjustment step size: Set weight adjustment step sizes Δα1 and Δα2 to control the magnitude of weight updates. For example, you can set Δα1 = 0.1 and Δα2 = 0.1, which means that the weight parameters are adjusted in steps of 0.1 each time the weights are updated.

[0224] S202. Calculate the difference between the predicted concentration of total tear IgE and the measured concentration of total tear IgE, and update the weight parameters according to the Bayesian adaptive filtering algorithm.

[0225] In this embodiment, the difference between the non-invasive predicted concentration and the measured concentration can be calculated using either absolute error (AE) or relative error (RE). Assuming the predicted concentration is... If the measured concentration is real, then the absolute error is... The relative error is

[0226] Update the weighting parameters: If AE is greater than the preset error threshold (e.g., the error threshold is set to 5 IU / mL), decrease the value of α1 to increase dependence on the measured concentration. The update formula is α1 = α1 - Δα1. If AE is less than or equal to the preset error threshold, appropriately increase the value of α1 to increase dependence on the predicted concentration. The update formula is α1 = α1 + Δα1. Simultaneously, ensure that the value of α1 is within the range [0,1]. If α1 exceeds 1, set it to 1; if α1 is less than 0, set it to 0.

[0227] S203. The predicted concentration of total tear IgE and the measured concentration of total tear IgE are weighted and summed to obtain a fusion concentration value. The fusion concentration value is then combined with the baseline IgE value and weighted to obtain a final reported value, thus obtaining the total tear IgE value.

[0228] In this embodiment, the non-invasive predicted concentration and the measured concentration are weighted and summed, and the calculation formula is as follows:

[0229] Obtain the baseline IgE value from historical medical data and sum it with the fusion concentration value using a weighted average. The calculation formula is: Final reported value = α2 × fusion concentration value + (1-α2) × baseline IgE value.

[0230] The final reported value is the total tear IgE value, which takes into account non-invasive prediction, measured concentration and historical baseline information, and can more accurately reflect the patient's total tear IgE level.

[0231] In this embodiment, the total IgE value of tears refers to the numerical value obtained through the above-described fusion calculation, which comprehensively reflects the actual level of IgE in the patient's tears, and is expressed in IU / mL. This value not only includes real-time prediction and measured information but also takes into account the patient's historical data, thereby providing a more accurate and personalized reference for clinical diagnosis.

[0232] In another embodiment, the method described above further includes steps S210 to S220.

[0233] S210. Expand the training dataset using the total IGE value of the tear fluid and the corresponding multimodal data.

[0234] In this embodiment, after completing the total tear IgE detection, the final reported value (total tear IgE value), non-invasive predicted concentration, measured concentration, historical baseline value, and corresponding multimodal data are collected, including image patch features, spatiotemporal attention weights, physiological parameters, environmental allergen data, and historical medical data.

[0235] The collected data is organized into a format consistent with the original training dataset to ensure data compatibility and consistency.

[0236] The final reported value is used as the label for the data sample, which is the actual value of total tear IgE.

[0237] Multimodal data should be labeled accordingly, such as the labeling of image patch features and the labeling of spatiotemporal attention weights, so as to facilitate the model's learning and understanding of the semantic information of the data.

[0238] The labeled data samples are integrated into the original training dataset to form the expanded training dataset.

[0239] Check whether the expanded dataset is balanced, that is, whether the number of data samples across different risk levels and concentration ranges is reasonably distributed. If the data is imbalanced, oversampling or undersampling can be used to adjust it, thereby improving the model's training performance and generalization ability.

[0240] S220. Use the expanded training dataset to fine-tune the spatiotemporal attention weights, image-text alignment parameters, etc. online.

[0241] Load a lightweight Transformer model and its related parameters, including the weights and biases of the input embedding layer, Transformer encoder, and output layer.

[0242] Choose an appropriate optimizer, such as the Adam optimizer, and set hyperparameters such as the learning rate to prepare for model fine-tuning.

[0243] Load data samples from the expanded training dataset and perform necessary preprocessing, such as normalization and one-hot encoding.

[0244] The preprocessed data is input into the model for forward propagation calculation to obtain the model's predicted output.

[0245] Calculate the loss function value between the predicted output and the actual label, such as mean squared error (MSE) for calculating the loss of predicted concentration, and cross-entropy loss for calculating the loss of allergy risk score.

[0246] The gradient is calculated using the backpropagation algorithm, and the optimizer is used to update the model parameters, including the spatiotemporal attention weights and the image-text alignment parameters.

[0247] Repeat the above steps until the set number of training rounds is completed or the convergence condition is met.

[0248] A portion of the expanded training dataset is used as a validation set to evaluate the model's performance.

[0249] Use appropriate evaluation metrics, such as root mean square error (RMSE), to assess the accuracy of predicted concentrations, and precision, recall, and F1 score to evaluate the performance of allergy risk prediction.

[0250] During training, the model is periodically validated on the validation set to monitor performance changes and prevent overfitting.

[0251] Once the model has been trained and achieved satisfactory performance, save the model and its parameters.

[0252] The updated model is deployed to the actual detection system to replace the original model, thereby improving the system's detection performance and accuracy.

[0253] As more data accumulates, the model is continuously fine-tuned and updated using new data, enabling it to be constantly optimized and adapted to new data distributions.

[0254] Through the above steps S210 and S220, the newly acquired total tear IgE value and its corresponding multimodal data can be fully utilized to expand the training dataset and fine-tune the model online, thereby continuously improving the model's performance and adaptability, and providing more accurate and reliable prediction results for total tear IgE detection.

[0255] For example: Take 5 RGB images of the conjunctiva using the front-facing camera of a mobile phone;

[0256] The Bluetooth tear sensor measured a BUT of 4 seconds and a secretion volume of 0.06 μL / min.

[0257] The environmental module synchronously read 120 pollen grains / m³ 3 PM 2.5 38μg / m 3Temperature 26℃, humidity 45%;

[0258] HIS end baseline IgE = 25 IU / mL, allergy history "vernal catarrhal conjunctivitis".

[0259] Images are gamma-corrected and Z-score is applied. Physiological / environmental data are normalized, and all data are aligned to the millisecond level with UTC timestamps.

[0260] The chief complaint was "severe itching in both eyes → rubbing eyes → congestion" → NLP extracted the keyword chain [itching - congestion - increased tear production], and the knowledge graph was expanded to [eosinophilic infiltration - mast cell degranulation - conjunctival congestion++].

[0261] ViT cuts a 2048×1536 image into 64×64 blocks → 512-dimensional vector × 64;

[0262] BERT encodes the symptom chain as a 768-dimensional text vector;

[0263] Calculate cosine similarity and retain 12 image patches (congested areas) with a similarity > 0.88.

[0264] Centered on the girl, 18 children within 5km were included, with a weight w = 0.8 × exp(-d / 2) + 0.2 × pollen similarity.

[0265] The second-level BUT decreases, while the minute-level IgE increases. After Granger criterion p < 0.05, it is written into the causality matrix.

[0266] Pollen 24h sequence → main frequency 0.042Hz (daily cycle), corresponding gain G(f) = 1.3.

[0267] Data for the girl was insufficient (only 3 historical records), so the transfer weight TW = exp(-1.2 × 0.3) = 0.70, and the attention weight was transferred from similar children.

[0268] Concatenate [12×512,128,10,20,30] → Combined vector 6700 dimensions → Embedded 512 → 2-layer encoder → Output layer:

[0269]

[0270] Allergy risk score S = 0.81 (extremely high);

[0271] S>0.75 and The rate of change (42.3-25) / 25 = 69% > 30% → triggers microfluidic sampling.

[0272] S1800.5μL sampling;

[0273] The 0.5μL capillary tube inside the disposable chip is filled with tears in 5 seconds.

[0274] The results of the fluorescence immunoassay were obtained in 2 minutes and 20 seconds: C_real = 45.7 IU / mL.

[0275] Initially, α1 = 0.5; |42.3-45.7| = 3.4 IU / mL > 5 IU / mL threshold → α1 decreases to 0.3;

[0276] Fusion concentration = 0.3 × 42.3 + 0.7 × 45.7 = 44.7 IU / mL;

[0277] Baseline weight α2 = 0.2 → Final reported value = 0.8 × 44.7 + 0.2 × 25 = 40.8 IU / mL.

[0278] Add (40.8, corresponding to multimodal data) to the training library; Adam fine-tuned the spatiotemporal attention weights and image-text alignment parameters for 3 epochs, reducing RMSE from 4.1 to 2.6 IU / mL.

[0279] The app displayed "extremely high risk, immediate medical attention recommended." The parents took the girl to the doctor, who adjusted the dosage of antihistamine eye drops based on a dose of 40.8 IU / mL. The symptoms were significantly relieved the next day.

[0280] The aforementioned method for detecting total IgE in tears based on multimodal data fusion simultaneously acquires ocular surface images, physiological parameters, and environmental allergen data, and combines this with historical medical data from the hospital information system. Preprocessing and feature extraction techniques are used to identify image features related to symptoms. Frequency domain analysis and progressive transfer learning are employed to dynamically adjust spatiotemporal attention weights to enhance the understanding of environmental allergen data. Multiple data points are input into a lightweight Transformer model to predict total IgE concentration and allergy risk score. Based on the prediction results, a micro-volume tear sampling procedure (requiring only 0.5 μL sample) is initiated. Actual measurement is performed using a disposable microfluidic chip, and results are obtained within minutes based on fluorescence immunoassay or surface plasmon resonance technology. Finally, a Bayesian adaptive filtering algorithm is applied to integrate all information and calculate the final reported value. This method achieves extremely small sampling volume, rapid detection time, dynamic correction for environmental interference, and output of absolute concentration values, meeting the dual needs of precise clinical diagnosis and home monitoring.

[0281] Figure 2 This is a schematic block diagram of a tear total IGE detection system 300 based on multimodal data fusion provided in an embodiment of the present invention. Figure 2As shown, corresponding to the above-described method for detecting total IGE of tears based on multimodal data fusion, this invention also provides a system 300 for detecting total IGE of tears based on multimodal data fusion. This system 300 includes a unit for performing the above-described method for detecting total IGE of tears based on multimodal data fusion, and can be configured in a desktop computer, tablet computer, laptop computer, or other terminal. For details, please refer to... Figure 2 The tear total IGE detection system 300 based on multimodal data fusion includes a multimodal data acquisition unit 301, a preprocessing unit 302, a text processing unit 303, an image block feature calculation unit 304, an adjustment unit 305, a prediction unit 306, a judgment unit 307, a micro-sampling unit 308, a detection unit 309, and a final value calculation unit 310.

[0282] The multimodal data acquisition unit 301 is used to simultaneously acquire ocular surface images, physiological parameters, environmental allergen data, and retrieve historical medical data of users from the hospital information system using a camera, tear sensor, and environmental sensor to obtain multimodal data; the preprocessing unit 302 is used to preprocess and align the multimodal data to obtain preprocessing results; the text processing unit 303 is used to acquire patient symptom descriptions, extract keywords from the patient symptom descriptions using natural language processing technology and construct symptom keyword chains, and simultaneously perform semantic expansion using a medical knowledge graph to obtain symptom text; the image patch feature calculation unit 304 is used to calculate the cosine similarity between the ocular surface image patches and the symptom text feature vectors in the preprocessing results to identify image patch features related to symptoms; the adjustment unit 305 is used to dynamically adjust the spatiotemporal attention weights by combining spatial association graphs and causal relationship matrices through frequency domain analysis and progressive transfer learning to enhance the understanding of environmental allergen data; the prediction unit 306 is used to combine the image patch features, spatiotemporal attention weights, and physiological parameters... The system integrates the data on environmental allergens and historical medical data, inputs it into a lightweight Transformer model, generates a predicted concentration of total tear IgE and an allergy risk score, and classifies the risk level. A judgment unit 307 determines whether to trigger a micro-tear sampling command based on the predicted concentration of total tear IgE and the allergy risk score. A micro-sampling unit 308, when triggered, activates a disposable microfluidic chip to collect a 0.5 μL tear sample via a capillary glass tube or tear test strip adsorption device. A detection unit 309 imports the collected tear sample into the microfluidic chip and performs detection based on a fluorescent immunoassay or surface plasmon resonance technique, reading the measured concentration of total tear IgE within several minutes. A final value calculation unit 310 dynamically adjusts the spatiotemporal attention weights based on a Bayesian adaptive filtering algorithm, combining the predicted concentration of total tear IgE, the measured concentration of total tear IgE, and the baseline IgE value from the historical medical data, to calculate the final reported value, thus obtaining the total tear IgE value.

[0283] In one embodiment, the image patch feature calculation subunit includes:

[0284] The image encoding subunit is used to segment and encode the preprocessed ocular surface image using a visual Transformer to obtain encoded image patches; the text encoding subunit is used to encode the symptom text using a BERT model to obtain text feature vectors; and the similarity calculation subunit is used to calculate the cosine similarity between the encoded image patches and the text feature vectors, and to filter out image patches that meet the matching requirements related to symptoms to obtain symptom-related image patch features.

[0285] In one embodiment, the adjustment unit 305 is used to construct a patient spatial association map and a cross-scale causal relationship matrix, perform a fast Fourier transform on the pollen concentration sequence in the environmental allergen data to generate frequency domain attention gain, and construct a progressive transfer weight function for users whose data volume does not meet the requirements, and dynamically calculate the spatiotemporal attention weight.

[0286] In one embodiment, the adjustment unit 305 includes:

[0287] The first construction subunit is used to acquire patient geographic location information, calculate the geographic distance and pollen exposure similarity for each pair of patients, and construct a patient spatial association map. The second construction subunit is used to define time scale levels, extract feature data at each time scale, use Granger causality test to identify causal relationships between features at different time scales, and construct a cross-scale causal relationship matrix. The gain calculation subunit is used to extract pollen concentration time series from environmental allergen data, perform preprocessing, apply Fast Fourier Transform to transform it to the frequency domain, calculate the attention gain corresponding to different frequency components to obtain frequency domain attention gain. The function design subunit is used to set data volume thresholds, calculate user similarity matrices, determine the data sufficiency index for users whose data volume does not meet the requirements, and design a progressive transfer weight function. The fusion subunit is used to fuse information from the patient spatial association map and the cross-scale causal relationship matrix, integrate frequency domain attention gain, and apply a progressive transfer weight function to adjust spatiotemporal attention weights.

[0288] In one embodiment, the prediction unit 306 includes:

[0289] The splicing and fusion subunit is used to convert image patch features, spatiotemporal attention weights, physiological parameters, environmental allergen data, and historical medical data into fixed-length vectors, and then splice and fuse them in sequence to form a comprehensive feature vector. The transformation subunit is used to convert the comprehensive feature vector into a fixed-dimensional embedding vector. The self-attention subunit is used to calculate the correlation between the features in the embedding vector, generate self-attention weights, and perform nonlinear transformation on the features output by the self-attention mechanism to obtain the encoder output. The generation subunit is used to generate the predicted concentration of total tear IgE and the allergy risk score based on the encoder output, and classify the risk level.

[0290] In one embodiment, the final value calculation unit 310 includes:

[0291] The initialization subunit initializes the weight parameters to balance the influence of non-invasive predicted concentration, measured concentration, and historical baseline value. The difference calculation subunit calculates the difference between the predicted concentration and the measured concentration of total tear IgE, and updates the weight parameters according to the Bayesian adaptive filtering algorithm. The summation subunit sums the predicted concentration and the measured concentration of total tear IgE with the weights to obtain the fused concentration value, and then sums it with the baseline IgE value to obtain the final reported value, thus obtaining the total tear IgE value.

[0292] In one embodiment, the system further includes:

[0293] The expansion unit is used to expand the training dataset using the total IGE value of the tear fluid and the corresponding multimodal data; the fine-tuning unit is used to fine-tune the spatiotemporal attention weights, image-text alignment parameters, etc. online using the expanded training dataset.

[0294] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned tear total IGE detection system 300 based on multimodal data fusion and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0295] The aforementioned tear total IGE detection system 300 based on multimodal data fusion can be implemented as a computer program, which can be used in, for example... Figure 3 It runs on the computer device shown.

[0296] Please see Figure 3 , Figure 3 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0297] See Figure 3 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0298] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform a tear total IGE detection method based on multimodal data fusion.

[0299] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0300] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a method for total tear gas IGE detection based on multimodal data fusion.

[0301] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0302] The processor 502 is used to run the computer program 5032 stored in the memory to implement all the steps of the tear total IGE detection method based on multimodal data fusion.

[0303] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0304] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0305] Therefore, the present invention also provides a storage medium. This storage medium may be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform all steps of the tear total IGE detection method based on multimodal data fusion.

[0306] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0307] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0308] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of each unit is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0309] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the system of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0310] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0311] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for detecting total IGE in tears based on multi-modal data fusion, characterized in that, The method comprises the following steps: Synchronously acquiring ocular surface images, physiological parameters, and environmental allergen data by using a camera, a tear sensor, and an environmental sensor, and calling historical medical data of a user from a hospital information system to obtain multi-modal data; Preprocessing and aligning the multi-modal data to obtain a preprocessing result; Acquiring a patient symptom description, extracting keywords in the patient symptom description by natural language processing technology, constructing a symptom keyword chain, and expanding semantics by combining a medical knowledge graph to obtain a symptom text; Calculating the cosine similarity between image blocks in the preprocessing result and symptom text feature vectors to identify image block features related to symptoms; Combining a spatial correlation graph and a causal relationship matrix, dynamically adjusting the spatiotemporal attention weight through frequency domain analysis and progressive transfer learning, and enhancing the understanding of environmental allergen data; Fusing the image block features, the spatiotemporal attention weight, the physiological parameters, the environmental allergen data, and the historical medical data, inputting them into a lightweight Transformer model, generating a predicted concentration of total IgE in tears and an allergy risk score, and dividing the risk level; Determining whether to trigger a micro-amount tear sampling instruction according to the predicted concentration of total IgE in tears and the allergy risk score; When the micro-amount tear sampling instruction is triggered, a disposable microfluidic chip is started, and 0.5 μL of tear sample is collected by a capillary glass tube or a tear test paper adsorption device; The collected tear sample is introduced into the microfluidic chip, and detection is performed based on a fluorescence immunoassay or a surface plasmon resonance technology, and the measured concentration of total IgE in tears is read within a few minutes; According to the Bayesian adaptive filtering algorithm, combining the predicted concentration of total IgE in tears, the measured concentration of total IgE in tears, and the baseline IgE value in the historical medical data, dynamically adjusting the spatiotemporal attention weight, and calculating the final report value to obtain the total IgE value in tears.

2. The method for detecting total IGE in tears based on multi-modal data fusion according to claim 1, characterized in that, The method for calculating the cosine similarity between the image blocks in the preprocessing result and the symptom text feature vectors to identify the image block features related to the symptoms comprises the following steps: Segmenting and encoding the ocular surface images of the preprocessing result using a visual Transformer to obtain encoded image blocks; Encoding the symptom text using a BERT model to obtain a text feature vector; Calculating the cosine similarity between the encoded image blocks and the text feature vector, and screening out image blocks with a matching degree meeting the requirements related to the symptoms to obtain the image block features related to the symptoms.

3. The method for detecting total IGE in tears based on multi-modal data fusion according to claim 2, characterized in that, The method for combining the spatial correlation graph and the causal relationship matrix, dynamically adjusting the spatiotemporal attention weight through frequency domain analysis and progressive transfer learning, and enhancing the understanding of environmental allergen data comprises the following steps: Constructing a patient spatial correlation graph and a cross-scale causal relationship matrix, performing fast Fourier transform on the pollen concentration sequence in the environmental allergen data to generate a frequency domain attention gain, and constructing a progressive transfer weight function for users with insufficient data volume to dynamically calculate the spatiotemporal attention weight.

4. The method for detecting total IGE in tears based on multi-modal data fusion according to claim 3, characterized in that, The process involves constructing a patient spatial association map and a cross-scale causal relationship matrix, performing a fast Fourier transform on the pollen concentration sequence within the environmental allergen data to generate frequency domain attention gain, and constructing a progressive transfer weight function for users with insufficient data volume to dynamically calculate spatiotemporal attention weights, including: Obtain patient geographic location information, calculate the geographic distance and pollen exposure similarity for each pair of patients, and construct a patient spatial association map; Define time scale hierarchy, extract feature data of each time scale, use Granger causality test to identify causal relationships between features of different time scales, and construct cross-scale causal relationship matrix; Pollen concentration time series was extracted from environmental allergen data, preprocessed, and then converted to the frequency domain using Fast Fourier Transform. Attention gain corresponding to different frequency components was calculated to obtain frequency domain attention gain. Set a data volume threshold, calculate the user similarity matrix, determine the data sufficiency index of users whose data volume does not meet the requirements, and design a progressive migration weight function. By integrating information from patient spatial correlation maps and cross-scale causal relationship matrices, frequency domain attention gain is combined, and a progressive transfer weighting function is applied to adjust the spatiotemporal attention weights.

5. The method for detecting total IGE in tears based on multi-modal data fusion according to claim 4, characterized in that, The lightweight Transformer model includes an input embedding layer, multiple lightweight Transformer encoders, and an output layer. The input embedding layer converts the comprehensive feature vector into a fixed-dimensional embedding vector. The multiple lightweight Transformer encoders process the embedding vectors sequentially. Each encoder contains a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism is used to capture the correlation between features, and the feedforward neural network is used for nonlinear transformation. The output layer is based on the results output by the encoders.

6. The method for detecting total IGE in tears based on multi-modal data fusion according to claim 5, characterized in that, The process involves fusing the image patch features, spatiotemporal attention weights, physiological parameters, environmental allergen data, and historical medical data, inputting these into a lightweight Transformer model to generate a predicted concentration of total tear IgE and an allergy risk score, and classifying the risk level, including: Image patch features, spatiotemporal attention weights, physiological parameters, environmental allergen data, and historical medical data are each converted into fixed-length vectors and then concatenated and fused in sequence to form a comprehensive feature vector. Convert the composite feature vector into a fixed-dimensional embedding vector; The correlation between features in the embedding vector is calculated, self-attention weights are generated, and nonlinear transformation is performed on the features output by the self-attention mechanism to obtain the encoder output. Based on the encoder output, the predicted concentration of total tear IgE and the allergy risk score are generated, and the risk level is classified.

7. The method for detecting total IGE in tears based on multi-modal data fusion according to claim 1, wherein, The process involves using a Bayesian adaptive filtering algorithm, combining the predicted concentration of total tear IgE, the measured concentration of total tear IgE, and the baseline IgE value from historical medical data, to dynamically adjust the spatiotemporal attention weights and calculate the final reported value to obtain the total tear IgE value. This includes: Initialize weight parameters to balance the effects of non-invasive predicted concentrations, measured concentrations, and historical baseline values; The difference between the predicted concentration of total tear IgE and the measured concentration of total tear IgE is calculated, and the weight parameters are updated according to the Bayesian adaptive filtering algorithm. The predicted concentration of total tear IgE and the measured concentration of total tear IgE are weighted and summed to obtain a fusion concentration value. This fusion concentration value is then combined with the baseline IgE value and weighted to obtain a final reported value, which is the total tear IgE value.

8. The method for detecting total IGE in tears based on multi-modal data fusion according to claim 1, wherein, The step of dynamically adjusting the spatiotemporal attention weights based on the Bayesian adaptive filtering algorithm, combining the predicted concentration of total tear IgE, the measured concentration of total tear IgE, and the baseline IgE value in historical medical data, to calculate the final reported value and obtain the total tear IgE value, further includes: The training dataset was expanded using the total IGE value of the tear fluid and the corresponding multimodal data. We used the expanded training dataset to fine-tune the spatiotemporal attention weights, image-text alignment parameters, and other parameters online.

9. A total IGE detection system based on multi-modal data fusion, characterized in that, include: The multimodal data acquisition unit is used to simultaneously acquire ocular surface images, physiological parameters, environmental allergen data, and retrieve users' historical medical data from the hospital information system to obtain multimodal data using a camera, tear sensor, and environmental sensor. A preprocessing unit is used to preprocess and align the multimodal data to obtain preprocessing results; The text processing unit is used to obtain patient symptom descriptions, extract keywords from the patient symptom descriptions using natural language processing technology and construct symptom keyword chains, and simultaneously perform semantic expansion by combining medical knowledge graphs to obtain symptom text. The image patch feature calculation unit is used to calculate the cosine similarity between the ocular surface image patch and the symptom text feature vector in the preprocessing result, so as to identify image patch features related to symptoms; The adjustment unit is used to combine spatial correlation maps and causal relationship matrices to dynamically adjust spatiotemporal attention weights through frequency domain analysis and progressive transfer learning, thereby enhancing the understanding of environmental allergen data. The prediction unit is used to fuse the image patch features, spatiotemporal attention weights, physiological parameters, environmental allergen data, and historical medical data, input them into a lightweight Transformer model, generate the predicted concentration of total tear IgE and allergy risk score, and classify the risk level. The judgment unit is used to determine whether to trigger a micro-tear sampling command based on the predicted concentration of total tear IgE and the allergy risk score. The micro-sampling unit is used to activate a disposable microfluidic chip when a micro-tear sampling command is triggered, and collect a 0.5μL tear sample through a capillary glass tube or tear test paper adsorption device. The detection unit is used to import the collected tear samples into the microfluidic chip and perform detection based on fluorescence immunoassay or surface plasmon resonance technology, and read the measured concentration of total IgE in the tear within a few minutes; The final value calculation unit is used to dynamically adjust the spatiotemporal attention weights based on the Bayesian adaptive filtering algorithm, combined with the predicted concentration of total tear IgE, the measured concentration of total tear IgE, and the baseline IgE value in the historical medical data, to calculate the final reported value and obtain the total tear IgE value.

10. A storage medium, characterized by The storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 8.

Images