Disease remote diagnosis system based on artificial intelligence driving

Abstract

The invention relates to the technical field of artificial intelligence medical diagnosis, and discloses a remote disease diagnosis system based on artificial intelligence driving. The system comprises five functional modules, a real-time medical data acquisition module acquires multi-modal physiological data of a patient from a remote terminal in real time, and the multi-modal physiological data comprises vital sign time sequence data, medical image data and symptom description text data; the cross-modal correlation analysis module carries out feature fusion on the data, extracts a correlation mode between modals and generates an abnormal feature data set with a weight mark; establishing a dynamic diagnosis baseline by the dynamic diagnosis baseline model, and adapting to individual physiological feature changes by updating parameters; the abnormal behavior detection module is used for comparing the real-time data with a baseline, identifying abnormal indexes and generating a marking result; the deep pathogenesis tracing module analyzes a potential disease mode and traces a pathogenesis path in combination with a patient's historical medical record. The system can optimize the remote diagnosis effect and assist clinical diagnosis and treatment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence medical diagnosis, in particular to a disease remote diagnosis system based on artificial intelligence driving. BACKGROUND

[0002] In the modern medical system, remote diagnosis has become an important way to solve the uneven distribution of medical resources and improve the accessibility of medical services, especially for patients in remote areas, the elderly and the disabled, and medical support in public health emergencies. However, the current mainstream disease remote diagnosis scheme still has many limitations, which is difficult to meet the accuracy and real-time needs of clinical diagnosis.

[0003] From the perspective of data collection, existing remote diagnosis systems mostly focus on the collection of single type of physiological data, such as only obtaining the vital sign data or single medical image data of patients, and cannot realize the synchronous collection of multi-dimensional physiological information. This single data collection mode makes the information available for reference in the diagnosis process more one-sided, and it is difficult to fully reflect the physical state of the patient, which affects the integrity of the diagnosis result. At the same time, although some systems try to collect multi-modal data, the data collection process lacks real-time, and there is often a problem of data delay. For diseases that need to dynamically monitor the changes of physiological indicators, such as cardiovascular diseases and intensive care scenarios, it is impossible to capture key physiological change information in time, which may lead to delay in diagnosis opportunity.

[0004] In the aspect of data processing and analysis, the existing systems lack the processing ability of multi-modal physiological data. Most systems can only analyze different types of data separately, and fail to establish the association between different data modalities, such as being unable to combine and analyze the abnormalities found in medical images with the fluctuations in vital sign time series data, resulting in the potential diagnostic information contained in the data being missed. In addition, the existing diagnosis models are mostly constructed based on general population data, lacking consideration of individual physiological characteristic differences, and the model parameters are fixed, which cannot be dynamically adjusted according to individual factors such as patient's age, gender, underlying diseases, physiological indicator change trend, etc. Such general models are prone to diagnosis bias when facing patients with large individual differences, especially for patients with rare diseases or special physiological characteristics, the diagnosis accuracy is difficult to guarantee.

[0005] In terms of anomaly detection and cause tracing, the anomaly detection function of existing remote diagnosis systems is mostly implemented based on fixed thresholds, that is, when a physiological indicator exceeds the preset range, it is determined to be abnormal. However, the normal physiological indicator range of different individuals is different, and fixed thresholds are difficult to adapt to individual differences, which may lead to false positive or false negative results. At the same time, even if the abnormal indicator is detected, the existing system is difficult to realize in-depth analysis of the potential disease mode and tracing of the cause path, often only indicating the abnormal indicator, and cannot provide effective reference information for doctors about the cause of the abnormality, possible related diseases and disease development path, resulting in that doctors still need to rely on a large amount of additional manual analysis work when formulating treatment plans, increasing the diagnosis and treatment burden and reducing the efficiency of remote diagnosis.

[0006] In actual clinical application, the existing remote diagnosis system has low degree of combination with patient historical medical record data, and cannot fully utilize the past diagnosis and treatment information of the patient to assist current diagnosis. The information such as disease history, medication history and treatment response contained in the historical medical record has important reference value for the diagnosis and cause analysis of the current disease, and the neglect of this part of information by the existing system further limits the accuracy and comprehensiveness of diagnosis. SUMMARY

[0007] The purpose of the present application is to provide an artificial intelligence driven disease remote diagnosis system to solve the problems raised in the background art.

[0008] To achieve the above purpose, the present application provides an artificial intelligence driven disease remote diagnosis system, which comprises:

[0009] A real-time medical data acquisition module is configured to acquire multi-modal physiological data of a patient in real time from a remote terminal device, wherein the multi-modal physiological data comprises time series data of vital signs, medical image data and symptom description text data.

[0010] A cross-modal correlation analysis module is configured to perform cross-modal feature fusion on the multi-modal physiological data, extract correlation patterns between different data modalities, and generate an abnormal feature data set with weight labels.

[0011] A dynamic diagnosis baseline model is configured to establish a dynamic diagnosis baseline according to the multi-modal physiological data and the abnormal feature data set, and adapt to the physiological characteristic changes of individual patients by updating model parameters in real time.

[0012] An abnormal behavior detection module is configured to compare the real-time acquired physiological data with the dynamic diagnosis baseline model, identify abnormal physiological indicators deviating from the normal range, and generate abnormal behavior labeling results.

[0013] A deep cause tracing module is configured to analyze potential disease patterns and trace cause paths based on the abnormal behavior labeling results and in combination with patient historical medical record data.

[0014] Preferably, the real-time medical data acquisition module comprises:

[0015] Capturing patient vital sign time series data from remote terminal devices and standardizing the data;

[0016] Parsing medical image data, extracting image features and generating structured image description data;

[0017] Integrating symptom description text data, performing semantic analysis and generating symptom feature vectors;

[0018] Fusing standardized vital sign data, structured image description data and symptom feature vectors into a unified format of multi-modal physiological data stream.

[0019] Preferably, the cross-modal correlation analysis module comprises:

[0020] Performing hierarchical feature extraction on the multi-modal physiological data stream to identify potential correlations between different modal data;

[0021] Analyzing abnormal interaction patterns based on context information, generating preliminary abnormal signals and recording abnormal types and severity;

[0022] Assigning weights according to the characteristics of abnormal signals and outputting abnormal feature data sets with weights.

[0023] Preferably, the dynamic diagnostic baseline model comprises:

[0024] Training an initial diagnostic model based on the multi-modal physiological data stream and the abnormal feature data set;

[0025] Adjusting model parameters dynamically through real-time data feedback to optimize the accuracy of the diagnostic baseline;

[0026] Generating a personalized diagnostic baseline in combination with individual patient differences.

[0027] Preferably, the abnormal behavior detection module comprises:

[0028] Inputting real-time collected physiological data into the dynamic diagnostic baseline model to calculate deviation values from normal behavior patterns;

[0029] Labeling deviation values that exceed threshold values to generate abnormal behavior labeling results;

[0030] Recording context information of abnormal behaviors, including timestamps, abnormal types and deviation amplitudes.

[0031] Preferably, the deep cause tracing module comprises:

[0032] Retrieving patient historical medical record data, extracting diagnosis records related to the abnormal behavior marker results;

[0033] Analyzing potential disease patterns and identifying possible etiological pathways;

[0034] Generating a comprehensive diagnosis report containing etiological analysis results.

[0035] Preferably, the system further comprises:

[0036] An environmental factor compensation module for adjusting the diagnosis baseline according to patient environmental data, including temperature, humidity, and air quality indicators;

[0037] Combining environmental data to correct abnormal physiological indicators and improve diagnosis accuracy.

[0038] Preferably, the environmental factor compensation module comprises:

[0039] Collecting real-time environmental data of the patient and calculating the influence coefficient of environmental factors on physiological indicators;

[0040] Adjusting the parameters of the dynamic diagnosis baseline model based on the influence coefficient;

[0041] Generating a corrected diagnosis result after environmental compensation.

[0042] Preferably, the system further comprises:

[0043] A model optimization feedback module for adjusting model parameters based on the deviation between actual diagnosis results and predicted results;

[0044] Optimizing the prediction ability of the dynamic diagnosis baseline model through a continuous learning mechanism.

[0045] Preferably, the model optimization feedback module comprises:

[0046] Collecting actual diagnosis data, calculating prediction errors, and generating error correction coefficients;

[0047] Updating the weight parameters of the dynamic diagnosis baseline model based on the error correction coefficients;

[0048] Outputting the optimized diagnosis model for subsequent diagnosis tasks.

[0049] Compared with the prior art, the beneficial effects of the present application are:

[0050] From the perspective of data collection and processing, the system sets up a real-time medical data collection module, which can realize the real-time synchronous collection of multi-modal physiological data of patients, covering time series data of vital signs, medical image data and symptom description text data and other multiple dimensions. This multi-modal data collection method breaks the limitations of traditional single data collection systems and can comprehensively collect various information reflecting the patient's physical condition, providing a rich and complete data basis for subsequent diagnosis. At the same time, the real-time collection feature ensures that data can be analyzed in a timely manner, which is particularly important for disease scenarios that require dynamic monitoring of physiological indicators. The system can quickly capture key physiological changes and avoid delays in diagnosis due to data delays.

[0051] The design of the cross-modal correlation analysis module realizes the deep fusion and correlation mining of multi-modal physiological data. Instead of processing different types of data separately, this module actively extracts the correlation patterns between different data modalities and generates an abnormal feature dataset with weight labels. This processing method can fully exploit the potential relationships between data, such as correlating the features of the lesion area in medical images with abnormal fluctuations in vital signs data. This allows the integration of diagnostic information that was previously scattered across different data sources, reducing information loss and providing more valuable reference for subsequent diagnosis, which helps to improve the comprehensiveness and reliability of the diagnosis results.

[0052] The construction of the dynamic diagnosis baseline model overcomes the limitations of traditional general diagnosis models that cannot adapt to individual differences. This model can establish a personalized dynamic diagnosis baseline for each patient based on the collected multi-modal physiological data and abnormal feature dataset, and continuously adapt to changes in the patient's physiological characteristics by updating the model parameters in real time. For example, for elderly patients, the model adjusts the baseline range based on their age-related physiological indicator characteristics. For patients with underlying diseases, the model dynamically optimizes the diagnosis criteria by taking into account the impact of the underlying disease on physiological indicators. This personalized diagnosis baseline setting makes the diagnosis process more tailored to the individual patient, effectively reducing the diagnostic bias caused by general models and improving the accuracy of the diagnosis. This is particularly important for patients with special physiological characteristics or rare diseases, as it can provide more accurate diagnosis results that are tailored to the individual's actual situation.

[0053] The abnormal behavior detection module compares real-time physiological data with the dynamic diagnosis baseline model, replacing the traditional fixed threshold detection method. Since the dynamic diagnosis baseline takes into account individual differences, this module can more accurately identify abnormal physiological indicators that deviate from the patient's normal range, significantly reducing the probability of false positive and false negative results. At the same time, the abnormal behavior label generated by the module can clearly indicate the specific type and deviation degree of the abnormal indicators, providing clear guidance for doctors to quickly locate patient physical abnormalities and reducing the workload of manual analysis during the abnormality identification process, thereby improving the efficiency of diagnosis.

[0054] The application of the deep etiology tracing module further expands the diagnostic assistance capability of the system. Based on the abnormal behavior marking result, the module combines patient historical medical record data to deeply analyze potential disease patterns and trace the etiology path. By integrating the patient's past disease history, medication history, treatment response, and other information, the module can associate the current abnormal indicators with historical diagnosis and treatment information, and mine the possible disease causes behind the abnormality, such as determining whether a physiological indicator abnormality is related to the patient's past underlying disease, or whether it may be caused by a specific drug side effect. At the same time, the module can also sort out possible disease development paths to provide a reference for doctors to understand the disease evolution process, help doctors more comprehensively grasp the patient's condition, and thus develop more targeted treatment plans, reduce the doctor's diagnosis and treatment burden, and improve the support of remote diagnosis for clinical treatment. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 a timing diagram of the disease remote diagnosis system based on artificial intelligence driving according to the present application;

[0056] Figure 2 a flowchart of the real-time medical data acquisition module;

[0057] Figure 3 a flowchart of the abnormal behavior detection module. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0059] Please refer to Figure 1 The present application provides a disease remote diagnosis system based on artificial intelligence driving, which comprises a real-time medical data acquisition module, a cross-modal correlation analysis module, a dynamic diagnosis baseline model, an abnormal behavior detection module, and a deep etiology tracing module.

[0060] The real-time medical data acquisition module is responsible for continuously acquiring time series data of vital signs, medical image data, and symptom description text data actively reported by patients from remote terminal devices deployed at the patient end. These data constitute the multi-modal basis for intelligent diagnosis. The cross-modal correlation analysis module receives the above-mentioned multi-modal data stream, and its core task is to break down the barriers between different data modalities, mine the deep correlations between vital signs, medical images, and text symptoms through advanced feature fusion techniques, and generate a set of abnormal feature data with weight labels. The dynamic diagnosis baseline model is not a static reference standard, but an intelligent agent that can continuously evolve. It establishes and continuously updates a personalized health baseline for the patient based on the individual's multi-modal physiological data history and the identified abnormal features, thereby accurately capturing the dynamic changes in their physiological characteristics. The abnormal behavior detection module acts as a "sentinel" for the system, comparing real-time incoming physiological data with the dynamic diagnosis baseline in real time, and acutely identifying any abnormal physiological indicators that deviate from the normal range, and generating detailed labeling results for each abnormal event. The deep cause tracing module is the "brain" of the system, which comprehensively considers the abnormal behavior labels detected in real time and the patient's rich historical medical record data, uses knowledge graphs and causal reasoning models to analyze potential disease development patterns, and trace the most likely cause path, providing in-depth insights for clinical decision-making.

[0061] Embodiment 1: refer to Figure 2 The real-time medical data acquisition module, as the data entry of the system, is responsible for interacting with various remote terminal devices and parsing multi-format raw data. The module internally builds a multi-protocol adaptation layer, which can support Bluetooth, ZigBee, Wi-Fi, and wired interfaces, ensuring that physiological parameter monitoring devices of different brands and models can seamlessly access the system. After the device connects through a secure handshake protocol, the module reads raw data packets from the device buffer at a preset time interval or triggered by events. These data packets usually contain unprocessed sensor readings, device status information, and time stamps accurate to milliseconds. For time series data of vital signs, such as ECG waveforms, pulse waves, non-invasive blood pressure values, and blood oxygen saturation readings, the module first calls a signal processing subroutine and applies wavelet transform or adaptive filtering-based algorithms to perform noise reduction on the raw signal, eliminating motion artifacts and environmental electromagnetic interference. Subsequently, the data standardization engine begins to work, which uniformly converts readings with different units (such as mmHg, mV, bpm) to internal standard measurement units and resamples signals with different sampling frequencies to align them to a unified time axis, ultimately generating a regular time series data set with consistent dimensions and sampling rates, laying the foundation for subsequent time series analysis.

[0062] The processing of medical image data follows another parallel but eventually merged data pipeline. This module is equipped with a DICOM-compliant communication service, which can securely receive image files containing rich metadata from remote imaging devices such as portable ultrasound, digital dermatoscope, and fundus camera. The image preprocessing unit first calibrates the received raw images, such as adjusting window width and window level to optimize display, or performing geometric correction to eliminate device-induced distortion. Next, the feature extraction engine is activated, which contains deep convolutional neural network models pre-trained on large-scale medical image datasets. These network models can automatically identify anatomical structure boundaries, abnormal density areas, and specific texture patterns in images. For example, for a lung X-ray, the network will automatically outline the heart shadow contour and lung field area, and detect whether there are ground glass opacities, consolidation, or nodular shadows. The extracted high-dimensional feature vector and the corresponding spatial location information are jointly encapsulated to generate a structured image description data, which is recorded in structured formats such as JSON or ProtocolBuffers, clearly describing "what part", "what feature" and "its quantitative attributes", so that the image information is elevated from the pixel level to the semantic level.

[0063] The processing of symptom description text data needs to deal with the inherent ambiguity and diversity of natural language. The system receives symptom descriptions input by patients or medical staff through a graphical interface or a voice-to-text interface. These descriptions are usually short and colloquial phrases or sentences, such as "feeling hot in the afternoon, accompanied by dry cough". The natural language processing submodule first performs word segmentation, part-of-speech tagging, and syntax analysis on the input text to build a basic text structure tree. Then, the entity recognition model scans the text to accurately locate symptom entities, body parts, severity modifiers, and time descriptions. The relationship extraction component further analyzes the syntactic dependency relationships between these entities to determine the temporal relationship between "fever" and "afternoon" and the "companion" relationship of "dry cough". Finally, all this information is mapped to a standardized medical terminology ontology (such as SNOMED CT) and converted into a fixed-dimensional symptom feature vector that not only contains the presence of symptoms but also encodes their intensity, duration, and association strength with other symptoms.

[0064] After the independent processing of the three types of data described above, the data fusion sub-module begins to work, which aligns and packs the three information streams of standardized vital signs data, structured image description data and symptom feature vectors in time and space. The fusion process strictly follows the time stamp of the data generation, and associates different types of data generated in the same time window into a multi-modal data instance. For example, the sudden increase in heart rate data at a certain time, the increased heart shadow feature detected in the chest image collected in the same period, and the patient's report of "sudden palpitation" text description are integrated into a data package with inherent logical relationship. Finally, these data packages are serialized into a unified format of multi-modal physiological data stream, which is continuously delivered to the downstream cross-modal association analysis module through a high-throughput data bus. After receiving the data stream, the cross-modal association analysis module's first task is to further mine the non-obvious statistical association and clinical significance between different modal data. The module uses a hierarchical feature extraction strategy to perform multi-scale and multi-granularity feature abstraction on the input data stream. For time series data of vital signs, the module not only focuses on its instantaneous value, but also uses long short-term memory network or time series convolution network to extract its short-term fluctuation pattern, long-term trend change and periodic rhythm feature. For structured image description data, the module further analyzes the importance of each dimension in the high-dimensional feature vector, focusing on those semantic features with higher correlation with clinical manifestations. For symptom feature vectors, the commonality of symptom clusters is understood through proximity analysis in embedding space.

[0065] After obtaining the deep feature representations of each modality, the module initiates a cross-modality correlation analysis engine. The core of this engine is a neural network structure based on attention mechanism, which can dynamically calculate the correlation weights between different modalities. For example, when the system analyzes a suspected heart failure case, the engine learns that there is a strong correlation between the text symptom feature of "nighttime paroxysmal dyspnea" and the vital sign feature of "nighttime elevated mean pulmonary artery pressure", and also pays attention to the pattern of the co-occurrence of the feature of "pulmonary markings thickening, Kerley B lines" in the chest image and the above two features. Through this cross-attention, the module can identify abnormal interaction patterns that single modality analysis cannot discover. Whenever such a potential abnormal pattern is detected, the system generates a preliminary abnormal signal, which is classified into a preset abnormal type and preliminarily evaluates its severity according to the significance and consistency of the involved features. The weight allocation component evaluates the importance of each preliminary abnormal signal. The evaluation criteria include but are not limited to: the matching degree of the abnormal pattern with the known disease atlas, the amplitude of the deviation of the abnormality involved physiological parameters from the individual baseline, and the mutual support degree between different modalities. An abnormal signal strongly supported by multiple modalities will be given a high weight, while a signal prompted by only a single modality weak evidence will obtain a lower weight. The output generates an abnormal feature data set with precise weight marking, which not only points out "where the problem may be", but also quantifies "how much attention this problem deserves", providing valuable target-oriented information for the establishment and update of the dynamic diagnosis baseline model. The whole process from raw data acquisition to the generation of weighted abnormal feature set forms an efficient and automated data processing closed loop, laying a solid foundation for high-level intelligent diagnosis.

[0066] Example 2: Initialization and construction of the dynamic diagnostic baseline model is a process that relies on the accumulation of historical data. During the startup phase, the system requests access to the multi-modal physiological data stream of the target individual over a period of time that should cover as much as possible the physiological fluctuations of the individual in a relatively healthy state. These data are pre-processed via the real-time medical data acquisition module and form a regular historical data sequence. At the same time, the cross-modal correlation analysis module scans this historical data to identify possible minor abnormalities or intrinsic patterns and generates a corresponding historical abnormal feature dataset. The training algorithm of the initial diagnostic model, for example, an ensemble learning algorithm based on gradient boosting or a deep autoencoder, takes as main input features the multi-modal physiological data stream in a healthy state and as regularization constraints or auxiliary learning objectives the abnormal feature dataset. The learning task of the model is to construct a complex mapping function that describes the intrinsic coordination between the individual's physiological parameters in the absence of significant disease and their normal dynamic range over time, thus establishing an initial health baseline profile for this particular individual. This baseline is not a single line but a multi-dimensional dynamic space that allows for normal fluctuations.

[0067] After the model is put into operation, its dynamic updating mechanism is activated, which is a closed-loop learning process realized through a real-time data feedback loop. The system continuously inputs the latest multi-modal physiological data stream collected and standardized into the current version of the dynamic diagnostic baseline model. Instead of passively receiving data, the model actively compares new data with its current expectations, calculates the prediction error. When a new data point is verified by subsequent clinical information as belonging to the individual's normal physiological variation, this data point is marked as a valid sample for model updating. The model uses online learning strategies, such as using small batch gradient descent or Bayesian update rules, to use these continuously flowing valid new samples to make small, incremental adjustments to the model's internal parameters. This adjustment allows the model to slowly track the individual's physiological characteristics over time, such as the decline in resting heart rate due to aging, the improvement in blood pressure due to regular exercise, or the periodic physiological rhythm changes due to seasonal changes. In this way, the diagnostic baseline avoids gradually deviating from the individual's true physiological state due to fixedness, thus maintaining its timeliness and accuracy.

[0068] Personalization is another essence of dynamic baseline modeling, and the model incorporates the mechanism of considering individual differences in the architecture design. In the whole life cycle of model training and updating, the individual's static attribute information, such as age, gender, genetic susceptibility, history of diagnosed chronic diseases, and dynamic lifestyle information, are introduced into the model as conditional variables or modal adjustment factors. These factors do not directly participate in the prediction of physiological data, but as conditional parameters, they affect the model's judgment of the normal fluctuation range of physiological data. For example, for an individual with a long history of hypertension, the model will adjust the expected baseline level when setting the normal range of his blood pressure fluctuation, making it more consistent with the individual's actual situation rather than the ideal range of the healthy population. For example, for an athlete individual, the model will learn his unique heart rate variability characteristics and lower resting heart rate, and identify it as a normal pattern rather than bradycardia. The realization of this deep personalization enables the model to establish a unique health benchmark for each user, greatly reducing false positives caused by individual differences, and improving the recognition sensitivity of real health threats.

[0069] The maintenance and visualization of the internal state of the model are also important aspects of implementation. The system will record the version change history of the key parameters of the diagnostic baseline model, forming a "health record" of the model itself. When the model updates its internal weights due to receiving new data, the summary information of the update, such as the adjustment amplitude, the physiological parameters most affected, and the update timestamp, will be safely recorded. This traceability allows system administrators or authorized medical professionals to review the evolution trajectory of the baseline model and understand how the model adapts to individual changes. In some application scenarios, the system can also provide a simplified visualization view of the current state of the baseline model, such as displaying the slow change trend of the individualized normal range of a key physiological parameter (such as systolic blood pressure) in the past few months in the form of a trend chart, which provides a visual reference for users to understand the long-term evolution of their physiological state.

[0070] The stability and safety mechanism of the dynamic baseline model also needs to be carefully designed. In order to prevent abnormal data or temporary data quality problems from polluting the model, a series of guardian conditions are included in the model update logic. For example, only physiological data that is normal for a continuous period of time and verified by multiple data sources will trigger a substantial update of model parameters. For sudden, isolated, and large amplitude outliers, the system will take a conservative approach and temporarily isolate them, not immediately used for model updating, but trigger an additional manual confirmation or in-depth inspection process. The model also has a rollback mechanism. If the system detects an abnormal increase in the alarm rate of subsequent data after an update, it can automatically revert to the model parameters of the last stable version, ensuring the robustness of the system. The operation of the entire dynamic baseline model is like an tireless private health consultant, constantly observing, learning, and adjusting to accurately depict the individual's evolving health blueprint and provide a solid and timely judgment standard for abnormal behavior detection.

[0071] Example 3: refer to Figure 3 The abnormal behavior detection module continuously receives the latest standardized multi-modal physiological data stream from the data acquisition front end. These data streams are completely consistent with the data format used when constructing the baseline, ensuring comparability. Inside the module, there is a high-efficiency data alignment engine that can match real-time streaming single data points or data slices with the expected value range at the corresponding time point in the dynamic baseline model based on accurate timestamps. For time series data of vital signs, the matching is done for each sampling point, while for image feature vectors and symptom vectors, instantaneous comparison is made at the time of their generation. The comparison process is not a simple threshold judgment, but a statistical deviation between the real-time observation value and the baseline expected distribution. This deviation is quantified as a comprehensive anomaly score. The calculation of this anomaly score considers multiple dimensions. For example, for vital sign parameters that follow a Gaussian distribution, the anomaly score may be related to the number of standard deviations that the observed value deviates from the expected mean. The calculation method can be represented as:

[0072] ;

[0073] Where the symbols have specific meanings: represents the final calculated comprehensive anomaly score, whose value reflects the degree of deviation of the current multi-modal data from the baseline as a whole; the symbol is an index variable used to traverse the physiological feature parameters contained in the current data instance; each specific physiological parameter, such as heart rate, systolic pressure, or the value of a certain image feature, is represented by ; the expected normal value center position of the th parameter provided by the dynamic baseline model for the individual at the current time is represented by Characterization; Parameter The baseline model is described as follows: The expected fluctuation range of each parameter under normal conditions, i.e., the individualized standard deviation; weighting coefficients. This is pre-defined by the cross-modal correlation analysis module, which indicates the first... The differences in clinical importance among the features are as follows: important features are given higher weight and have a greater influence on the overall score.

[0074] The system has a preset configurable anomaly score threshold. When the calculated anomaly score threshold is reached... When this threshold is exceeded, an abnormal behavior event is determined to have occurred. The module then generates a structured abnormal behavior labeling result. This result is a rich data object that not only contains a Boolean abnormality flag but also records in detail the specific timestamp that triggered the abnormality, the abnormal physiological indicator identifier mapped according to the feature index, the calculated original abnormality score, the threshold level exceeded, and a list of the specific deviations of all physiological parameters involved in the abnormal event. This detailed contextual information is fully encapsulated in the labeling result, providing ample data support for subsequent etiological analysis. After receiving the abnormal behavior labeling result, the deep etiological tracing module initiates its analysis process. The primary task of this module is to perform efficient and accurate relevant data retrieval. Based on the abnormality type identifier and time range provided in the labeling result, it automatically initiates a query request to the connected medical information system. The query aims to obtain patient historical medical record data that may be related to the abnormality, including but not limited to past diagnostic records, inpatient medical record summaries, laboratory test results, imaging reports, medication history, and surgical records. The retrieval process sets a reasonable time backtracking window and prioritizes extracting historical information most closely related to the current abnormal physiological system.

[0075] After integrating real-time anomaly markers with rich historical data, the module invokes its built-in disease pattern analysis engine. This engine is typically built on graph neural networks or probabilistic graphical models, internally encoding a wealth of medical knowledge, including disease-symptom associations, disease progression pathways, and etiological networks. The engine works by jointly reasoning with currently observed anomaly patterns and patterns implicit in historical data. For example, when the system detects that an elderly patient has persistent microalbuminuria (anomaly marker 1) combined with newly developed, mild controlled hypertension (anomaly marker 2), the engine immediately activates nodes in the knowledge graph related to "early diabetic nephropathy" and "hypertensive nephropathy." It also retrieves the patient's historical data; if it finds a long history of type II diabetes but no history of hypertension, the reasoning will tend to consider hypertension as a secondary manifestation, thus pointing the etiological path to the progression of diabetic complications rather than primary hypertension.

[0076] The tracing of etiological pathways is the deepening of the analysis process, the engine attempts to construct one or more logical chains that trace back from the current abnormal performance to the possible root cause, it evaluates the likelihood of each potential path, considering factors including the consistency of the path with known medical consensus, the continuity of each link in the path in time, and whether the path can most fully explain all the observed abnormal markers and historical information. For complex cases, the system may generate multiple alternative etiological pathways and calculate a confidence score for each path. Finally, all analysis results are summarized and structured into a comprehensive diagnostic report, which usually contains several main parts: first is the "summary of this abnormal finding", clearly listing all the significant abnormal physiological indicators detected and their severity; second is the "correlation analysis", which explains the possible internal relationship between these abnormal indicators and their common pathophysiological significance; the core part is the "potential etiology inference", listing the most likely etiological hypotheses in descending order of confidence, and briefly explaining the supporting evidence and reasoning logic for each hypothesis; the report may also contain a "suggested follow-up examination" section, which proposes laboratory or imaging examination items that may be needed to further clarify the diagnosis. This report is not a final diagnosis, but a powerful decision support document to help medical professionals quickly focus on the core of the problem and improve diagnostic efficiency and accuracy. The entire process from abnormal detection to etiological tracing reflects the system's deep ability to transform data into clinical insights.

[0077] Example 4: The design of the environmental factor compensation module is to identify and quantify the potential impact of external environmental conditions on physiological measurements, so as to distinguish between pathological changes and physiological adaptations. The module continuously collects data through a network of microsensors deployed in the patient's living environment or on wearable devices, which measure environmental temperature, relative humidity, atmospheric pressure, and air quality indicators such as fine particulate matter concentration and total volatile organic content, all of which are time-stamped with high precision synchronized with physiological data. The environmental physiology model is the core of the module, which embeds knowledge rules induced from a large number of environmental medical studies, which describe how common environmental variables affect the short-term physiological response of the human body. For example, in a high temperature and humidity environment, the body will cause skin vasodilation, compensatory heart rate acceleration, and increased sweating to enhance heat dissipation, which may cause a certain degree of increase in heart rate and core body temperature measurements.

[0078] The workflow of the environmental factor compensation module begins with real-time parsing of environmental data and calculation of influence coefficients. Upon receiving the latest environmental data package, the module immediately calls the built-in environmental-physiological response function for processing. This function takes into account the combined effects of multiple environmental parameters and calculates a dynamic influence coefficient for different physiological systems. For example, for the cardiovascular system, the module may calculate a "heart rate environmental influence coefficient" and a "blood pressure environmental influence coefficient". These coefficients are usually normalized values between -1 and 1. A positive value indicates that the environmental conditions tend to increase the physiological parameter, while a negative value indicates a decreasing effect. The absolute value of the coefficient reflects the strength of the influence. The calculation process refers to standard environmental physiology models and allows for fine-tuning based on individual baseline health conditions (such as whether there is a history of cardiovascular disease), making the estimation of influence coefficients more personalized.

[0079] To illustrate how environmental factors are quantified, consider a simple example scenario that demonstrates preset influence coefficients for the heart rate parameter under different environmental conditions. These coefficients are one of the bases for the model's calculations, as shown in Table 1.

[0080] Table 1: Heart rate environmental influence coefficients under different environmental combinations

[0081] After obtaining the real-time environmental impact coefficient, the module begins to adjust the parameters of the dynamic diagnostic baseline model or compensate and correct the real-time physiological readings. The specific strategy depends on the system design. One strategy is to dynamically adjust the boundaries of the diagnostic baseline. For example, when the system detects a sudden rise in environmental temperature and calculates a positive heart rate impact coefficient, it does not directly change the real-time heart rate reading, but temporarily and moderately raises the upper threshold of the normal heart rate range of the dynamic diagnostic baseline model. This means that in this specific environment, a relatively high heart rate reading may still be determined to be within the "compensated normal" range, thereby avoiding false positives of physiologically tachycardia caused by heat stress as a cardiac abnormality. Another strategy is to compensate at the source before data comparison, that is, to use the impact coefficient to reverse correct the original physiological data collected, estimate its equivalent value under standard environmental conditions, and then compare it with the baseline. The output of the environmental factor compensation module is an environmental modified diagnostic result that clearly distinguishes the original measurement of physiological indicators and the contribution of environmental factors. The report will indicate the current environmental conditions and their estimated impact on specific physiological parameters, and indicate whether the parameter still deviates from its individualized baseline after environmental compensation. For example, a modified report may state: "The current environmental conditions (temperature 32°C, humidity 75%) are expected to increase the resting heart rate by about 8-12 beats per minute. The patient's current measured heart rate is 92 beats per minute, and the equivalent heart rate after environmental compensation is about 80-84 beats per minute, which is within the normal fluctuation range of your individual baseline. This monitoring did not find pathological tachycardia." This processing greatly reduces false positive alarms caused by environmental interference.

[0082] The long-term learning ability of the module cannot be ignored. It can fine-tune its internal environmental-physiological response function parameters by comparing the physiological responses of different individuals under similar environmental conditions (under the premise of anonymization) and the responses of the same individual under the same environment multiple times. For example, if the system continuously observes that the actual increase in heart rate of a certain user under moderate temperature is always slightly higher than the initial prediction of the model, it can gradually and slightly increase the heart rate environmental impact coefficient for that user, making future compensation more accurate. The introduction of the environmental factor compensation module enables the remote diagnostic system to understand individuals in their real-life scenarios, significantly improving the reliability and practicality of health status assessment in complex and changing environments, and making the diagnostic conclusion more close to the physiological truth.

[0083] Example 5: The starting condition of the model optimization feedback module is based on the comparison of the diagnostic output produced by the system prelude module with the real-world clinical outcome. When the patient goes to a medical institution for final diagnosis according to the system prompt, or obtains a clear medical conclusion through gold standard methods such as laboratory examination, imaging examination, etc., these final and verified diagnostic information will be structured into the system. The model optimization feedback module actively collects these key data, including the name of the diagnosed disease, the severity classification, the related key physiological indicator abnormality confirmation, and the treatment response information. These data are accurately spatio-temporally aligned with the system's previous prediction records for this case, including abnormal behavior marker results, deep cause tracing reports, and their corresponding confidence levels. After data alignment, the module starts the prediction error calculation process, which is the core step of quantifying the system's performance. Error calculation is not a simple binary judgment (correct or incorrect), but a multi-dimensional, fine-grained error measurement for different levels of prediction output. For classification prediction, such as the system's inference of the underlying cause being "community-acquired pneumonia" instead of the final diagnosis of "acute bronchitis", the module will calculate the classification error, possibly using cross-entropy loss to quantify the difference between the prediction probability distribution and the true one-hot encoded label. For numerical prediction, such as the system's estimate of the abnormal deviation amplitude of a certain physiological indicator, the module will calculate the absolute error or mean square error between its predicted value and the final measured value. Even for the timing characteristics of anomaly detection, such as the prediction bias of the abnormal start time, it will also be included in the error calculation category. Through these calculations, the module generates a set of error correction coefficients for each case with final diagnosis results, which specifically quantify the size and direction of the system's bias in various prediction tasks.

[0084] After generating the error correction coefficients, the module enters the model parameter update phase, which is a key step to inject clinical feedback knowledge into the model. The update process usually uses incremental learning algorithms to avoid catastrophic forgetting and adapt to the continuous evolution of medical knowledge. These error correction coefficients are used as part of the loss signal, and their gradients on the millions of weight parameters in the dynamic diagnosis baseline model are calculated through backpropagation algorithms. The update strategy focuses on robustness, and for the gradient calculated from a single case, a small learning rate is usually applied for parameter update to avoid drastic fluctuations in the model due to the particularity of a single case. At the same time, the module may use a small batch update strategy, that is, after accumulating a certain number of clinical feedback cases, the parameters are adjusted in batches, so that the update direction is closer to the general rule learned from group feedback, thereby improving the generalization ability of the model.

[0085] Model version management and rollback mechanism is an important part of ensuring the safe and reliable operation of the system. After each substantial parameter update of the dynamic diagnosis baseline model, the system will automatically create a new model version snapshot and archive the old version model parameters. The new version of the model will enter a limited pilot application stage, during which the system will closely monitor its performance indicators, such as the sensitivity, specificity and consistency with clinical diagnosis in subsequent cases. If the performance of the new version model is significantly reduced during the pilot application, such as an abnormal increase in false positive rate, the system can automatically trigger a rollback procedure to quickly switch to the last stable version of the model, ensuring the continuity of the diagnosis service. This mechanism provides a safety cushion for bold iteration of the model.

[0086] The effect of model optimization is ultimately reflected in subsequent diagnosis tasks. The internal parameters of the dynamic diagnosis baseline model optimized by clinical feedback data contain more discriminative features learned from real cases. For example, in the case of misjudging "viral pneumonia" as "bacterial pneumonia" before, the error signal will prompt the model that there is a bias in the weight allocation when distinguishing certain image features (such as the distribution pattern of ground glass shadow) and specific inflammation indicators (such as the downward trend of lymphocytes). After parameter update, when the model encounters similar feature combinations again, it will give higher attention weight to more discriminative features. This continuous, evidence-based optimization enables the system to gradually identify more subtle disease patterns and reduce the likelihood of confusing different causes of similar symptoms. The output of the system for subsequent diagnosis tasks will theoretically be closer to the clinical reality. The model optimization feedback module works silently, using every real clinical interaction as an opportunity to improve its capabilities, and pushing the entire system forward along the track of autonomous optimization.

[0087] It should be noted that the relational terms such as first and second, and the like, are used solely to distinguish one from another entity or action, without necessarily requiring or implying any actual such relationship or order between or among such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0088] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1.A disease remote diagnosis system based on artificial intelligence driving, characterized in that, The system comprises: a real-time medical data acquisition module for acquiring multi-modal physiological data of a patient in real time from a remote terminal device, the multi-modal physiological data including vital sign time series data, medical image data, and symptom description text data; a cross-modal correlation analysis module for performing cross-modal feature fusion on the multi-modal physiological data, extracting correlation patterns between different data modalities, and generating an abnormal feature dataset with weight labels; a dynamic diagnosis baseline model for establishing a dynamic diagnosis baseline based on the multi-modal physiological data and the abnormal feature dataset, and adapting to physiological feature changes of individual patients by updating model parameters in real time; an abnormal behavior detection module for comparing real-time physiological data with the dynamic diagnosis baseline model, identifying abnormal physiological indicators deviating from normal ranges, and generating abnormal behavior labeling results; a deep cause tracing module for analyzing potential disease patterns and tracing cause paths based on the abnormal behavior labeling results and in combination with patient historical medical record data. 2.The AI-driven disease remote diagnosis system according to claim 1, wherein, The real-time medical data acquisition module comprises: capturing patient vital sign time series data from a remote terminal device and performing standardization processing on the data; analyzing medical image data, extracting image features, and generating structured image description data; integrating symptom description text data, performing semantic analysis, and generating symptom feature vectors; fusing standardized vital sign data, structured image description data, and symptom feature vectors into a unified format of multi-modal physiological data stream. 3.The AI-driven disease remote diagnosis system according to claim 2, characterized in that, The cross-modal correlation analysis module comprises: performing hierarchical feature extraction on the multi-modal physiological data stream to identify potential correlations between different modal data; analyzing abnormal interaction patterns based on context information, generating preliminary abnormal signals, and recording abnormal types and severity; allocating weights according to features of abnormal signals, and outputting an abnormal feature dataset with weights. 4.The AI-driven disease remote diagnosis system according to claim 3, characterized in that, The dynamic diagnosis baseline model comprises: training an initial diagnosis model based on the multi-modal physiological data stream and the abnormal feature dataset; dynamically adjusting model parameters through real-time data feedback to optimize the accuracy of the diagnosis baseline; generating a personalized diagnosis baseline in combination with individual differences of patients. 5.The AI-driven disease remote diagnosis system according to claim 4, characterized in that, The abnormal behavior detection module comprises: inputting real-time physiological data into the dynamic diagnosis baseline model to calculate deviation values from normal behavior patterns; labeling deviation values exceeding a threshold to generate abnormal behavior labeling results; recording context information of abnormal behaviors, including timestamps, abnormal types, and deviation amplitudes. 6.The AI-driven disease remote diagnosis system according to claim 5, wherein, The deep cause tracing module comprises: searching patient historical medical record data to extract diagnosis records related to the abnormal behavior labeling results; analyzing potential disease patterns to identify possible cause paths; generating a comprehensive diagnosis report containing cause analysis results. 7.The AI-driven disease remote diagnosis system according to claim 6, characterized in that, The system further comprises: an environmental factor compensation module for adjusting the diagnosis baseline based on patient environment data, including temperature, humidity, and air quality indicators; combining environmental data to correct abnormal physiological indicators to improve diagnosis accuracy. 8.The AI-driven disease remote diagnosis system according to claim 7, wherein, The environmental factor compensation module comprises: acquiring real-time environmental data of a patient to calculate influence coefficients of environmental factors on physiological indicators; adjusting parameters of the dynamic diagnosis baseline model based on the influence coefficients; A corrected diagnosis result after environmental compensation is generated. 9.The AI-driven disease remote diagnosis system according to claim 8, wherein, The system further comprises: a model optimization feedback module for adjusting model parameters according to the deviation between actual diagnosis results and predicted results; The prediction ability of the dynamic diagnosis baseline model is optimized through a continuous learning mechanism. 10.The AI-driven disease remote diagnosis system according to claim 9, characterized in that, The model optimization feedback module comprises: collecting actual diagnosis data, calculating prediction errors and generating error correction coefficients; updating the weight parameters of the dynamic diagnosis baseline model based on the error correction coefficients; outputting the optimized diagnosis model for subsequent diagnosis tasks.

Images