Medical data management system based on artificial intelligence

Abstract

The invention discloses a medical data management system based on artificial intelligence, and the system comprises the following steps: a sliding window detection module which is used for analyzing a monitoring sequence of medical data, and capturing a continuous missing section; the wavelet fractal decomposition module is used for executing wavelet packet transformation and self-similar fractal decomposition on the missing section to generate a fractal embedded tensor; the differential feature generation module is used for generating a joint feature flow; the three-branch coding module is used for respectively inputting the joint feature flow into a context branch, a fractal branch and a physiological prior branch, and fusing to generate a comprehensive latent code; the controlled diffusion generation module is used for generating reconstruction missing fragments; the confidence evaluation module is used for calculating a confidence score; the dynamic access management module is used for dynamically processing and reconstructing missing fragments according to the confidence score; and the explanatory report module is used for generating an explanatory report. According to the invention, sliding detection, fractal analysis and diffusion generation methods are combined, and medical data missing is intelligently repaired.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and medical data management, and in particular to a medical data management system based on artificial intelligence. Background Art

[0002] With the continuous advancement of medical information technology, medical monitoring equipment is widely used in various fields, including clinical treatment, rehabilitation management, and personal health management. By continuously monitoring key physiological indicators such as heart rate, blood pressure, respiratory rate, and blood oxygen saturation, medical systems can understand the patient's condition in real time, assisting in early disease detection and treatment decision-making. However, due to factors such as equipment failure, sensor detachment, wireless transmission interruption, or data collection anomalies, medical monitoring data often contains missing fragments. If these missing fragments are not effectively corrected, they will not only reduce the accuracy of data analysis but may also lead to diagnostic bias, decision delays, and even medical safety risks. Therefore, intelligent processing technology to address the problem of missing medical data has become an important research direction in the field of medical big data management.

[0003] In the prior art, conventional methods for dealing with missing medical monitoring data mainly include simple interpolation, statistical regression, and prediction and filling methods based on machine learning. Simple interpolation methods usually use linear interpolation, spline interpolation, or polynomial fitting, and estimate and fill in the missing data by using the data points before and after the missing point. The method is simple to implement, but when dealing with complex medical physiological fluctuations, it often cannot accurately capture nonlinear change trends and is prone to introduce systematic errors. The statistical regression method predicts missing data by establishing a regression model, which can take into account multivariate associations to a certain extent. However, when faced with medical signals of different scales and frequencies, the model fitting ability is insufficient, especially during extreme physiological changes or abnormal events. Although prediction methods based on machine learning, such as using support vector machines, random forests, or basic neural networks for reconstruction, have achieved better performance than traditional statistical methods in some tasks, they still have problems such as strong model dependence, sensitivity to abnormal data, lack of physiological rationality constraints, and poor interpretability of results, which have limited their application in clinical practice.

[0004] Furthermore, existing medical data management systems generally have the following technical bottlenecks. First, the detection mechanism for missing data is relatively rough, mostly based on fixed thresholds or simple rules, and cannot dynamically and accurately capture continuous missing segments, especially when the data fluctuates slightly or is intermittently missing, which can easily lead to missed detections or false detections. Secondly, existing reconstruction methods usually fail to fully tap the potential multi-scale fractal characteristics in medical monitoring data, and cannot effectively model local trend changes, roughness evolution and long-term dependencies, resulting in significant differences in the structure and dynamic characteristics of the reconstruction results from the real physiological data. In addition, most current medical data reconstruction methods use fixed-step differential modeling, which ignores changes in data fluctuation amplitude and is difficult to adapt to data characteristics under different physiological states, resulting in single differential feature extraction and insufficient expression capabilities.

[0005] More importantly, during the reconstruction of missing data, existing technologies often lack dynamic guidance for physiological boundary conditions. Although many reconstruction results numerically conform to trends, they may exceed reasonable physiological ranges, such as generating abnormal hypertension values or heart rate sequences that do not conform to physiological rhythms. This deficiency significantly reduces the clinical usability and security of the data. Furthermore, most existing systems use a unified standard to handle missing reconstruction results and lack a dynamic decision-making mechanism based on reconstruction quality. They are unable to classify and process reconstructed fragments based on their reliability, such as automatic access, manual review, or exception recording. This leads to rigid data management strategies and poor system adaptability.

[0006] When it comes to assessing the quality of reconstructed data, existing methods generally lack systematic confidence modeling, often relying on crude error statistics or manual empirical judgment, and lack a quantitative confidence assessment system based on physiological boundary consistency calculations. This not only affects the level of intelligence in data management, but also limits the reasoning ability of subsequent intelligent decision-making systems based on data reliability. Furthermore, during the manual review process, existing medical data management systems essentially only provide the reconstructed values themselves and lack auxiliary explanatory analysis, such as a lack of comprehensive reports on changes in fractal features, conformity with physiological parameters, and statistical differences. This increases the burden of manual review and reduces review efficiency and accuracy.

[0007] Therefore, how to provide a medical data management system based on artificial intelligence is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of the present invention is to propose an artificial intelligence-based medical data management system. This system combines sliding window detection, wavelet fractal decomposition, differential feature generation, and controlled time diffusion generation methods to intelligently identify and restore continuously missing segments in medical data with high fidelity. A three-branch encoder guides feature extraction, dynamically guides the reconstruction process based on physiological boundaries, and combines confidence assessment with dynamic access management to achieve continuous optimization of data quality. This system offers the advantages of accurate detection, detailed reconstruction, high physiological plausibility, and intelligent management, effectively improving the integrity and reliability of medical data.

[0009] An artificial intelligence-based medical data management system according to an embodiment of the present invention includes the following modules:

[0010] Sliding window detection module, which is used to analyze the monitoring sequence of input medical data, capture continuous missing segments and record boundary points and timestamps;

[0011] Wavelet fractal decomposition module, used to perform wavelet packet transform and self-similar fractal decomposition on the missing segments, extract local trends, roughness index and Hurst index, and generate fractal embedding tensors;

[0012] The differential feature generation module is used to calculate the first-order differential vector and the second-order differential vector, and concatenate them with the fractal embedding tensor in the channel dimension to generate a joint feature flow;

[0013] The three-branch encoding module is used to input the joint feature stream into the context branch, fractal branch and physiological prior branch respectively, and fuse them to generate a comprehensive latent code;

[0014] A controlled diffusion generation module is used to take the comprehensive latent code as a conditional input, inject weighted noise in the forward direction, and dynamically guide the reverse direction by combining physiological boundary features to generate and reconstruct the missing fragments;

[0015] A confidence evaluation module is used to calculate a confidence score based on the distance between the reconstructed missing segment and the physiological boundary;

[0016] Dynamic access management module, used to dynamically decide whether to reconstruct missing segments and access them to the main data warehouse, push them for manual review, or retain the original missing segments based on the confidence score;

[0017] The explanatory report module is used to generate explanatory reports to assist manual review.

[0018] According to an embodiment of the present invention, a medical data management method based on artificial intelligence includes the following steps:

[0019] S1. Analyze the monitoring sequence of input medical data based on the sliding window detection algorithm, capture the continuous missing segments in real time, and record the observation boundary points and timestamps before and after the missing segments;

[0020] S2. Performing wavelet packet transform and self-similar fractal decomposition on the complete monitoring sequence of the missing section, extracting the local trend, roughness index and Hurst index at each scale, and generating a fractal embedding tensor;

[0021] S3. Based on the observation data of the missing segment neighborhood, calculate the first-order difference vector and the second-order difference vector, and splice them with the fractal embedding tensor in the channel dimension to generate a joint feature flow;

[0022] S4. Input the joint feature stream into a three-branch encoder, wherein the three-branch encoder includes a context branch, a fractal branch, and a physiological prior branch, and fuse the outputs of the three branches to generate a comprehensive latent code;

[0023] S5. The comprehensive latent code is used as a conditional vector and input into the controlled time diffusion model. Weighted noise is injected in the forward process. In the reverse process, it is dynamically guided according to the physiological boundary characteristics to generate and reconstruct the missing fragments and complete the medical data.

[0024] S6. Calculating a confidence score of the reconstructed missing segment based on the physiological boundary distance;

[0025] S7. Based on the confidence score, dynamically decide whether to connect the reconstructed missing segment to the main data warehouse, push it for manual review, or retain the original missing segment and generate an exception record to achieve continuous optimization management of medical data.

[0026] Optionally, the S1 specifically includes:

[0027] S11, setting the sliding window length and initializing the window starting point to the first data point in the monitoring sequence of the input medical data;

[0028] S12. In each window, count the number of valid data points and calculate the missing rate. If the missing rate exceeds the preset missing threshold, mark it as an abnormal window.

[0029] S13, merging consecutive abnormal windows into a preliminary missing segment, and recording the starting position and the ending position of the preliminary missing segment;

[0030] S14, expanding forward and backward at the boundary of the preliminary missing segment, detecting a change in the missing rate in a surrounding window, and determining a final boundary when the changing trend continuously decreases to below a restoration threshold;

[0031] S15. Extract observation points before and after the missing segment and their corresponding timestamps based on the final missing segment's front and back positions to generate missing segment boundary information.

[0032] S16. Number and standardize all detected missing segments to generate a complete one-dimensional missing segment mask.

[0033] Optionally, the S2 specifically includes:

[0034] S21, obtaining a complete monitoring sequence of the missing segment and performing preprocessing, including trend item elimination and stabilization processing, to eliminate long-term drift and mutation interference;

[0035] S22. Use the wavelet packet transform method to perform recursive multi-layer decomposition on the complete monitoring sequence after preprocessing, select the optimal wavelet basis function based on the entropy criterion, and obtain the decomposed sub-band signal set:

[0036]

[0037] Where X(t) represents the original input signal, represents the i-th subband signal under the n-th layer decomposition, and n represents the number of layers of wavelet packet decomposition;

[0038] S23. Extract local trend and calculate roughness index for each sub-band signal. The local trend change rate is expressed as The calculation formula of roughness index is:

[0039]

[0040] in, represents the roughness index of the ith subband signal under the nth layer decomposition, τ represents the interval length between the two observation time points, represents the value of the i-th subband signal at time point t+τ under the n-th layer decomposition, Represents the value of the i-th subband signal under the n-th layer decomposition at time point t;

[0041] S24. Based on the long-range correlation of sub-band signals, the Hurst exponent is calculated using the range standard deviation method:

[0042]

[0043] in, represents the Hurst exponent of the ith subband signal under the nth layer decomposition, R(m) represents the range when the subband signal length is m, S(m) represents the standard deviation when the subband signal length is m, and m represents the total length of the subband signal;

[0044] S25, extract the local trend change rate from each sub-band signal Roughness Index and Hurst index Perform feature fusion to form a scale feature vector;

[0045] S26. Stack and combine all scale feature vectors in the scale dimension to generate a fractal embedding tensor.

[0046] Optionally, the S3 specifically includes:

[0047] S31. Extract the observed data sequence of the missing segment neighborhood Based on the variable step difference method, the time interval is adaptively adjusted and the dynamic time step set {Δt k}, where Δt k represents the kth dynamic time step;

[0048] S32. Calculate the first-order difference vector based on the dynamic time step:

[0049] D1(t)=ω1(t)·(y(t+Δt1)-y(t))+ω2(t)·(y(t+

[0050] Δt2)-y(t));

[0051] Where D1(t) represents the first-order difference vector at time t, ω1(t) and ω2(t) represent the difference weights, y(t+Δt1) represents the observation value at the new moment after the dynamic time step Δt1, y(t) represents the observation value at the current moment t, and y(t+Δt2) represents the observation value at the new moment after the dynamic time step Δt2;

[0052] S33. Based on the first-order difference vector D1(t), the second-order difference vector D2(t) is calculated using a recursive residual structure:

[0053] D2(t)=D1(t+Δt1)-D1(t)+α·(D1(t+Δt2)-D1(t));

[0054] Where D2(t) represents the second-order difference vector at time t, D1(t+Δt1) represents the first-order difference vector at time t+Δt1, D1(t+Δt2) represents the first-order difference vector at time t+Δt2, and α represents the difference enhancement factor;

[0055] S34, performing statistical feature enhancement on the first-order difference vector D1(t) and the second-order difference vector D2(t), extracting the local mean, local standard deviation, and extreme value change rate as auxiliary features, and constructing an extended differential feature flow;

[0056] S35. Adaptively splicing the extended differential feature flow and the fractal embedding tensor in the channel dimension, and adjusting the splicing order according to feature correlation, to finally generate a joint feature flow.

[0057] Optionally, the S4 specifically includes:

[0058] S41, sending the joint feature stream to a three-branch encoder, wherein the three-branch encoder includes a context branch, a fractal branch, and a physiological prior branch, and each branch performs feature extraction and processing operations in parallel;

[0059] S42, the context branch encodes the time step position of the joint feature stream, adds a temporal feature description, and inputs the Transformer structure, wherein the Transformer structure includes a multi-head self-attention module and a feedforward neural network module. The multi-head self-attention module captures the long-range dependency in the joint feature stream, and the feedforward neural network module performs nonlinear mapping, and finally outputs the context feature representation;

[0060] S43, the fractal branch sends the joint feature stream to the gated fractal recursive unit. The gated fractal recursive unit introduces a gating mechanism at each recursive layer based on multi-scale recursive expansion to dynamically control the direction of information flow and recursive depth. The fractal pattern evolution relationship in the joint feature stream is extracted through cross-scale information integration, and finally a fractal feature representation is output.

[0061] S44, the physiological prior branch maps the joint feature stream into a physiological indicator node representation, performs relational reasoning based on the graph convolutional network, infers the physiological parameter boundary range of the joint feature stream on heart rate, blood pressure, and respiratory rate, and performs a priori consistency correction on the joint feature stream based on the reasoning result, and finally outputs the physiological prior feature representation;

[0062] S45. Perform feature splicing and fusion on the context feature representation, fractal feature representation, and physiological prior feature representation in the channel dimension to generate a comprehensive latent code.

[0063] Optionally, the S5 specifically includes:

[0064] S51, inputting the comprehensive latent code as a conditional vector into a controlled time diffusion model, wherein the controlled time diffusion model includes a forward diffusion process and a reverse generation process, and the physiological prior boundary information is synchronously loaded when the controlled time diffusion model is initialized;

[0065] S52. During the forward diffusion process, for each time step, the integrated latent code is superimposed with piecewise weighted noise, and the noise injection is dynamically set with weighting coefficients according to feature importance, so as to gradually perturb the integrated latent code into a standard normal distribution.

[0066] S53, at each forward diffusion time step, recording the disturbance trajectory and characteristic deviation degree during the noise injection process to form a disturbance scheduling table;

[0067] S54. During the reverse generation process, according to the perturbation schedule and the preliminary estimate of the integrated latent code, a denoising subnetwork is applied at each reverse time step to gradually remove noise, and physiological boundary features are dynamically introduced as guiding conditions to perform physiological rationality correction on the generation results at each step.

[0068] S55. During the reverse generation process, dynamically adjust the guidance intensity according to the difference between the current generation state and the physiological prior. When the generation result deviates from the physiological prior boundary, the guidance is enhanced; when the generation result approaches the neighborhood set by the physiological prior boundary, the guidance is reduced.

[0069] S56. After the reverse generation process is completed, the reconstructed missing fragments are output and added to the medical data sequence, completing the recovery of the missing fragments and the completeness supplement of the medical data.

[0070] Optionally, the S6 specifically includes:

[0071] After the reconstructed missing segments are generated, the confidence score is calculated based on the physiological boundary distance:

[0072]

[0073] Among them, S conf represents the confidence score of reconstructing the missing segment, β represents the adjustment coefficient, N represents the total number of physiological parameters, and x p represents the value of the pth physiological parameter in the reconstructed missing segment, u p represents the upper limit of the pth physiological parameter, l p represents the lower bound of the p-th physiological parameter, and max(·) represents the maximum value function.

[0074] Optionally, the S7 specifically includes:

[0075] S71, confidence score S based on the reconstruction of the missing fragment conf Perform classification judgment and set the high-level threshold θ high and the low-level threshold θ low ;

[0076] S72, if S conf ≥θ high , then the reconstructed missing segment is determined to be a high-confidence segment and is directly archived into the main data warehouse as a formal medical data record;

[0077] S73, if θ low ≤S conf <θ high , the reconstructed missing segment is determined to be a medium confidence segment, a manual review task is generated, and the reconstructed missing segment and the corresponding confidence index are pushed to the data administrator review system;

[0078] S74. Generate an explanatory report during the push process to assist manual decision-making;

[0079] S75, if S conf <θ low , the reconstructed missing segment is determined to be a low-confidence segment, the reconstructed missing segment is not archived, the original missing segment is retained, and the abnormal data processing log is recorded.

[0080] Optionally, the S74 specifically includes: based on the feature analysis results of the reconstructed missing fragments, extracting and recording local trend changes, roughness index changes and Hurst index changes; based on the physiological boundary matching analysis results, recording the degree of consistency deviation between the reconstructed data and the upper and lower boundaries of each physiological parameter; based on the statistical difference analysis, calculating the relative difference percentage between the reconstructed fragments and the historical real data in each statistical indicator, and dynamically generating an explanatory report containing charts, text descriptions and confidence score explanations.

[0081] The beneficial effects of the present invention are:

[0082] First, the present invention utilizes a sliding window detection module to capture consecutive missing segments in input medical monitoring sequences in real time. Compared to existing methods that rely on fixed thresholds or simple rules, sliding window detection dynamically identifies anomalous segments based on the local missingness rate of the data. Furthermore, through an extended detection mechanism, it accurately determines the missingness boundaries, effectively reducing false positives and missed detections, and laying a solid foundation for subsequent intelligent reconstruction of missing segments.

[0083] Secondly, this invention introduces wavelet packet transform and self-similar fractal decomposition techniques, combined with a differential feature generation module, to fully exploit the multi-scale time series characteristics and differential variation patterns of medical monitoring data. By extracting fractal features such as local trends, roughness index, and Hurst exponent, combined with dynamic step-size differential characteristics, this system enables a more detailed and rich description of medical data at the feature level, significantly improving the expressive power and fitting accuracy of the missing data reconstruction process, overcoming the limitations of existing reconstruction models, which suffer from a single feature and limited expressive power.

[0084] Furthermore, the present invention designs a three-branch encoder structure consisting of a contextual branch, a fractal branch, and a physiological prior branch. This structure extracts comprehensive features from three perspectives: global temporal dependency, cross-scale fractal evolution, and medical knowledge reasoning. This structure then generates a comprehensive latent code through an adaptive fusion mechanism. This structure not only improves the reconstruction model's ability to model complex dynamic features but also ensures the medical plausibility of the generated results through physiological prior guidance, avoiding the problem of physiological outliers often generated by missing reconstructions in traditional methods.

[0085] Furthermore, the present invention proposes a controlled time diffusion generation mechanism. During the forward diffusion process, piecewise weighted noise is injected based on feature importance. During the reverse diffusion process, physiological boundary features are dynamically incorporated to guide the generated trajectory. Generation deviations are flexibly adjusted, ultimately producing high-fidelity reconstructed segments that conform to physiological boundary constraints. Compared with existing direct regression or simple noise inverse recovery methods, this mechanism significantly improves the quality and reliability of missing segment reconstruction, ensuring continuity and physiological plausibility during the data recovery process.

[0086] Finally, the present invention uses a confidence assessment module to perform a quantitative credibility assessment on each reconstructed missing segment and dynamically determines the data access strategy based on the confidence score. This results in an intelligent processing flow in which high-confidence data is automatically stored, medium-confidence data is pushed for manual review, and low-confidence data is retained as missing and an exception record is generated. Furthermore, during the manual review phase, the present invention automatically generates an explanatory report that integrates fractal feature change analysis, physiological boundary matching analysis, and statistical difference analysis, greatly improving the efficiency and accuracy of manual review and enhancing the controllability and maintainability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0088] Figure 1 This is a schematic diagram of the structure of an artificial intelligence-based medical data management system proposed by the present invention;

[0089] Figure 2 This is an overall flow chart of the medical data management method based on artificial intelligence proposed by the present invention. DETAILED DESCRIPTION

[0090] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0091] refer to Figure 1 , an artificial intelligence-based medical data management system, including the following modules:

[0092] Sliding window detection module, which is used to analyze the monitoring sequence of input medical data, capture continuous missing segments and record boundary points and timestamps;

[0093] Wavelet fractal decomposition module, used to perform wavelet packet transform and self-similar fractal decomposition on the missing segments, extract local trends, roughness index and Hurst index, and generate fractal embedding tensors;

[0094] The differential feature generation module is used to calculate the first-order differential vector and the second-order differential vector, and concatenate them with the fractal embedding tensor in the channel dimension to generate a joint feature flow;

[0095] The three-branch encoding module is used to input the joint feature stream into the context branch, fractal branch and physiological prior branch respectively, and fuse them to generate a comprehensive latent code;

[0096] A controlled diffusion generation module is used to take the comprehensive latent code as a conditional input, inject weighted noise in the forward direction, and dynamically guide the reverse direction by combining physiological boundary features to generate and reconstruct the missing fragments;

[0097] A confidence evaluation module is used to calculate a confidence score based on the distance between the reconstructed missing segment and the physiological boundary;

[0098] Dynamic access management module, used to dynamically decide whether to reconstruct missing segments and access them to the main data warehouse, push them for manual review, or retain the original missing segments based on the confidence score;

[0099] The explanatory report module is used to generate explanatory reports to assist manual review.

[0100] The present invention forms a complete closed-loop medical data management system by setting up a sliding window detection module, a wavelet fractal decomposition module, a differential feature generation module, a three-branch encoding module, a controlled diffusion generation module, a confidence assessment module, a dynamic access management module and an explanatory report module. The system can accurately capture the missing segments in the monitoring data, extract multi-scale fractal features, combine dynamic differential features for modeling, and achieve high-fidelity reconstruction of missing data through a controlled diffusion mechanism. At the same time, the system introduces a physiological rationality dynamic guidance mechanism to ensure that the reconstructed data is reasonably distributed within the physiological range, and implements intelligent access management based on confidence scores, and further supports manual review through explanatory reports. The overall architecture has the advantages of accurate missing detection, rich feature extraction, high reconstruction reliability, and intelligent data management. It can significantly improve the integrity, credibility and application value of medical monitoring data, and has broad application prospects and promotion value in actual medical scenarios.

[0101] refer to Figure 2 , a medical data management method based on artificial intelligence, comprising the following steps:

[0102] S1. Analyze the monitoring sequence of input medical data based on the sliding window detection algorithm, capture the continuous missing segments in real time, and record the observation boundary points and timestamps before and after the missing segments;

[0103] S2. Performing wavelet packet transform and self-similar fractal decomposition on the complete monitoring sequence of the missing section, extracting the local trend, roughness index and Hurst index at each scale, and generating a fractal embedding tensor;

[0104] S3. Based on the observation data of the missing segment neighborhood, calculate the first-order difference vector and the second-order difference vector, and splice them with the fractal embedding tensor in the channel dimension to generate a joint feature flow;

[0105] S4. Input the joint feature stream into a three-branch encoder, wherein the three-branch encoder includes a context branch, a fractal branch, and a physiological prior branch, and fuse the outputs of the three branches to generate a comprehensive latent code;

[0106] S5. The comprehensive latent code is used as a conditional vector and input into the controlled time diffusion model. Weighted noise is injected in the forward process. In the reverse process, it is dynamically guided according to the physiological boundary characteristics to generate and reconstruct the missing fragments and complete the medical data.

[0107] S6. Calculating a confidence score of the reconstructed missing segment based on the physiological boundary distance;

[0108] S7. Based on the confidence score, dynamically decide whether to connect the reconstructed missing segment to the main data warehouse, push it for manual review, or retain the original missing segment and generate an exception record to achieve continuous optimization management of medical data.

[0109] The present invention forms a unified and coherent processing chain from sliding window detection to controlled diffusion reconstruction, and then to confidence assessment and dynamic access management by setting a detailed inter-module interaction method. By gradually processing missing data and integrating physiological boundary priors, it can effectively solve the problems of rough missing segment detection, single reconstruction features, and insufficient physiological rationality constraints in existing medical data processing. In particular, the adoption of a phased processing strategy not only ensures the high fidelity of the data reconstruction process, but also takes into account physiological rationality and flexibility in data management. The overall method has clear logic and rigorous steps, which can effectively improve the intelligence level and data recovery effect of the medical data management system, while improving the system's adaptability to complex physiological changes, and significantly enhancing the usability and reliability of the medical data management system in clinical applications and health monitoring.

[0110] In this embodiment, S1 specifically includes:

[0111] S11, setting the sliding window length and initializing the window starting point to the first data point in the monitoring sequence of the input medical data;

[0112] S12. In each window, count the number of valid data points and calculate the missing rate. If the missing rate exceeds the preset missing threshold, mark it as an abnormal window.

[0113] S13, merging consecutive abnormal windows into a preliminary missing segment, and recording the starting position and the ending position of the preliminary missing segment;

[0114] S14, expanding forward and backward at the boundary of the preliminary missing segment, detecting a change in the missing rate in a surrounding window, and determining a final boundary when the changing trend continuously decreases to below a restoration threshold;

[0115] S15. Extract observation points before and after the missing segment and their corresponding timestamps based on the final missing segment's front and back positions to generate missing segment boundary information.

[0116] S16. Number and standardize all detected missing segments to generate a complete one-dimensional missing segment mask.

[0117] By setting the specific implementation steps of the sliding window detection algorithm in detail, the present invention can dynamically capture continuous missing segments in the input medical monitoring sequence and accurately locate the boundary points and timestamps of the missing segments. Compared with the traditional detection method based on fixed thresholds, this technical solution uses the statistical missing rate and the trend of the recovery threshold change to achieve more sensitive and accurate missing detection, effectively avoiding the problem of missed detection caused by slight data fluctuations. At the same time, the missing mask is generated through standardized processing, providing a high-quality input foundation for the subsequent wavelet fractal decomposition and differential feature generation, greatly improving the robustness and accuracy of the overall system in dealing with complex missing patterns, and providing reliable protection for the complete recovery of medical data.

[0118] In this embodiment, S2 specifically includes:

[0119] S21, obtaining a complete monitoring sequence of the missing segment and performing preprocessing, including trend item elimination and stabilization processing, to eliminate long-term drift and mutation interference;

[0120] S22. Use the wavelet packet transform method to perform recursive multi-layer decomposition on the complete monitoring sequence after preprocessing, select the optimal wavelet basis function based on the entropy criterion, and obtain the decomposed sub-band signal set:

[0121]

[0122] Where X(t) represents the original input signal, represents the i-th subband signal under the n-th layer decomposition, and n represents the number of layers of wavelet packet decomposition;

[0123] S23. Extract local trend and calculate roughness index for each sub-band signal. The local trend change rate is expressed as The calculation formula of roughness index is:

[0124]

[0125] in, represents the roughness index of the ith subband signal under the nth layer decomposition, τ represents the interval length between the two observation time points, represents the value of the i-th subband signal at time point t+τ under the n-th layer decomposition, Represents the value of the i-th subband signal under the n-th layer decomposition at time point t;

[0126] S24. Based on the long-range correlation of sub-band signals, the Hurst exponent is calculated using the range standard deviation method:

[0127]

[0128] in, represents the Hurst exponent of the ith subband signal under the nth layer decomposition, R(m) represents the range when the subband signal length is m, S(m) represents the standard deviation when the subband signal length is m, and m represents the total length of the subband signal;

[0129] S25, extract the local trend change rate from each sub-band signal Roughness Index and Hurst index Perform feature fusion to form a scale feature vector;

[0130] S26. Stack and combine all scale feature vectors in the scale dimension to generate a fractal embedding tensor.

[0131] The present invention performs wavelet packet transform and self-similar fractal decomposition on the complete monitoring sequence of the missing section, fully extracting key fractal features such as local trends, roughness index and Hurst index at each scale, and effectively capturing the multi-scale dynamic changes and nonlinear evolution characteristics in medical monitoring data. Compared with traditional single-scale modeling or simple feature extraction methods, this technical solution can better reveal the complex structural patterns and changing trends within medical data. By forming a fractal embedding tensor, it provides rich information support for subsequent differential feature fusion and joint feature modeling, significantly improving the accuracy and physiological rationality of missing data reconstruction, and effectively solving the problems of single feature representation and insufficient reconstruction accuracy of existing methods.

[0132] In this embodiment, S3 specifically includes:

[0133] S31. Extract the observed data sequence of the missing segment neighborhood Based on the variable step difference method, the time interval is adaptively adjusted and the dynamic time step set {Δt k}, where Δt k represents the kth dynamic time step;

[0134] S32. Calculate the first-order difference vector based on the dynamic time step:

[0135] D1(t)=ω1(t)·(y(t+Δt1)-y(t))+ω2(t)·(y(t+

[0136] Δt2)-y(t));

[0137] Where D1(t) represents the first-order difference vector at time t, ω1(t) and ω2(t) represent the difference weights, y(t+Δt1) represents the observation value at the new moment after the dynamic time step Δt1, y(t) represents the observation value at the current moment t, and y(t+Δt2) represents the observation value at the new moment after the dynamic time step Δt2;

[0138] S33. Based on the first-order difference vector D1(t), the second-order difference vector D2(t) is calculated using a recursive residual structure:

[0139] D2(t)=D1(t+Δt1)-D1(t)+α·(D1(t+Δt2)-D1(t));

[0140] Where D2(t) represents the second-order difference vector at time t, D1(t+Δt1) represents the first-order difference vector at time t+Δt1, D1(t+Δt2) represents the first-order difference vector at time t+Δt2, and α represents the difference enhancement factor;

[0141] S34, performing statistical feature enhancement on the first-order difference vector D1(t) and the second-order difference vector D2(t), extracting the local mean, local standard deviation, and extreme value change rate as auxiliary features, and constructing an extended differential feature flow;

[0142] S35. Adaptively splicing the extended differential feature flow and the fractal embedding tensor in the channel dimension, and adjusting the splicing order according to feature correlation, to finally generate a joint feature flow.

[0143] The present invention introduces a variable-step differential mechanism and a recursive residual structure in the differential feature generation process, adaptively adjusts the time step and strengthens the local statistical characteristics of the differential features, significantly enhancing the expressive power and dynamic response characteristics of the feature stream. By enhancing the first-order and second-order differential features and adaptively splicing them with fractal features, a joint feature stream with a more compact structure and stronger representational capabilities is formed. This feature stream can more comprehensively reflect the local change trends and global dynamic characteristics of medical data, providing richer and more accurate input information for subsequent reconstruction models, effectively improving the quality and physiological rationality of the reconstruction of missing fragments, and overcoming the problems of thin features and severe reconstruction distortion in the existing technology.

[0144] In this embodiment, the S4 specifically includes:

[0145] S41, sending the joint feature stream to a three-branch encoder, wherein the three-branch encoder includes a context branch, a fractal branch, and a physiological prior branch, and each branch performs feature extraction and processing operations in parallel;

[0146] S42, the context branch encodes the time step position of the joint feature stream, adds a temporal feature description, and inputs the Transformer structure, wherein the Transformer structure includes a multi-head self-attention module and a feedforward neural network module. The multi-head self-attention module captures the long-range dependency in the joint feature stream, and the feedforward neural network module performs nonlinear mapping, and finally outputs the context feature representation;

[0147] S43, the fractal branch sends the joint feature stream to the gated fractal recursive unit. The gated fractal recursive unit introduces a gating mechanism at each recursive layer based on multi-scale recursive expansion to dynamically control the direction of information flow and recursive depth. The fractal pattern evolution relationship in the joint feature stream is extracted through cross-scale information integration, and finally a fractal feature representation is output.

[0148] S44, the physiological prior branch maps the joint feature stream into a physiological indicator node representation, performs relational reasoning based on the graph convolutional network, infers the physiological parameter boundary range of the joint feature stream on heart rate, blood pressure, and respiratory rate, and performs a priori consistency correction on the joint feature stream based on the reasoning result, and finally outputs the physiological prior feature representation;

[0149] S45. Perform feature splicing and fusion on the context feature representation, fractal feature representation, and physiological prior feature representation in the channel dimension to generate a comprehensive latent code.

[0150] By designing a three-branch encoding structure, the present invention extracts multi-level feature information from three perspectives: contextual dependency, fractal evolution, and physiological priors during the processing of the joint feature stream. The context branch uses the Transformer structure to extract long-range dependencies, the fractal branch captures cross-scale fractal dynamic features through a gated fractal recursive unit, and the physiological prior branch infers physiological parameter boundaries based on a graph convolutional network. The three-way features are fused to form a comprehensive latent code, which takes into account dynamic change features, local detail features, and physiological rationality constraints, effectively improving the modeling capability of the subsequent controlled diffusion generation module and significantly enhancing the accuracy and physiological compliance of missing data reconstruction.

[0151] In this embodiment, the S5 specifically includes:

[0152] S51, inputting the comprehensive latent code as a conditional vector into a controlled time diffusion model, wherein the controlled time diffusion model includes a forward diffusion process and a reverse generation process, and the physiological prior boundary information is synchronously loaded when the controlled time diffusion model is initialized;

[0153] S52. During the forward diffusion process, for each time step, the integrated latent code is superimposed with piecewise weighted noise, and the noise injection is dynamically set with weighting coefficients according to feature importance, so as to gradually perturb the integrated latent code into a standard normal distribution.

[0154] S53, at each forward diffusion time step, recording the disturbance trajectory and characteristic deviation degree during the noise injection process to form a disturbance scheduling table;

[0155] S54. During the reverse generation process, according to the perturbation schedule and the preliminary estimate of the integrated latent code, a denoising subnetwork is applied at each reverse time step to gradually remove noise, and physiological boundary features are dynamically introduced as guiding conditions to perform physiological rationality correction on the generation results at each step.

[0156] S55. During the reverse generation process, dynamically adjust the guidance intensity according to the difference between the current generation state and the physiological prior. When the generation result deviates from the physiological prior boundary, the guidance is enhanced; when the generation result approaches the neighborhood set by the physiological prior boundary, the guidance is reduced.

[0157] S56. After the reverse generation process is completed, the reconstructed missing fragments are output and added to the medical data sequence, completing the recovery of the missing fragments and the completeness supplement of the medical data.

[0158] The controlled time diffusion generation module proposed in this paper achieves high-fidelity restoration of missing fragments by injecting weighted noise perturbations during the forward diffusion process and dynamically guiding the generation trajectory incorporating physiological boundary characteristics during the reverse generation process. Unlike traditional unconstrained diffusion methods, this technical solution can flexibly control the degree of deviation in the generated results during the restoration process, ensuring that the generated fragments maintain realistic dynamic characteristics while strictly complying with physiological boundary constraints. This module significantly improves the quality and reliability of missing medical data repair, providing a solid data foundation for subsequent medical analysis and intelligent decision-making.

[0159] In this embodiment, S6 specifically includes:

[0160] After the reconstructed missing segments are generated, the confidence score is calculated based on the physiological boundary distance:

[0161]

[0162] Among them, S conf represents the confidence score of reconstructing the missing segment, β represents the adjustment coefficient, N represents the total number of physiological parameters, and x p represents the value of the pth physiological parameter in the reconstructed missing segment, u p represents the upper limit of the pth physiological parameter, l prepresents the lower bound of the p-th physiological parameter, and max(·) represents the maximum value function.

[0163] By constructing a confidence score calculation mechanism based on physiological boundary distance, the present invention achieves a quantitative assessment of the credibility of reconstructed missing fragments. Compared to traditional methods that rely on mean square error or manual experience, the physiological boundary distance method can more directly reflect the consistency of reconstructed data with the physiologically reasonable range, providing a scientific and accurate basis for data access decisions. This mechanism significantly improves the system's ability to manage the quality of reconstructed data, can dynamically adapt to the data credibility requirements in different scenarios, and enhances the flexibility and reliability of medical data management systems in practical applications.

[0164] In this embodiment, the S7 specifically includes:

[0165] S71, confidence score S based on the reconstruction of the missing fragment conf Perform classification judgment and set the high-level threshold θ high and the low-level threshold θ low ;

[0166] S72, if S conf ≥θ high , then the reconstructed missing segment is determined to be a high-confidence segment and is directly archived into the main data warehouse as a formal medical data record;

[0167] S73, if θ low ≤S conf <θ high , the reconstructed missing segment is determined to be a medium confidence segment, a manual review task is generated, and the reconstructed missing segment and the corresponding confidence index are pushed to the data administrator review system;

[0168] S74. Generate an explanatory report during the push process to assist manual decision-making;

[0169] S75, if S conf <θ low , the reconstructed missing segment is determined to be a low-confidence segment, the reconstructed missing segment is not archived, the original missing segment is retained, and the abnormal data processing log is recorded.

[0170] The dynamic access management module designed in this invention achieves refined and intelligent medical data management by setting high, medium, and low confidence levels and dynamically determining the processing strategy for reconstructed fragments based on the confidence scores. Based on the fragment's credibility, the system automatically decides whether to archive, push for manual review, or retain abnormal records. This avoids the one-size-fits-all data processing approach used in existing methods and significantly improves data management efficiency and security. Furthermore, through differentiated processing strategies, it optimizes data storage structure and subsequent analysis accuracy, providing a higher-quality data foundation for medical decision-making systems.

[0171] In this embodiment, the S74 specifically includes: based on the feature analysis results of the reconstructed missing fragments, extracting and recording local trend changes, roughness index changes and Hurst index changes; based on the physiological boundary matching analysis results, recording the degree of consistency deviation between the reconstructed data and the upper and lower boundaries of each physiological parameter; based on the statistical difference analysis, calculating the relative difference percentage between the reconstructed fragments and the historical real data in each statistical indicator, and dynamically generating an explanatory report containing charts, text descriptions and confidence score explanations.

[0172] During manual review, this invention provides data managers with intuitive and rich decision-making support by generating explanatory reports that include fractal feature changes, physiological compatibility analysis, and statistical difference analysis. Unlike traditional review processes that rely solely on simple numerical comparisons, this explanatory report comprehensively evaluates and reconstructs data from multiple dimensions, including structural changes, physiological rationality, and statistical stability, significantly improving the efficiency and accuracy of manual review. By enhancing review transparency and information support, the system's data quality control level and intelligent management capabilities are further enhanced.

[0173] Example 1:

[0174] In order to verify the feasibility of the present invention in implementation, the present invention was applied to the remote dynamic ECG monitoring data management project of the cardiology department of a large tertiary hospital. The hospital collected a total of about 180,000 24-hour dynamic ECG monitoring records in 2024, with a data volume of more than 500TB. In actual application, there is a common problem of continuous loss of ECG data due to displacement of wearable devices, detachment of sensor electrodes, loss of wireless signals, etc. Preliminary statistics show that about 11.3% of the records have obvious missing fragments, of which the data with a missing length of more than 30 minutes accounts for 2.7%. The traditional processing method based on linear interpolation and mean filling has obvious distortion when repairing data that is continuously missing for more than 10 minutes, and is prone to unreasonable heart rate fluctuations and false signal jumps, which not only affects the doctor's judgment, but also reduces the accuracy of subsequent intelligent screening for arrhythmias.

[0175] In this embodiment, the hospital introduced the artificial intelligence-based medical data management system proposed in this invention to systematically address the issue of missing Holter monitoring data. First, during the data access phase, the system uses a sliding window detection module to analyze the input ECG monitoring sequence in real time, dynamically capturing consecutive missing segments and recording the start and end timestamps of the missing segments and information about surrounding valid observation points. Compared to the static threshold detection system of the original system, the sliding detection of the present invention can accurately capture local missing segments when the missing rate is below 15%, reducing the missed detection rate by 21.4%.

[0176] Subsequently, wavelet packet transform and self-similar fractal decomposition are performed on the complete sequence of detected missing segments to extract local trends, roughness indices, and Hurst exponents at each scale, and construct a fractal embedding tensor. Simultaneously, first-order and second-order difference vectors are adaptively calculated based on the data in the missing neighborhood and concatenated with the fractal features in the channel dimension to generate a joint feature stream. This feature construction strategy enables the system to capture subtle dynamic features of ECG signals at different scales and time domains, significantly improving the richness of feature representation.

[0177] During the feature extraction phase, the joint feature stream is fed into a three-branch encoder. The context branch models global temporal dependencies, the fractal branch extracts cross-scale dynamic evolution features, and the physiological prior branch infers reasonable intervals for key physiological parameters such as heart rate, PR interval, and QT interval. The three-branch feature fusion forms a comprehensive latent code, providing deep feature support for intelligent reconstruction of missing segments.

[0178] During the data reconstruction process, the integrated latent code is input as a conditional vector into the controlled time diffusion generation module. During the forward diffusion process, the system dynamically injects weighted noise based on feature importance. During the reverse generation process, physiological boundary guidance is introduced in real time to flexibly adjust the reconstruction trajectory, ultimately generating high-fidelity reconstructed segments that conform to physiological plausibility. By introducing physiological prior dynamic guidance, the system avoids artifacts such as sudden heart rate changes and rhythm abnormalities during the reconstruction process. The reconstructed mean heart rate is consistent with the normal physiological range. There are no statistically significant differences between the reconstructed segments and the real samples in terms of PR interval, QRS interval, and QT interval (p>0.05).

[0179] After reconstruction is complete, the system uses a confidence assessment module to calculate a confidence score for each segment based on physiological boundary consistency. High-confidence reconstructed segments are automatically added to the master data warehouse, medium-confidence segments are sent for manual review, and low-confidence segments are marked as abnormal. During the manual review process, the system automatically generates an explanatory report that analyzes the fractal feature changes, physiological boundary matching, and statistical differences of the reconstructed segments, significantly reducing the review burden on data administrators.

[0180] In actual testing, the system processed 18,200 missing Holter monitoring records, of which 71.2% were high-confidence segments, 23.5% were medium-confidence segments, and 5.3% were low-confidence segments. After review, it was confirmed that the reconstruction accuracy of high-confidence segments reached 98.7%, and the review pass rate for medium-confidence segments reached 84.5%. Compared with traditional interpolation methods, the proposed system reduced the mean square error of reconstruction for data segments with a length of more than 10 minutes by 35.8%, and improved the physiological plausibility of the PR interval by 28.4%.

[0181] This embodiment verifies that the artificial intelligence-based medical data management system of the present invention has significant advantages in continuous medical monitoring data missing detection, feature extraction, intelligent reconstruction, quality assessment and dynamic access management, and can effectively improve the integrity, reliability and availability of medical big data, providing a high-quality data foundation for intelligent medical analysis and auxiliary decision-making.

[0182] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A medical data management system based on artificial intelligence, characterized in that: Includes the following modules: Sliding window detection module, which is used to analyze the monitoring sequence of input medical data, capture continuous missing segments and record boundary points and timestamps; Wavelet fractal decomposition module, used to perform wavelet packet transform and self-similar fractal decomposition on the missing segments, extract local trends, roughness index and Hurst index, and generate fractal embedding tensors; The differential feature generation module is used to calculate the first-order differential vector and the second-order differential vector, and concatenate them with the fractal embedding tensor in the channel dimension to generate a joint feature flow; The three-branch encoding module is used to input the joint feature stream into the context branch, fractal branch and physiological prior branch respectively, and fuse them to generate a comprehensive latent code; A controlled diffusion generation module is used to take the comprehensive latent code as a conditional input, inject weighted noise in the forward direction, and dynamically guide the reverse direction by combining physiological boundary features to generate and reconstruct the missing fragments; A confidence evaluation module is used to calculate a confidence score based on the distance between the reconstructed missing segment and the physiological boundary; Dynamic access management module, used to dynamically decide whether to reconstruct missing segments and access them to the main data warehouse, push them for manual review, or retain the original missing segments based on the confidence score; The explanatory report module is used to generate explanatory reports to assist manual review.

2. The medical data management system based on artificial intelligence according to claim 1, characterized in that: The modules are implemented as follows: S1. Analyze the monitoring sequence of input medical data based on the sliding window detection algorithm, capture the continuous missing segments in real time, and record the observation boundary points and timestamps before and after the missing segments; S2. Performing wavelet packet transform and self-similar fractal decomposition on the complete monitoring sequence of the missing section, extracting the local trend, roughness index and Hurst index at each scale, and generating a fractal embedding tensor; S3. Based on the observation data of the missing segment neighborhood, calculate the first-order difference vector and the second-order difference vector, and splice them with the fractal embedding tensor in the channel dimension to generate a joint feature flow; S4. Input the joint feature stream into a three-branch encoder, wherein the three-branch encoder includes a context branch, a fractal branch, and a physiological prior branch, and fuse the outputs of the three branches to generate a comprehensive latent code; S5. The comprehensive latent code is used as a conditional vector and input into the controlled time diffusion model. Weighted noise is injected in the forward process. In the reverse process, it is dynamically guided according to the physiological boundary characteristics to generate and reconstruct the missing fragments and complete the medical data. S6. Calculating a confidence score of the reconstructed missing segment based on the physiological boundary distance; S7. Based on the confidence score, dynamically decide whether to connect the reconstructed missing segment to the main data warehouse, push it for manual review, or retain the original missing segment and generate an exception record to achieve continuous optimization management of medical data.

3. The medical data management system based on artificial intelligence according to claim 2, characterized in that: Said S1 specifically includes: S11, setting the sliding window length and initializing the window starting point to the first data point in the monitoring sequence of the input medical data; S12. In each window, count the number of valid data points and calculate the missing rate. If the missing rate exceeds the preset missing threshold, mark it as an abnormal window. S13, merging consecutive abnormal windows into a preliminary missing segment, and recording the starting position and the ending position of the preliminary missing segment; S14, expanding forward and backward at the boundary of the preliminary missing segment, detecting a change in the missing rate in a surrounding window, and determining a final boundary when the changing trend continuously decreases to below a restoration threshold; S15. Extract observation points before and after the missing segment and their corresponding timestamps based on the final missing segment's front and back positions to generate missing segment boundary information. S16. Number and standardize all detected missing segments to generate a complete one-dimensional missing segment mask.

4. The artificial intelligence-based medical data management system according to claim 2, characterized in that: The S2 specifically includes: S21, obtaining a complete monitoring sequence of the missing segment and performing preprocessing, including trend item elimination and stabilization processing, to eliminate long-term drift and mutation interference; S22. Use the wavelet packet transform method to perform recursive multi-layer decomposition on the complete monitoring sequence after preprocessing, select the optimal wavelet basis function based on the entropy criterion, and obtain the decomposed sub-band signal set: Where X(t) represents the original input signal, represents the i-th subband signal under the n-th layer decomposition, and n represents the number of layers of wavelet packet decomposition; S23. Extract local trend and calculate roughness index for each sub-band signal. The local trend change rate is expressed as The calculation formula of roughness index is: in, represents the roughness index of the ith subband signal under the nth layer decomposition, τ represents the interval length between the two observation time points, represents the value of the i-th subband signal under the n-th layer decomposition at time point t+τ, Represents the value of the i-th subband signal under the n-th layer decomposition at time point t; S24. Based on the long-range correlation of sub-band signals, the Hurst exponent is calculated using the range standard deviation method: in, represents the Hurst exponent of the ith subband signal under the nth layer decomposition, R(m) represents the range when the subband signal length is m, S(m) represents the standard deviation when the subband signal length is m, and m represents the total length of the subband signal; S25, extract the local trend change rate from each sub-band signal Roughness Index and Hurst index Perform feature fusion to form a scale feature vector; S26. Stack and combine all scale feature vectors in the scale dimension to generate a fractal embedding tensor.

5. The artificial intelligence-based medical data management system according to claim 2, characterized in that: The S3 specifically includes: S31. Extract the observed data sequence of the missing segment neighborhood Based on the variable step difference method, the time interval is adaptively adjusted and the dynamic time step set {Δt k }, where Δt k represents the kth dynamic time step; S32. Calculate the first-order difference vector based on the dynamic time step: D1(t)=ω1(t)·(y(t+Δt1)-y(t))+ω2(t)·(y(t+ Δt2)-y(t)); Where D1(t) represents the first-order difference vector at time t, ω1(t) and ω2(t) represent the difference weights, y(t+Δt1) represents the observation value at the new moment after the dynamic time step Δt1, y(t) represents the observation value at the current moment t, and y(t+Δt2) represents the observation value at the new moment after the dynamic time step Δt2; S33. Based on the first-order difference vector D1(t), a recursive residual structure is used to calculate the second-order difference vector D2(t): D2(t)=D1(t+Δt1)-D1(t)+α·(D1(t+Δt2)-D1(t)); Where D2(t) represents the second-order difference vector at time t, D1(t+Δt1) represents the first-order difference vector at time t+Δt1, D1(t+Δt2) represents the first-order difference vector at time t+Δt2, and α represents the difference enhancement factor; S34, performing statistical feature enhancement on the first-order difference vector D1(t) and the second-order difference vector D2(t), extracting the local mean, local standard deviation, and extreme value change rate as auxiliary features, and constructing an extended differential feature flow; S35. Adaptively splicing the extended differential feature flow and the fractal embedding tensor in the channel dimension, and adjusting the splicing order according to feature correlation, to finally generate a joint feature flow.

6. The artificial intelligence-based medical data management system according to claim 2, characterized in that: The S4 specifically includes: S41, sending the joint feature stream to a three-branch encoder, wherein the three-branch encoder includes a context branch, a fractal branch, and a physiological prior branch, and each branch performs feature extraction and processing operations in parallel; S42, the context branch encodes the time step position of the joint feature stream, adds a temporal feature description, and inputs the Transformer structure, wherein the Transformer structure includes a multi-head self-attention module and a feedforward neural network module. The multi-head self-attention module captures the long-range dependency in the joint feature stream, and the feedforward neural network module performs nonlinear mapping, and finally outputs the context feature representation; S43, the fractal branch sends the joint feature stream to the gated fractal recursive unit. The gated fractal recursive unit introduces a gating mechanism at each recursive layer based on multi-scale recursive expansion to dynamically control the direction of information flow and recursive depth. The fractal pattern evolution relationship in the joint feature stream is extracted through cross-scale information integration, and finally a fractal feature representation is output. S44, the physiological prior branch maps the joint feature stream into a physiological indicator node representation, performs relational reasoning based on the graph convolutional network, infers the physiological parameter boundary range of the joint feature stream on heart rate, blood pressure, and respiratory rate, and performs a priori consistency correction on the joint feature stream based on the reasoning result, and finally outputs the physiological prior feature representation; S45. Perform feature splicing and fusion on the context feature representation, fractal feature representation, and physiological prior feature representation in the channel dimension to generate a comprehensive latent code.

7. The artificial intelligence-based medical data management system according to claim 2, characterized in that: The S5 specifically includes: S51, inputting the comprehensive latent code as a conditional vector into a controlled time diffusion model, wherein the controlled time diffusion model includes a forward diffusion process and a reverse generation process, and the physiological prior boundary information is synchronously loaded when the controlled time diffusion model is initialized; S52. During the forward diffusion process, for each time step, the integrated latent code is superimposed with piecewise weighted noise, and the noise injection is dynamically set with weighting coefficients according to feature importance, so as to gradually perturb the integrated latent code into a standard normal distribution. S53, at each forward diffusion time step, recording the disturbance trajectory and characteristic deviation degree during the noise injection process to form a disturbance scheduling table; S54. During the reverse generation process, according to the perturbation schedule and the preliminary estimate of the integrated latent code, a denoising subnetwork is applied at each reverse time step to gradually remove noise, and physiological boundary features are dynamically introduced as guiding conditions to perform physiological rationality correction on the generation results at each step. S55. During the reverse generation process, dynamically adjust the guidance intensity according to the difference between the current generation state and the physiological prior. When the generation result deviates from the physiological prior boundary, the guidance is enhanced; when the generation result approaches the neighborhood set by the physiological prior boundary, the guidance is reduced. S56. After the reverse generation process is completed, the reconstructed missing fragments are output and added to the medical data sequence, completing the recovery of the missing fragments and the completeness supplement of the medical data.

8. The artificial intelligence-based medical data management system according to claim 2, characterized in that: The S6 specifically includes: After the reconstructed missing segments are generated, the confidence score is calculated based on the physiological boundary distance: Among them, S conf represents the confidence score of reconstructing the missing segment, β represents the adjustment coefficient, N represents the total number of physiological parameters, and x p represents the value of the pth physiological parameter in the reconstructed missing segment, u p represents the upper limit of the pth physiological parameter, l p represents the lower bound of the p-th physiological parameter, and max(·) represents the maximum value function.

9. The artificial intelligence-based medical data management system according to claim 2, characterized in that: The S7 specifically includes: S71, confidence score S based on the reconstruction of the missing fragment conf Perform classification judgment and set the high-level threshold θ high and the low-level threshold θ low ; S72, if S conf ≥θ high , then the reconstructed missing segment is determined to be a high-confidence segment and is directly archived into the main data warehouse as a formal medical data record; S73, if θ low ≤S conf <θ high , the reconstructed missing segment is determined to be a medium confidence segment, a manual review task is generated, and the reconstructed missing segment and the corresponding confidence index are pushed to the data administrator review system; S74. Generate an explanatory report during the push process to assist manual decision-making; S75, if S conf <θ low , the reconstructed missing segment is determined to be a low-confidence segment, the reconstructed missing segment is not archived, the original missing segment is retained, and the abnormal data processing log is recorded.

10. The artificial intelligence-based medical data management system according to claim 9, characterized in that: The S74 specifically includes: based on the feature analysis results of the reconstructed missing fragments, extracting and recording local trend changes, roughness index changes and Hurst index changes; based on the physiological boundary matching analysis results, recording the degree of consistency deviation between the reconstructed data and the upper and lower boundaries of each physiological parameter; based on the statistical difference analysis, calculating the relative difference percentage between the reconstructed fragments and the historical real data in various statistical indicators, and dynamically generating an explanatory report containing charts, text descriptions and confidence score explanations.

Images