Artificial intelligence-driven medical diagnosis and treatment data processing method and system

Abstract

The invention relates to the technical field of medical data processing systems, in particular to an artificial intelligence-driven medical diagnosis and treatment data processing method and system. The method comprises the steps that a multi-modal medical data acquisition module acquires and processes multi-source heterogeneous medical data of a patient, and a standardized data set is generated; a medical feature depth extraction module performs multi-dimensional feature extraction on the data set, and constructs a dynamic evolution feature matrix; a multi-dimensional health state space construction module constructs a patient health state multi-dimensional space according to the matrix and determines a key medical early warning index set; the real-time medical data fusion module maps real-time data to the space to generate real-time health risk factors; and the medical risk prediction and decision-making module establishes a personalized model, outputs a disease occurrence probability and generates personalized treatment suggestions. The system solves the problems that medical data processing is difficult, diagnosis analysis is not comprehensive, and a treatment scheme lacks personality, and diagnosis accuracy and treatment pertinence are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical data processing systems, in particular to a data processing method and system for medical diagnosis and treatment driven by artificial intelligence. BACKGROUND

[0002] In today's medical field, medical data is showing an explosive growth trend. With the popularization of various medical devices and the in-depth application of information technology in the medical industry, from patient medical records, test reports, to real-time physiological data collected by wearable devices, a large amount of medical data resources has been formed. However, these data have the characteristics of multi-source heterogeneity, and the data format and standards generated by different medical institutions and different devices are different, which brings great difficulties to the effective use of medical data.

[0003] Traditional medical diagnosis and treatment mainly rely on the experience and professional knowledge of doctors. When facing complex conditions, doctors need to analyze the patient's information from multiple aspects, including symptom description, medical history, and various test results. However, due to the complexity and dispersion of data, doctors are difficult to fully and accurately grasp the patient's health status, and are prone to misdiagnosis or missed diagnosis. For example, when diagnosing some rare diseases or complex diseases, it may be necessary to refer to a large number of medical literature and past cases, but doctors are difficult to do comprehensive retrieval and analysis within a limited time, resulting in low diagnosis efficiency.

[0004] In the development of treatment plans, the commonly used treatment strategies are often adopted, without fully considering the individual differences of each patient. Due to the differences in genes, living habits, and physical conditions, different patients may have different reactions to the same treatment method. For example, some drugs may have significant effects on some patients, but may produce serious adverse reactions in other patients. Moreover, with the continuous progress of medical technology, new treatment methods and drugs emerge in an endless stream, and doctors are difficult to quickly and accurately screen out the most suitable treatment plan for each patient.

[0005] In the process of processing medical data, the quality of data cannot be ignored. There are often noise, missing values, and incorrect labels in the original medical data, which seriously affects the analysis and utilization value of the data. Moreover, due to the lack of effective data cleaning and standardization means, data from different sources is difficult to integrate and compare, which limits the application of medical data in clinical decision-making, medical research, and other aspects. Under such background, it is of urgent practical need to develop a data processing system that can efficiently process multi-source heterogeneous medical data and realize accurate diagnosis and personalized treatment. SUMMARY

[0006] The main purpose of the present application is to provide an artificial intelligence driven medical diagnosis and treatment data processing method and system, aiming to solve the technical problems in the prior art.

[0007] The present application proposes an artificial intelligence driven medical diagnosis and treatment data processing system, comprising:

[0008] A multi-modal medical data acquisition module is configured to acquire multi-source heterogeneous medical data of a patient within a preset time period, and perform data cleaning and standardization processing on the multi-source heterogeneous medical data to generate a standardized multi-modal medical data set;

[0009] A medical feature deep extraction module is configured to perform multi-dimensional feature extraction on the standardized multi-modal medical data set to obtain an initial medical feature set, and construct a dynamic evolution feature matrix according to the variation law of each feature in the initial medical feature set over time;

[0010] A multi-dimensional health state space construction module is configured to construct a multi-dimensional space representation of the patient's health state according to the dynamic evolution feature matrix, calculate the distribution density of historical medical events in the multi-dimensional space, and determine a set of key medical warning indicators according to the distribution density;

[0011] A real-time medical data fusion module is configured to receive real-time acquired patient medical data, extract real-time medical features and form a real-time feature vector, map the real-time feature vector to the multi-dimensional health state space, and generate a real-time health risk factor;

[0012] A medical risk prediction and decision module is configured to establish a personalized health risk prediction model according to the real-time health risk factor, the initial medical feature set and the dynamic evolution feature matrix, output a disease occurrence probability and generate an individualized treatment recommendation scheme.

[0013] Preferably, the multi-modal medical data acquisition module is specifically configured to:

[0014] Acquire multi-source heterogeneous medical data including imaging data, physiological parameter time series data, laboratory test data and clinical text data;

[0015] Perform outlier rejection, noise filtering and missing value interpolation processing on the multi-source heterogeneous medical data to complete data quality enhancement;

[0016] Align and normalize the quality-enhanced multi-source heterogeneous medical data according to data modalities to generate a standardized multi-modal medical data set.

[0017] Preferably, the medical feature deep extraction module is specifically configured to:

[0018] extracting multi-scale medical features from the standardized multi-modal medical data set by using a deep convolutional neural network;

[0019] fusing and splicing medical features of different scales to form an initial medical feature set;

[0020] calculating the change rate and trend of each feature in the initial medical feature set within a specified time window;

[0021] organizing the change rate and trend of each feature in time sequence to construct a dynamic evolution feature matrix.

[0022] As a preferred, the multi-dimensional health state space construction module is specifically used for:

[0023] determining the number of dimensions of the multi-dimensional health state space according to the number of feature dimensions of the dynamic evolution feature matrix;

[0024] calculating the distribution density of the corresponding feature in the multi-dimensional health state space when the historical diagnosis event occurs;

[0025] setting a density threshold according to the distribution density, and screening features with distribution density exceeding the density threshold as a key medical early warning indicator set.

[0026] As a preferred, the real-time medical data fusion module is specifically used for:

[0027] extracting features from the real-time collected patient medical data to generate real-time medical features;

[0028] constructing the real-time medical features into a real-time feature vector according to the organization format of the dynamic evolution feature matrix;

[0029] projecting the real-time feature vector to the multi-dimensional health state space and calculating its spatial distance with the key medical early warning indicator set;

[0030] generating a real-time health risk factor according to the spatial distance.

[0031] As a preferred, the medical risk prediction and decision module establishes a personalized health risk prediction model, which is specifically:

[0032] fusing the real-time health risk factor with the initial medical feature set;

[0033] using a time sequence attention mechanism to weight and reconstruct the dynamic evolution feature matrix;

[0034] training a risk prediction network based on the fused features and the reconstructed feature matrix;

[0035] outputting a probability value of patient disease occurrence and generating an individualized treatment recommendation scheme.

[0036] As preferred, the system further comprises:

[0037] a medical knowledge graph construction module for integrating a medical knowledge base and a clinical guideline to construct a medical knowledge graph containing relationships of diseases, symptoms, drugs and treatment plans;

[0038] The medical risk prediction and decision module is further configured to match and verify the individualized treatment recommendation plan with the medical knowledge graph.

[0039] As preferred, the system further comprises:

[0040] a multi-center medical data collaboration module for aggregating standardized multi-modal medical data sets of multiple medical institutions, performing federated learning training, and updating the personalized health risk prediction model;

[0041] The medical risk prediction and decision module is further configured to perform prediction using the updated personalized health risk prediction model.

[0042] As preferred, the system further comprises:

[0043] a diagnosis and treatment process tracing module for recording the implementation process and effect feedback of the individualized treatment recommendation plan to form a closed-loop medical data;

[0044] The medical risk prediction and decision module is further configured to continuously optimize the personalized health risk prediction model according to the closed-loop medical data.

[0045] As preferred, the present application further comprises a data processing method for artificial intelligence driven medical diagnosis and treatment, which comprises all the modules and method processes of the above-mentioned artificial intelligence driven medical diagnosis and treatment data processing system.

[0046] The present application has the following advantages:

[0047] The system can obtain multi-source heterogeneous medical data of a patient within a preset time period through the multi-modal medical data acquisition module, and perform data cleaning and standardization processing to generate a standardized multi-modal medical data set. This solves the problem that traditional medical data is difficult to integrate and utilize due to inconsistent formats and standards. For example, in the past, the formats of medical records of different hospitals were different, and the units and reference ranges of test indicators were also inconsistent, making it extremely difficult for doctors to analyze patient cross-institutional medical data. However, the present system can unify and standardize these complex and diverse data, providing a reliable basis for subsequent analysis, and enabling medical data to be truly seamlessly connected and efficiently utilized.

[0048] The medical feature deep extraction module performs multi-dimensional feature extraction on the standardized multi-modal medical data set to obtain an initial medical feature set, and constructs a dynamic evolution feature matrix according to the change law of the features over time. Compared with traditional single-dimensional or static feature analysis, it can more comprehensively and dynamically reflect the health status of the patient. For example, in monitoring the development process of a patient's chronic disease, the traditional method may only focus on fixed indicators at a few time points, while the system can clearly present the continuous change trend of each indicator over time through the dynamic evolution feature matrix, so that doctors can more accurately judge the development stage and change law of the disease and discover potential health risks in advance.

[0049] The multi-dimensional health status space construction module constructs a multi-dimensional space representation of the patient's health status based on the dynamic evolution feature matrix, calculates the distribution density of historical medical events in the multi-dimensional space, and determines a set of key medical warning indicators. This innovative multi-dimensional space construction method breaks through the limitations of traditional two-dimensional or simple dimensional analysis. Taking the diagnosis of cardiovascular disease as an example, the traditional method may only consider a few indicators such as blood pressure and heart rate, while the system comprehensively considers many factors such as the patient's age, family medical history, lifestyle, multiple blood indicators, and the dynamic changes of these indicators in the multi-dimensional space, greatly improving the accuracy and comprehensiveness of diagnosis, and enabling more accurate identification of key medical warning indicators, providing strong support for early warning of diseases.

[0050] The real-time medical data fusion module receives real-time collected patient medical data, extracts real-time medical features and forms real-time feature vectors, and maps them to the multi-dimensional health status space to generate real-time health risk factors. This enables doctors to timely grasp the patient's current health risk situation. For example, during a patient's hospitalization, various monitoring devices collect data in real time, and the system can quickly convert these data into real-time health risk factors, allowing doctors to adjust treatment plans in a timely manner to respond to sudden changes in health status and prevent disease deterioration.

[0051] The medical risk prediction and decision-making module establishes a personalized health risk prediction model based on the real-time health risk factors, the initial medical feature set, and the dynamic evolution feature matrix, outputs the disease occurrence probability, and generates an individualized treatment recommendation plan. Compared with traditional general treatment plans, the system fully considers the unique situation of each patient. For example, in cancer treatment, different patients have different genetic characteristics and physical tolerance levels, and traditional treatment plans often "one size fits all", resulting in poor effectiveness. However, the system can accurately calculate the disease occurrence probability for each cancer patient through the personalized health risk prediction model and develop a treatment plan that is most suitable for the individual, greatly improving treatment effectiveness, reducing unnecessary waste of medical resources, and reducing the pain and side effects experienced by patients due to unsuitable treatment plans. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 A timing diagram of the artificial intelligence driven medical diagnosis and treatment data processing system of the present application;

[0053] Figure 2 A flowchart for feature extraction and evolution matrix construction.

[0054] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0055] It should be understood that the specific embodiments described herein are merely illustrative of the present application and do not limit the present application.

[0056] As shown in Figure 1 The present application provides an artificial intelligence driven medical diagnosis and treatment data processing system, which includes: a multi-modal medical data acquisition module that acquires multi-source heterogeneous medical data of a patient within a preset time period, and the module performs data cleaning and standardization processing to generate a standardized multi-modal medical data set. A medical feature deep extraction module then performs multi-dimensional feature extraction on the data set to produce an initial medical feature set, and constructs a dynamic evolution feature matrix based on the change law of each feature over time. A multi-dimensional health state space construction module uses the matrix to establish a multi-dimensional space representation of the patient's health state, and determines a key medical warning indicator set by calculating the distribution density of historical medical events in the space. A real-time medical data fusion module receives real-time acquired patient medical data, extracts real-time medical features to form a real-time feature vector, and maps the vector to the multi-dimensional health state space to generate a real-time health risk factor. A medical risk prediction and decision module combines the real-time health risk factor, the initial medical feature set and the dynamic evolution feature matrix to establish a personalized health risk prediction model, outputs the disease occurrence probability and generates an individualized treatment recommendation scheme. The entire system realizes efficient data processing and intelligent decision support through modular design.

[0057] Example 1: see Figure 2, the multimodal medical data acquisition module performs comprehensive acquisition and preprocessing of patient data. The module connects CT / MRI devices, electrocardiogram monitors, laboratory information systems, and electronic medical record systems through digital interfaces of medical institutions to collect multi-source heterogeneous medical data covering imaging data, physiological parameter time series data, laboratory test data, and clinical text data. Imaging data includes DICOM format tomographic images and three-dimensional reconstruction data, physiological parameter time series data involves continuous monitoring of heart rate, blood pressure, blood oxygen saturation, and other vital sign waveforms, laboratory test data covers blood biochemical indicators and microbial culture results, and clinical text data includes physician-written medical history records and nursing documents. The data acquisition sets a time period of 12 months of patient history records to ensure clinical relevance of the time span. After data acquisition, a data cleaning process is immediately started: outlier rejection uses a quantile-based outlier detection algorithm to automatically mark and remove data points that exceed the reasonable range of medical data; noise filtering applies a Butterworth low-pass filter to physiological signals to eliminate power frequency interference and motion artifacts; missing value imputation uses a K-nearest neighbor algorithm to find similar patterns in similar patient data for intelligent filling. After completing data quality enhancement, the standardized stage is entered: time alignment operation takes the R-wave peak of the electrocardiogram signal as the reference point to synchronize the timestamps of other modal data; normalization processing performs window width and window level adjustment on imaging data and log conversion on laboratory data to eliminate dimension differences, and finally generates a standardized multimodal medical data set with unified dimensions stored in a medical database.

[0058] The medical feature deep extraction module performs deep analysis on the standardized dataset. The module is configured with a multi-stream convolutional neural network architecture. The image data processing branch uses a 3DResNet-50 network to extract ground glass shadow features and lesion volume changes in lung CT images. The physiological signal branch uses a 1D CNN combined with an LSTM layer to capture abnormal RR interval patterns in heart rate variability. The test data branch identifies key indicators such as the albumin / globulin ratio through a fully connected network. The text branch uses a BERT model to extract entity features such as "dyspnea" and "fever" from clinical records. The feature extraction process produces four scale feature maps: 128x128 pixel macro-organizational features, 64x64 pixel vascular distribution features, 32x32 pixel cellular-level features, and 16x16 pixel microstructure features. Multi-scale feature fusion uses a channel attention mechanism to calculate channel weights using an SE module and then performs weighted concatenation to form an initial medical feature set containing image texture features, physiological fluctuation features, test value features, and text semantic features. The feature set is further analyzed for dynamic evolution: for time-series features such as blood glucose, the 24-hour sliding window rate of change (first derivative) and trend (second derivative) are calculated; for tumor volume features, a 72-hour window is used to analyze the growth rate. The dynamic properties of all features are organized into a dynamic evolution feature matrix with minute-level timestamps. The matrix row vectors represent hourly feature snapshots, and the column vectors store three-tuples of feature values, rates of change, and trend directions. A 1200x150-dimensional spatiotemporal feature matrix is finally constructed to comprehensively characterize the disease evolution trajectory.

[0059] The construction process of the feature matrix includes strict medical logic verification. The dynamic analysis of image features is cross-checked with radiologists' lesion progression reports, and the trend calculation of physiological features refers to the fluctuation threshold standards published by the American Heart Association. The laboratory feature rate of change is calibrated using the total allowable error range of the International Clinical Laboratory Standards Committee (CLSI) to ensure that the feature extraction results meet the requirements of evidence-based medicine. The dynamic evolution feature matrix is transmitted to the hospital's private cloud platform through a medical encryption protocol, and is updated and retrieved in real time in the memory computing framework, providing high-precision spatiotemporal feature support for subsequent health state modeling. Strict adherence to the HIPAA medical privacy protection specification, patient identity information has been anonymized before feature extraction. The data processing server is deployed in the hospital's local computer room, and physical isolation measures are used with external networks. All data transmissions use AES-256 encryption algorithm to ensure information security. The training data for the feature extraction model comes from a de-identified medical dataset approved by the ethics committee, and the model weights are iteratively updated every quarter based on the latest clinical research results to ensure the scientificity and timeliness of the medical features.

[0060] Example 2: The multi-dimensional health state space construction module processes the dynamic evolution feature matrix, which contains 150 feature dimensions corresponding to the definition of the spatial coordinate axes. The space construction adopts the Euclidean geometry model, and each coordinate axis corresponds to the evolution trajectory of a specific medical feature, for example, the X-axis maps the blood glucose change rate trend, the Y-axis represents the tumor volume growth rate, and the Z-axis is associated with the inflammation factor fluctuation amplitude. The mapping process of historical medical events extracts timestamps from the diagnostic record database of the hospital information system, including the occurrence time of critical events such as myocardial infarction and diabetic ketoacidosis, and locates the feature vector at the corresponding time point in the dynamic evolution feature matrix as a spatial coordinate point. The distribution density calculation deploys the Gaussian kernel density estimation algorithm to generate a probability density distribution field centered on each historical event coordinate point, and the bandwidth parameter is adaptively adjusted according to the standard deviation of the feature value. The density calculation covers the entire 150-dimensional space to form a probability cloud map.

[0061] The spatial density analysis adopts a hierarchical scanning strategy, dividing the 150-dimensional space into 15 clinical subsystems such as cardiovascular, metabolic, and tumor according to medical significance, and calculating the density distribution of each subsystem independently. The density threshold setting uses the three-part statistical method, taking the 75th percentile of the density value of each subsystem as the critical line, and triggering an early warning marker when the density of a specific region exceeds the critical value of the same subsystem. The selection of key medical early warning indicators combines with the clinical expert rule base, for example, the feature combination of "ST segment elevation density > threshold" and "troponin rising slope density > threshold" in the cardiovascular subsystem is included in the early warning set. The early warning indicator set adopts a tree structure for storage, with the root node being the disease classification (such as respiratory system disease), and the leaf node storing specific feature threshold conditions (such as blood oxygen saturation change rate > 5% / hour). The space construction process introduces a real-time verification mechanism, which automatically triggers space reconstruction when new historical events continue to occur: the new event coordinates are added to the point set to recalculate the density distribution, and if it causes the density field form of a certain subsystem to change more than a preset tolerance (such as density peak shift > 10%), the subsystem boundary is re-divided and the threshold is updated. Dynamic maintenance of the early warning indicator set is achieved through version control, generating a new indicator set version after each space reconstruction, and the old version is retained in the medical archive for efficacy backtracking analysis. All spatial coordinate data is stored using double-precision floating-point numbers, and the density field is compressed and archived in the form of voxel grid, supporting spatiotemporal range retrieval through medical special query language.

[0062] The clinical validation of the multi-dimensional health state space interfaces with the hospital clinical decision support system, automatically triggering the order review process when the space model identifies a high-risk area. For example, the tumor subsystem detects that a patient's feature vector is located in the high-density area of bone metastasis, and the system will push the CT bone window review suggestion to the attending physician's workstation. The space model is evaluated by the clinical audit committee every quarter, and the density algorithm parameters are adjusted by comparing the consistency of the warning indicator set and the actual incidence record to ensure that the model prediction meets the requirements of evidence-based medicine. The space data visualization component generates a three-dimensional heat map for multidisciplinary consultation, and physicians can observe the relative position of the patient's feature point in the historical event density field through virtual reality equipment. The allocation of computing resources for this module uses a medical priority scheduling strategy, automatically preempting GPU cluster resources for real-time space mapping when emergency patient data is input. Historical data processing uses night idle computing power for batch space reconstruction, and all operation processes are recorded in medical blockchain nodes to achieve operation traceability. The space model update package is distributed to each clinical terminal through the hospital intranet security channel, maintaining the continuity of the warning service during updates to ensure that real-time monitoring in the intensive care unit is not affected by system upgrades.

[0063] A 58-year-old male patient with a history of hypertension for 10 years and diabetes for 5 years was admitted to the Department of Cardiology in a certain third-grade Class-A hospital due to persistent chest pain for 2 hours. The medical feature depth extraction module generated a dynamic evolution feature matrix from the patient's multi-modal data within 72 hours of admission: containing 125-dimensional feature time series, data collected every 30 minutes, forming a feature matrix of 150 time points. The matrix includes key indicators such as ST segment elevation amplitude of electrocardiogram (feature dimensions 1-5), troponin I rise slope (dimensions 6-10), heart rate variability SDNN value (dimensions 11-15), blood pressure fluctuation coefficient (dimensions 16-20), etc.

[0064] The multi-dimensional health state space construction module first parses the dimension structure of the feature matrix to determine the construction of a 125-dimensional health state space. Each dimension corresponds to a time evolution trajectory of a medical feature, for example, dimension 1 is dedicated to mapping the ST segment elevation amplitude of the anterior wall lead, and dimension 6 records the relative hourly change rate of troponin I. The system extracts data from the hospital cardiology database for patients diagnosed with acute myocardial infarction in the past three years, and maps the feature values corresponding to 217 confirmed events into the 125-dimensional space: each event takes the 125 feature values at the corresponding time point in the feature matrix as the spatial coordinate point according to its occurrence time point. The space density calculation uses adaptive bandwidth Gaussian kernel density estimation, and the density distribution is calculated separately for the anterior wall infarction sub-group (89 cases) in the myocardial infarction feature set. The system detects that in the feature subspace composed of dimension 1 (ST segment elevation amplitude) and dimension 6 (troponin I rising slope), there is a clear high-density area: when the ST segment elevation is ≥2mm and the troponin I rises by ≥0.2ng / mL per hour, the case distribution density reaches a peak. The density threshold setting uses the percentile method, taking the 80th percentile of the density value of this subspace as the critical threshold, and the feature combinations exceeding this threshold are included in the key medical early warning indicator set.

[0065] A clinical expert validation mechanism was introduced in the early warning indicator set generation process: the director of the cardiology department reviewed the feature combinations screened by the system and confirmed that the combination of ST segment elevation, troponin I rising slope, and ventricular arrhythmia frequency had the highest clinical relevance. The final myocardial infarction early warning rule generated by the system contains multiple layers of judgment logic: the primary condition is ST segment elevation ≥ 2mm combined with troponin I rising rate ≥ 0.2ng / mL / h, and the secondary condition includes heart rate variability drop of more than 30% or frequent ventricular premature beats. After the early warning system was put into clinical use, when the real-time data of new inpatients was input, the system mapped its feature vector to a 125-dimensional health state space. In a certain monitoring, a 63-year-old female patient's feature coordinates were found to be located in a high-density area: ST segment elevation 2.3mm (dimension 1 value), troponin I rising rate 0.25ng / mL / h (dimension 6 value), and heart rate variability decreased by 35% compared with the baseline (dimension 11 value). The system immediately triggered a red alert and sent an emergency consultation request to the cardiovascular intensive care unit. After the doctor received the case, he confirmed that the patient had acute anterior myocardial infarction and immediately initiated the catheter room for emergency intervention treatment. The space model continues to be dynamically updated: when the hospital admits new myocardial infarction variant cases (such as inferior wall combined with right ventricular infarction), the system automatically adds the feature coordinates of the new cases to the space point set. The space is reconstructed once a month, the density distribution is recalculated, and the early warning threshold is adjusted. After a certain reconstruction, it was found that the diabetic myocardial infarction patient group showed a unique feature distribution: the ST segment elevation amplitude was often not obvious but the troponin rising rate was faster, and the system generated a specific early warning threshold for diabetic patients accordingly, improving the diagnostic sensitivity of this population. All space operation data storage uses a medical special encryption format, and access permission is limited to the cardiovascular specialist team. The space visualization interface supports three-dimensional projection viewing, and doctors can choose any three feature dimensions to observe the case distribution pattern. The early warning history record is automatically compared with the final diagnosis result to generate an early warning accuracy report for the medical quality improvement committee to analyze.

[0066] Example 3: Real-time medical data fusion module continuously receives streaming data from bedside monitoring devices, which is configured to process hundreds of physiological parameter readings per second generated by devices such as electrocardiogram monitors, ventilators, and intravenous infusion pumps. A real-time feature extraction engine is deployed on a medical edge computing node, which uses a lightweight convolutional neural network to analyze ST segment shift features in electrocardiogram waveforms, a recurrent neural network to calculate the instantaneous variation coefficient of respiratory rate, and a sliding window statistical method to extract fluctuation entropy values of blood pressure sequences. The generation of real-time medical features follows millisecond-level timing constraints, and each feature vector is attached with a timestamp accurate to milliseconds and a device source identifier. The feature vector construction process strictly aligns with the dimension specification of the dynamically evolving feature matrix, and the 150-dimensional feature slots are filled according to the preset mapping table: dimensions 1-30 are assigned to electrocardiogram features (such as QRS complex area), dimensions 31-60 store respiratory parameters (such as inspiration time ratio), dimensions 61-90 record hemodynamic indicators (such as arterial compliance), and the remaining dimensions are reserved for laboratory emergency results (such as blood gas analysis values).

[0067] The projection of real-time feature vectors into a multi-dimensional health state space uses an orthogonal decomposition algorithm to project 150-dimensional vectors into a feature subspace composed of principal components of historical data. The spatial distance calculation uses an improved Mahalanobis distance metric to avoid computational bias caused by different feature dimensions:

[0068] ;

[0069] where: represents the real-time risk distance scalar, represents the real-time feature vector at time t, is the mean vector of the set of key medical warning indicators in the corresponding dimension, is the inverse matrix of the covariance matrix of the set of warning indicators. This distance value is converted to a health risk factor in the range of 0-1 through a sigmoid function, and the risk factor is updated every second and written to a medical real-time database. The clinical interpretation of the risk factor is classified as follows: the 0-0.3 interval is marked as a green low-risk state, the 0.3-0.7 interval triggers a yellow moderate warning, and a red high-risk state above 0.7 immediately starts an alarm process.

[0070] The model construction of the medical risk prediction and decision module adopts a multi-modal fusion architecture. The fusion of real-time health risk factors and the initial medical feature set is achieved through a gating attention mechanism. Risk factors act as gating signals to weight and filter relevant patterns in historical features. The time series attention mechanism reconstructs the dynamic evolution feature matrix. The cosine similarity between each historical time point feature vector and the current real-time state is calculated as the attention weight. The weighted sum generates a context-aware feature representation. The risk prediction network adopts a deep residual structure. The input layer receives the fused feature tensor. The hidden layer contains three 512-unit LSTM layers to capture temporal dependencies. The output layer generates a disease probability distribution through the softmax function. The generation of individualized treatment recommendations combines a clinical guideline rule engine. For coronary heart disease risk prediction, when the probability value exceeds 0.75, the ACC / AHA drug treatment rule tree is automatically invoked. For diabetes risk prediction, the ADA guidelines are used to generate insulin dose adjustment recommendations. All treatment recommendations are pushed to the physician workstation through the HL7 protocol for confirmation and execution. The model training process uses a leave-time cross-validation strategy. The first 80% of the patient's time data is used to train the network, and the last 20% of the data is used to verify the prediction accuracy. The adaptive moment estimation algorithm is used for network parameter optimization. The learning rate is dynamically adjusted according to the validation set loss. Containerization technology is used for model deployment to ensure consistency of the computing environment. The prediction service response time is controlled within 200 milliseconds to meet the clinical real-time requirements. The treatment recommendation generation module has a built-in safety checker that filters suggestions with drug risks by comparing with the drug contraindication database, ensuring that the output scheme meets the individual contraindications of the patient.

[0071] A 67-year-old male patient with sepsis was admitted to the intensive care unit of a certain general hospital. The patient developed symptoms of fever, tachycardia, and hypotension after abdominal infection surgery. The real-time medical data fusion module continuously receives vital sign data from the bedside monitor: heart rate, blood pressure, and blood oxygen saturation parameters are collected every 2 seconds, mean arterial pressure readings are obtained from the arterial catheter every 5 minutes, and body temperature trends are recorded every hour. The feature extraction engine synchronously processes multi-source data streams: the low-to-high frequency power ratio of heart rate variability is extracted from the electrocardiogram signal, the vascular tone index is calculated from the blood pressure waveform, and the fever rate indicator is derived from the body temperature curve.

[0072] The real-time feature vector construction strictly follows the dimensional specifications trained by historical data. The LF / HF ratio of heart rate variability is filled into the 23rd dimension of the vector, the vascular tone index is filled into the 47th dimension, the fever rate is filled into the 89th dimension, and the latest procalcitonin test value is filled into the 112th dimension. Vector projection operation maps the 125-dimensional features at the current time to the sepsis warning space, which is constructed from the data of sepsis patients admitted in the past three years. The space distance calculation identifies that the current vector is only 0.31 Mahalanobis distance units away from the center point of the sepsis deterioration cluster, which is significantly lower than the warning threshold of 1.5. The system immediately generates a health risk factor of 0.83, triggering a red high-risk alert. The medical risk prediction and decision-making module initiates multi-dimensional analysis: the real-time risk factor is fused with the initial feature set at the patient's admission: the initial features include basic immune function indicators, chronic disease history codes, and surgical trauma grade scores. The time series attention mechanism focuses on the feature change trajectory of the patient in the last 6 hours, especially focusing on the downward trend of mean arterial pressure from 72 mmHg to 65 mmHg and the upward process of lactate value from 2.1 mmol / L to 3.8 mmol / L. The risk prediction network outputs a sepsis shock occurrence probability of 0.79, exceeding the intervention threshold of 0.7. The individualized treatment recommendation module refers to the sepsis rescue guideline: first, it recommends a crystalloid fluid resuscitation scheme of 30 ml / kg, and suggests intravenous injection of norepinephrine to maintain mean arterial pressure ≥65 mmHg, and recommends the use of a broad-spectrum antibiotic regimen of meropenem combined with vancomycin. The system automatically checks the drug contraindication library: confirms that the patient has no history of vancomycin allergy and the renal function indicators are within the allowed range. The treatment plan is immediately implemented after being confirmed by the attending physician, and the nurse accurately controls the infusion rate of vasoactive drugs through the intelligent infusion pump. The real-time monitoring system continuously tracks the treatment response: central venous pressure is measured every 15 minutes after fluid resuscitation, and capillary refill time is continuously monitored after the use of vasoactive drugs. When the patient's lactate level is detected to decrease from 3.8 mmol / L to 2.4 mmol / L within 4 hours and the heart rate decreases from 125 beats / min to 98 beats / min, the system automatically adjusts the risk factor to 0.45. The warning level is downgraded to yellow moderate risk, but the continuous recommendation of antibiotic treatment is still maintained. All treatment process data is recorded and transmitted to the hospital data center in real time, forming a complete digital trajectory of sepsis treatment.

[0073] Example 4: The medical knowledge graph construction module integrates multi-source medical knowledge bases. This module extracts disease diagnosis and treatment standards from the UpToDate clinical decision system, obtains drug interaction information from the FDA drug instruction database, and constructs a disease-symptom association network from the NCBI medical subject headings (MeSH). Knowledge extraction uses a BERT-based entity recognition model to analyze medical literature and identify clinical entities such as "acute myocardial infarction," "aspirin," and "percutaneous coronary intervention." Relationship extraction uses a graph neural network to mine semantic relationships between entities and establish association rules such as "disease-treatment plan," "drug-contraindication," and "symptom-examination method." The graph storage uses a Neo4j graph database, and the node attributes include disease ICD-10 codes, drug ATC classification codes, and treatment plan evidence levels.

[0074] The matching verification of individualized treatment recommendation schemes and knowledge graphs performs a two-way checking process: forward checking to check if the drug treatment in the scheme matches the disease guidelines. When the system generates the "acute ST segment elevation myocardial infarction patient uses ticagrelor" recommendation, it automatically queries the graph for the treatment association between the disease and P2Y12 inhibitors; reverse checking to check for contraindication conflicts. If the patient has a history of active bleeding, the system traverses the association path between "ticagrelor" and "bleeding risk" in the graph and triggers a warning. The matching verification result generates a structured report, marking the compliant schemes that pass the verification and the risk schemes that have conflicts. The multi-center medical data collaboration module connects three hospital data centers and aggregates standardized multi-modal medical data sets. Data aggregation follows the FHIR standard conversion format, and patient identity information is transmitted anonymously through hash encryption. The federated learning architecture uses a horizontal partitioning mode, and each hospital locally trains the neural network weights of the risk prediction model, exchanging gradient updates through secure multi-party computation protocols. The model aggregation server performs a weighted average algorithm: according to the data volume of each center (center A: 45%, center B: 30%, center C: 25%), the weight coefficients are allocated, and the global model parameters are updated monthly. The updated individualized health risk prediction model is protected by differential privacy technology, and after being reviewed by the medical ethics committee, it is deployed to the reasoning server of each hospital. Key indicators are monitored during the multi-center data collaboration process, as shown in Table 1.

[0075] Table 1: Multi-center federated learning operation index monitoring table

[0076] The medical risk prediction and decision-making module performs version compatibility checks when loading the updated model, ensuring that the input and output interfaces seamlessly integrate with the existing hospital systems. The prediction service uses an A / B testing strategy to gradually switch to the new model, first comparing the prediction results for 10% of the patient flow, and then fully enabling it after the clinical doctors confirm that there are no errors. The model inference process records complete audit logs, including input feature values, prediction probability calculation paths, treatment plan generation logic, and other key steps, for regular review by the medical quality management department. The continuous updating mechanism of the knowledge graph listens to the latest research findings in medical journals, and when a new clinical trial of an anticoagulant drug is published in The New England Journal of Medicine, the system automatically extracts the key conclusions and updates the treatment rules in the graph. The multi-center data collaboration establishes an abnormal data detection mechanism, and when the data distribution uploaded by a hospital deviates from the overall benchmark by more than 3 standard deviations, it automatically suspends the node from participating in federated learning and triggers the data quality review process. The transmission and storage of all medical data comply with HIPAA security specifications, and the encryption keys are centrally managed by the hospital information security department, ensuring the safe and reliable operation of the medical intelligent system.

[0077] Example 5: The diagnosis and treatment process tracing module automatically collects treatment implementation data through medical Internet of Things devices. This module connects with intelligent infusion pumps to obtain drug infusion rate and duration parameters, extracts physician order execution records from the electronic medical record system, and collects physiological parameter responses during treatment through bedside monitors. Taking the example of anticoagulant therapy for patients with coronary heart disease after percutaneous coronary intervention (PCI) surgery: the system records the administration time and dose of aspirin 100mg once daily and ticagrelor 90mg twice daily, while monitoring laboratory indicators such as platelet aggregation rate and bleeding time. Effect feedback data includes daily ECG ST segment fall amplitude, myocardial enzyme spectrum dynamic change curve, and patient complaint of chest pain relief within 30 days after surgery. All data are stored in the medical data warehouse in chronological order to form a structured diagnosis and treatment trajectory.

[0078] The integration of closed-loop medical data adopts a time-series database architecture, with each record containing four-dimensional attributes: a timestamp accurate to the millisecond, a treatment item code using the LOINC standard, a physiological parameter value with a measurement device identifier, and an efficacy evaluation marked with a clinician verification status. The data correlation engine automatically matches treatment measures with corresponding effect feedback, such as correlating the administration time point of ticagrelor with subsequent platelet inhibition rate test results to establish a drug-effect pharmacokinetic response curve. Data quality verification rules reject records with logical conflicts. For example, if the bleeding time is significantly prolonged without recording the adjustment of anticoagulant drugs, the data review process will be triggered. The model optimization of the medical risk prediction and decision-making module uses an incremental learning framework, and the preprocessed closed-loop medical data are converted into model training samples. The implementation effect of the treatment recommendation plan is quantified as a reward signal: the positive sample reward value is +1 for complete compliance with the plan and significant therapeutic effect, and the negative sample reward value is -1 for partial compliance with the plan and adverse reactions. The model parameter update uses the stochastic gradient descent algorithm, focusing on adjusting the weight parameters that significantly affect the prediction of treatment effects. The optimization process preserves the original knowledge structure of the model and prevents overfitting to recent small samples through regularization terms.

[0079] Taking insulin dose adjustment for diabetic patients as an example: the system initially recommends 12 units of insulin aspart for subcutaneous injection before meals, and the retrospective module records the subsequent 72-hour blood glucose monitoring values to form an effect feedback dataset. When it is detected that the blood glucose is consistently higher than 13.9 mmol / L after lunch and a hypoglycemic event occurs before breakfast, the incremental learning algorithm adjusts the dose prediction model: it reduces the weight coefficient of the breakfast dose and increases the dose sensitivity parameter after lunch. The optimized model generates a new plan to change to 10 units before breakfast, 14 units before lunch, and 11 units before dinner, which is reviewed by endocrinologists and put into clinical use. The diagnosis and treatment retrospective data support multi-dimensional efficacy analysis, and clinicians can query the response rate statistics of a specific treatment plan in different populations. The system provides a treatment path backtracking function, which can trace back to all treatment operations in the last 30 days when patients experience adverse reactions, assisting the medical team in analyzing the cause and effect. All data operation records comply with the FDA 21 CFR Part 11 electronic record specification, and the audit tracking log details the data modification time, operator identity, and modification reason.

[0080] The model optimization process establishes clinical safety boundaries, and any parameter adjustment must not exceed the maximum range specified in the drug instructions. When the optimized recommendation significantly differs from the clinical guidelines, the system automatically initiates a multidisciplinary consultation request, which is evaluated by a team of medical experts for the rationality of the algorithm's recommendation. The optimized model version is run through a shadow mode, comparing its prediction results with actual clinical decisions, and continuously collecting efficacy data until statistical significance is achieved before being officially enabled. The entire retrospective and optimization cycle is incorporated into the hospital quality management system, generating a medical artificial intelligence system performance report every month, which is submitted to the medical technology committee for clinical utility evaluation.

[0081] The above merely provides the preferred embodiments of the present application, and is not intended to limit the patent scope of the present application. Any equivalent structure or equivalent flowchart transformation, or direct or indirect application in other related technical fields, which is made based on the contents of the present application specification and drawings, shall be included in the patent protection scope of the present application.

Claims

1. A data processing system for medical diagnosis and treatment driven by artificial intelligence, characterized in that, include: The multimodal medical data acquisition module is used to acquire multi-source heterogeneous medical data of patients within a preset time period, and to perform data cleaning and standardization processing on the multi-source heterogeneous medical data to generate a standardized multimodal medical dataset. The medical feature deep extraction module is used to extract multi-dimensional features from the standardized multimodal medical dataset to obtain an initial medical feature set, and to construct a dynamic evolution feature matrix based on the time-varying patterns of each feature in the initial medical feature set. The multidimensional health status space construction module is used to construct a multidimensional spatial representation of the patient's health status based on the dynamic evolution feature matrix, calculate the distribution density of historical medical events in the multidimensional space, and determine a set of key medical early warning indicators based on the distribution density. The real-time medical data fusion module is used to receive real-time collected patient medical data, extract real-time medical features and form real-time feature vectors, map the real-time feature vectors to the multi-dimensional health state space, and generate real-time health risk factors. The medical risk prediction and decision-making module is used to establish a personalized health risk prediction model based on the real-time health risk factors, the initial medical feature set, and the dynamic evolution feature matrix, output the probability of disease occurrence, and generate individualized treatment suggestions.

2. The AI-driven medical diagnosis and treatment data processing system according to claim 1, characterized in that, The multimodal medical data acquisition module is specifically used for: Collect multi-source heterogeneous medical data, including imaging data, time-series physiological parameter data, laboratory test data, and clinical text data; The multi-source heterogeneous medical data is subjected to outlier removal, noise filtering, and missing value imputation to enhance data quality. The enhanced multi-source heterogeneous medical data is aligned and normalized according to data modality to generate a standardized multimodal medical dataset.

3. The data processing system for artificial intelligence-driven medical diagnosis and treatment according to claim 2, characterized in that, The medical feature deep extraction module is specifically used for: A deep convolutional neural network is used to extract multi-scale medical features from the standardized multimodal medical dataset; Medical features at different scales are fused and spliced ​​together to form an initial medical feature set; Calculate the rate of change and trend of change of each feature in the initial medical feature set within a specified time window; The rate of change and trend of change of each feature are organized according to time series to construct a dynamic evolution feature matrix.

4. The data processing system for artificial intelligence-driven medical diagnosis and treatment according to claim 3, characterized in that, The multidimensional health state space construction module is specifically used for: The number of dimensions of the multidimensional health state space is determined based on the number of feature dimensions of the dynamic evolution feature matrix. Calculate the distribution density of the corresponding features at the time of historical diagnostic events in the multidimensional health state space; A density threshold is set based on the distribution density, and features whose distribution density exceeds the density threshold are selected as the key medical early warning indicator set.

5. The AI-driven medical diagnosis and treatment data processing system according to claim 4, characterized in that, The real-time medical data fusion module is specifically used for: Feature extraction is performed on real-time collected patient medical data to generate real-time medical features; The real-time medical features are constructed into real-time feature vectors according to the organization format of the dynamic evolution feature matrix; The real-time feature vector is projected onto the multidimensional health state space, and its spatial distance with the set of key medical early warning indicators is calculated. Real-time health risk factors are generated based on the spatial distance.

6. The data processing system for artificial intelligence-driven medical diagnosis and treatment according to claim 5, characterized in that, The medical risk prediction and decision-making module establishes a personalized health risk prediction model as follows: The real-time health risk factors are fused with the initial medical feature set; The dynamic evolution feature matrix is ​​reconstructed using a temporal attention mechanism with weights. A risk prediction network is trained based on the fused features and the reconstructed feature matrix. Output the probability value of the patient's disease occurrence and generate a personalized treatment suggestion plan.

7. The AI-driven medical diagnosis and treatment data processing system according to claim 6, characterized in that, Also includes: The medical knowledge graph construction module is used to integrate medical knowledge bases and clinical guidelines to build a medical knowledge graph that includes the relationships between diseases, symptoms, drugs, and treatment plans. The medical risk prediction and decision-making module is also used to match and verify the individualized treatment recommendations with the medical knowledge graph.

8. The data processing system for artificial intelligence-driven medical diagnosis and treatment according to claim 7, characterized in that, Also includes: The multi-center medical data collaboration module is used to aggregate standardized multimodal medical datasets from multiple medical institutions, perform federated learning training, and update the personalized health risk prediction model. The medical risk prediction and decision-making module is also used to make predictions using an updated personalized health risk prediction model.

9. The AI-driven medical diagnosis and treatment data processing system according to claim 8, characterized in that, Also includes: The diagnosis and treatment process traceability module is used to record the implementation process and effect feedback of the individualized treatment recommendations, forming closed-loop medical data; The medical risk prediction and decision-making module is also used to continuously optimize the personalized health risk prediction model based on the closed-loop medical data.

10. A data processing method for medical diagnosis and treatment driven by artificial intelligence, characterized in that, It includes all modules and method flows of the AI-driven medical diagnosis and treatment data processing system as described in any one of claims 1 to 9.

Images