Auxiliary system based on artificial intelligence medical data analysis

Abstract

The invention belongs to the technical field of medical artificial intelligence, and discloses an auxiliary system based on artificial intelligence medical data analysis. A hidden trend turning point in time sequence physiological index data is analyzed and identified through a micro-period sliding window, time axis accurate alignment of multi-modal medical data is realized based on the turning point, and a time-space synchronization matrix is constructed to quantify a time coupling relationship among different modal data. And extracting a curvature distribution and tangent angle sequence by adopting geometric characteristic analysis of a feature vector change trajectory, and constructing a multi-branch path prediction model of illness state evolution. The system monitors the fit degree change rate of the patient state and the prediction path in real time, scientifically identifies the switching time of the illness state evolution branch, foresight marks the optimal intervention opportunity, and generates personalized clinical intervention suggestions. According to the invention, the change from passive monitoring to active prediction is realized, the timeliness and accuracy of clinical intervention are improved, and the medical resource configuration is optimized.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, and more specifically, to an auxiliary system based on artificial intelligence medical data analysis. Background Technology

[0002] In the current clinical setting, patients' temporal physiological indicators, imaging examinations, and laboratory test data are often processed in isolation. Due to inconsistent data collection times and the varying physiological response delays of different data types, it is difficult to construct a unified spatiotemporal representation of the patient's condition. Existing monitoring systems rely excessively on threshold triggering mechanisms, leading to situations where "data appears normal while the patient's condition has subtly deteriorated" in the monitoring of critically ill patients. For example, in patients with heart failure, subtle changes in heart rate variability and respiratory rate may exist 24-48 hours before the clinical manifestation of pulmonary edema, but these changes are often overlooked because they do not exceed warning thresholds. Furthermore, clinical prediction models often employ simple linear or single-path prediction logic, failing to reflect the reality that complex diseases such as acute kidney injury may simultaneously exhibit multiple evolutionary directions—recovery, stabilization, or deterioration. Physicians rely primarily on experience to determine the timing of intervention, lacking dynamic quantitative analysis of the disease's evolutionary trajectory. This is particularly challenging in complex cases such as multiple organ dysfunction syndromes, where determining the optimal treatment window is even more difficult. Current feature engineering primarily focuses on static statistical characteristics, such as the mean and standard deviation of vital signs, while neglecting the dynamic geometric characteristics of disease evolution. This results in limited predictive accuracy in clinical scenarios such as epileptic seizure prediction and arrhythmia risk assessment. This threshold-based, passive, reactive early warning model often leads to delayed intervention in clinical practice, missing optimal treatment opportunities. This delay is particularly problematic for high-risk patient groups such as the elderly and those with immunosuppression. Such delays can directly translate into higher mortality rates and longer hospital stays, while also increasing the consumption of medical resources and the workload of healthcare workers.

[0003] In view of this, the present invention proposes an auxiliary system based on artificial intelligence medical data analysis to solve the above problems. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: an auxiliary system based on artificial intelligence medical data analysis, comprising: The data acquisition module is used to acquire patients' time-series physiological indicators, imaging data, and laboratory test data; The micro-period analysis module is used to perform micro-period sliding window analysis on time-series physiological index data, extract the data fluctuation direction and fluctuation amplitude within each micro-period, and identify the implicit trend turning points in the time-series physiological index data based on the number of turning points in the fluctuation direction and the second derivative of the fluctuation amplitude between adjacent micro-periods. The time axis alignment module is used to align imaging data and laboratory test data on the time axis based on the timestamps of the latent trend inflection points. During the alignment process, the time offset compensation amount of each modality data relative to the latent trend inflection point is calculated according to the acquisition delay characteristics of each modality data. The matrix construction module is used to construct the spatiotemporal synchronization matrix of multimodal data based on the time offset compensation amount. The element values ​​of the spatiotemporal synchronization matrix represent the temporal coupling strength of different modal data when the same physiological event occurs. The feature trajectory analysis module is used to extract the feature vector change trajectory of different modal data before and after the latent trend inflection point based on the spatiotemporal synchronization matrix, and to calculate the curvature distribution and tangent angle sequence of the feature vector change trajectory. The model building module is used to construct a multi-branch path prediction model for disease evolution based on curvature distribution and tangent angle sequence. The multi-branch path prediction model includes the transition probability of each evolution branch and the boundary conditions of the feature vector corresponding to each branch. The fit calculation module is used to calculate the fit between the feature vector of the incremental physiological index data and the boundary conditions of each branch in the multi-branch path prediction model by collecting the incremental physiological index data of the patient in real time. The intervention timing marking module is used to identify the switching time of the disease evolution branch based on the rate of change of fit, and mark the optimal intervention time within a preset time window before the switching time. The intervention suggestion generation module is used to generate personalized clinical intervention suggestions based on the optimal intervention timing and the corresponding evolutionary branch characteristics. The modules are connected via wired and / or wireless means to enable data transmission between them.

[0005] The technical effects and advantages of the artificial intelligence-based medical data analysis auxiliary system of the present invention are as follows: This invention's proactive early warning capabilities in critical care management provide healthcare teams with a valuable intervention window, improving treatment success rates. This early warning can translate into a substantial survival advantage, especially for rapidly progressing critical illnesses. The system-generated personalized intervention recommendations precisely match the patient's disease progression path, improving the accuracy of clinical decision-making and enabling physicians to make more confident and scientific judgments in complex clinical situations. For chronic disease management, the system's multi-pathway prediction function provides clinicians with a more comprehensive risk assessment perspective, shifting management from periodic adjustments to continuous optimization. In environments with strained medical resources, this invention optimizes resource allocation efficiency and reduces medical costs by reducing unnecessary examinations, avoiding overtreatment and treatment delays, while also alleviating the cognitive burden on healthcare staff, allowing them to focus more on complex medical decisions requiring human judgment. Attached Figure Description

[0006] Figure 1 This is a schematic diagram of an auxiliary system for medical data analysis based on artificial intelligence, according to the present invention. Detailed Implementation

[0007] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0008] This application provides an auxiliary system based on artificial intelligence medical data analysis. The system's execution entities include, but are not limited to, medical decision support platforms, clinical early warning systems, multimodal medical data analysis centers, and personalized treatment planning platforms, which can be considered general computing nodes in this application. The auxiliary system includes, but is not limited to, at least one cloud-based medical data analysis engine, distributed medical data processing system, and intelligent clinical decision support device.

[0009] Please see Figure 1 In this embodiment of the invention, an auxiliary system based on artificial intelligence medical data analysis includes: The data acquisition module is used to acquire patients' time-series physiological indicators, imaging data, and laboratory test data. Time-series physiological indicators include continuously monitored vital signs, electrocardiogram data, blood oxygen saturation, and other frequently collected physiological indicators; imaging data includes various medical images such as CT, MRI, and ultrasound; laboratory test data includes various laboratory examination data such as blood biochemical indicators, immunological parameters, and microbiological test results. This multimodal data is acquired in real time through the hospital information system, continuous monitoring equipment, and laboratory information system, providing comprehensive raw data for subsequent analysis and ensuring that the system can capture subtle changes in the patient's condition from multiple dimensions.

[0010] The micro-period analysis module performs micro-period sliding window analysis on time-series physiological indicator data, extracting the direction and amplitude of data fluctuations within each micro-period. Based on the number of turning points in the direction of fluctuation between adjacent micro-periods and the second derivative of the fluctuation amplitude, it identifies latent trend turning points in the time-series physiological indicator data. This module extracts local features from continuous physiological indicators through finely divided time windows, capturing subtle trends in indicator changes. The direction of fluctuation reflects the upward or downward trend of the indicator, and the fluctuation amplitude quantifies the degree of change. By analyzing the turning features between adjacent windows, the system can identify latent trend turning points that are easily overlooked in traditional monitoring, providing crucial signals for early disease warning.

[0011] The timeline alignment module aligns imaging and laboratory test data based on the timestamps of latent trend inflection points. During alignment, it calculates the time offset compensation amount for each modality of data relative to the latent trend inflection point, taking into account the acquisition delay characteristics of each modality. This module addresses the time asynchrony issue of different modalities of medical data. By establishing a unified time reference system based on latent trend inflection points and considering the inherent physiological response delays and equipment acquisition delays of different examination methods, it calculates precise time offset compensation amounts, achieving accurate alignment of multimodal data on the same physiological event timeline, laying the foundation for subsequent cross-modal analysis.

[0012] The matrix construction module constructs a spatiotemporal synchronization matrix for multimodal data based on the time offset compensation. The elements of this matrix represent the temporal coupling strength of different modalities during the same physiological event. By calculating the cross-correlation function between the corrected modalities, this module quantifies the degree of temporal synchronization of different data types during the same physiological event, forming a complete spatiotemporal synchronization network. The spatiotemporal synchronization matrix visually demonstrates the temporal correlation strength between various medical data, providing a mathematical expression for understanding the interactions between different physiological systems and offering a basis for modal weight allocation in feature extraction.

[0013] The Feature Trajectory Analysis module, based on a spatiotemporal synchronization matrix, extracts the trajectory of feature vector changes in different modalities before and after latent trend inflection points, and calculates the curvature distribution and tangent angle sequence of these trajectory changes. This module maps multimodal data to a high-dimensional feature space, tracks the movement of feature vectors within this space over time, and captures dynamic feature change patterns in disease evolution by calculating the geometric properties of these trajectories, such as curvature distribution and tangent angle sequence. These geometric properties reflect the rate, direction, and acceleration of disease progression, providing a dynamic perspective for predicting disease development.

[0014] The model building module constructs a multi-branch path prediction model for disease evolution based on curvature distribution and tangent angle sequences. This model includes the transition probability of each evolutionary branch and the corresponding feature vector boundary conditions. By clustering similar patient trajectories in historical cases, this module identifies different possible disease development paths, calculates the probability distribution of each path, and extracts the feature boundary conditions for each path. This multi-branch model surpasses traditional linear prediction methods, simulating the complexity and diversity of disease development and providing more comprehensive risk assessment and prognostic prediction for clinical decision-making.

[0015] The fit calculation module is used to calculate the fit between the feature vectors of incremental physiological index data collected in real time and the boundary conditions of each branch in the multi-branch path prediction model. This module continuously collects the patient's latest physiological indicators, converts them into feature vectors, and calculates the matching degree with each possible development path in the prediction model, thus assessing the most likely evolutionary path corresponding to the patient's current state in real time. The rate of change in fit reflects the trend of the patient's state moving closer to or away from a specific path, providing a dynamic basis for timely adjustments to treatment strategies.

[0016] The intervention timing marker module identifies the switching moments of disease progression branches based on the rate of change in fit, and marks the optimal intervention time within a preset time window before the switching moment. This module precisely captures the critical moments when a patient's condition shifts from one evolutionary path to another by monitoring jumps in the rate of change in fit, and identifies the best time window for intervention before the switching occurs. This proactive identification mechanism enables clinicians to implement interventions in the early stages of disease deterioration, improving treatment outcomes and reducing adverse consequences.

[0017] The intervention suggestion generation module generates personalized clinical intervention recommendations based on the optimal intervention timing and corresponding evolutionary branch characteristics. This module combines the patient's specific situation, the predicted disease progression path, and the optimal intervention timing to generate targeted treatment plan suggestions, including medication adjustments, treatment method selection, and monitoring strategy optimization. These suggestions take into account individual patient differences and the dynamic evolution of the disease, providing clinicians with precision medicine decision support.

[0018] The modules are connected via wired and / or wireless means to enable data transmission between them.

[0019] In this embodiment of the invention, the detailed implementation steps for performing micro-period sliding window analysis on time-series physiological index data include: The sliding window length is set to an integer multiple of the time-series physiological indicator data acquisition period, and the data is segmented into sliding windows with a step size as a single acquisition period. The sliding window is a fundamental method for analyzing continuous data, capturing local features by setting an appropriate window length. Setting the window length to an integer multiple of the acquisition period (usually 10-30 periods) ensures sufficient data points for feature extraction while avoiding the overloading of local features due to excessively long windows. Choosing a step size of a single acquisition period ensures smooth window sliding, without missing any data changes. For example, for ECG monitoring data (the acquisition period is usually 1 second), the window length can be set to 15 seconds, moving forward 1 second each time to form a new window. This fine-grained windowing strategy lays the foundation for capturing minute changes in physiological indicators.

[0020] Linear fitting is performed on the data points within each sliding window to obtain the slope of the fitted line. The sign of the slope determines the direction of data fluctuation within the sliding window. Linear fitting is a simple and effective method for extracting local trends in data. The best-fit line is calculated using the least squares method. The fitting process can be represented as: ;in, and These are the time points and their corresponding physiological indicator values. The slope This is the intercept.

[0021] The slope 'a' obtained by solving this minimization problem directly reflects the overall trend of data change within the window; a positive value indicates an upward trend, a negative value indicates a downward trend, and a value close to zero indicates a stable state. This slope-based method for determining the direction of fluctuation is simple, intuitive, and computationally efficient, making it suitable for the needs of real-time monitoring systems.

[0022] The ratio of the range to the standard deviation of data points within each sliding window is calculated as the volatility of that window. Volatility quantifies the drastic nature of data changes and is an important indicator for assessing the stability of physiological indicators. The range (the difference between the maximum and minimum values) reflects the absolute range of data variation, while the standard deviation measures the dispersion of data points around the mean. The ratio of the two eliminates the influence of differences in the units of measurement of different indicators, providing a standardized indicator of volatility.

[0023] A larger ratio indicates a significant jump in the data within a short period of time, rather than a gradual change within the normal fluctuation range, which may indicate a sudden change in physiological state.

[0024] The system counts the number of times the fluctuation direction changes within consecutive sliding windows. When this number exceeds a preset turning point threshold within a preset observation window, the center moment of that observation window is recorded. Changes in fluctuation direction are a key indicator for identifying trend reversals; frequent changes in direction may indicate instability or state transitions in a physiological system. The system sets a larger observation window (containing multiple sliding windows) and counts the number of fluctuation direction changes within it. When the number of changes exceeds a preset threshold (typically 30%-40% of the observation window length), the center moment is marked as a potential turning point candidate. This detection method based on the frequency of direction reversals can identify oscillation turning points that are easily overlooked in traditional monitoring, providing signals for early warning.

[0025] The second-order difference of the fluctuation amplitude is calculated. Points where the absolute value of the second-order difference is greater than a preset acceleration threshold and the time interval between the second-order difference and the center time is less than a preset interval are marked as latent trend inflection points. Second-order difference analysis is an effective method for detecting accelerated changes in data, capable of capturing changes in the rate of change of physiological indicators, i.e., the acceleration of change. The calculation process first performs first-order difference on the fluctuation amplitude sequence to obtain the rate of change of the fluctuation amplitude; then, it performs first-order difference on the rate of change again to obtain the acceleration of the fluctuation amplitude. The formula for calculating the second-order difference is: ;in, For the first The second difference value of the fluctuation amplitude of each window. For the first The fluctuation range of each window.

[0026] When the absolute value of the second-order difference exceeds a preset acceleration threshold, and the time of occurrence is close to the previously identified center of directional inflection, that moment is ultimately confirmed as a latent trend inflection point. This dual verification mechanism, combining directional inflection and acceleration analysis, significantly improves the accuracy of inflection point identification, reduces false positives, and provides a reliable time reference point for subsequent analysis.

[0027] In this embodiment of the invention, the detailed implementation steps for aligning imaging data and laboratory test data along the timeline based on the timestamps of latent trend inflection points include: The system acquires the capture timestamps of imaging data and the sampling timestamps of laboratory test data. Timestamps are fundamental information for data alignment, precisely recording the exact point in time when the data was acquired. The system extracts the capture times of various medical images and the sampling times of laboratory tests through a hospital information system interface, ensuring time accuracy down to the minute level. For imaging data, timestamps are typically included in the DICOM header file; for laboratory data, timestamps come from records in the laboratory information system. These raw timestamps provide a reference benchmark for subsequent time calibration, but due to differences in the acquisition process and response latency of different data types, further processing is required to achieve true physiological time alignment.

[0028] Based on the physiological response delay models for each modality of data, the theoretical delay time from the occurrence of a physiological event to the presentation of changes in each modality of data is calculated. This theoretical delay time includes physiological conduction delay and equipment acquisition delay. The physiological response delay model is crucial for achieving accurate alignment of different modalities of data, considering the entire process delay from the occurrence of a physiological event to the presentation of data changes. Physiological conduction delay reflects the time required for signal transmission and metabolic transformation within the body; for example, changes in blood biochemical indicators typically lag behind the occurrence of pathological changes. Equipment acquisition delay includes the time consumed in sample collection, processing, and detection. The system has established typical delay models for various indicators based on large-scale clinical data. For instance, changes in the inflammatory marker CRP typically lag 4-6 hours after the onset of infection, while changes in renal function indicators may lag 12-24 hours after kidney injury. These theoretical delay parameters provide a scientific basis for time calibration, ensuring that different modalities of data can be aligned on the true timeline of physiological events.

[0029] Using the timestamp of the latent trend inflection point as the baseline, the system calculates the first time difference between the imaging data's capture timestamp and the baseline, and the second time difference between the laboratory test data's sampling timestamp and the baseline. Using the latent trend inflection point as the baseline is the core strategy for multimodal data alignment, unifying all data into the same reference frame. Inflection points typically represent key moments of change in physiological states and are clinically significant time anchors. The system calculates the difference between the timestamp of each modality's data and the baseline, obtaining the raw time bias. For data occurring before the inflection point, the time difference is negative; for data occurring after the inflection point, the time difference is positive. This calculation of relative time differences lays the foundation for subsequent delay correction, enabling data from different acquisition frequencies and time points to be compared and analyzed on a unified time axis.

[0030] Subtracting the theoretical delay time corresponding to the imaging data from the first time difference yields the time offset compensation amount for the imaging data; subtracting the theoretical delay time corresponding to the laboratory test data from the second time difference yields the time offset compensation amount for the laboratory test data. Time offset compensation is the final step in achieving true physiological time alignment. By introducing a correction for the theoretical delay, it eliminates the time deviation introduced during the acquisition process. The formula for calculating the compensation amount is: ;in, For modality Time offset compensation amount This is the time difference between the timestamp of the modal data and the reference time. This is the theoretical delay time for this mode.

[0031] This compensation mechanism takes into account the inherent time lag of different data types. For example, laboratory blood glucose test results may show the sampling time point, but actually reflect the physiological state at an earlier time; imaging examinations may be performed after symptoms appear, but the pathological changes they reflect may have existed for several days. Through theoretical delay correction, the system achieves precise alignment of different modal data on the real physiological timeline, providing a time consistency guarantee for the comprehensive analysis of multimodal data.

[0032] In this embodiment of the invention, the detailed implementation steps for constructing a multi-branch path prediction model include: In the historical case database, a set of historical cases with an initial feature vector similarity greater than a preset similarity threshold to the current patient is retrieved. Similar case retrieval is a fundamental step in personalized prediction, providing empirical evidence for prediction by finding historical cases similar to the current patient's condition. The retrieval process first converts the current patient's multimodal data into a high-dimensional feature vector as the query condition; then, in a database containing a large number of standardized cases, the similarity between each historical case and the query vector is calculated; finally, a set of cases with a similarity exceeding a preset threshold (typically 0.75-0.85) is selected. The similarity calculation uses a weighted cosine similarity method, assigning higher weights to key features, as shown in the formula: ;in, and These are the feature vectors for the current patient and historical cases, respectively. For the first The weight coefficients of each feature.

[0033] This similarity-based case retrieval method fully leverages the empirical value of medical big data, providing a high-quality sample set for subsequent trajectory clustering and path prediction.

[0034] Cluster analysis is performed on the feature vector change trajectories of each case in the historical case set, grouping trajectories with similar curvature distribution patterns into the same evolutionary branch. Trajectory clustering is a key step in discovering typical disease development paths, revealing potential disease evolution patterns by identifying the geometric similarity of trajectories in the feature space. The clustering process first extracts the curvature distribution features of each case's trajectory, calculates the curvature values ​​of the trajectory at each time point, forming a curvature sequence; then, the Dynamic Time Warping (DTW) algorithm is applied to the curvature sequence to calculate the distance between sequences; finally, density clustering methods (such as DBSCAN) are used to automatically identify trajectory clusters. The curvature calculation uses a three-point approximation method; for discrete time series, the formula is: ;in, For the first Curvature at a given time point and For the feature vector in and The first derivative of the component (velocity). and It is the second derivative (acceleration), which can be approximated by finite difference.

[0035] Curvature describes the degree of bending of a trajectory, reflecting the severity and directionality of disease state changes, and is a key feature for trajectory classification. Through clustering, the system identifies several typical disease evolution paths, serving as the foundational branches of the multi-branch prediction model.

[0036] The proportion of cases in each evolutionary branch to the total number of historical cases is used as the metastasis probability of that branch. Metastasis probability quantifies the likelihood of a patient progressing along a specific path and is an important basis for risk assessment. The calculation process is based on simple frequency statistics, and the formula is: ;in, For branches The transition probability, Belongs to branch The number of cases, This represents the total number of similar cases.

[0037] Metastasis probability reflects the probability distribution of disease progression in different directions under similar initial conditions, directly influencing the risk assessment of treatment decisions. For example, if the metastasis probability of the deterioration path is as high as 60%, while that of the improvement path is only 20%, it suggests a high risk of disease deterioration in the current state, and a more aggressive intervention strategy should be considered. This probability estimation based on historical data provides a quantitative risk reference for clinical decision-making, compensating for the shortcomings of traditional experience-based judgment.

[0038] Envelopes are extracted from the feature vector change trajectories within each evolutionary branch to obtain the upper and lower bounds of the feature vectors for that branch. These upper and lower bounds serve as the boundary conditions for the feature vectors of that branch. Boundary conditions define the activity range of each evolutionary branch in the feature space and are crucial for determining the attribution of new data. The extraction process first aligns all trajectories within each branch along the time dimension; then, at each time point, the distribution statistics of all case feature vectors are calculated to extract the upper and lower boundaries; finally, a smoothing process is used to generate continuous boundary envelopes. Boundary calculation typically employs the quantile method, with the upper bound set at the 95th quantile and the lower bound at the 5th quantile, ensuring that 90% of the normal variation within the branch is included while excluding outliers. This boundary definition method based on statistical distribution considers both the natural variability of disease evolution and provides clear criteria for branch attribution, offering a quantitative basis for path matching in real-time monitoring.

[0039] Based on the transition probabilities and eigenvector boundary conditions of each evolutionary branch, a multi-branch path prediction model is constructed. This model represents the probability distribution and feature constraints of the evolution from the current disease state to different disease outcomes. The multi-branch prediction model is the core algorithm component of the system, integrating the results of path identification, probability estimation, and boundary definition to form a complete prediction framework. The model adopts a directed graph structure, with the current state as the root node, each evolutionary branch as a possible path, transition probabilities as edge weights, and boundary conditions as path constraints. The mathematical expression of the model is a triple: ;in, For prediction models, It is a set of states (including the initial state and each branch state). The transition probability matrix, This is the set of boundary conditions.

[0040] This multi-branch model surpasses traditional linear prediction methods, simulating the complexity and diversity of disease development, capturing key features of state transitions, and providing comprehensive risk assessment and prognostic prediction for clinical decision-making. The model supports real-time updates, dynamically adjusting the current state and prediction path as new data is input, enabling continuous risk monitoring and early warning.

[0041] In this embodiment of the invention, the detailed implementation steps for identifying the switching moment of the disease progression branch based on the rate of change of fit, and marking the optimal intervention time within a preset time window before the switching moment, include: The distance between the feature vectors of incremental physiological index data and the boundary conditions of each evolutionary branch is calculated. The evolutionary branch with the smallest distance is taken as the current best-fitting branch, and the reciprocal of the distance is used as the fit degree. Fit degree calculation is a key step in assessing the degree of matching between the patient's current state and the predicted path, quantifying the matching relationship through a distance metric in the feature space. The calculation process first converts the real-time acquired incremental physiological indices into feature vectors of the same dimension as the prediction model; then, it calculates the distance between this vector and the boundary conditions of each branch; finally, it selects the branch with the smallest distance as the current best-fitting branch, and uses the reciprocal of the distance as the fit degree index. The distance calculation uses a modified Mahalanobis distance, considering the correlation between features, and the formula is: ;in, For feature vectors To branch Distance to the boundary For points on the boundary, Let be the characteristic covariance matrix.

[0042] Fit is the reciprocal of distance. A higher fit value indicates that the patient's current state is closer to the typical characteristics of that branch, and the more likely they are to progress along that path. By continuously monitoring changes in fit, the system can track the dynamic evolution of the patient's state and promptly detect signs of path transition.

[0043] Incremental physiological index data are collected at fixed time intervals. The change in fit over consecutive time periods is calculated, and this change is divided by the fixed time interval to obtain the rate of change in fit. Rate of change analysis is an effective method for capturing rapid changes in condition, reflecting the trend of change by calculating the time derivative of fit. The system collects new data at fixed intervals (usually 5-15 minutes) to ensure the consistency of the rate of change calculation. The formula for calculating the rate of change is: ;in, for Time Branch The rate of change in fit for The fit at all times This represents the data collection time interval.

[0044] The sign of the rate of change reflects the trend of the patient's condition moving closer to or away from a specific branch, while the absolute value of the rate of change reflects the speed of the condition change. By monitoring abnormal fluctuations in the rate of change, the system can sensitively capture the acceleration or deceleration of disease progression, providing a time window for early intervention.

[0045] When the absolute value of the rate of change in the fit exceeds a preset threshold, and the best-fit branch switches, the moment of the switch is marked as the switch moment. Identifying the switch moment is a crucial step in capturing disease turning points, signifying a shift in the patient's condition from one evolutionary path to another. The identification process combines quantitative analysis of the rate of change with qualitative judgment of the best-fit branch; a switch is confirmed only when both conditions are met simultaneously. The rate of change threshold is dynamically set based on different disease types and monitoring indicators, typically 3-5 times the normal fluctuation range. This dual-verification mechanism effectively reduces false alarms and ensures that the identified switch moments are clinically significant. Accurate identification of switch moments is crucial for disease prognosis and treatment adjustments, especially for acute illnesses and critically ill patients; timely detection of disease turning points can be life-saving.

[0046] Starting from the switching point, the system traces back a preset time window to find the moment when the rate of change in fit first exceeds a preset warning threshold. This moment is marked as the optimal intervention time, with the warning threshold being less than the rate of change threshold. Marking the optimal intervention time is the core value of the system, identifying the golden time window for intervention through prospective analysis. The marking process employs a backtracking strategy, searching for early warning signals from the confirmed switching point. The warning threshold is set at 50%-70% of the rate of change threshold, providing earlier warnings while ensuring signal reliability. The time window length is set according to the disease progression rate; for acute diseases, it may be 2-6 hours, and for chronic diseases, it may be 12-24 hours. The optimal intervention time is typically at the stage where the rate of change in fit begins to rise significantly but has not yet reached a drastic change. Intervention at this point may block or slow the disease's progression down an undesirable path, achieving the best intervention effect. This automatic identification of the prospective intervention time provides clinicians with precise decision support, helping to improve treatment outcomes and reduce the waste of medical resources.

[0047] In this embodiment of the invention, the detailed implementation steps for constructing the spatiotemporal synchronization matrix of multimodal data based on the time offset compensation amount include: Based on the time offset compensation, the timestamps of each modality's data are offset corrected to align data points corresponding to the same physiological event across different modalities on the time axis. Time correction is the first step in constructing the spatiotemporal synchronization matrix. By applying the calculated time offset compensation, precise alignment of different modalities' data is achieved. The correction process applies compensation adjustments to the original timestamp of each data point to reflect the actual physiological event time, rather than the data acquisition time. For data point i, the formula for calculating its corrected timestamp is: ;in, This is the corrected timestamp. This is the original timestamp. This is the time offset compensation amount for the mode m to which this data point belongs.

[0048] The corrected timestamps reflect the actual time when the corresponding physiological state occurred, eliminating the impact of time delays during the data acquisition process for different modalities. This time correction mechanism based on physiological models enables data points that were originally scattered in terms of acquisition time to be aligned on the true physiological timeline, laying the foundation for subsequent cross-correlation analysis.

[0049] Calculate the cross-correlation function between any two corrected modal data points, and obtain the peak position and peak amplitude of the cross-correlation function. Cross-correlation analysis is an effective method for quantifying the time synchronization of different modal data. By calculating the similarity of signals under different time delays, the optimal matching point is identified. The analysis process first converts the corrected modal data into time series with a uniform sampling rate; then, it calculates the cross-correlation function between any two modes; finally, it extracts the peak position and amplitude of the cross-correlation function. For discrete time series, the formula for calculating the cross-correlation function is: ;in, For modality and Time delay The cross-correlation value under the following conditions and for Data values ​​for the two modalities at time points.

[0050] The peak position represents the time delay when two modes reach optimal matching, ideally close to zero; the peak amplitude represents the degree of similarity at optimal matching, reflecting the synergy of the changes in the two modes. This correlation analysis based on signal processing can accurately quantify the temporal correlation of responses in different physiological systems, providing core data for constructing a spatiotemporal synchronization matrix.

[0051] When the time offset of the peak position is less than a preset synchronization threshold, the peak amplitude is used as the temporal coupling strength between the two modalities. Temporal coupling strength is a key indicator for measuring the synchronicity of different modalities, reflecting the degree of coordination between physiological systems. Only when the peak position is close to zero (time offset less than the preset threshold, typically 10%-20% of the expected physiological delay) is the peak amplitude considered to truly reflect the synchronization relationship between modalities, rather than a random correlation. This threshold-based screening mechanism ensures the reliability of coupling strength calculation and avoids interference from spurious correlations. Coupling strength directly affects the weight allocation of different modalities in subsequent analysis; modal combinations with high coupling strength receive higher weights in feature fusion, reflecting their synergistic importance in disease development.

[0052] A matrix is ​​constructed with each modality's data as its row and column indices. The calculated temporal coupling strength is then filled into the corresponding matrix elements to form a spatiotemporal synchronization matrix. The spatiotemporal synchronization matrix is ​​a mathematical expression of the temporal relationships between multimodal data, visually demonstrating the synchronization network between various medical data. The matrix construction process first creates an n×n square matrix (n being the number of modalities); then, the calculated temporal coupling strength between each pair of modalities is filled into the corresponding positions; for modal pairs that have not reached the synchronization threshold, a value of zero or a low baseline is set. Each element in the matrix... Indicates from modality To mode The temporal coupling strength reflects the degree of correlation between two data types in the time dimension. This matrix representation not only facilitates computation and storage but also enables subsequent visualization analysis and pattern recognition, providing a mathematical tool for understanding the synergistic mechanisms of multi-system diseases.

[0053] The spatiotemporal synchronization matrix is ​​normalized so that the values ​​of the matrix elements range from 0 to 1. Larger values ​​indicate stronger temporal synchronization between the corresponding modalities. Normalization is a necessary step to ensure comparability between different modal combinations, mapping the original coupling strength to a unified range through linear transformation. The normalization process uses a min-max normalization method. The normalized matrix provides an intuitive representation of synchronization strength; values ​​close to 1 indicate high synchronization, while values ​​close to 0 indicate almost no correlation. This standardized representation facilitates setting uniform thresholds and weights, simplifying subsequent feature fusion and pattern recognition algorithms. The normalized spatiotemporal synchronization matrix is ​​an important tool for understanding the synergistic mechanisms of multi-system diseases, revealing the temporal coupling relationships between different physiological systems in disease development and providing a systemic perspective for precision treatment.

[0054] In this embodiment of the invention, the detailed implementation steps for extracting the feature vector change trajectories of different modal data before and after the latent trend inflection point based on the spatiotemporal synchronization matrix, and calculating the curvature distribution and tangent angle sequence of the feature vector change trajectories include: Using the latent trend inflection point as the time center, data segments within symmetrical time intervals before and after this inflection point are extracted for each modality. Symmetrical data extraction is a fundamental step in trajectory analysis, capturing the complete process of state change by setting a reasonable time window. The extraction process centers on the identified latent trend inflection point, extending equal time lengths before and after it (typically 24-48 hours before and after the inflection point) to ensure coverage of the entire state change process. For different modalities, considering their sampling frequency and change characteristics, different window lengths may be set, but symmetry is maintained throughout. This symmetrical extraction strategy centered on the inflection point enables the system to comprehensively capture the complete process of disease state from stability to change and then to a new steady state, providing a sufficient data foundation for trajectory feature extraction.

[0055] Feature extraction is performed on each modal data segment to obtain a sequence of feature vectors for that modality over time. These feature vectors include statistical and frequency domain features. Feature extraction is a crucial step in transforming the raw data into a high-dimensional feature space. By calculating various statistical measures and frequency domain characteristics, the multidimensional features of the data are comprehensively captured. Statistical features include indicators describing distribution characteristics such as mean, standard deviation, skewness, kurtosis, and quantiles. Frequency domain features are obtained through Fast Fourier Transform (FFT) or wavelet analysis, including indicators reflecting the periodicity of the data such as energy distribution, dominant frequency, and harmonic ratio. For time-varying characteristics, a sliding window method is used to calculate feature values ​​at continuous time points, forming a time series of feature vectors. This multidimensional feature extraction method considers both the static distribution characteristics of the data and captures dynamic change patterns, providing a comprehensive feature representation for subsequent trajectory analysis.

[0056] Based on the coupling strength of each modality in the spatiotemporal synchronization matrix, the feature vectors of each modality are weighted and fused to form a time series of multimodal fused feature vectors. Feature fusion is a key step in integrating multimodal information, constructing a unified feature representation by considering the temporal coupling relationship between modalities. The fusion process first aligns and normalizes the dimensions of each modality's feature vector; then, it calculates the fusion weights based on the coupling strength in the spatiotemporal synchronization matrix; finally, it obtains the fused feature vector through weighted summation. The fusion weights reflect the synchronicity strength between modalities, with modalities with higher coupling strength gaining greater influence in the fusion. This adaptive weight allocation mechanism based on coupling strength ensures the preservation of key information and the suppression of noise during the fusion process, improving the quality of feature representation and disease specificity.

[0057] This method maps time series data of multimodal fused feature vectors to a high-dimensional feature space, fitting trajectory curves of the feature vector changes within this space. High-dimensional mapping is an effective method for analyzing the dynamics of complex systems, revealing the evolutionary patterns of system states by treating time series as trajectories in a high-dimensional space. The mapping process preserves the original dimensionality of the feature vectors or projects them into a visualized low-dimensional space using nonlinear dimensionality reduction techniques such as t-SNE. Trajectory fitting employs spline interpolation or local multinomial regression to construct smooth, continuous curves between discrete data points. The fitted trajectory curves visually demonstrate the patient's state's movement path in the feature space, reflecting the transition of the disease from one steady state to another. This geometric perspective of trajectory analysis provides a new dimension for understanding disease dynamics, surpassing traditional static feature comparison methods.

[0058] The curvature values ​​of the trajectory curve at each time point are calculated to form a curvature distribution; the angles between the tangent vectors of the trajectory curve at each time point and the reference axis are calculated to form a tangent angle sequence. Geometric feature extraction is a key step in quantifying the dynamic characteristics of the trajectory. By calculating curvature and tangent angles, the rate and direction information of state changes are captured. Curvature is calculated using the three-point method or differential geometry formulas mentioned earlier; the tangent angle is calculated using the vector directions of adjacent points, with the following formula: ;in, For the first Tangent angle at each time point and These represent the changes in the y and x directions of the point in the feature space, respectively.

[0059] Curvature distribution reflects the severity of trajectory bending, with high curvature regions typically corresponding to rapid state changes or turning points; tangent angle sequences reflect the directionality of change, with abrupt changes in angle indicating a shift in system behavior. These geometric features directly map the dynamics of disease progression, providing key inputs to predictive models in terms of dynamic dimensions and helping to capture early change patterns that are difficult to detect with traditional static features.

[0060] This invention achieves precise monitoring and prospective intervention for disease development through micro-period analysis of time-series physiological indicators, spatiotemporal alignment of multimodal data, geometric analysis of feature trajectories, and multi-branch path prediction. The multimodal integration and latent trend identification method of this invention can capture early change signals that are difficult to detect with traditional monitoring, improving the timeliness and accuracy of clinical intervention, providing patients with personalized treatment strategies, and significantly improving treatment outcomes and the efficiency of medical resource utilization.

[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0062] It should be noted that all formulas in this manual are calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0063] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. An auxiliary system for medical data analysis based on artificial intelligence, characterized in that, include: The data acquisition module is used to acquire patients' time-series physiological indicators, imaging data, and laboratory test data; The micro-period analysis module is used to perform micro-period sliding window analysis on the time-series physiological index data, extract the data fluctuation direction and fluctuation amplitude within each micro-period, and identify the implicit trend turning points in the time-series physiological index data based on the number of turning points in the fluctuation direction and the second derivative of the fluctuation amplitude between adjacent micro-periods. The time axis alignment module is used to align the imaging data and the laboratory test data on the time axis based on the timestamp of the latent trend inflection point. During the alignment process, the time offset compensation amount of each modality data relative to the latent trend inflection point is calculated according to the acquisition delay characteristics of each modality data. The matrix construction module is used to construct a spatiotemporal synchronization matrix for multimodal data based on the time offset compensation amount. The feature trajectory analysis module is used to extract the feature vector change trajectory of different modal data before and after the latent trend inflection point based on the spatiotemporal synchronization matrix, and to calculate the curvature distribution and tangent angle sequence of the feature vector change trajectory. The model building module is used to construct a multi-branch path prediction model for disease evolution based on the curvature distribution and the tangent angle sequence. The fit calculation module is used to calculate the fit between the feature vector of the incremental physiological index data and the boundary conditions of each branch in the multi-branch path prediction model by collecting the incremental physiological index data of the patient in real time. The intervention timing marking module is used to identify the switching time of the disease evolution branch based on the rate of change of the fit, and mark the optimal intervention time within a preset time window before the switching time; The intervention suggestion generation module is used to generate personalized clinical intervention suggestions based on the optimal intervention timing and the corresponding evolutionary branch characteristics.

2. The auxiliary system for medical data analysis based on artificial intelligence according to claim 1, characterized in that, The step of performing micro-period sliding window analysis on the time-series physiological index data, extracting the data fluctuation direction and amplitude within each micro-period, and identifying latent trend turning points in the time-series physiological index data based on the number of turning points in the fluctuation direction and the second derivative of the fluctuation amplitude between adjacent micro-periods, includes: The sliding window length is set to an integer multiple of the time-series physiological index data acquisition period, and the time-series physiological index data is divided into sliding window segments with a step size as a single acquisition period; Linear fitting is performed on the data points within each sliding window to obtain the slope of the fitted line, and the direction of data fluctuation within the sliding window is determined based on the sign of the slope. Calculate the ratio of the range to the standard deviation of the data points within each sliding window, and use this as the fluctuation range corresponding to the sliding window; The number of times the fluctuation direction changes between consecutive sliding windows is counted. When the number of changes exceeds a preset turning threshold within a preset observation window, the center time of that observation window is recorded. Calculate the second-order difference value of the fluctuation amplitude, and mark the moment when the absolute value of the second-order difference value is greater than a preset acceleration threshold and the time interval between the second-order difference value and the center moment is less than a preset interval as the implicit trend turning point.

3. The auxiliary system for medical data analysis based on artificial intelligence according to claim 1, characterized in that, The imaging data and laboratory test data are time-axis aligned based on the timestamp of the latent trend inflection point. During the alignment process, the time offset compensation amount of each modality data relative to the latent trend inflection point is calculated according to the acquisition delay characteristics of each modality data, including: Obtain the capture timestamp of the imaging data and the sampling timestamp of the laboratory test data; Based on the physiological response delay model of each modality of data, the theoretical delay time from the occurrence of a physiological event to the presentation of changes in each modality of data is calculated. The theoretical delay time includes physiological conduction delay and device acquisition delay. Using the timestamp of the latent trend inflection point as the reference time, calculate the first time difference between the imaging data capture timestamp and the reference time, and the second time difference between the laboratory test data sampling timestamp and the reference time; Subtracting the theoretical delay time corresponding to the imaging data from the first time difference yields the time offset compensation amount of the imaging data; subtracting the theoretical delay time corresponding to the laboratory test data from the second time difference yields the time offset compensation amount of the laboratory test data.

4. The auxiliary system for medical data analysis based on artificial intelligence according to claim 1, characterized in that, The step of constructing a multi-branch path prediction model for disease evolution based on the curvature distribution and the tangent angle sequence includes: In the historical case database, retrieve a set of historical cases whose initial feature vectors are more similar to the current patient than a preset similarity threshold; Cluster analysis was performed on the feature vector change trajectory of each case in the historical case set, and trajectories with similar curvature distribution patterns were grouped into the same evolutionary branch; The proportion of cases in each evolutionary branch to the total number of historical cases is used as the transition probability of that evolutionary branch. Envelopes are extracted from the feature vector change trajectories within each evolutionary branch to obtain the upper and lower bounds of the feature vectors of that evolutionary branch. The upper and lower bounds are then used as the feature vector boundary conditions of that evolutionary branch. The multi-branch path prediction model is constructed based on the transition probabilities and eigenvector boundary conditions of each evolutionary branch.

5. The auxiliary system for medical data analysis based on artificial intelligence according to claim 1, characterized in that, The step of identifying the switching moment of the disease evolution branch based on the rate of change of the fit, and marking the optimal intervention time within a preset time window before the switching moment, includes: Calculate the distance between the feature vector of the incremental physiological index data and the boundary conditions of each evolutionary branch, take the evolutionary branch with the smallest distance as the current best-fitting branch, and take the reciprocal of the distance as the fit degree. The incremental physiological index data are collected at fixed time intervals, the change in fit over consecutive time periods is calculated, and the change in fit is divided by the fixed time interval to obtain the rate of change in fit. When the absolute value of the rate of change of the fit is greater than a preset rate of change threshold, and the best fit branch switches, the moment of the switch is marked as the switching moment. Using the switching moment as the endpoint, trace back to the length of a preset time window, and within this time window, find the moment when the fitting change rate first exceeds a preset warning threshold. Mark this moment as the optimal intervention time, where the warning threshold is less than the change rate threshold.

6. The auxiliary system for medical data analysis based on artificial intelligence according to claim 1, characterized in that, The step of constructing a spatiotemporal synchronization matrix for multimodal data based on the time offset compensation amount includes: Based on the time offset compensation amount, the timestamps of each modality data are offset corrected so that the data points of different modal data corresponding to the same physiological event are aligned on the time axis. Calculate the cross-correlation function between any two modal data after correction, and obtain the peak position and peak amplitude of the cross-correlation function; When the time offset of the peak position is less than the preset synchronization threshold, the peak amplitude is used as the temporal coupling strength between the two modal data. A matrix is ​​constructed with each modal data as its row and column index, and the calculated temporal coupling strength is filled into the corresponding matrix element positions to form the spatiotemporal synchronization matrix.

7. The auxiliary system for medical data analysis based on artificial intelligence according to claim 1, characterized in that, Based on the spatiotemporal synchronization matrix, the feature vector change trajectories of different modal data before and after the latent trend inflection point are extracted, and the curvature distribution and tangent angle sequence of the feature vector change trajectories are calculated, including: Using the latent trend inflection point as the time center, extract the data segments of each modality data within the symmetrical time interval before and after the inflection point; For each modal data segment, feature extraction is performed to obtain the feature vector sequence of the modal data in the time series. The feature vector includes statistical features and frequency domain features. Based on the coupling strength corresponding to each mode in the spatiotemporal synchronization matrix, the feature vectors of each mode are weighted and fused to form a time series of multimodal fusion feature vectors; The time series of the multimodal fused feature vector is mapped to a high-dimensional feature space, and the change trajectory curve of the feature vector is fitted in this space; Calculate the curvature value of the trajectory curve at each time point to form a curvature distribution; calculate the angle between the tangent vector of the trajectory curve at each time point and the reference axis to form a tangent angle sequence.

Images