Multi-source medical data mining and health management system based on AI agent driving

Abstract

The invention relates to the technical field of data mining, in particular to a multi-source medical data mining and health management system based on AI (artificial intelligence) agent driving, which solves the problem of time sequence isomerism of multi-source medical data through a sensing layer dynamic alignment module, constructs a telescopic time window to realize physiological cycle alignment, and improves the accuracy of health management. The data fusion confidence is improved by using a conflict resolution gradient pool; a decision-making layer evolution engine module generates an individualized dynamic health baseline through a bimodal prediction network, and realizes accurate risk early warning based on real-time deviation degree calculation; the execution layer closed loop verification module generates a personalized intervention scheme according to the early warning signal, and continuously optimizes a confidence coefficient weight of a conflict resolution gradient pool in the perception layer dynamic alignment module and parameters of a bimodal prediction network in the decision layer evolution engine module through feedback data; finally, an intelligent health management system capable of coevolution is constructed, and the problems that multi-source data fusion is low in reliability, a health baseline is static and the system lacks self-learning ability are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of data mining technology, specifically to a multi-source medical data mining and health management system driven by AI intelligent agents. Background Technology

[0002] With the widespread adoption of the Internet of Things (IoT) in healthcare and wearable devices, real-time acquisition of multi-source, heterogeneous medical data has become a fundamental capability in the field of health management. However, existing technologies face several challenges: data generated by devices with different frequencies and precision exhibits temporal asynchrony and physiological rhythm differences, leading to significant biases during data fusion; conflicting data lacks dynamic weighting and resolution mechanisms, relying on static weights or manual rules, making it difficult to adapt to device error fluctuations and user specificity; health baseline models are mostly static thresholds, failing to reflect the dynamic changes and tolerance boundaries of individual physiological states; and the lack of closed-loop feedback between early warning systems and intervention plans results in lagging model iteration and insufficient generalization ability. These issues collectively limit the effectiveness of health management systems in terms of real-time performance, personalization, and self-evolution capabilities. Summary of the Invention

[0003] The purpose of this invention is to provide a multi-source medical data mining and health management system driven by AI intelligent agents to solve the problems mentioned in the background art. The specific technical problems include how to achieve dynamic temporal alignment and conflict resolution of multi-source medical data, and to build a health management system that can co-evolve based on bimodal prediction and closed-loop feedback mechanism to solve the problems of low confidence in heterogeneous data fusion, static health baseline and lack of adaptive optimization capability of the system.

[0004] To achieve the above objectives, the present invention provides the following technical solution: This AI-driven multi-source medical data mining and health management system includes a perception layer dynamic alignment module, a decision layer evolution engine module, and an execution layer closed-loop verification module, among which: The dynamic alignment module of the perception layer first receives multi-source heterogeneous medical data; this module performs time-series consistency alignment operation, which constructs a scalable time window based on the characteristics of human metabolic cycle, compresses high-frequency second-level data into hour-level feature blocks through physiological rhythm cycle folding, and aligns it with low-frequency data such as weekly data under the same physiological cycle benchmark, thereby solving the problem that different frequency data sources are difficult to integrate due to time-series asynchrony. Subsequently, the dynamic alignment module of the perception layer performs dynamic weighted resolution of conflict data collected from multiple devices through its conflict resolution gradient pool. This process sets initial weights based on the device type registration library and adjusts these initial weights in real time with reference to the error rate of similar devices in the historical verification library, ultimately outputting a high-confidence fusion feature vector. This vector contains three key data elements: steady-state parameters, dynamic fluctuation factors, and user historical health events. When generating this vector, fluctuation suppression processing is performed on steady-state parameters to preserve their long-term statistical characteristics, and an adaptive window scaling algorithm is used for dynamic fluctuation factors to maintain the high resolution of their original data. At the same time, error compensation coefficients are loaded into the confidence weights of the conflict resolution gradient pool for user historical health events through a dedicated event response channel. This series of operations ensures the temporal consistency of multi-source heterogeneous data before fusion and the high reliability during the fusion process, providing a high-quality data foundation for subsequent decision-making.

[0005] The decision-making layer evolutionary engine module runs a bimodal prediction network to utilize cleaned and fused data. Its first prediction modality receives a high-confidence fused feature vector output from the perception layer, analyzes it through a deep spatiotemporal feature extraction network, and uses an adaptive state transition matrix to separate the steady-state parameters and dynamic fluctuation factors. Subsequently, an individual-specific modeling layer is used to construct the user's physiological tolerance boundary. Finally, a dynamic baseline generation layer integrates the steady-state parameters and tolerance boundary to generate an individualized dynamic health baseline that includes dynamic floating thresholds. This makes the health baseline no longer a static value, but a dynamic model that can reflect the user's unique physiological state and change patterns.

[0006] The second prediction mode in the dual-modal prediction network receives real-time streaming data and an individualized dynamic health baseline output by the first prediction mode. Through a high-precision deviation calculation engine, it calculates the instantaneous deviation of the real-time data based on the dynamic floating threshold in the baseline, generating a structured risk quantification matrix. Then, a risk pattern decoupler performs real-time pattern matching and trend extrapolation to trigger a health risk warning signal. This signal carries key metadata such as the specific risk level, signal confidence level, impact dimension, and reference time period, providing sufficient basis for precise intervention at the execution layer.

[0007] The generation unit in the execution layer closed-loop verification module generates individualized health intervention plans based on health risk warning signals. Based on the risk level and impact dimensions identified in the health risk warning signals, combined with the user's current physiological rhythm state and historical health baseline characteristics, it performs pattern matching and dynamic adaptation in the preset health intervention plan knowledge base to generate and execute highly customized intervention plans. Subsequently, the update unit in the execution layer closed-loop verification module collects feedback data on the user's physiological state based on the execution results of the intervention plan, including dynamic changes in physiological indicators and behavioral response characteristics. This feedback data is transmitted back to the perception layer dynamic alignment module and the decision layer evolution engine module through an integrated communication channel. In the perception layer dynamic alignment module, the initial weights in the conflict resolution gradient pool are recalibrated according to the actual performance of the device, and dynamic optimization is performed with reference to the error patterns in the historical verification library. In the decision layer evolution engine module, the feedback data is synchronously input, and the parameters of the bimodal prediction network are updated through the gradient optimization process. This closed-loop feedback mechanism drives the co-evolution of the entire agent, enabling the system to learn from each intervention, continuously optimize the accuracy of its data fusion, risk prediction, and intervention recommendation, and ultimately achieve personalized, continuous self-evolving health management.

[0008] Compared with the prior art, the beneficial effects of the present invention are: By employing the dynamic temporal alignment and conflict resolution mechanism of the perception layer dynamic alignment module, the confidence and reliability of multi-source heterogeneous medical data fusion are improved. With the help of the dual-modal prediction network of the decision layer evolutionary engine module, a health baseline that accurately reflects individual physiological dynamics and tolerance boundaries is generated, enabling early and accurate warnings of health risks. Finally, through the closed-loop verification and feedback of the execution layer closed-loop verification module, the system parameters and strategies are continuously self-optimized, thereby constructing a highly adaptive, personalized, and collaboratively evolving intelligent health management system, comprehensively improving the accuracy, real-time performance, and long-term effectiveness of health management. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the overall modules of the present invention; Figure 2 This is a schematic diagram of the execution layer closed-loop verification module unit of the present invention; Figure 3 This is a schematic diagram of the feedback closed-loop process of the present invention.

[0010] In the diagram: 100, Perception Layer Dynamic Alignment Module; 200, Decision Layer Evolution Engine Module; 300, Execution Layer Closed-Loop Verification Module; 301, Generation Unit; 302, Update Unit. Detailed Implementation

[0011] The technical solutions in 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.

[0012] Next, please refer to Figure 1 The present invention provides a technical solution: a multi-source medical data mining and health management system based on AI intelligent agent driving, including a perception layer dynamic alignment module 100, a decision layer evolution engine module 200 and an execution layer closed-loop verification module 300.

[0013] The technical actions of the perception layer dynamic alignment module 100 are guided by the first prediction mode of the service decision layer evolution engine module 200, and the specific process includes: First, it receives multi-source heterogeneous medical data. To address the time-series fragmentation problem in multi-source acquisition systems such as wearable devices and clinical instruments, it performs time-series consistency alignment. Based on the characteristics of human metabolic cycles (such as circadian rhythms), it constructs a scalable time window. Through physiological rhythm cycle folding, it compresses second-level pulse data into hour-level feature blocks, aligning them with weekly physical examination reports under the same physiological cycle benchmark, ensuring that the first predictive modality receives a data structure that conforms to the laws of biological rhythms. Subsequently, to address numerical conflicts caused by differences in device accuracy (such as measurement discrepancies between home blood pressure monitors and hospital monitors), a conflict resolution gradient pool was activated to perform dynamic weighted resolution. First, initial weights were assigned to medical devices from different sources based on the device type registry. Then, historical error records of similar devices under the same physiological conditions were queried in real time from the historical verification database. The initial weights were dynamically adjusted based on the error rate. Finally, the adjusted weights were used to perform weighted fusion calculations on conflicting data collected from multiple devices to generate a high-confidence fusion feature vector that eliminates device bias.

[0014] The final generated high-confidence fusion feature vector is not a general data product. During the generation of the high-confidence fusion feature vector, the perception layer dynamic alignment module 100 explicitly carries three key data elements required by the subsequent decision layer: first, steady-state parameters characterizing long-term stable physiological characteristics, including basal metabolic rate and mean resting heart rate; second, dynamic fluctuation factors reflecting short-term physiological changes, including post-exercise heart rate recovery rate and blood glucose fluctuation coefficient; and third, time-series data recording the user's past health events, including hypoglycemia attack time and arrhythmia records. A differentiated processing strategy is employed when generating the high-confidence fusion feature vector, specifically including: For steady-state parameters, a moving average filter is used to suppress fluctuations and preserve their long-term statistical characteristics. For dynamic fluctuation factors, an adaptive window scaling algorithm is used to maintain the original sampling frequency during critical physiological events. For users' historical health events, the corresponding error compensation coefficients are loaded into the confidence weights of the conflict resolution gradient pool through the event response channel to realize the dynamic influence of historical events on the current data weighting.

[0015] The decision-making layer evolutionary engine module 200 realizes data value transformation based on a dual-modal prediction network. The first prediction modality is dedicated to parsing the data structure preprocessed by the perception layer and receives high-confidence fused feature vectors from the perception layer dynamic alignment module 100. The first prediction modality decodes the individual metabolic patterns (such as blood glucose tolerance) implicit in the high-confidence fused feature vectors into continuous dynamic health parameters through a deep spatiotemporal feature extraction network, generating an individualized dynamic health baseline (e.g., a user-specific resting heart rate range). This baseline is not a static threshold but a dynamic benchmark that evolves in real time with the user's physiological state. The specific process of generating the individualized dynamic health baseline includes: First, a high-confidence fusion feature vector is received from the dynamic alignment module 100 of the perception layer. This vector contains time-series data aligned by physiological rhythm cycles and multi-source fusion features after conflict resolution gradient pool weighted resolution. The first prediction mode analyzes the intrinsic correlation of vectors through an adaptive state transition matrix and uses a metabolic mode decoupling layer to separate the steady-state parameters and dynamic fluctuation factors in the high-confidence fused feature vectors. The baseline values ​​of health indicators are determined based on steady-state parameters (such as basal metabolic rate and mean resting heart rate) in the high-confidence fusion feature vector provided by the perception layer. The physiological limit range of each health parameter is determined by combining the tolerance boundary constraints generated by the individual-specific modeling layer based on the user's historical health events (such as hypoglycemia attack records and arrhythmia events). The real-time change characteristics of dynamic fluctuation factors (such as heart rate recovery rate after exercise and blood glucose fluctuation coefficient) are introduced, and the fluctuation characteristics of recent physiological parameters are calculated by using the sliding time window statistical method. The percentile method is used to superimpose an adaptive adjustment based on recent behavioral patterns on the baseline values ​​to generate a dynamic floating threshold that evolves in real time with the user's physiological state. This threshold can be dynamically adjusted according to the user's recent activity level, physiological rhythm state, and health event history. The final output includes a personalized dynamic health baseline with dynamically floating thresholds. This personalized dynamic health baseline is also presented as a multi-dimensional dynamic parameter space, which includes, for example, a base baseline value (based on the median feature of the user's historical data) and a risk evolution vector (identifying potential paths that deviate from the normal trajectory). The real-time update mechanism of this baseline relies on a high-confidence fusion feature vector of new input from the perception layer to ensure that the baseline evolves in tandem with the user's health status.

[0016] The core technical action of the second prediction modality in the bimodal prediction network is to use an individualized dynamic health baseline as a benchmark to perform real-time risk quantification and early warning triggering on real-time streaming data. Specifically, this includes: The second predictive modality continuously receives real-time streaming data from the front-end sensing device, and simultaneously acquires the individualized dynamic health baseline (which includes a multi-dimensional dynamic parameter space, with each dimension defined by a basic baseline value, a dynamic floating threshold, and a risk evolution vector) generated and continuously evolved by the first predictive modality. The early warning process begins with a high-precision deviation calculation engine. This engine first extracts key features from the incoming real-time streaming data to ensure data structure compatibility. Then, it strictly uses the dynamic floating threshold defined in the individualized dynamic health baseline as the dynamic benchmark for real-time comparison, calculating the instantaneous deviation between the current value and the elastic boundary dimension by dimension. The deviation calculation result is then superimposed with a confidence decay factor. The deviation vector generated by this process is not a simple scalar difference, but a structured risk quantification matrix that integrates the instantaneous deviation amplitude, physiological rhythm state, and the credibility of the data source. The structured risk quantification matrix is ​​input into a preset risk pattern decoupler. The decoupler performs real-time pattern matching and trend extrapolation based on the potential deviation path pattern identified by the risk evolution vector. If the current deviation trajectory is detected to match a certain risk evolution vector preset by the baseline, or if the deviation magnitude continues to exceed the dynamic floating threshold and the confidence level meets the standard, the corresponding level of health risk warning signal will be automatically triggered. The final output health risk warning signal carries key metadata, including the specific risk level (based on the degree of deviation and pattern matching results), signal confidence (calculated based on the resolution gradient label), impact dimension (identifying the breakthrough parameter dimension), and reference time period (associated time sequence alignment marker).

[0017] Please see Figure 2 Upon receiving a health risk warning signal from the second prediction mode output of the decision-making layer evolutionary engine module 200, which carries key metadata (specific risk level, signal confidence level, impact dimension, and reference time period), the execution layer closed-loop verification module 300 immediately initiates a closed-loop response process, specifically including: The generation unit 301 in the execution layer closed-loop verification module 300 first performs pattern matching and dynamic adaptation in the preset health intervention plan knowledge base based on the specific risk level and impact dimension identified in the health risk warning signal, combined with the user's current physiological rhythm state and historical health baseline characteristics, to generate a highly customized individual health intervention plan and execute it. The update unit 302 in the execution layer closed-loop verification module 300 obtains the individualized intervention effect based on the execution results of the individualized health intervention plan, and collects feedback data on the user's physiological state to trigger the coordinated operation of the confidence weights of the conflict resolution gradient pool in the perception layer dynamic alignment module 100 and the parameters of the bimodal prediction network in the decision layer evolutionary engine module 200. Specifically, this includes: The update unit 302 simultaneously collects feedback data after intervention, including dynamic changes in physiological indicators and behavioral response characteristics. The collected feedback data is transmitted to the dynamic alignment module 100 of the perception layer through the integrated communication channel to update the confidence weights in the conflict resolution gradient pool. The method is to recalibrate the initial weights based on the actual performance of the device and dynamically optimize them with reference to the error patterns in the historical verification library. Simultaneously, the feedback data is synchronously input into the decision layer evolution engine module 200, which updates the parameters of the dual-modal prediction network through the gradient optimization process, including the parameters of the deep spatiotemporal feature extraction network of the first prediction mode and the high-precision deviation calculation engine of the second prediction mode, thereby improving the accuracy of baseline generation and risk quantification. This dual update mechanism effectively drives the agent's collaborative evolution mechanism, ensuring that the perception layer, decision-making layer, and execution layer achieve dynamic collaborative evolution through continuous feedback.

[0018] Please see Figure 3 The perception layer dynamic alignment module 100 first fuses and processes the multi-source raw data through its conflict resolution gradient pool to resolve potential data conflicts, ultimately outputting a high-confidence fused feature vector. This refined data is then passed to the middle decision layer evolutionary engine module 200. The core of this module is a bimodal prediction network, which, through the collaborative work of the first and second prediction modes, performs in-depth analysis, calculation, and generates decision signals (such as health risk warnings) on the received high-confidence fused feature vector. This decision signal is finally sent to the bottom execution layer closed-loop verification module 300, which is responsible for executing decisions and collecting feedback data. This feedback data is crucial; it flows back upwards to continuously optimize and adjust the conflict resolution strategy of the perception layer dynamic alignment module 100 and the parameters of the first and second prediction modes of the bimodal prediction network in the decision layer evolutionary engine module 200, thus forming a self-iteratory and continuously evolving closed-loop system.

[0019] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-source medical data mining and health management system driven by AI intelligent agents, characterized in that, It includes a perception layer dynamic alignment module (100), a decision layer evolution engine module (200), and an execution layer closed-loop verification module (300), wherein: The dynamic alignment module (100) of the perception layer is configured to receive multi-source heterogeneous medical data, construct a scalable time window by performing a time-series consistency alignment operation and fold and align different frequency data sources based on physiological rhythm cycles, and perform dynamic weighted resolution on conflict data collected by multi-source devices through a conflict resolution gradient pool to output a high-confidence fusion feature vector. The decision layer evolution engine module (200) is configured to run a dual-modal prediction network, wherein the first prediction modality receives the high-confidence fusion feature vector output by the perception layer dynamic alignment module (100) and generates an individualized dynamic health baseline, and the second prediction modality receives real-time streaming data and the individualized dynamic health baseline output by the first prediction modality, and generates a health risk warning signal by calculating the deviation between the real-time streaming data and the baseline. The execution layer closed-loop verification module (300) generates an individualized health intervention plan based on the health risk warning signal and collects feedback data after intervention. It updates the confidence weight of the conflict resolution gradient pool in the perception layer dynamic alignment module (100) and the parameters of the bimodal prediction network in the decision layer evolution engine module (200) to drive the agent's collaborative evolution mechanism.

2. The AI-driven multi-source medical data mining and health management system according to claim 1, characterized in that, When the dynamic alignment module (100) of the perception layer performs the time-series consistency alignment operation, it constructs a scalable time window based on the characteristics of human metabolic cycle, compresses second-level data into hour-level feature blocks through physiological rhythm cycle folding, and aligns them with week-level data under the same physiological cycle benchmark.

3. The AI-driven multi-source medical data mining and health management system according to claim 2, characterized in that, When the dynamic alignment module (100) of the perception layer performs dynamic weighted resolution through the conflict resolution gradient pool, it sets the initial weights based on the device type registration library and adjusts the initial weights in real time with reference to the error rate of similar devices in the historical verification library to generate the high confidence fusion feature vector.

4. The AI-driven multi-source medical data mining and health management system according to claim 3, characterized in that, The high-confidence fusion feature vector output by the dynamic alignment module (100) of the perception layer includes three key data elements: steady-state parameters, dynamic fluctuation factors, and user historical health events.

5. The AI-driven multi-source medical data mining and health management system according to claim 4, characterized in that, When generating a high-confidence fusion feature vector, the perception layer dynamic alignment module (100) performs fluctuation suppression processing on the steady-state parameters to retain long-term statistical features, enables an adaptive window scaling algorithm on the dynamic fluctuation factor to maintain the original data resolution, and loads the error compensation coefficient of the user's historical health events into the confidence weight of the conflict resolution gradient pool through the event response channel.

6. The multi-source medical data mining and health management system based on AI intelligent agent driven according to claim 1, characterized in that, The first prediction mode of the decision layer evolution engine module (200) parses the high-confidence fusion feature vector through a deep spatiotemporal feature extraction network, separates the steady-state parameters and dynamic fluctuation factors using an adaptive state transition matrix, constructs the tolerance boundary using an individual-specific modeling layer, and integrates the steady-state parameters and tolerance boundary using a dynamic benchmark generation layer to generate an individualized dynamic health baseline containing dynamic floating thresholds.

7. The multi-source medical data mining and health management system based on AI intelligent agent driven according to claim 1, characterized in that, The second prediction mode calculates the instantaneous deviation of real-time streaming data based on the dynamic floating threshold in the individualized dynamic health baseline using a high-precision deviation calculation engine, generates a structured risk quantification matrix, and performs real-time pattern matching and trend extrapolation through a risk pattern decoupler to trigger a health risk warning signal.

8. The multi-source medical data mining and health management system based on AI intelligent agent driven according to claim 7, characterized in that, The health risk warning signal carries key information, including the specific risk level, signal confidence level, impact dimension and reference time period, which is used to drive the execution layer closed-loop verification module (300) to generate an intervention plan.

9. The multi-source medical data mining and health management system based on AI intelligent agent driven according to claim 1, characterized in that, The execution layer closed-loop verification module (300) includes a generation unit (301). The generation unit (301) is configured to generate and execute a highly customized individualized health intervention plan by performing pattern matching and dynamic adaptation in a preset health intervention plan knowledge base based on the specific risk level and impact dimension identified in the health risk warning signal, combined with the user's current physiological rhythm state and historical health baseline characteristics.

10. The AI-driven multi-source medical data mining and health management system according to claim 1, characterized in that, The execution layer closed-loop verification module (300) includes an update unit (302). The update unit (302) obtains the individualized intervention effect based on the execution result of the individualized health intervention plan and collects feedback data of the user's physiological state. The feedback data includes dynamic changes in physiological indicators and behavioral response characteristics. The update unit (302) transmits the feedback data to the perception layer dynamic alignment module (100) through an integrated communication channel. This is used to recalibrate the initial weights based on the actual performance of the device and dynamically optimize them by referring to the error patterns in the historical verification library, so as to update the confidence weights in the conflict resolution gradient pool. Simultaneously, the feedback data is input into the decision-making layer evolution engine module (200), and the parameters of the dual-modal prediction network are updated through the gradient optimization process, thereby driving the agent's collaborative evolution mechanism.

Images