Chronic disease early detection method and system based on multi-mode large model

Abstract

The invention discloses a chronic disease early detection method and system based on a multi-modal large model, and relates to the technical field of intelligent medical treatment and artificial intelligence, and the method comprises the steps: obtaining a multi-modal data stream of a target user in a target time window from a pathology database, and generating an original multi-modal data set; performing timestamp unification and numerical value standardization processing on the original multi-modal data set to obtain a time sequence feature sequence; inputting the time sequence feature sequence to the multi-modal large model to obtain an abnormal symptom feature; calculating the similarity between the abnormal symptom features and feature vectors of marked cases in a historical case library, and determining matched cases; a diagnosis result and a development process of the matched case are extracted, a disease risk level and a development trend corresponding to the original multi-modal data set are determined in combination with the medical knowledge graph, and a pathology assessment result is obtained; and generating an early warning signal containing the risk type and the intervention suggestion according to the pathological assessment result. By implementing the application, the accuracy of early detection of chronic diseases can be improved.

Description

Technical Field

[0001] This application relates to the fields of intelligent healthcare and artificial intelligence technologies, and particularly to a method and system for early detection of chronic diseases based on a multimodal large model. Background Art

[0002] With the acceleration of the aging population process and the change of lifestyle, the incidence of chronic diseases shows a continuous upward trend. Chronic diseases are characterized by a long course, slow development, and non-obvious early symptoms. Once they develop to the middle and late stages, they often bring serious health damage and economic burden to patients. Therefore, early detection and intervention of chronic diseases are of great significance for improving the quality of life of patients and reducing medical costs.

[0003] In related technologies, medical institutions mainly carry out chronic disease screening by means of regular physical examinations and single-disease monitoring. During the physical examination, medical staff will conduct routine examinations and laboratory tests on patients, obtain basic physiological indicators such as blood pressure and blood sugar, and combine the symptoms described by the patients for diagnosis and analysis by specialist doctors. For patients with diagnosed chronic diseases, specific indicators are dynamically monitored through wearable devices such as smart bracelets, and follow-up reviews are carried out regularly.

[0004] However, in related technologies, wearable devices and the like are used, and the detection accuracy of relevant data monitoring of personnel is often insufficient due to the use of a fixed threshold alarm mechanism. Summary of the Invention

[0005] This application provides a method and system for early detection of chronic diseases based on a multimodal large model, which is used to improve the accuracy of early detection of chronic diseases.

[0006] In a first aspect, this application provides a method for early detection of chronic diseases based on a multimodal large model, which is applied to a pathological monitoring system. The method includes: obtaining physiological signal data stream, clinical test data stream, and symptom description data stream of a target user within a target time window from a pathological database to generate an original multimodal data set; performing timestamp unification and numerical standardization processing on the original multimodal data set to obtain a time series feature sequence; inputting the time series feature sequence into a pre-trained multimodal large model to obtain abnormal symptom features; calculating the similarity between the abnormal symptom features and the feature vectors of the labeled cases in the historical case database, and selecting the cases with similarity exceeding a preset similarity threshold as matching cases; extracting the diagnosis results and development processes of the matching cases, and combining with a medical knowledge graph to determine the disease risk level and development trend corresponding to the original multimodal data set to obtain a pathological evaluation result; generating a warning signal including risk types and intervention suggestions according to the pathological evaluation result.

[0007] In the above embodiments, the pathological monitoring system acquires multimodal data and combines temporal feature processing, multimodal large model analysis, and medical knowledge graphs to achieve early and accurate detection of chronic diseases; it can capture weak abnormal signs, predict the disease development trend through similarity matching with historical cases, and generate targeted early warnings and intervention suggestions, improving the user experience.

[0008] Combined with some embodiments of the first aspect, in some embodiments, before the step of inputting the temporal feature sequence into the pre-trained multimodal large model to obtain the abnormal sign features, the method further includes: receiving the local training parameters of multiple terminal nodes to obtain an initial model parameter set; performing weighted processing on the initial model parameter set to obtain a fused model parameter; performing a gradient clipping operation on the fused model parameter to obtain the clipped parameter; adding Gaussian noise to the clipped parameter to obtain the final training parameter; constructing a neural network structure based on the final training parameter to obtain the multimodal large model.

[0009] In the above embodiments, the pathological monitoring system performs weighted fusion, gradient clipping, and noise addition processing on the training parameters of multiple terminal nodes to construct a more robust multimodal large model; this distributed training method can make full use of the data resources of different medical institutions while protecting data privacy, improving the generalization ability and anti-interference ability of the model.

[0010] Combined with some embodiments of the first aspect, in some embodiments, after the step of generating a warning signal including risk type and intervention suggestions according to the pathological evaluation result, the method further includes: extracting the abnormal index features in the pathological evaluation result to obtain the target monitoring index; obtaining the historical change data of the target user corresponding to the target monitoring index to obtain the benchmark parameter range; detecting the fluctuation value of the target monitoring index based on the benchmark parameter range to obtain the index change feature; determining the pathological intervention parameter according to the index change feature and generating an intervention correction plan according to the pathological intervention parameter.

[0011] In the above embodiments, the pathological monitoring system analyzes the abnormal indexes in the pathological evaluation result and establishes a personalized benchmark parameter range in combination with the user's historical data, realizing the dynamic adjustment and accurate intervention of the monitoring index; this adaptive intervention mechanism can better adapt to the individual differences of different patients, improving the pertinence and effectiveness of the intervention plan.

[0012] Combined with some embodiments of the first aspect, in some embodiments, the step of unifying the timestamps and normalizing the numerical values of the original multimodal dataset to obtain the temporal feature sequence specifically includes: interpolating and complementing the timestamps in the original multimodal dataset to obtain an equally spaced sampling sequence; performing a normalization operation on the numerical features in the equally spaced sampling sequence to obtain the normalized features; arranging the normalized features in time sequence to generate the temporal feature sequence.

[0013] In the above embodiments, the pathological monitoring system performs timestamp interpolation and numerical standardization on the original multi-modal data, solving the problem of temporal alignment of multi-source heterogeneous data; this preprocessing method ensures the comparability and consistency of different modal data, laying a foundation for subsequent feature extraction and model analysis.

[0014] Combined with some embodiments of the first aspect, in some embodiments, the step of generating a warning signal including a risk type and an intervention suggestion according to the pathological evaluation result specifically includes: extracting an abnormal index combination in the pathological evaluation result to obtain a risk feature set; determining intervention parameters corresponding to the risk feature set according to a medical knowledge graph and constructing a parameter matrix; inputting the parameter matrix into a preset intervention generation model to obtain an intervention plan set; and generating a warning signal according to the risk feature set and the intervention plan set.

[0015] In the above embodiments, the pathological monitoring system analyzes the abnormal index combination in the pathological evaluation result and constructs an intervention parameter matrix in combination with the medical knowledge graph, realizing the automatic generation of warning signals; this knowledge-driven intervention plan generation mechanism not only ensures the professionalism of the intervention suggestions but also improves the interpretability of the warning information.

[0016] Combined with some embodiments of the first aspect, in some embodiments, before the step of generating a warning signal including a risk type and an intervention suggestion according to the pathological evaluation result, the method further includes: obtaining a change curve of the detection index of the target user to generate a dynamic feature sequence; determining an index threshold set according to the reference index range in the medical knowledge graph; calculating the deviation degree between the dynamic feature sequence and the index threshold set to obtain a deviation feature; and correcting the pathological evaluation result based on the deviation feature.

[0017] In the above embodiments, the pathological monitoring system introduces a dynamic feature sequence and the reference index range of the medical knowledge graph to realize real-time correction of the pathological evaluation result; this dynamic evaluation mechanism can timely capture the changing trend of the condition, improving the timeliness and accuracy of the evaluation result.

[0018] Combined with some embodiments of the first aspect, in some embodiments, the step of determining an index threshold set according to the reference index range in the medical knowledge graph specifically includes: extracting the time-series index data of the matching cases to generate an evolution feature sequence; calculating the index change trend based on the evolution feature sequence to obtain a trend parameter; adjusting the reference index range in the medical knowledge graph according to the trend parameter to obtain a corrected threshold; and determining the index threshold set according to the corrected threshold and the dynamic feature sequence.

[0019] In the above embodiment, the pathology monitoring system analyzes the evolution characteristics of the time series indicators of the matching cases, dynamically adjusts the reference indicator range in the medical knowledge graph, and realizes the adaptive update of the threshold; this threshold optimization method based on case evolution improves the sensitivity and specificity of indicator monitoring.

[0020] In a second aspect, an embodiment of the present application provides a pathology monitoring system, comprising: one or more processors and a memory; the memory is coupled to the one or more processors, the memory being used to store computer program code, the computer program code comprising computer instructions, the one or more processors calling the computer instructions to cause the pathology monitoring system to perform the method described in the first aspect and any possible implementation of the first aspect.

[0021] In a third aspect, an embodiment of the present application provides a computer program product comprising instructions, which, when run on a pathology monitoring system, enables the pathology monitoring system to perform the method described in the first aspect and any possible implementation of the first aspect.

[0022] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium comprising instructions. When the instructions are executed on a pathology monitoring system, the pathology monitoring system executes the method described in the first aspect and any possible implementation of the first aspect.

[0023] It is understandable that the pathology monitoring system provided in the second aspect, the computer program product provided in the third aspect, and the computer storage medium provided in the fourth aspect are all used to execute the methods provided in the embodiments of the present application. Therefore, the beneficial effects that can be achieved can be referred to the beneficial effects of the corresponding methods and will not be repeated here.

[0024] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. Due to the adoption of feature extraction and multimodal large model analysis methods based on multimodal data streams, combined with the evaluation mechanism of historical case matching and medical knowledge graphs, it is able to comprehensively capture multi-dimensional information such as patients' physiological signals, clinical tests and symptom descriptions, and discover potential abnormal signs through time series feature analysis, effectively solving the problem of insufficient accuracy of single indicator monitoring and inability to effectively identify early weak symptoms in existing technologies, thereby realizing early and accurate detection and risk warning of chronic diseases, improving the accuracy and timeliness of disease prediction, and providing a reliable decision-making basis for clinical intervention.

[0025] 2. Since the method for extracting target monitoring indicators based on abnormal indicator features is adopted, and personalized benchmark parameter ranges are established by combining with the user's historical data, and the change characteristics of indicators are determined through fluctuation value analysis, it is possible to realize the dynamic adjustment of monitoring indicators and the setting of personalized thresholds, effectively solving the problems that the fixed-threshold alarm mechanism in the prior art is difficult to adapt to individual differences and the intervention plan lacks pertinence. Furthermore, it realizes more accurate disease monitoring and the generation of personalized intervention plans, improving the effect of chronic disease management and patient compliance.

[0026] 3. Since the method for dynamic feature sequence analysis based on the change curve of detection indicators is adopted, and deviation analysis and evaluation result correction are carried out by combining with the reference indicator range of the medical knowledge graph, it is possible to capture the disease change trend in real time and perform dynamic evaluation and adjustment, effectively solving the problems that the evaluation result in the prior art lags behind and cannot reflect the disease change in time. Furthermore, it realizes the real-time optimization and dynamic update of the pathological evaluation result, improving the real-time performance and accuracy of disease monitoring, and providing a more timely and reliable basis for clinical intervention decision-making. Brief Description of the Drawings

[0027] Figure 1 is a schematic flowchart of a method for early detection of chronic diseases based on a multi-modal large model in an embodiment of the present application; Figure 2 is another schematic flowchart of a method for early detection of chronic diseases based on a multi-modal large model in an embodiment of the present application; Figure 3 is a schematic structural diagram of an entity device of a pathological monitoring system in an embodiment of the present application. Detailed Embodiments

[0028] The terms used in the following embodiments of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification of the present application, the singular forms "a", "an", "the above", "the", and "this" are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in the present application refers to any or all possible combinations including one or more of the listed items.

[0029] Hereinafter, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as implying or suggesting relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the embodiments of the present application, unless otherwise specified, the meaning of "a plurality" is two or more.

[0030] For ease of understanding, the application scenarios of the embodiments of the present application are introduced below.

[0031] The amount of health monitoring data processed by a certain medical information system has been continuously increasing every year. The data analysis in 2024 shows that in the multi-modal data collected by the system, the recognition accuracy of abnormal indicators is relatively low, and the timeliness of the data analysis results is insufficient. Especially when dealing with multi-source heterogeneous data, due to problems such as difficult temporal matching between different data modalities and insufficient feature extraction, it is difficult for the system to timely detect weak abnormal patterns in the data. This reflects the technical bottlenecks existing in the existing data analysis system in multi-modal data fusion and abnormal pattern recognition.

[0032] In related technologies, the detection of data anomalies can be achieved by adopting a single-modal data analysis method with a fixed threshold. The following introduces the scenario of using the early chronic disease detection method based on a multi-modal large model in related technologies.

[0033] A certain data analysis system mainly relies on a single type of sensor device to collect data. For example, the system only monitors a single physiological indicator through a specific sensor and triggers an alarm when the value exceeds a preset threshold. However, in actual applications, due to differences in the basic parameters of different users, the fixed threshold standard often cannot accurately reflect the actual status of an individual. For example, although multiple indicators of a certain user do not exceed the warning line preset by the system, the subtle change trends and combined characteristics of these indicators actually already show abnormal signs. This single-dimensional and static threshold monitoring method is difficult to capture complex patterns and potential associations in the data.

[0034] By adopting the early chronic disease detection method based on a multi-modal large model in the embodiments of the present application, through multi-source data fusion and dynamic feature extraction, the accurate recognition of weak abnormal patterns in the data is realized. It not only improves the accuracy of anomaly detection but also realizes the intelligent generation of warning signals. The following introduces the scenario of using the early chronic disease detection method based on the multi-modal large model in the present application.

[0035] After a certain medical data monitoring center introduces the multi-modal data analysis system of the present application, it conducts intelligent analysis on the collected multi-source data. The system simultaneously processes multi-dimensional information such as various sensor signals, test data, and text descriptions. Through multi-modal large model analysis, when all indicators are within the reference range, the system can timely identify potential abnormal patterns by analyzing the fluctuation patterns, change trends, and feature associations of the data. Combining the evolution laws of similar cases in the historical database, the system can predict the development trend of abnormal patterns and provide a decision-making basis for early intervention.

[0036] It can be seen that by adopting the multi-modal data analysis method in the embodiments of the present application, while realizing data anomaly detection, it can also effectively solve the problems of difficult data fusion and insufficient feature extraction, and thus achieve more accurate data analysis and warning.

[0037] For ease of understanding, the following describes the process of the method provided by this implementation in combination with the above scenario. Figure 1 , which is a flow chart of the method for early detection of chronic diseases based on a multimodal large model in an embodiment of the present application.

[0038] S101. Acquire the physiological signal data stream, clinical test data stream, and symptom description data stream of a target user within a target time window from a pathology database to generate an original multimodal data set.

[0039] Among them, the pathology database refers to a structured data storage system that stores patient health-related data, including various physiological signals, test results and clinical records; the target time window refers to a specific continuous observation period, such as the last 30 days or 3 months; the physiological signal data stream refers to continuously collected vital signs data, such as heart rate, blood pressure, respiratory rate, etc.; the clinical test data stream refers to the time series of various laboratory test results, such as blood routine, biochemical indicators, etc.; the symptom description data stream refers to the text data of the patient's main symptoms and clinical observation records; the original multimodal dataset refers to a comprehensive data set that integrates the above-mentioned different types of data sources.

[0040] When a pathology monitoring system receives a health monitoring request for a target user, it must first obtain sufficient raw data as a basis for analysis. Specifically, the pathology monitoring system first determines the time range for data collection and then extracts various types of data about the target user within that time range from different data tables in the pathology database. For physiological signal data, the system obtains all available continuous monitoring data; for clinical test data, the system collects all completed test reports; and for symptom description data, the system extracts relevant medical records and follow-up records. Finally, the system integrates these heterogeneous data into a complete multimodal dataset according to a unified data structure.

[0041] In some embodiments, the acquisition and integration of multimodal data can be achieved in a variety of ways: Optionally, the pathology monitoring system can adopt a hierarchical query strategy, first establish a user identification index, then preliminarily filter the data according to the timestamp, then use the data type label to classify and extract, and finally organize the various types of data into a unified data structure; Optionally, the pathology monitoring system can adopt a streaming processing solution, by establishing a data pipeline, receiving and processing data streams from different sources in real time, using a cache mechanism to temporarily store data, and using a synchronizer to ensure the temporal integrity of the data, and finally merging the processed data into a unified format. It is understandable that other methods can also be used to achieve the acquisition and integration process of multimodal data, which are not limited here.

[0042] S102: Perform timestamp unification and numerical standardization on the original multimodal dataset to obtain a time series feature sequence.

[0043] Among them, timestamp unification means aligning the sampling time points of different data sources to a unified time base; numerical normalization refers to the normalization conversion of numerical features of different dimensions and ranges; time series feature sequence refers to a sequence of feature data arranged in chronological order and standardized; the original multimodal dataset refers to the initial data set containing multiple types of data.

[0044] After acquiring the raw multimodal data, the pathology monitoring system needs to pre-process the data to ensure the accuracy of subsequent analysis. Specifically, the pathology monitoring system first uniformly calibrates the timestamps of all data to solve the problem of time asynchrony between different devices and systems. Then, the system will identify missing values in the data and supplement them. Linear interpolation, spline interpolation and other methods can be used to fill in the missing time points. For numerical features, the system will calculate the statistical distribution of each feature and select appropriate standardization methods based on the nature of the feature, such as minimum-maximum standardization, Z-score standardization, etc. Finally, the system will reorganize the processed features in chronological order to form a regular time series feature sequence.

[0045] During numerical standardization, the system will adopt corresponding standardization strategies for different types of numerical features. For continuous features, the mean μ and standard deviation σ of the feature within the time window are first calculated, and then the Z-score standardization formula (x-μ) / σ is used to map the value to a distribution with a mean of 0 and a standard deviation of 1. For features with a certain range, the minimum-maximum standardization formula (x-min) / (max-min) is used to map the value to the interval [0, 1]. For features with long-tail distributions, a logarithmic transformation is first performed to reduce the skewness of the data, and then standardization is performed. The system also detects outliers and uses the quantile constraint method to truncate the value when the value exceeds the range of μ±3σ.

[0046] In some embodiments, data standardization and time series alignment can be achieved in a variety of ways: Optionally, the pathology monitoring system can adopt an adaptive time window method, dynamically adjust the alignment window size according to the sampling frequency of different types of data, calculate feature statistics through a sliding window, and use a kernel density estimation method to smooth data and handle outliers; Optionally, the pathology monitoring system can adopt a multi-level standardization strategy, first performing preliminary standardization within a local time window, then performing secondary calibration based on global statistical features, and finally applying time series correlation analysis to ensure the continuity and consistency of the feature sequence. It is understandable that other methods can also be used to achieve data standardization and time series processing, which are not limited here.

[0047] S103. Input the time-series feature sequence into the pre-trained multi-modal large model to obtain abnormal symptom features.

[0048] Among them, the pre-trained multi-modal large model refers to a deep learning model pre-trained with large-scale data, which can process various types of input data; the abnormal symptom features refer to the feature combinations recognized by the model that may indicate early signs of diseases; the time-series feature sequence represents pre-processed multi-modal data arranged in chronological order.

[0049] After the pathological monitoring system completes data pre-processing, it needs to use a deep learning model for feature extraction and anomaly detection. Specifically, the pathological monitoring system first organizes the time-series feature sequence according to the input format required by the model, including grouping features, supplementing time-series position encoding, etc. Then, the system extracts features from the input data through the encoder network of the multi-modal large model to obtain high-dimensional feature representations. The attention mechanism of the model will automatically focus on the correlation relationships between different time points and different modal features to identify potential abnormal patterns. Finally, the system maps the high-dimensional features to the abnormal symptom feature space through the decoder network of the model to obtain feature representations with clinical interpretability.

[0050] It should be noted that the multi-modal large model collects training parameters from multiple terminal nodes through federated learning, performs weighted fusion, gradient clipping, and Gaussian noise processing on the parameters. The training data includes multi-modal data such as physiological signal data streams, clinical test data streams, and symptom description data streams. The training objective is to minimize the loss function of the model in the abnormal symptom recognition task while ensuring the model's processing ability for different modal data and feature extraction effects. The training process adopts a distributed method. Each terminal uploads the parameters after training based on local data, and the central node performs parameter aggregation. The multi-modal large model adopts a deep neural network structure, including feature extraction sub-networks for different data modalities, cross-modal attention mechanisms, and multi-head self-attention networks. The model can process time-series feature sequences, capture long-range dependencies, and achieve the fusion and feature extraction of multi-modal data. The model has an encoder-decoder structure. The encoder is responsible for feature extraction, and the decoder is responsible for mapping to the abnormal symptom feature space. The system inputs the pre-processed time-series feature sequence, which contains standardized data of multiple modalities. The model processes the data of each modality through the feature extraction network, analyzes the correlations between features using the attention mechanism, and finally outputs an abnormal symptom feature vector. These feature vectors are used for subsequent similarity matching and risk assessment.

[0051] The multi-modal large model adopts a hierarchical feature extraction architecture. At the bottom layer, a one-dimensional convolutional neural network is used to extract time-frequency features for physiological signals, a fully connected layer is used to extract statistical features for clinical test data, and a Transformer encoder is used to extract semantic features for text descriptions. At the middle layer, the cross-modal attention mechanism is used to calculate the correlation weights between different modal features to achieve adaptive fusion of features. Specifically, for feature vectors vi and vj, the dynamic weight assignment between features is realized by calculating the attention score aij = softmax(vi·Wq + vj·Wk + T), where Wq and Wk are learnable parameter matrices. At the high layer, a multi-head self-attention network is used to capture long-range dependencies, and each attention head is responsible for extracting temporal patterns at different scales.

[0052] In some embodiments, feature extraction and anomaly detection can be implemented in various ways: Optionally, the pathological monitoring system can adopt a hierarchical feature extraction strategy. First, a dedicated sub-network is used to process data of different modalities, then the features of each modality are fused through a cross-modal attention mechanism, and finally, a multi-head self-attention network is used to capture long-range temporal dependencies; Optionally, the pathological monitoring system can adopt a self-supervised learning method, and the internal structure of the data is learned by designing pre-training tasks, such as temporal prediction, anomaly reconstruction, etc., so as to improve the sensitivity of the model to abnormal patterns. It can be understood that other ways can also be adopted to implement feature extraction and anomaly detection, which are not limited here.

[0053] S104. Calculate the similarity between the abnormal symptom features and the feature vectors of the labeled cases in the historical case database, and select the cases with similarity exceeding the preset similarity threshold as matching cases.

[0054] Among them, the historical case database refers to a case database storing complete diagnosis and treatment records and annotation information; the feature vector refers to a high-dimensional numerical vector representing the case features; the similarity represents the distance metric between two feature vectors; the preset similarity threshold refers to the standard value for determining case similarity; the matching case refers to a historical case with a similar feature pattern to the current case.

[0055] After obtaining the abnormal symptom features, the pathological monitoring system needs to find similar cases in the historical case database to assist in diagnosis. Specifically, the pathological monitoring system first converts the abnormal symptom features into a standardized feature vector representation. Then, the system will traverse all the cases in the historical case database and calculate the similarity between their feature vectors and the current case. Various measurement methods such as cosine similarity and Euclidean distance can be used. The system will screen out the cases with higher similarity according to the preset similarity threshold and sort them according to the similarity size. Finally, the system extracts the selected matching cases and their complete clinical information as the reference basis for subsequent analysis.

[0056] It should be noted that the pathological monitoring system adopts a multi-level similarity calculation method. First, the cosine similarity cos(θ) = (a·b) / (||a||·||b||) is calculated for the feature vectors to obtain the overall similarity score. Then, the dynamic time warping (DTW) algorithm is used to calculate the alignment distance of the time series pattern. This algorithm finds the optimal alignment path between two time series through dynamic programming and can handle time scale and phase differences. Finally, the system assigns weight coefficients wi to different features based on professional knowledge and calculates the comprehensive similarity score through weighted summation Σwi·si, where si is each similarity index.

[0057] In some embodiments, case matching and similarity calculation can be implemented in various ways: Optionally, the pathological monitoring system can adopt a multi-level similarity calculation method. First, coarse-grained matching is performed based on the disease type and basic features, then the similarity of detailed features is calculated in the candidate set, and finally, the comprehensive similarity scores of multiple dimensions are comprehensively sorted; Optionally, the pathological monitoring system can adopt a dynamic weighting strategy, dynamically adjust the feature weights according to the importance of different features in disease diagnosis, and optimize the similarity calculation results in combination with expert knowledge. It can be understood that other ways can also be adopted to implement case matching and similarity analysis, which are not limited here.

[0058] S105. Extract the diagnosis results and development processes of the matching cases, and combine with the medical knowledge graph to determine the disease risk level and development trend corresponding to the original multi-modal data set, and obtain the pathological evaluation result.

[0059] Among them, the diagnosis result represents the disease diagnosis information confirmed clinically; the development process refers to the complete evolution process of the disease from the early signs to the diagnosis; the medical knowledge graph represents a structured knowledge base containing medical concepts such as diseases, symptoms, treatments, and their relationships; the disease risk level refers to the hierarchical assessment of the possibility of disease development; the development trend represents the possible evolution direction and speed of the disease; the pathological evaluation result refers to the disease risk assessment report obtained through comprehensive analysis.

[0060] After obtaining the matching cases, the pathological monitoring system needs to conduct in-depth pathological analysis and risk assessment. Specifically, the pathological monitoring system first extracts the complete diagnosis and treatment process information from the matching cases, including the symptoms at the first visit, the change trends of various examination results, treatment plans and their effects, etc. Then, the system queries the medical knowledge graph to obtain the typical development paths, risk factors, and prognosis characteristics of related diseases. The system matches the characteristics of the current case with the disease development patterns in the knowledge graph, and at the same time refers to the actual development process of the matching cases to comprehensively evaluate the disease risk. Finally, the system determines the risk level according to the evaluation results and predicts the possible development trend to generate a detailed pathological evaluation report.

[0061] It should be noted that the pathological monitoring system constructs a risk assessment framework based on a probabilistic graphical model. First, a Bayesian network model is established, where the nodes represent various indicators and the edges represent the conditional dependence relationships between the indicators. By calculating the conditional probability distribution P(Y|X), the risk probabilities corresponding to different combinations of indicators are evaluated, where X is the observed abnormal feature and Y is the risk level. The system uses variational inference algorithms to solve the posterior probability distribution and divides the risk levels into multiple levels according to probability thresholds. At the same time, the system also considers the time factor and uses an exponential decay function to adjust the weights of historical data, so that recent abnormal features have a higher influence.

[0062] In some embodiments, pathological evaluation and risk prediction can be achieved in various ways: Optionally, the pathological monitoring system can adopt an evaluation method based on knowledge reasoning. By constructing a Bayesian network model of disease development and combining the conditional probability relationships in the medical knowledge graph, the impacts of different risk factors are quantitatively analyzed, and finally the risk assessment results are obtained through probability reasoning; Optionally, the pathological monitoring system can adopt a time series pattern mining method. By analyzing the index change sequences in matching cases, typical disease development patterns are identified, a time series prediction model is established, and the possible development trajectory of the current case is predicted in combination with the characteristics of the current case. It can be understood that other ways can also be adopted to achieve pathological evaluation and risk prediction, which are not limited herein.

[0063] S106. Generate a warning signal including the risk type and intervention suggestions according to the pathological evaluation result.

[0064] Among them, the risk type represents the specific category and severity of the potential disease risk; the intervention suggestions refer to the preventive and treatment measures proposed for the identified risks; the warning signal represents the risk prompt and intervention guidance information generated by the system; the pathological evaluation result refers to the comprehensive evaluation report obtained from the previous analysis.

[0065] After the pathological monitoring system completes the pathological evaluation, it needs to generate warning information with practical guiding significance. Specifically, the pathological monitoring system first classifies the pathological evaluation results to identify the main risk types and influencing factors. Then, the system formulates targeted intervention suggestions according to the characteristics of different risk types, in combination with the treatment guidelines and expert experience in the medical knowledge graph. The system takes into account the individual characteristics of the patient (such as age, underlying diseases, etc.) and objective conditions (such as the accessibility of medical resources) to ensure the feasibility of the intervention suggestions. Finally, the system integrates the risk prompt and intervention suggestions into a warning signal and sets different warning levels according to the urgency.

[0066] In some embodiments, the generation of warning signals can be achieved in various ways: Optionally, the pathological monitoring system can adopt a hierarchical warning strategy, designing different forms of warning information for different risk levels, including real-time monitoring reminders, regular assessment reports, and long-term management suggestions, and dynamically adjusting the warning frequency and content according to risk changes; Optionally, the pathological monitoring system can adopt a personalized suggestion generation method, generating intervention suggestions more in line with personal characteristics and providing specific and operable implementation steps by establishing a patient portrait model and analyzing factors such as their living habits and compliance. It can be understood that other methods can also be used to generate and push warning signals, which are not limited here.

[0067] In the above embodiments, through multi-modal data fusion and deep learning analysis, the detection accuracy of the system is improved. In practical applications, the system can adaptively adjust analysis parameters to achieve more accurate abnormal pattern recognition. The following supplements the scenarios of this embodiment.

[0068] During continuous operation, the analysis mechanism of the pathological monitoring system is continuously optimized. For example, for long-term collected user data, the system can establish a dynamic reference range based on historical data and adaptively adjust analysis parameters according to conditions such as environmental factors and activity patterns. When the system detects abnormal fluctuations in the data, it will analyze the causes of the anomalies in combination with the knowledge graph and generate corresponding warning signals. For example, when detecting minor changes in certain indicators, the system will comprehensively analyze multi-dimensional data features to provide accurate risk assessment results, achieving more intelligent and personalized data analysis and warning.

[0069] After combining the above scenarios, the following further describes the more specific process of the method provided in this embodiment. Please refer to Figure 2 , which is another process schematic diagram of the early detection method for chronic diseases based on a multi-modal large model in the embodiments of the present application.

[0070] S201. Obtain the physiological signal data stream, clinical test data stream, and symptom description data stream of the target user within the target time window from the pathological database to generate an original multi-modal data set.

[0071] Referring to step S101, the pathological monitoring system will obtain the original multi-modal data set.

[0072] It should be noted that the pathology monitoring system utilizes a multimodal data processing framework based on the Transformer architecture to fuse data from multiple sensor sources, including millimeter-wave radar chips, photoelectric sensors, and temperature and humidity sensors. First, a time alignment algorithm is used to unify data with different sampling frequencies to the same time scale, and missing values are handled using linear or spline interpolation. Feature extraction is then performed on each modality, for example, by performing time-frequency analysis to extract dynamic features from millimeter-wave radar signals, waveform features from photoelectric sensor data, and statistical features from temperature and humidity data. Finally, the features of each modality are mapped using the formula e^(m) = W^(m) * Z^(m) + b^(m). A unified feature representation is generated using a weighted fusion mechanism Z = Σ(α^(m) * e^(m)), where α^(m) is calculated using the softmax function and reflects the importance of each modality.

[0073] S202: Perform timestamp unification and numerical standardization on the original multimodal dataset to obtain a time series feature sequence.

[0074] Referring to step S102 , the pathology monitoring system generates a time series feature sequence.

[0075] In some embodiments, the pathology monitoring system will perform numerical completion and normalization, that is, the pathology monitoring system will interpolate and complete the timestamps in the original multimodal data set to obtain an equally spaced sampling sequence; perform normalization operations on the numerical features in the equally spaced sampling sequence to obtain standardized features; and arrange the standardized features in time series to generate a time series feature sequence.

[0076] Among them, timestamp interpolation completion means estimating and filling the data of missing time points; equally spaced sampling sequence refers to a data sequence with a uniform time interval; normalization operation is used to represent the mathematical processing of converting data of different dimensions to a uniform scale; standardized features represent unified format data after normalization; time series feature sequence refers to a standardized feature data stream arranged in chronological order.

[0077] After acquiring the raw data, the pathology monitoring system needs to perform data preprocessing to ensure the accuracy of subsequent analysis. Specifically, the pathology monitoring system first analyzes the time distribution characteristics of the raw data and identifies the time points where the data is missing. For different types of data, the pathology monitoring system adopts different interpolation strategies: spline interpolation is used for continuous data, and nearest neighbor interpolation is used for discrete data. Then, the pathology monitoring system normalizes different features, using minimum-maximum normalization or Z-score normalization for continuous features, and encoding categorical features. Finally, the pathology monitoring system organizes all standardized features in chronological order to ensure temporal alignment between features.

[0078] In some embodiments, data preprocessing and standardization can be achieved in various ways: Optionally, the pathological monitoring system can adopt an adaptive interpolation strategy. First, a time series decomposition model is constructed to identify the periodicity and trend of the data. Then, the most suitable interpolation method is selected based on the local data pattern. Finally, the interpolation result is optimized through a sliding window smoothing technique. Optionally, the pathological monitoring system can adopt a multi-stage standardization scheme. By calculating the statistical distribution characteristics of the features, a feature-related transformation function is designed, and a robustness process is carried out in combination with an outlier detection mechanism. It can be understood that other ways can also be adopted to achieve data preprocessing and standardization, which are not limited here.

[0079] S203. Input the time series feature sequence into the pre-trained multi-modal large model to obtain abnormal symptom features.

[0080] Referring to step S103, the pathological monitoring system will determine the abnormal symptom features based on the model.

[0081] In some embodiments, before this step S203, the pathological monitoring system will perform model construction, that is, the pathological monitoring system will receive the local training parameters of multiple terminal nodes to obtain an initial model parameter set; perform a weighting process on the initial model parameter set to obtain a fused model parameter; perform a gradient clipping operation on the fused model parameter to obtain the clipped parameter; add Gaussian noise to the clipped parameter to obtain the final training parameter; construct a neural network structure based on the final training parameter to obtain a multi-modal large model.

[0082] Among them, the terminal node represents each data source device participating in federated learning; the local training parameter refers to the model weights obtained by each terminal based on local data training; the initial model parameter set represents the aggregated original training parameter set; the fused model parameter refers to the unified parameter after weighted integration; the gradient clipping operation is used to represent the mathematical process of restricting the parameter update amplitude; the Gaussian noise refers to the random perturbation obeying the normal distribution; the final training parameter represents the parameter set used for model construction after all processes are completed; the multi-modal large model refers to a deep neural network model capable of processing various types of data.

[0083] When the pathological monitoring system performs distributed model training, it is necessary to integrate the training results of multiple terminals and ensure data security. Specifically, the pathological monitoring system first receives the local training parameters uploaded by each terminal through a secure communication channel and performs integrity verification. Then, the pathological monitoring system assigns weight coefficients according to the data quality and credibility of each terminal and fuses the parameters through a weighted average method. To prevent the model from being unstable due to excessive parameter updates, the pathological monitoring system will set a gradient threshold for clipping. To enhance the privacy protection ability of the model, the pathological monitoring system adds random noise that conforms to a specific distribution to the parameters. Finally, the pathological monitoring system initializes each layer of the neural network with the processed parameters to complete the construction of the multi-modal large model.

[0084] The core of the multi-modal large model adopts the Transformer architecture, and feature extraction is achieved through the self-attention mechanism Attention(Q, K, V) = softmax(QK^T / √d_k)V. Among them, Q, K, and V represent Query, Key, and Value respectively, and d k is the dimension size of the key. For the input multi-modal data z^(m), first, the feature representations of each modality are obtained through independent Transformer encoders, z^(m) = Transformer(X^(m); θ). In the feature fusion stage, a cross-modal attention mechanism is used to calculate the correlation weights between different modality features to achieve adaptive feature fusion. The specific calculation formulas include: e (m) = W (m) Z (m) + b (m) ; The model improves its feature extraction ability through self-supervised pre-training, and at the same time adopts a federated learning framework for distributed training to improve the generalization performance of the model while protecting data privacy.

[0085] It should be noted that the pathological monitoring system adopts a strict differential privacy protection mechanism, and its mathematical definition is P(M(D) ∈ S) ≤ e^(∈) * P(M(D′) ∈ S) + δ. Among them, D and D' represent any two adjacent data sets (i.e., data sets with only one piece of data different), M represents the differential privacy mechanism, ∈ is the privacy budget parameter used to control the privacy protection intensity, and δ is the acceptable failure probability. In practical applications, a balance can be achieved between data utility and privacy protection by adjusting the ε and δ parameters.

[0086] The pathological monitoring system also adopts a gradient clipping and noise addition strategy during the parameter update process, and its calculation formula is Among them, g_i represents the gradient of a single sample, C is the gradient clipping threshold used to limit the maximum norm of the gradient, L is the number of samples in the batch, δ is the noise coefficient used to control the intensity of the added noise, and N(0, δ2C2I) represents Gaussian noise with a mean of 0 and a variance of δ2C2. In this way, both gradient explosion can be prevented and the privacy of the training data can be protected through noise addition.

[0087] In some embodiments, parameter fusion and privacy protection in federated learning can be achieved in a variety of ways: Optionally, the pathology monitoring system can adopt an adaptive aggregation strategy, first perform similarity analysis on the parameters of each terminal, establish a parameter clustering model, and then dynamically adjust the fusion weights based on the clustering results. Finally, the weighted median method is used for parameter integration to effectively reduce the impact of abnormal parameters; Optionally, the pathology monitoring system can adopt a differential privacy protection scheme, determine the noise intensity by analyzing the model sensitivity, add perturbations using the Laplace mechanism or the exponential mechanism, and balance the privacy protection strength and model performance in combination with a dynamic budget allocation strategy. It is understandable that other methods can also be used to achieve parameter fusion and privacy protection processing, which are not limited here.

[0088] S204: Calculate the similarity between the abnormal sign features and the feature vectors of the marked cases in the historical case database, and select cases whose similarity exceeds a preset similarity threshold as matching cases.

[0089] Referring to step S104 , the pathology monitoring system determines matching cases.

[0090] It should be noted that the pathology monitoring system uses a weighted similarity approach when calculating feature vector similarity, taking into account the characteristics of multimodal data. First, the cosine similarity is calculated for the feature vectors of each modality. Then, based on the importance weights α^(m) learned during the pre-training phase for each modality, a weighted summation is used to obtain a comprehensive similarity score.

[0091] S205. Extract the diagnostic results and development process of the matching cases, and determine the disease risk level and development trend corresponding to the original multimodal dataset in combination with the medical knowledge graph to obtain the pathological assessment results.

[0092] Referring to step S105 , the pathology monitoring system generates a pathology assessment result.

[0093] S206: Obtain a detection index change curve of the target user and generate a dynamic feature sequence.

[0094] Among them, the detection indicator change curve represents the trend graph of various physiological indicators changing over time; the dynamic feature sequence refers to the time series data structure that describes the indicator change pattern; the target user represents the specific individual who needs health monitoring; the change curve refers to a continuous function that reflects the law of data change over time.

[0095] After obtaining the preliminary pathological assessment results, the pathological monitoring system needs to perform dynamic feature analysis to improve the accuracy of the assessment. Specifically, the pathological monitoring system first extracts all relevant detection indicators from the multi-modal dataset and arranges them in chronological order to form a time series. For continuously monitored indicators, the system will use interpolation algorithms to construct smooth change curves; for discretely sampled indicators, the system will reconstruct the continuous change process through curve fitting methods. Then, the system extracts features from these change curves, including calculating the change rate, fluctuation amplitude, periodic features, etc., and finally generates sequence data that can characterize the dynamic change features of the indicators.

[0096] In some embodiments, the generation of the dynamic feature sequence can be achieved in multiple ways: Optionally, the pathological monitoring system can adopt a time series decomposition method. First, it decomposes the original time series into trend components, seasonal components, and random components, then extracts features from each component separately, and finally obtains a complete dynamic feature representation through feature fusion; Optionally, the pathological monitoring system can adopt a sliding window-based feature extraction strategy. By setting time windows of different scales, it calculates statistical features, morphological features, and frequency domain features within the window to construct multi-scale dynamic feature descriptions. It can be understood that other methods can also be used to achieve the dynamic feature extraction of detection indicators, which is not limited here.

[0097] S207. Determine the index threshold set according to the reference index range in the medical knowledge graph.

[0098] Among them, the reference index range represents the normal value interval of various physiological indicators recognized in medicine; the index threshold set refers to the set of dynamic boundary values used for abnormality determination; the medical knowledge graph refers to a structured knowledge base system containing medical professional knowledge.

[0099] After obtaining the dynamic feature sequence, the pathological monitoring system needs to establish a scientific judgment criterion. Specifically, the pathological monitoring system first retrieves the standard reference range of the target monitoring indicator from the medical knowledge graph, including stratification criteria such as age and gender. Then, the system will make personalized adjustments to the standard range in combination with the individual characteristics and basic conditions of the patient. The system will also consider the dynamic change characteristics of the indicator and set threshold values for different alarm levels, such as multiple levels including mild abnormality, moderate abnormality, and severe abnormality. Finally, the system integrates these threshold values into a complete threshold determination system.

[0100] In some embodiments, the determination of the metric threshold can be achieved in various ways: Optionally, the pathological monitoring system can adopt an adaptive threshold method based on population statistics. By analyzing the metric distribution characteristics of a large number of healthy people and patient groups, a dynamic probability model is established, and the threshold range is dynamically adjusted according to the fluctuation law of the metrics. Optionally, the pathological monitoring system can adopt a threshold determination strategy that combines multiple factors, consider the mutual influence relationship between metrics, construct a multi-dimensional threshold space, and optimize the threshold boundary through machine learning methods. It can be understood that other ways can also be adopted to determine the metric threshold, which is not limited here.

[0101] In some embodiments, the pathological monitoring system will perform precise matching of the metric threshold based on matching cases, that is, the pathological monitoring system will extract the time-series metric data of the matching cases to generate an evolution feature sequence; calculate the metric change trend based on the evolution feature sequence to obtain a trend parameter; adjust the reference metric range in the medical knowledge graph according to the trend parameter to obtain a corrected threshold; and determine the metric threshold set according to the corrected threshold and the dynamic feature sequence.

[0102] Among them, the time-series metric data represents the continuous monitoring data during the disease development process; the evolution feature sequence refers to the time series that describes the disease evolution law; the trend parameter represents the parameter that quantitatively describes the change direction and speed of the metric; the reference metric range refers to the normal value range specified in medical knowledge; the corrected threshold represents the judgment standard adjusted according to the actual situation; and the metric threshold set refers to the set of dynamic boundary values used for anomaly determination.

[0103] After obtaining the matching cases, the pathological monitoring system needs to analyze the disease development law to optimize the monitoring standard. Specifically, the pathological monitoring system first extracts the complete monitoring data from the matching cases, focusing on the change process of the metric values. Then, the pathological monitoring system calculates the change rate, acceleration, and fluctuation characteristics of each metric through time series analysis methods to construct a trend description model. The system compares these trend parameters with the standard range in the medical knowledge base and dynamically adjusts the warning threshold according to the actually observed change law. Finally, the system determines the personalized multi-level warning threshold in combination with the characteristics of the current monitoring data.

[0104] In some embodiments, the dynamic adjustment of the threshold can be achieved in various ways: Optionally, the pathological monitoring system can adopt a threshold optimization strategy based on the temporal pattern. By analyzing the index change patterns in historical cases, a dynamic probability model is established to predict the possible change range of the index, and an adaptive warning threshold is set according to the risk tolerance; Optionally, the pathological monitoring system can adopt a multi-dimensional joint threshold method. By constructing a correlation network between indexes, analyzing the co-variation characteristics of the index group, and establishing a multi-dimensional threshold determination space, more accurate anomaly detection can be realized. It can be understood that other ways can also be adopted to achieve the dynamic optimization and adjustment of the threshold, which are not limited herein.

[0105] S208. Calculate the deviation degree between the dynamic feature sequence and the index threshold set to obtain the deviation feature.

[0106] Among them, the deviation degree represents a quantitative index of the deviation of the actual index value from the threshold range; the deviation feature refers to the feature vector describing the degree and duration of the anomaly; the dynamic feature sequence represents the temporal data reflecting the index change; the index threshold set refers to the set of reference boundary values for determining the anomaly.

[0107] After the pathological monitoring system determines the index threshold, it needs to perform deviation analysis on the actual index value. Specifically, the pathological monitoring system first compares the dynamic feature sequence with the corresponding threshold range and calculates the deviation degree at each time point. The system will consider the direction of the deviation (exceeding the upper limit or the lower limit), the amplitude (the degree of deviation from the normal range), and the duration. For the combination of multiple indexes, the system will calculate the comprehensive deviation score and analyze the mutual influence between the indexes. Finally, the system integrates these deviation information into a structured deviation feature for subsequent evaluation and correction.

[0108] In some embodiments, the calculation of the deviation feature can be achieved in various ways: Optionally, the pathological monitoring system can adopt a hierarchical deviation calculation method. First, calculate the deviation statistics of a single index on different time scales, then analyze the co-deviation pattern between the indexes, and finally obtain the overall deviation feature through weighted fusion; Optionally, the pathological monitoring system can adopt a deviation evaluation strategy based on fuzzy logic. By defining a fuzzy membership function to describe the membership degree of the index value to different anomaly degrees, and using a fuzzy rule inference system to comprehensively evaluate the multi-dimensional deviation feature. It can be understood that other ways can also be adopted to achieve the calculation and analysis of the deviation feature, which are not limited herein.

[0109] S209. Modify the pathological evaluation result based on the deviation feature.

[0110] Among them, the pathological evaluation result represents the preliminary disease risk assessment report; the modification refers to adjusting the evaluation result according to the new deviation information; the deviation feature refers to the quantitative anomaly degree index.

[0111] After obtaining the deviation features, the pathological monitoring system needs to optimize and adjust the preliminary evaluation results. Specifically, the pathological monitoring system first analyzes the consistency between the deviation features and the preliminary evaluation results to identify the evaluation items that may have deviations. Then, the system makes corresponding adjustments to the risk level and development trend prediction according to the nature and degree of the deviation. For the indicators with deviations, the system will focus on analyzing their clinical significance and updating the evaluation conclusion in combination with medical knowledge. Finally, the system generates a revised evaluation report to ensure that the evaluation results can more accurately reflect the actual situation of the patient.

[0112] In some embodiments, the correction of the evaluation results can be achieved in various ways: Optionally, the pathological monitoring system can adopt a correction method based on the weight of evidence. By calculating the reliability score of the deviation features, the weights of different correction factors are determined, and then the evaluation results are adjusted in a weighted manner; Optionally, the pathological monitoring system can adopt a dynamic feedback correction strategy. By establishing a closed-loop system of evaluation - feedback - correction, the parameters of the evaluation model are continuously optimized to improve the accuracy of the evaluation results. It can be understood that other ways can also be used to achieve the correction of the evaluation results, which are not limited here.

[0113] S210. Generate a warning signal including the risk type and intervention suggestions according to the pathological evaluation results.

[0114] Referring to step S106, the pathological monitoring system will generate a warning signal.

[0115] In some embodiments, the pathological monitoring system will perform risk intervention, that is, the pathological monitoring system will extract the abnormal index combination in the pathological evaluation results to obtain a risk feature set; determine the intervention parameters corresponding to the risk feature set according to the medical knowledge graph, and construct a parameter matrix; input the parameter matrix into a preset intervention generation model to obtain an intervention plan set; generate a warning signal according to the risk feature set and the intervention plan set.

[0116] Among them, the abnormal index combination represents multiple related abnormal health indicators; the risk feature set refers to a set of feature vectors describing health risks; the intervention parameter represents a quantitative index for guiding intervention measures; the parameter matrix refers to the structured representation of multi-dimensional intervention parameters; the intervention plan set represents a set of feasible intervention measure suggestions; the warning signal refers to the risk prompt and intervention guidance information generated by the system.

[0117] After the pathological monitoring system completes the assessment, it needs to generate specific intervention suggestions. Specifically, the pathological monitoring system first analyzes the abnormal indicators in the assessment results, identifies the combinations of indicators with clinical relevance, and constructs a risk feature vector. Then, the pathological monitoring system queries the medical knowledge graph to obtain the intervention reference values corresponding to each risk feature, including treatment goals, intervention intensity, monitoring frequency, etc. The system organizes these parameters into a structured matrix form as the input for generating intervention plans. Through a pre-trained intervention generation model, the system generates a series of feasible intervention plans. Finally, the system generates a hierarchical warning signal based on the risk level and the urgency of the intervention plan.

[0118] It should be noted that the intervention generation model is trained based on the intervention parameters in the medical knowledge graph and the historical intervention effect data. The input data includes structured data such as risk feature sets and intervention parameter matrices. The training goal is to optimize the feasibility and effectiveness of the intervention plans generated by the model, and adjust the model parameters through expert feedback and historical data verification. The training adopts a supervised learning method, using the labeled successful intervention cases as the standard. The model constructs a state-action value model based on the deep reinforcement learning framework. It includes a feature encoding layer, a policy network, and a value network. The model can generate the optimal intervention strategy according to the input state information (risk features and parameter matrices). The system inputs the risk feature set and the parameter matrix, and the model generates a series of feasible intervention plans through the policy network and evaluates the expected effects of each plan through the value network. Finally, it outputs an optimized set of intervention plans for generating warning signals.

[0119] The pathological monitoring system continuously monitors the changes in the health status through a real-time feature analysis algorithm. When an abnormal signal is detected, it immediately activates a hierarchical warning mechanism. The generation of warning information is based on multi-dimensional feature analysis, including the degree of abnormality of physiological indicators, changes in behavior patterns, and the influence of environmental factors, etc. The system automatically generates personalized intervention suggestions according to the warning level, including lifestyle adjustment suggestions, precise medication management reminders, and targeted health guidance programs. At the same time, the system will continuously track the intervention effect and dynamically adjust the intervention strategy through a feedback mechanism to ensure the effectiveness of the intervention measures.

[0120] In some embodiments, the generation of intervention plans can be achieved in various ways: Optionally, the pathological monitoring system can adopt a knowledge reasoning-driven plan generation strategy. By constructing a medical rule reasoning engine, combining the severity of risk features and clinical guidelines, it derives a suitable combination of intervention parameters and optimizes the feasibility of the intervention plan through causal relationship analysis; Optionally, the pathological monitoring system can adopt the deep reinforcement learning method. By establishing a state-action value model, it learns the optimal intervention strategy and tunes the generated plan in combination with expert experience. It can be understood that other ways can also be used to achieve the generation of intervention plans and the generation of warning signals, which are not limited here.

[0121] S211. Extract abnormal indicator features from the pathological assessment results to obtain target monitoring indicators.

[0122] Among them, abnormal indicator characteristics refer to health indicator characteristics that exceed the normal range or show abnormal change trends; target monitoring indicators refer to key health indicators that need to be focused on and tracked; pathological assessment results refer to the health status assessment report obtained after comprehensive analysis.

[0123] After completing the pathology assessment, the pathology monitoring system needs to determine the key indicators to monitor. Specifically, the system first analyzes the pathology assessment results, identifying all indicators that are abnormal or pose potential risks. The system then ranks the indicators based on their clinical importance, degree of abnormality, and intercorrelations. For highly correlated groups of indicators, the system selects the most representative indicators as monitoring targets. Finally, the system integrates these key indicators into a target monitoring indicator set.

[0124] In some embodiments, the determination of target monitoring indicators can be achieved in a variety of ways: Optionally, the pathology monitoring system can adopt a multi-level screening strategy, first performing an initial screening based on clinical importance, then analyzing the correlation between indicators to construct an indicator network, and finally determining key node indicators through centrality analysis, and combining medical knowledge to determine the final set of monitoring indicators; Optionally, the pathology monitoring system can adopt a dynamic priority method, by establishing an indicator importance scoring model, taking into account the abnormality, change trend and predictive value of the indicators, and dynamically adjusting the priority of the monitoring indicators. It is understandable that other methods can also be used to achieve the screening and determination of monitoring indicators, which are not limited here.

[0125] S212: Obtain historical change data of the target monitoring indicator corresponding to the target user to obtain a benchmark parameter range.

[0126] Among them, historical change data represents the target user's past health monitoring records; the baseline parameter range refers to the reference interval that reflects the indicator fluctuation range under the individual's normal state; the target monitoring indicator represents the set of health indicators that need to be focused on.

[0127] After determining the target monitoring indicators, the pathology monitoring system needs to establish personalized evaluation criteria. Specifically, the pathology monitoring system first retrieves the target user's historical monitoring data from the database, including routine examination records and continuous monitoring data. The system cleans and filters this historical data, removing data from abnormal periods and retaining data segments that represent the user's normal state. Then, the system uses statistical analysis methods to calculate the typical distribution characteristics of each indicator, including mean, standard deviation, and periodic changes. Finally, based on these statistical characteristics, the system determines a personalized baseline parameter range for each monitoring indicator.

[0128] In some embodiments, the determination of the reference parameter range can be achieved in various ways: Optionally, the pathological monitoring system can adopt an adaptive interval estimation method. First, time-weight the historical data, focusing on the recent stable data, then use kernel density estimation to calculate the probability distribution of the indicators, and finally determine the personalized parameter range based on the confidence interval theory; Optionally, the pathological monitoring system can adopt a multi-dimensional reference range construction strategy. By analyzing the co-variation relationship between indicators, establish a multi-dimensional joint distribution model, so as to obtain a more accurate reference parameter space. It can be understood that other ways can also be used to calculate and optimize the reference parameter range, which is not limited here.

[0129] S213. Detect the fluctuation value of the target monitoring indicator based on the reference parameter range to obtain the indicator change characteristics.

[0130] Among them, the fluctuation value represents the degree of deviation of the indicator from the reference range; the indicator change characteristics refer to the set of characteristics describing the dynamic change law of the indicator; the reference parameter range represents the indicator change interval in the individual's normal state.

[0131] After establishing the reference parameter range, the pathological monitoring system needs to analyze the real-time change characteristics of the indicators. Specifically, the pathological monitoring system first compares the current monitoring data with the reference parameter range and calculates the deviation degree at each time point. The system will analyze the time patterns of these deviation values, including the duration of deviation, change rate, and fluctuation frequency, etc. For multiple related indicators, the system will study the linkage relationship between them and identify potential abnormal patterns. Finally, the system integrates these dynamic characteristics into a structured description of the indicator change characteristics.

[0132] In some embodiments, the analysis of the indicator change characteristics can be achieved in various ways: Optionally, the pathological monitoring system can adopt a multi-scale time series analysis method. Through techniques such as wavelet transform or empirical mode decomposition, decompose the indicator fluctuation into components of different time scales, analyze the change characteristics at each scale respectively, and finally fuse to obtain a complete description of the change characteristics; Optionally, the pathological monitoring system can adopt a dynamic pattern recognition strategy. By establishing a state transition model, describe the conversion law of the indicator between different states, and combine sequence pattern mining techniques to identify typical change patterns. It can be understood that other ways can also be used to extract and analyze the indicator change characteristics, which is not limited here.

[0133] S214. Determine the pathological intervention parameters according to the indicator change characteristics, and generate an intervention correction plan according to the pathological intervention parameters.

[0134] Among them, pathological intervention parameters represent quantitative indicators used to guide intervention measures; intervention modification plans refer to targeted health intervention recommendations; indicator change characteristics represent a set of features that describe the dynamic change patterns of indicators; and modification plans represent intervention measures adjusted based on the latest monitoring results.

[0135] After identifying the characteristics of indicator changes, the pathology monitoring system needs to develop a scientific intervention plan. Specifically, the pathology monitoring system first assesses the urgency and intensity of intervention based on the characteristics of the indicator changes. The system then combines treatment guidelines from the medical knowledge base to determine the appropriate intervention parameters for each abnormal indicator, including the target value, adjustment step, and monitoring frequency. The system then comprehensively considers the intervention needs of multiple indicators, coordinates any potential conflicts in intervention measures, and formulates a unified intervention strategy. Finally, based on these intervention parameters, the system generates specific implementation recommendations, including lifestyle adjustments, medication recommendations, and follow-up appointments.

[0136] In some embodiments, the formulation of intervention plans can be achieved through a variety of methods: Optionally, the pathology monitoring system can adopt a feedback control strategy, by establishing an indicator response model, predicting the effects of different intervention measures, dynamically adjusting intervention parameters, achieving precise control of intervention intensity, and continuously optimizing the intervention plan based on actual feedback; Optionally, the pathology monitoring system can adopt a multi-objective optimization method, by establishing an intervention effect evaluation model, simultaneously considering multiple health goals, balancing the effects and risks of intervention measures, and generating the optimal combination of intervention parameters. It is understandable that other methods can also be used to achieve the generation and optimization of intervention plans, which are not limited here.

[0137] It should be noted that the indicator response model is trained using indicator change data and intervention records from historical monitoring data. Inputs include indicator time series data, intervention parameters, and effect feedback. The training goal is to accurately predict the impact of different interventions on the indicator. Using a time series prediction training paradigm, model parameters are optimized by minimizing prediction error. The model utilizes a recurrent neural network architecture, comprising a time series feature extraction layer and a prediction layer. The model can process time series data and predict the response trend of indicators to intervention measures. The system inputs the current indicator status and the intervention parameters to be evaluated, and the model predicts the indicator's change trend over a period of time. The output is a predicted response curve, which is used to dynamically adjust the intervention plan.

[0138] In the embodiments of the present application, due to the use of feature extraction and deep learning analysis methods based on multimodal data streams, combined with an intelligent evaluation mechanism of historical case matching and medical knowledge graphs, and a personalized dynamic threshold adjustment strategy, it is possible to comprehensively capture multi-dimensional information such as the patient's physiological signals, clinical tests, and symptom descriptions, discover potential abnormal signs through time series feature analysis, and achieve dynamic adjustment and precise intervention of monitoring indicators. This effectively solves the problem in the existing technology that single indicator monitoring and fixed threshold alarms are difficult to achieve early and precise detection, thereby achieving early warning and personalized intervention for chronic diseases, significantly improving the accuracy of disease prediction and the pertinence of intervention measures, and providing more reliable decision support for clinical practice.

[0139] The following describes the pathology monitoring system in the embodiment of the present invention from the perspective of hardware processing. Figure 3 , is a schematic diagram of the physical device structure of the pathology monitoring system in an embodiment of the present application.

[0140] It should be noted that Figure 3 The structure of the pathology monitoring system shown is only an example and should not limit the functions and scope of use of the embodiments of the present invention.

[0141] like Figure 3 As shown, the pathology monitoring system includes a CPU 301, which can perform various appropriate actions and processes based on programs stored in a ROM 302 or programs loaded from a storage unit 308 into a RAM 303, such as executing the methods described in the above embodiments. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An I / O interface 305 is also connected to the bus 304.

[0142] The following components are connected to the I / O interface 305: an input section 306 including an audio input device, push button switches, etc.; an output section 307 including a liquid crystal display (LCD), an audio output device, indicator lights, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN (Local Area Network) card or a modem. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to the I / O interface 305 as needed. Removable media 311, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed in the drive 310 as needed, so that computer programs read therefrom can be installed into the storage section 308 as needed.

[0143] In particular, according to an embodiment of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, an embodiment of the present invention includes a computer program product comprising a computer program carried on a computer-readable medium, the computer program including a computer program for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 309 and / or installed from the removable medium 311. When the computer program is executed by the CPU 301, the various functions defined in the present invention are performed.

[0144] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present invention. Each box in the flowchart or block diagram can represent a module, program segment, or part of the code, and the above-mentioned module, program segment, or part of the code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings.

[0145] Specifically, the pathology monitoring system of this embodiment includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, the early detection method of chronic diseases based on a multimodal large model provided in the above embodiment is implemented.

[0146] As another aspect, the present invention further provides a computer-readable storage medium, which may be included in the pathology monitoring system described in the above embodiments, or may exist independently and not be incorporated into the pathology monitoring system. The storage medium carries one or more computer programs, which, when executed by a processor of the pathology monitoring system, enable the pathology monitoring system to implement the multimodal large model-based chronic disease early detection method provided in the above embodiments.

[0147] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

[0148] As used in the above embodiments, the term “when…” may be interpreted to mean “if…” or “after…” or “in response to determining…” or “in response to detecting…”, depending on the context. Similarly, the phrases “upon determining…” or “if (stated condition or event) is detected” may be interpreted to mean “if determining…” or “in response to determining…” or “upon detecting (stated condition or event)” or “in response to detecting (stated condition or event)”, depending on the context.

Claims

1. A method for early detection of chronic diseases based on a multimodal large model, characterized in that Applied to a pathological monitoring system, the method includes: Obtain the physiological signal data stream, clinical test data stream, and symptom description data stream of the target user within the target time window from the pathological database, and generate an original multimodal dataset; Perform timestamp unification and numerical standardization processing on the original multimodal dataset to obtain a time series feature sequence; Input the time series feature sequence into a pre-trained multimodal large model to obtain abnormal symptom features; Calculate the similarity between the abnormal symptom features and the feature vectors of the labeled cases in the historical case database, and select the cases with similarity exceeding the preset similarity threshold as matching cases; Extract the diagnosis results and development processes of the matching cases, and combine with the medical knowledge graph to determine the disease risk level and development trend corresponding to the original multimodal dataset, and obtain a pathological evaluation result; Generate a warning signal including risk types and intervention suggestions according to the pathological evaluation result.

2. The method according to claim 1, wherein Before the step of inputting the time series feature sequence into a pre-trained multimodal large model to obtain abnormal symptom features, the method further includes: Receive the local training parameters of multiple terminal nodes to obtain an initial model parameter set; Perform weighted processing on the initial model parameter set to obtain a fused model parameter; Perform a gradient clipping operation on the fused model parameter to obtain the clipped parameter; Add Gaussian noise to the clipped parameter to obtain the final training parameter; Construct a neural network structure based on the final training parameter to obtain a multimodal large model.

3. The method according to claim 1, characterized in that After the step of generating a warning signal including risk types and intervention suggestions according to the pathological evaluation result, the method further includes: Extract the abnormal index features in the pathological evaluation result to obtain the target monitoring index; Obtain the historical change data of the target user corresponding to the target monitoring index to obtain a benchmark parameter range; Detect the fluctuation value of the target monitoring index based on the benchmark parameter range to obtain an index change feature; Determine the pathological intervention parameter according to the index change feature, and generate an intervention correction plan according to the pathological intervention parameter.

4. The method according to claim 1, wherein The step of performing timestamp unification and numerical standardization processing on the original multimodal dataset to obtain a time series feature sequence specifically includes: Perform interpolation and completion on the timestamps in the original multimodal dataset to obtain an equally spaced sampling sequence; Perform a normalization operation on the numerical features in the equally spaced sampling sequence to obtain standardized features; Arrange the standardized features in time series to generate a time series feature sequence.

5. The method according to claim 1, characterized in that, The step of generating a warning signal including risk types and intervention suggestions according to the pathological evaluation result specifically includes: Extract the abnormal index combination in the pathological evaluation result to obtain a risk feature set; Determine the intervention parameters corresponding to the risk feature set according to the medical knowledge graph and construct a parameter matrix; Input the parameter matrix into a preset intervention generation model to obtain an intervention plan set; Generate a warning signal according to the risk feature set and the intervention plan set.

6. The method according to claim 1, wherein Before the step of generating a warning signal including risk types and intervention suggestions according to the pathological evaluation result, the method further includes: Obtain the detection index change curve of the target user and generate a dynamic feature sequence; Determine an index threshold set according to the reference index range in the medical knowledge graph; Calculate the deviation degree between the dynamic feature sequence and the index threshold set to obtain a deviation feature; Correct the pathological evaluation result based on the deviation feature.

7. The method according to claim 6, wherein The step of determining an index threshold set according to the reference index range in the medical knowledge graph specifically includes: Extract the time-series index data of the matching cases and generate an evolution feature sequence; Calculate the index change trend based on the evolution feature sequence to obtain a trend parameter; Adjust the reference index range in the medical knowledge graph according to the trend parameter to obtain a corrected threshold; Determine an index threshold set according to the corrected threshold and the dynamic feature sequence.

8. A pathological monitoring system, characterized in that, The pathological monitoring system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to enable the pathological monitoring system to execute the method according to any one of claims 1-7.

9. A computer-readable storage medium, comprising instructions, characterized in that, When the instruction runs on the pathological monitoring system, the pathological monitoring system is enabled to execute the method according to any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product runs on the pathological monitoring system, the pathological monitoring system is enabled to execute the method according to any one of claims 1-7.

Images