A personalized risk prediction and data encryption method based on health big data

Abstract

The application discloses a kind of individualized risk prediction and data encryption method based on health big data, it is related to health big data processing and data security technical field, including acquisition health monitoring scene under the multi-source health big data, and according to the privacy sensitive grade of each type of data corresponding initial different encryption period is generated, and initial encryption data sample set is integrated to form;Its technical key points are: the residual sequence originally only used for error statistics in risk prediction process is directly introduced into encryption link, is transformed into chaotic key seed by multiple fractal feature extraction, to drive the generation of dynamic encryption key;Can let the fluctuation law of key and individual health risk depth binding, without additional deployment independent random source module, encryption strength can be dynamically adapted according to risk level, when individual health risk increases, the protection level of high-sensitive data such as diagnosis and treatment record, genetic information will be naturally strengthened, fill the problem of insufficient anti-cracking ability of traditional general key.

Description

Technical Field

[0001] This invention relates to the field of health big data processing and data security technology, specifically a personalized risk prediction and data encryption method based on health big data. Background Technology

[0002] In the current engineering implementation of health big data platforms, a common approach is to separate risk prediction and data encryption deployments. During actual operation, the process begins with the collection and preprocessing of multi-source data: data is collected from four sources—wearable device terminals, medical institution information systems, daily behavior monitoring platforms, and gene testing databases—collecting physiological signs, medical history, lifestyle behaviors, and genetic susceptibility. After cleaning and standardization, the data is stored separately. Next, risk prediction is performed: the preprocessed data is input into a pre-trained disease risk assessment model. Using a fixed historical baseline set during system deployment as a reference, an individual disease risk prediction value is calculated, and a risk assessment report is output. Finally, data encryption is executed: an independent security module uses a universal random number generation algorithm to generate encryption keys, and the entire health data is encrypted for storage and transmission according to a preset uniform encryption strength. Under this implementation process, the residual data generated in the risk prediction stage is only used for post-event error statistics and model acceptance, and will not enter the encryption stage to participate in key generation; the key entropy source of the encryption module relies entirely on a general random function, which is not related to the fluctuation characteristics of individual health data, and the key generation pattern is singular. For highly sensitive health data such as medical records and genetic information, there are obvious problems with anti-cracking capabilities. Meanwhile, the operating parameters of the entire system are all static values ​​preset at the time of deployment: the historical baseline of the risk assessment model will not be automatically adjusted as the system runs. As the platform continues to run, the accuracy drift of the front-end data acquisition device, changes in the characteristics of the user group, and the natural decay of the model parameters will gradually cause the baseline to shift, and the prediction error will continue to accumulate. The relevant parameters of the encryption process cannot be iteratively optimized by reverse using the verification information generated during operation. The overall performance of the system will not improve naturally with the running time and requires regular investment of manpower for manual calibration and maintenance. To this end, we disclose a personalized risk prediction and data encryption method based on health big data. Summary of the Invention

[0003] To achieve the above objectives, the present invention provides the following technical solution: A personalized risk prediction and data encryption method based on health big data, the method includes the following steps: Collect multi-source health big data covering four dimensions in health monitoring scenarios: physiological signs data, lifestyle behavior data, medical history data, and genetic susceptibility data. After multiple batches of data encryption, integrate them to form an initial encrypted data sample set. A multi-dimensional security assessment is conducted on the initial encrypted data sample set to obtain the initial encryption quality assessment value, which is then used as the data credibility weight in the weighted calculation of health features to output the initial personalized health risk prediction value. Within a preset sliding time window, the point-by-point difference between the initial personalized health risk prediction value and the historical baseline is calculated to generate a risk prediction residual sequence. Then, the risk prediction residual sequence is subjected to multifractal feature extraction to generate a fractal dimension sequence and a local extremum density sequence. After normalization and splicing, they are mapped to generate a chaotic key seed. The chaotic key seed is injected into the basic encryption key expansion function to generate a dynamic encryption key, and the corresponding batch of health data is re-encrypted in blocks; The integrity verification residuals of each data block after re-encryption are extracted, the distribution characteristics of the verification residuals are statistically analyzed, the fractal feature fidelity is inferred, and fractal parameter correction coefficients are generated to adjust the parameters for the next round of multifractal feature extraction. At the same time, the baseline drift correction coefficient is calculated based on the distribution of verification residuals to dynamically adjust the historical baseline of health risk prediction. After correction, the corrected personalized health risk prediction value is obtained.

[0004] Furthermore, the specific steps for collecting multi-source health big data are as follows: Multi-source health big data is collected synchronously through four types of access nodes: wearable device interface, medical information system interface, daily behavior log interface, and gene testing database interface. Extract the inherent periodic fluctuation characteristics of physiological signs data rate and respiratory rate, and add implicit time-series synchronization markers to all categories of data with the peak point of fluctuation as a unified time reference; The data were classified and determined based on four categories: physiological indicator threshold characteristics, behavioral event record characteristics, medical record identification characteristics, and gene locus sequence characteristics.

[0005] Furthermore, after data collection, an initial differentiated encryption period is generated based on the privacy sensitivity level corresponding to various types of data. The steps for generating the initial differentiated encryption period include: The encryption baseline period is determined for each of the four types of data, namely, physiological signs data combined with data update frequency and privacy sensitivity level, lifestyle behavior data combined with user behavior activity level, medical history data combined with business dependence on medical decisions, and genetic susceptibility data combined with data leakage risk sensitivity. Based on the average fluctuation frequency of the residual sequence of the previous risk prediction, the four types of encryption benchmark periods are dynamically corrected in the same direction. The higher the fluctuation frequency, the shorter the encryption period, thus obtaining the initial differentiated encryption period.

[0006] Furthermore, the steps for forming the initial encrypted data sample set specifically include: Based on the initial differentiated encryption cycle, an encryption task corresponding to each type of data is generated, and the corresponding encryption configuration parameters are matched with the original traceable data source. The encryption configuration parameters include encryption granularity, key length, and redundant padding bit length. An encryption sequence containing execution order, block rules, and key update interval is generated and encrypted in batches. During the encryption process, the first-order fractal dimension of the risk prediction residual of the current batch is extracted, converted into a fixed-length binary code, and embedded as a data block redundant padding bit as an implicit batch identifier. The timelines of the four types of encrypted data are aligned and integrated based on implicit time synchronization markers.

[0007] Furthermore, the calculation process for the initial encryption quality assessment value is as follows: Extract the data structure format and encrypted data content from the initial encrypted data sample set, calculate the encryption fit of three types of data (structured data, semi-structured data, and unstructured data) respectively, and integrate them to obtain the encryption continuity evaluation value. The content encryption effect is evaluated from six dimensions: key strength level, anti-cracking capability level, data integrity level, access controllability level, transmission security level, and storage reliability level, resulting in a basic content encryption score. The implicit batch identifier is decoded and compared with the actual residual fractal characteristics to obtain the risk residual matching degree. The consistency between the verification residual distribution and the preset fractal characteristics is calculated to obtain the verification residual fidelity. The initial encryption quality assessment value is obtained by allocating the weights of each indicator using the analytic hierarchy process (AHP) and then summing the weights, combined with the encryption continuity assessment value.

[0008] Furthermore, the output process of the initial personalized health risk prediction value is as follows: The weighting of four types of health data in the risk assessment of the target disease is pre-set. Combining the actual coverage and time span of each type of data, the initial encryption quality assessment value is used as the credibility correction coefficient to calculate the health feature contribution value of each type of data. After accumulating the total risk correction base, input it into the preset disease risk assessment model and output the initial personalized health risk prediction value; dynamically adjust the historical baseline of health risk prediction, wherein the historical baseline is adjusted frame by frame by superimposing the baseline drift correction coefficient onto the baseline parameters of the risk assessment model.

[0009] Furthermore, the process for generating the risk prediction residual sequence is as follows: Set a fixed-duration sliding time window, extract the initial personalized health risk prediction values ​​within the window according to the preset step size to form a time series risk value sequence, match the historical baseline values ​​of the corresponding window to form a time series baseline sequence, and perform point-by-point subtraction and adaptive denoising and smoothing to obtain a standardized risk prediction residual sequence. The process of generating the chaotic key seed is as follows: The multifractal spectrum of the residual sequence is calculated using the sliding box dimension method to obtain the pointwise fractal dimension sequence. At the same time, the number of local extreme points per unit time is counted to obtain the local extreme value density sequence. The two types of sequences are normalized and then spliced ​​together pointwise over time to form a combined feature sequence, which is then mapped to a preset numerical range to generate a chaotic key seed.

[0010] Furthermore, the process for generating the dynamic encryption key is as follows: The chaotic key seed is used as a perturbation parameter input to the key expansion function of the basic encryption key, so that the number of iterations of the expansion function is positively correlated with the value of the chaotic key seed. After multiple rounds of iteration to spread the randomness of the key, the dynamic encryption key corresponding to the time window is output. The steps for block re-encryption are as follows: Encryption operations are performed sequentially on each data block. The fractal dimension code corresponding to the current moment is written into the redundant padding bits of each data block to form an implicit verification identifier for fractal features.

[0011] Furthermore, the generation process of the fractal parameter correction coefficient and the baseline drift correction coefficient is as follows: Integrity verification is performed on each re-encrypted data block to obtain the corresponding verification residual value. The mean, variance, skewness and fractal dimension distribution characteristics of all residuals within the statistical window are then analyzed. The fractal feature implicit verification identifier in the decoded data block is compared with the actual fractal dimension of the verification residual to obtain the fractal feature fidelity. Based on this, fractal parameter correction coefficients are generated to dynamically adjust the window size of the sliding box dimension during the next round of multifractal feature extraction. The baseline drift correction coefficient is obtained by weighting and summing the residual mean, variance, and skewness according to preset weights. This coefficient is used to offset the baseline deviation caused by data acquisition drift and model parameter aging.

[0012] Furthermore, based on the obtained corrected personalized health risk prediction value, its corresponding risk level range is determined. The processing steps after determining the risk level range are as follows: For high-risk areas, identify the corresponding high-sensitivity data categories and data volumes, match the corresponding level of encryption granularity and key length, and use dynamic encryption keys to perform secondary encryption on high-sensitivity data such as medical history and genetic susceptibility. During secondary encryption, embed the current risk prediction residual code as an integrity check code, and simultaneously output a health intervention plan that includes risk factor analysis and graded adjustment suggestions. For low-risk ranges, the computational resource utilization rate of the current encryption strategy is calculated. If it is lower than the preset threshold, the current key and period remain unchanged. If it is higher than the preset threshold, the encryption granularity is reduced according to the fluctuation range of the risk prediction residual sequence. The residual feature lightweight encryption mode is adopted to balance security and resource consumption.

[0013] This invention provides a personalized risk prediction and data encryption method based on health big data, which has the following beneficial effects: The residual sequence, originally used only for error statistics in the risk prediction process, is directly introduced into the encryption link and transformed into a chaotic key seed through multifractal feature extraction, thereby driving the generation of dynamic encryption keys. This approach deeply binds the fluctuation pattern of the key with the inherent changing characteristics of individual health risks, eliminating the need for an additional independent random source module. The encryption strength can dynamically adapt to the risk level. When an individual's health risk increases, the protection level of highly sensitive data such as medical records and genetic information will naturally be strengthened, filling the gap in the anti-cracking capability of traditional general-purpose keys. Secondly, the verification residual generated after encryption is simultaneously fed back to both the risk prediction baseline and the fractal extraction parameters, performing baseline drift correction and fractal accuracy optimization during continuous operation. Furthermore, the accuracy of risk prediction will gradually offset the errors caused by data acquisition device drift and model parameter decay over time, and the adaptability of encryption features will also be dynamically adjusted according to data characteristics. The performance of the entire system will gradually improve iteratively as it runs, eliminating the need for regular manual calibration and maintenance. Meanwhile, in terms of data verification design, the redundant padding bits of the data block are directly used to carry the fractal verification identifier during the encryption execution phase. There is no need to generate an additional independent verification file, nor does it occupy the transmission bandwidth and storage space of the effective data. This enables the self-verification capability of encrypted data. When data anomalies occur, the corresponding batch can be quickly located through the identifier to complete the anomaly tracing. In addition, the entire mechanism dynamically matches the encryption granularity based on the risk level. In high-risk scenarios, resources are concentrated to strengthen the protection of core sensitive data, while in low-risk scenarios, a lightweight encryption mode is adapted. Under the premise of ensuring data security, computing resources are reasonably allocated, and it can be smoothly adapted to various application scenarios of different scales, such as community health monitoring, chronic disease management, and public health prevention and control. Attached Figure Description

[0014] Figure 1 This is an overall flowchart of the present invention; Figure 2 This is a diagram of the dynamic security encryption architecture based on fractal chaos in this invention; Figure 3 This is a dynamic trend chart of the 24-hour double correction coefficient in this invention; Figure 4 This is a risk level distribution map for different age groups in this invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0016] Based on the urban community chronic disease health management platform, any similar data management platform will suffice. This platform primarily targets high-risk groups for cardiovascular disease, conducting health data management and dynamic risk assessment. It connects to four data sources: the community health service center's diagnostic and treatment information system, residents' smart wearable devices, daily health behavior reporting mini-programs, and third-party gene testing institution databases. This covers the health data management needs of contracted residents within the jurisdiction. Figure 1 and Figure 2 As shown.

[0017] Traditional health big data platforms generally adopt an existing technical architecture that separates risk prediction and data encryption. That is, the risk prediction stage only calculates the risk based on the numerical value of the data itself, while the data encryption stage uses a general random number or timestamp to generate keys and performs uniform strength encryption. The two systems run independently. Under this architecture, the risk prediction residual is only used for error statistics and cannot be reused. In addition, the key entropy source is single and the anti-cracking ability is weak. Furthermore, the risk model baseline is fixed and will continue to drift as it runs, resulting in unreasonable resource allocation. To address this issue, we utilize risk prediction residuals to enable cross-domain encryption and verification residuals to perform inverse two-dimensional correction as our core approach. Through the reconstruction of data processing logic, we simultaneously achieve three core enhancements: first, deep binding of keys to individual health characteristics improves anti-cracking capabilities; second, continuous self-correction of risk prediction accuracy as the system runs, reducing prediction errors; and third, dynamic adaptation of encryption strength based on risk levels, reducing computational resource consumption in low-risk scenarios. Furthermore, the risk prediction residuals generated in the previous run are used to generate the initial encryption cycle, initial encryption implicit batch identifier, and chaotic encryption key for this round; while the verification residuals generated in this re-encryption phase are used to generate fractal parameter correction coefficients and baseline drift correction coefficients, which are applied to the fractal extraction process and risk prediction baseline in the next round. This iterative process achieves continuous performance iteration. A complete cold start process is also configured, ensuring normal operation even on the first run without historical data. Figure 3 As shown, the specific content is as follows: Conventional methods generally suffer from the problem of missing cold start. When there is no historical data for the first run, directly entering the closed-loop logic will lead to abnormal parameters. However, we have configured a complete cold start strategy. The first run uses preset parameters to complete the entire process. Once the switching conditions are met, it automatically switches to the normal closed-loop operation mode. The preset parameters for the cold start phase are all determined based on industry standards and large sample statistical values. In addition, they can be freely adjusted as needed. The initial encryption baseline period is preset according to the default attributes of four types of data, namely 5 minutes, 30 minutes, 2 hours and 24 hours. The implicit batch identifier for initial encryption adopts the system's preset 8-bit fixed code. The initial risk prediction baseline adopts the statistical values ​​of a large sample population in accordance with industry standards. The first round of encryption quality assessment does not introduce the residual matching degree index and only uses the conventional six dimensions for assessment. The first round of multifractal extraction uses the default box size parameters (box size range 2-16, step size 2).

[0018] The switching trigger condition is set so that when the system generates a complete 24-hour sliding window risk prediction residual sequence for the first time, and generates a complete re-encryption verification residual and double correction coefficient, it automatically switches from cold start mode to normal closed-loop mode.

[0019] I. Collection of Multi-Source Health Big Data: The target residents' full-dimensional health data are synchronously retrieved through four types of standardized access nodes. Each node uses the HTTPS encrypted transmission protocol to ensure the security of the data collection link. This part belongs to the existing mature technology and its function is to provide standardized multi-source raw data input. Among them, (1) the wearable device interface synchronizes the data of devices such as smart bracelets and smart blood pressure monitors through the Bluetooth Low Energy protocol, and collects minute-level physiological signs such as heart rate, systolic blood pressure, diastolic blood pressure, blood oxygen saturation, and respiratory rate in real time. The data is accompanied by the device's unique identifier and collection timestamp; (2) the diagnosis and treatment information system interface is connected through the standardized interface of the hospital information system to synchronize outpatient visits. (2) Records, inpatient medical records, medication prescriptions, test results and other medical history data. The data update trigger node is within 2 hours after the resident's visit. Each data is associated with a unique medical record number. (3) Daily behavior log interface collects daily steps, sleep duration, diet structure, exercise type and other life behavior data through the backend interface of the mini program. The data is automatically synchronized by the device or manually reported by the resident. The data is aggregated by hour. (4) Gene detection database interface connects to a third-party testing institution through a dedicated data line to retrieve the test results of 32 susceptibility gene loci related to cardiovascular diseases. It is a quarterly low-frequency update data. Each data is associated with a unique subject number.

[0020] Conventional multi-source data synchronization typically involves adding an extra timestamp field or sending separate synchronization signaling, which increases data transmission volume and system overhead. Unlike existing methods, this solution employs implicit time-series synchronization based on physiological characteristics. It leverages the inherent periodicity of physiological data to align multi-source data without requiring additional synchronization information transmission. Details are as follows: After data acquisition, multi-source time-series synchronization processing is first performed to extract the inherent periodic fluctuation characteristics of the heart rate data of physiological signs. A peak detection algorithm is used to identify the timestamp corresponding to the first heart rate fluctuation peak within an hour. Using this timestamp as a unified time reference, the other three types of data within the same time period are aligned by linear interpolation of the timestamps. After alignment, an implicit time-series synchronization mark is generated, which uses the lowest 2 bits of the data frame reserved bits as the encoding mark. It does not occupy a valid data field and does not require additional transmission of synchronization instructions. For example, if the heart rate peak occurs at 15 minutes and 20 seconds in a certain hour, then the time reference of all data in that hour is aligned to 15 minutes and 20 seconds, and the corresponding 2-bit mark is encoded as 01. All subsequent steps can quickly identify the time reference through this mark.

[0021] After synchronization, data is classified and determined according to feature rules. The classified data is stored in the corresponding data cache area, awaiting subsequent encryption and analysis. The classification rules are as follows: data with physiological indicator threshold ranges and continuous time series markers are classified as physiological sign data; data with behavioral type labels and occurrence timestamps are classified as lifestyle behavior data; data containing medical institution identifiers, document numbers, and treatment serial numbers are classified as treatment history data; and sequence data containing locus numbers and genotype information are classified as genetic susceptibility data. During the classification process, data format standardization is carried out simultaneously, uniformly adopting UTF-8 encoding and structured field naming to provide a unified data format for subsequent encryption and analysis.

[0022] II. Initial Differential Encryption Period: Initial encryption baseline periods were set for four types of data, taking into account three dimensions: data update frequency, privacy sensitivity level, and business dependence. Specifically, the baseline periods were set as follows: physiological vital signs data (high update frequency, medium privacy sensitivity level) had an initial encryption baseline period of 5 minutes; lifestyle behavior data (low update frequency, large fluctuations in user activity, low privacy sensitivity level) had an initial encryption baseline period of 30 minutes; medical history data (involving core personal medical privacy, high business dependence on medical decisions, high sensitivity level) had an initial encryption baseline period of 2 hours; and genetic susceptibility data (highly sensitive privacy data, highest risk of data leakage) had an initial encryption baseline period of 24 hours. Privacy sensitivity levels were determined according to industry data grading rules, and business dependence was assessed based on the data's support weight for clinical decision-making.

[0023] After the baseline period is determined, a dynamic correction is made in the same direction based on the average fluctuation frequency of the risk prediction residual sequence output from the previous cycle. That is, the encryption period adjustment ratio = residual fluctuation frequency change ratio × 0.8. Here, the baseline value is the preset baseline fluctuation frequency corresponding to each type of data. When the actual fluctuation frequency increases by more than 20% compared to the baseline value, it indicates that the residents' health status fluctuates greatly and the data timeliness requirement is high. The encryption baseline period for the four types is shortened by 20% simultaneously to improve the encryption update speed. For example, the baseline period for physiological data is 5 minutes, which is shortened by 20% to 4 minutes. When the fluctuation frequency decreases by more than 15% compared to the baseline value, it indicates that the residents' health status tends to be stable and the data changes slowly. The encryption baseline period is extended by 15% simultaneously to reduce system resource consumption. For example, the baseline period for lifestyle behavior data is 30 minutes, which is extended by 15% to 34.5 minutes. For ease of acquisition, the period can be rounded. The same-direction correction ensures that the encryption rhythm matches the data change rhythm, avoiding encryption lag or resource waste. Finally, an initial differentiated encryption period adapted to the current state is output and directly input into the encryption of the next stage.

[0024] III. Constructing the initial encrypted data sample set: Based on the initial differentiated encryption cycle, corresponding encryption tasks are generated for each type of data and scheduled uniformly according to priority. Each task corresponds to a data source type and matches the corresponding encryption configuration parameters with the original traceability data source. The encryption configuration parameters include three core components: First, the encryption granularity, which is divided according to the data sensitivity level. Highly sensitive data such as genetic and diagnostic data are divided into 128-byte blocks, moderately sensitive data such as physiological signs are divided into 256-byte blocks, and low-sensitivity data such as lifestyle data are divided into 512-byte blocks. The smaller the block size, the finer the encryption granularity and the higher the encryption strength. Second, the initial key length, which is uniformly configured to 128 bits. Third, the length of the redundant padding bits, with 8 bits reserved at the end of each data block to carry implicit identifiers. The original traceability data source is located through the unique data identifier and timestamp to ensure that the encrypted data can be traced back to the original collection node.

[0025] In addition, it is necessary to determine the encryption execution sequence according to the sensitivity level priority principle, encrypting data in descending order of sensitivity level to avoid highly sensitive data remaining in plaintext form in memory for a long time; the block division rule shall be executed according to the above granularity standard, and data that is less than a block shall be padded with the actual length before encryption; the key update interval shall be consistent with the encryption baseline period of the corresponding data; and the AES basic algorithm shall be used to encrypt data batch by batch according to the sequence. AES encryption is an existing mature cryptographic algorithm, and its function is to provide basic encryption computing capabilities.

[0026] The main approach is a two-layer encryption architecture. The initial encryption provides basic security protection for all data, with simple scheduling logic and high encryption efficiency, ensuring that data is encrypted as soon as it is stored. Subsequent re-encryption provides dynamic enhancement protection for core sensitive data, with keys deeply bound to individual health characteristics for higher security. The two-layer architecture balances encryption efficiency and deep security, avoiding the problem of excessive resource consumption caused by using highly complex dynamic encryption for all data.

[0027] Conventional batch management of encrypted data typically uses separate batch log files, which consumes additional storage resources and poses security risks. In contrast, we employ implicit batch identifiers, utilizing the redundant padding bits within the encrypted data blocks to carry batch identification information, eliminating the need for additional storage media. Details are as follows: During the encryption process, an implicit batch identifier is embedded synchronously. The first-order fractal dimension of the risk prediction residual output from the previous cycle is extracted, converted into an 8-bit binary code using a linear mapping, and written into the 8-bit redundant padding bits of each data block. Each batch identifier is uniquely bound to the corresponding time window and can be used for subsequent data traceability and integrity verification. There is no need to store a separate batch log file, making full use of the redundant bit space of the encrypted data block. For example, if the first-order fractal dimension of the residual from the previous cycle is 1.5, it is converted into the value 128 in the range of 0-255 through linear mapping, corresponding to the 8-bit binary code 10000000, which is written into the redundant padding bits of all data blocks. Subsequently, only this code needs to be decoded to quickly identify the corresponding batch.

[0028] After all batches of encryption are completed, the timeline is aligned according to the aforementioned implicit timing synchronization mark. The four types of encrypted data are then spliced ​​together in the order of timestamps using a timestamp matching algorithm to form multiple batches of initial encrypted data sample sets. Each sample set corresponds to a minimum time window of 5 minutes. The initial encrypted data sample sets will serve as the input for subsequent encryption quality evaluation.

[0029] IV. Initial Encryption Quality Assessment: We extract the data structure format and encrypted data content from the initial encrypted data sample set, and conduct multi-dimensional evaluation calculations. The initial encryption quality assessment value output from this step will be used as a data credibility correction coefficient input into the risk prediction process. Since existing encryption quality assessments typically only include two dimensions—key strength and data integrity—they cannot verify the feature consistency and source authenticity of the encrypted data. Therefore, we add two verification indicators—risk residual matching degree and verification residual fidelity—to the conventional assessment dimensions, thereby verifying the authenticity and tamper-proof state of the encrypted data at the feature level. Specifically: First, calculate the encryption continuity assessment value. For three different types of data, calculate the encryption fit: (1) Structured data, such as test index values ​​and medical record fields, calculate the fit based on the integrity of the encrypted fields. If the fields are complete and without missing data, the fit is 1. If they are missing data, the fit is reduced proportionally. (2) Semi-structured data, such as medical record texts and behavior logs, calculate the fit based on the retention of structural labels. If the key structural labels are complete, the fit is high. (3) Unstructured data, such as gene sequence fragments, calculate the fit based on the integrity of the data. If the sequence is not truncated, the fit is 1. The fit of these three types of data is weighted and summed with weights of 0.5, 0.3, and 0.2 respectively to obtain an encryption continuity assessment value with a maximum score of 1. The higher the value, the better the continuity and integrity of the encrypted data.

[0030] Secondly, a basic content encryption score is calculated, with scores ranging from 0 to 1 across six dimensions. Key strength level is scored based on key length and algorithm complexity; anti-cracking capability level is scored based on key entropy and resistance to brute-force attacks; data integrity level is scored based on data consistency after encryption and decryption; access controllability level is scored based on the granularity of permission control; transmission security level is scored based on the encryption strength of the transmission link; and storage security level is scored based on the encryption level of the storage medium. The arithmetic mean of these six dimensions is taken to obtain the basic content encryption score, comprehensively covering the security performance across the entire chain of encryption algorithms, transmission, and storage.

[0031] Based on this, two additional feature verification indicators are added. The first is the risk residual matching degree. This involves decoding the fractal dimension encoding in the implicit batch identifier and comparing it with the first-order fractal dimension calculated from the residual of the actual risk prediction in the previous round. The smaller the difference, the higher the matching degree, with a maximum score of 1 point. This is used to verify the fidelity of the residual features during the encryption process. For example, if the fractal dimension decoded from the identifier is 1.5 and the actual residual fractal dimension is 1.48, then the matching degree is 1 - |1.5 - 1.48| = 0.98. The second is the residual fidelity verification. This involves using the CRC32 algorithm to perform integrity verification on the initial encrypted data to obtain the initial verification residual. The degree of agreement between the distribution characteristics of the initial verification residual and the preset fractal distribution characteristics is calculated, with a maximum score of 1 point. This is used to verify the feature consistency of the encrypted data.

[0032] The weights of each indicator were allocated using the analytic hierarchy process (AHP), a three-level judgment matrix was constructed, and consistency checks were performed. The final weights were determined as follows: content encryption base score (0.6), risk residual matching degree (0.2), and verification residual fidelity (0.2). These weighted sums yielded the content encryption assessment value. The encryption continuity assessment value and the content encryption assessment value were then weighted a second time with weights of 0.3 and 0.7, respectively, to obtain the initial encryption quality assessment value. Anomaly linkage rules were set as follows: if data is missing, the fit score of the missing category is proportionally reduced, and the weighting of subsequent risk prediction steps is adjusted accordingly to ensure consistency in the assessment logic. For example, when genetic data is missing, the corresponding dimension score in the encryption quality assessment is halved, and the weight of genetic data in risk prediction is proportionally allocated to the other three data categories. When the missing percentage of a single data category exceeds 50%, the system automatically marks the sample as a low-confidence sample, and a reference prompt is added to the risk prediction result.

[0033] V. Initial Personalized Health Risk Prediction: For cardiovascular disease risk assessment, four types of health data are pre-defined with varying weights. These weights are determined based on clinical epidemiological research findings: physiological signs data have a weight of 0.4, lifestyle behavior data 0.3, medical history data 0.2, and genetic susceptibility data 0.1. The initial personalized health risk prediction value output from this step will serve as input for subsequent residual sequence generation. A mechanism for correcting encryption quality and reliability is introduced, using the encryption quality assessment result as a correction coefficient for data reliability. Higher encryption quality equates to higher reliability and a higher weight in the risk assessment. Details are as follows: The correction coefficient is calculated by combining the actual coverage and time span of various types of data: the coverage coefficient is calculated based on the complete proportion of data items, and missing items are deducted proportionally. For example, if a resident is missing gene data, the corresponding weight is proportionally allocated to the other three types of data; the time span coefficient is calculated based on the duration of continuous data coverage. The longer the coverage time, the higher the coefficient. The time span coefficient for data with continuous coverage for more than 1 year is 1. Using the initial encryption quality assessment value as the data credibility correction coefficient, the health feature contribution value of each data type is calculated separately. The calculation formula is: Single data type contribution value = Original data risk score × Corresponding weight ratio × Coverage coefficient × Time span coefficient × Initial encryption quality assessment value.

[0034] For example: If the original risk score of a resident's physiological characteristics data is 0.8, with a corresponding weight of 0.4, a data coverage coefficient of 0.9, a time span coefficient of 1, and an initial encryption quality assessment value of 0.9, then the health characteristic contribution value of the physiological data = 0.8 × 0.4 × 0.9 × 1 × 0.9 = 0.2592; the other three types of data are calculated using the same logic, and finally the health characteristic contribution values ​​of the four types of data are summed to obtain the total risk correction base.

[0035] The total risk correction base is input into the preset cardiovascular disease risk assessment model. The model uses a logistic regression algorithm, which is a mature existing risk modeling method. Its function is to provide standardized risk quantification calculation capabilities. The model outputs an initial personalized health risk prediction value in the range of 0 to 1. The higher the value, the higher the risk of the corresponding disease.

[0036] The model's initial historical baseline value is set based on a large sample population data. The baseline drift correction coefficient output in the previous round will be superimposed on the historical baseline parameters in this round to adjust the predicted baseline within the sliding window frame by frame, thereby offsetting the prediction error caused by the baseline shift. For example, if the baseline drift correction coefficient in the previous round was +0.015, then all baseline values ​​in this round will be uniformly increased by 0.015.

[0037] VI. Risk Prediction Residual Sequence and Chaos Key Seed: The sliding time window is set to 24 hours and the sliding step is 5 minutes. Each slide extracts all the initial personalized health risk prediction values ​​within the window and arranges them in chronological order to form a time-series risk value sequence. The corresponding window's historical baseline health risk value is matched to form a time-series baseline sequence. The risk prediction residual sequence output by this step is the core intermediate carrier of the entire closed-loop logic, connecting the risk prediction and encryption stages.

[0038] The initial risk residual sequence is obtained by subtracting the corresponding time points of the two sequences. A wavelet adaptive denoising algorithm is then used to remove high-frequency acquisition noise. The db4 wavelet basis is chosen, and the decomposition layer is three. After removing high-frequency noise components, the signal is reconstructed and then smoothed to obtain the standardized risk prediction residual sequence. Wavelet denoising is a mature signal processing technique used to remove the interference of acquisition noise on the residual features. The db4 wavelet basis is chosen because of its tight support and good symmetry, making it suitable for processing non-stationary time series such as physiological signals. This scheme focuses on generating chaotic keys from risk residuals across different boundaries. The risk prediction residuals, originally used only for error statistics, are transformed into an entropy source for encryption keys through multifractal feature extraction. The key features are deeply bound to individual health fluctuations, significantly improving unpredictability. Details are as follows: Multifractal feature extraction was performed on the standardized residual sequence using the sliding box dimension method. The box size was traversed sequentially from 2 to 16 with a step size of 2. The multifractal spectrum of the residual sequence within each sliding window was calculated to obtain the pointwise fractal dimension sequence, with a value range of 1 to 2, which is used to characterize the local complexity of the residual. At the same time, the total number of local maxima and minima of the residual sequence within a unit time (5-minute window) was counted to obtain the local extremum density sequence, which is used to characterize the intensity of residual fluctuation.

[0039] The fractal dimension sequence and the extreme value density sequence are normalized to the interval of 0 to 1 using the min-max normalization method. They are then concatenated point by point over time to form an 8-bit combined feature, where the high 4 bits correspond to the fractal dimension and the low 4 bits correspond to the extreme value density. The feature is then mapped to a preset value interval of 0 to 255 using a linear mapping function to generate a chaotic key seed.

[0040] For example: at a certain moment, the fractal dimension is 1.6, which corresponds to a value of 0.6 after normalization, and is 1001 in the high 4 bits of binary; at the same moment, the local extremum density corresponds to a value of 0.7 after normalization, and is 1011 in the low 4 bits of binary; after concatenation, the 8-bit combination is 10011011, which corresponds to the decimal value 155, which is the chaotic key seed value at that moment.

[0041] This seed is deeply bound to the individual's health fluctuation characteristics. The seeds differ for different residents and at different times, avoiding the homogenization problem of universal random seeds and significantly improving the unpredictability of the key.

[0042] VII. Dynamic Encryption Key Generation and Block Re-encryption: This embodiment uses the AES-128 algorithm as the basic encryption algorithm, with the basic encryption key being the system-preset 128-bit root key. The key expansion function uses the SHA-256 hash expansion function. Both the AES algorithm and SHA hash expansion are existing mature cryptographic technologies, serving to provide basic key expansion and encryption capabilities. The main approach is to dynamically match the number of iterations using the seed value. The more complex the health state and the greater the residual fluctuation, the more key expansion iterations are needed, resulting in higher security strength. Details are as follows: The chaotic key seed is used as a perturbation parameter input to the key expansion function. The number of iterations of the expansion function is positively correlated with the value of the chaotic key seed. The mapping rule is: number of iterations = 64 + value of chaotic key seed. For example, when the seed value is 128, the number of iterations is set to 192 rounds. The higher the seed value, the greater the residual fluctuation and the more complex the health state, requiring more rounds of iteration to diffuse the key, and the higher the security. After multiple rounds of iteration and diffusion, a 128-bit dynamic encryption key corresponding to the time window is generated.

[0043] The generated dynamic encryption key is used to re-encrypt the corresponding batch of health data in blocks. The encryption operation is completed sequentially according to the data block order. After each data block is encrypted, the fractal dimension code corresponding to the current time is written into the 8-bit redundant padding bits of the data block to form a fractal feature implicit verification identifier.

[0044] Conventional encryption data integrity verification typically requires the additional generation and transmission of a checksum, consuming valuable data bandwidth and storage resources. This fractal feature implicit verification identifier, however, serves a dual purpose: firstly, it verifies the fidelity of the fractal feature in subsequent verification stages; secondly, it can reverse-verify the key's legitimacy during data decryption. If the decoded fractal dimension does not match the actual fractal dimension of the residual for the corresponding time period, it indicates a key error or data tampering, achieving data tamper-proof verification without the need for an additional verification module. This implicit verification mechanism fully utilizes the redundancy within the encrypted data itself, completing both integrity and legitimacy verification without the need for additional verification information transmission. The re-encrypted data will then be input into subsequent verification and correction stages.

[0045] 8. Encryption Mechanism for Homogeneous Data Aggregation Based on Residual Fractal Similarity: The similarity threshold is set at a difference of less than 0.05 in the residual fractal dimension and less than 10% in the extreme value density of three adjacent time windows. Consecutive batches that meet the threshold are identified as data groups from the same source. A unified chaotic key seed base value is used for data groups from the same source. Subkeys are generated by superimposing small time perturbation factors within each group. Data within the group share the master key expansion parameters, and there is no need to execute a complete key expansion process for each batch separately.

[0046] This mechanism improves the encryption efficiency of consecutive similar batches while ensuring key uniqueness. At the same time, the aggregated residual features are richer, further improving key information entropy, thus optimizing encryption security and efficiency simultaneously.

[0047] Furthermore, the CRC32 checksum algorithm is used to perform integrity checks on each re-encrypted data block, calculating the re-encryption checksum residual for each data block. CRC32 is a mature existing checksum technology, providing standardized integrity check capabilities. The mean, variance, and skewness of all re-encryption checksum residuals within a 24-hour sliding window, as well as the fractal dimension distribution characteristics of the residual sequence itself, are statistically analyzed. A single-residue dual-parameter collaborative correction method is primarily employed, generating two correction coefficients based on the same re-encryption checksum residual. These coefficients are applied to the next round of fractal extraction and risk prediction, respectively, achieving dual optimization in a single calculation. Details are as follows: The process of generating fractal parameter correction coefficients based on re-encryption verification residuals includes: decoding the implicit verification identifier of fractal features in each data block, comparing it with the fractal dimension actually calculated from the re-encryption verification residuals, and calculating the fractal feature fidelity, where fidelity = 1 - |identified fractal dimension - actual fractal dimension|; if the fidelity is lower than 0.8, it indicates that the fractal extraction accuracy is insufficient under the current box size, and the window size of the sliding box dimension is increased accordingly, with an adjustment range of 2 units; if the fidelity is higher than 0.95, it indicates that the current accuracy meets the requirements, and the window size is appropriately reduced to improve computational efficiency, with an adjustment range of 2 units; finally, fractal parameter correction coefficients are generated, which are simultaneously applied to the next two fractal calculation stages: one is the calculation of the first-order fractal dimension of the implicit batch identifier in the initial encryption stage, and the other is the calculation of the multifractal spectrum in the chaotic key generation stage, ensuring unified dynamic optimization of fractal calculation parameters throughout the entire process.

[0048] Simultaneously, a baseline drift correction coefficient is generated based on the re-encryption and verification residuals: the residual mean weight is set to 0.5, the residual variance weight is set to 0.3, and the residual skewness weight is set to 0.2. The baseline drift correction coefficient is obtained by weighted summation, with a coefficient range of -0.05 to 0.05. For example, if the residual mean is 0.02, the residual variance is 0.01, and the residual skewness is 0.01 within a certain window, then the baseline drift correction coefficient = 0.02 × 0.5 + 0.01 × 0.3 + 0.01 × 0.2 = 0.015.

[0049] This coefficient will be superimposed on the historical baseline parameters of the next round of cardiovascular disease risk assessment model, and the predicted baseline within the next sliding window will be adjusted frame by frame to offset the baseline deviation caused by the accuracy drift of the front-end acquisition equipment, the natural aging of model parameters, and changes in population characteristics.

[0050] Both correction coefficients are calculated using the same re-encryption verification residual, eliminating the need for additional data collection. A single calculation simultaneously improves the performance of both fractal extraction accuracy optimization and risk prediction baseline calibration, forming a closed-loop self-optimization logic. When a high-risk secondary encryption is triggered, a secondary encryption verification residual is generated after the secondary encryption is completed. This residual replaces the original re-encryption verification residual to generate the two types of correction coefficients, thus incorporating the encryption quality of high-risk scenarios into the closed-loop self-optimization system.

[0051] IX. Anomaly data tracing and hierarchical circuit breaker mechanism based on fractal feature fingerprints: Conventional anomaly data investigation requires data decryption, batch log comparison, and data integrity verification, a complex and time-consuming process, and the decryption process itself poses security risks. This solution, however, extends this by using a fractal fingerprint mechanism. It leverages the implicit verification identifiers embedded in each data block, combined with the residual fractal baseline features of the corresponding time window, to construct the fractal fingerprint of the data. This enables rapid anomaly tracing and tiered handling without requiring complete data decryption. Details are as follows: A standard fractal fingerprint database for each time window is pre-established, storing the standard fractal dimension range, extreme value density range, and characteristic distribution pattern of the corresponding window. When data anomalies are detected during verification, only the fractal code in the redundant padding bits of the data block is extracted and compared with the standard fingerprint database to locate the time window, corresponding data batch, and data category where the anomaly occurred. The location accuracy can reach the level of 5 minutes, without the need to decrypt the data content.

[0052] Simultaneously, a three-tiered circuit breaker strategy is implemented: Level 1, minor anomalies (fractal deviation less than 5%), automatically triggers data retransmission and re-verification without interrupting business; Level 2, moderate anomalies (fractal deviation 5%-15%), blocks access to the batch of data, triggers an alarm, and automatically re-encrypts; Level 3, severe anomalies (fractal deviation greater than 15%), locks all data access permissions for the corresponding user, and triggers manual verification and security audit. This mechanism improves anomaly detection efficiency by more than 60% and avoids the risk of data decryption during anomaly detection.

[0053] A preset risk level range is established: 0 to 0.3 is low risk, 0.3 to 0.7 is medium risk, and 0.7 to 1 is high risk. The corresponding range is matched to the initial personalized health risk prediction value for this round and then graded. After processing, the risk prediction residuals, fractal parameter correction coefficients, and baseline drift correction coefficients generated in this round are all stored in a cache and input into the next loop. This process is repeated continuously to achieve iterative optimization of risk prediction accuracy and encryption quality. Figure 4 As shown, the details are as follows: For high-risk areas, the two most sensitive data categories, medical history and genetic susceptibility, were identified. The encryption granularity was upgraded to 64-byte blocks, and the key length was upgraded to 256 bits. On the basis of the dynamic re-encryption, residual embedded secondary encryption was completed. During the secondary encryption, the risk prediction residual code at the current moment was embedded in the reserved bits of the data block as an additional integrity check code. At the same time, a corresponding level of health intervention plan was generated based on the risk cause analysis. It includes dietary adjustment suggestions, exercise intervention plans and follow-up medical reminders, and is pushed to the resident's mini-program and community doctor's workstation to realize the linkage between risk warning and intervention.

[0054] For medium-risk areas, a tiered encryption mode is adopted. Core physiological indicators such as heart rate and blood pressure and key medical records maintain standard encryption strength, while non-core behavioral data such as exercise statistics and sleep duration are encrypted with lightweight encryption. Under the premise of matching risk and safety, resource allocation efficiency is further optimized.

[0055] For low-risk ranges, calculate the CPU and memory usage of the current encryption strategy. If the total resource usage is less than 30%, maintain the current key length and encryption cycle. If the total resource usage is greater than 50%, adjust the encryption granularity based on the fluctuation range of the risk prediction residual sequence. For example, adjust from 256-byte blocks to 512-byte blocks, adopt a lightweight encryption mode with residual features, and replace some complex verification logic with fractal identifiers to reduce computing resource consumption while ensuring data security level.

[0056] In actual operation, a full-scenario anomaly adaptation mechanism is also configured. When a certain type of data is missing, the corresponding weight will be proportionally allocated to other valid data categories. At the same time, the corresponding dimension score in the encryption quality assessment will be proportionally converted to ensure that the risk assessment and encryption assessment run synchronously and continuously. When the missing proportion of a single type of data exceeds 50%, the system will automatically mark it as a low-confidence sample and add a reference prompt. When network fluctuations cause data transmission interruption, the encrypted data will retain fractal feature identifiers. After the network is restored, the broken data can be quickly filled in through the identifiers without the need for full retransmission. When the equipment is replaced and the collection features change, the baseline drift correction coefficient will automatically adapt to the new collection features without the need for manual recalibration of the baseline.

[0057] 10. Lightweight risk verification mechanism for encrypted fields: A mapping library corresponding to risk levels and fractal characteristics is pre-established, clarifying the range of fractal dimension and extreme value density corresponding to different risk intervals. The terminal side only needs to extract the fractal code in the data block and compare it with the mapping library to quickly determine whether the risk level of the current data has changed significantly. The judgment time is only about 10% of the time required for complete decryption risk calculation. Complete decryption and accurate risk calculation are only triggered when the risk is determined to have changed significantly, which greatly reduces the computing power consumption of edge terminals and is suitable for low computing power scenarios such as wearable devices and community terminals.

[0058] The entire solution can be smoothly adapted to application scenarios of different scales, such as community health monitoring, chronic disease management, and public health prevention and control. It can also be extended to health data management scenarios for other chronic diseases such as diabetes, respiratory diseases, and chronic kidney disease. Only the weight parameters of the risk assessment model and disease-related gene loci need to be adjusted. The core residual cross-border encryption and double closed-loop correction logic do not need to be changed. The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0059] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0060] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A personalized risk prediction and data encryption method based on health big data, characterized in that, The method includes the following steps: Collect multi-source health big data covering four dimensions in health monitoring scenarios: physiological signs data, lifestyle behavior data, medical history data, and genetic susceptibility data. After multiple batches of data encryption, integrate them to form an initial encrypted data sample set. A multi-dimensional security assessment is conducted on the initial encrypted data sample set to obtain the initial encryption quality assessment value, which is then used as the data credibility weight in the weighted calculation of health features to output the initial personalized health risk prediction value. Within a preset sliding time window, the point-by-point difference between the initial personalized health risk prediction value and the historical baseline is calculated to generate a risk prediction residual sequence. Then, the risk prediction residual sequence is subjected to multifractal feature extraction to generate a fractal dimension sequence and a local extremum density sequence. After normalization and splicing, they are mapped to generate a chaotic key seed. The chaotic key seed is injected into the basic encryption key expansion function to generate a dynamic encryption key, and the corresponding batch of health data is re-encrypted in blocks; The integrity verification residuals of each data block after re-encryption are extracted, the distribution characteristics of the verification residuals are statistically analyzed, the fractal feature fidelity is inferred, and fractal parameter correction coefficients are generated to adjust the parameters for the next round of multifractal feature extraction. At the same time, the baseline drift correction coefficient is calculated based on the distribution of verification residuals to dynamically adjust the historical baseline of health risk prediction. After correction, the corrected personalized health risk prediction value is obtained.

2. The personalized risk prediction and data encryption method based on health big data according to claim 1, characterized in that, The specific steps for collecting multi-source health big data are as follows: Multi-source health big data is collected synchronously through four types of access nodes: wearable device interface, medical information system interface, daily behavior log interface, and gene testing database interface. Extract the inherent periodic fluctuation characteristics of physiological signs data rate and respiratory rate, and add implicit time-series synchronization markers to all categories of data with the peak point of fluctuation as a unified time reference; The data were classified and determined based on four categories: physiological indicator threshold characteristics, behavioral event record characteristics, medical record identification characteristics, and gene locus sequence characteristics.

3. The personalized risk prediction and data encryption method based on health big data according to claim 1, characterized in that, After data collection, an initial differentiated encryption period is generated based on the privacy sensitivity level corresponding to various types of data. The steps for generating the initial differentiated encryption period include: The encryption baseline period is determined for each of the four types of data, namely, physiological signs data combined with data update frequency and privacy sensitivity level, lifestyle behavior data combined with user behavior activity level, medical history data combined with business dependence on medical decisions, and genetic susceptibility data combined with data leakage risk sensitivity. Based on the average fluctuation frequency of the residual sequence of the previous risk prediction, the four types of encryption benchmark periods are dynamically corrected in the same direction. The higher the fluctuation frequency, the shorter the encryption period, thus obtaining the initial differentiated encryption period.

4. The personalized risk prediction and data encryption method based on health big data according to claim 1, characterized in that, The steps for forming the initial encrypted data sample set specifically include: Based on the initial differentiated encryption cycle, an encryption task corresponding to each type of data is generated, and the corresponding encryption configuration parameters are matched with the original traceable data source. The encryption configuration parameters include encryption granularity, key length, and redundant padding bit length. An encryption sequence containing execution order, block rules, and key update interval is generated and encrypted in batches. During the encryption process, the first-order fractal dimension of the risk prediction residual of the current batch is extracted, converted into a fixed-length binary code, and embedded as a data block redundant padding bit as an implicit batch identifier. The timelines of the four types of encrypted data are aligned and integrated based on implicit time synchronization markers.

5. The personalized risk prediction and data encryption method based on health big data according to claim 1, characterized in that, The calculation process for the initial encryption quality assessment value is as follows: Extract the data structure format and encrypted data content from the initial encrypted data sample set, calculate the encryption fit of three types of data (structured data, semi-structured data, and unstructured data) respectively, and integrate them to obtain the encryption continuity evaluation value. The content encryption effect is evaluated from six dimensions: key strength level, anti-cracking capability level, data integrity level, access controllability level, transmission security level, and storage reliability level, resulting in a basic content encryption score. The implicit batch identifier is decoded and compared with the actual residual fractal characteristics to obtain the risk residual matching degree. The consistency between the verification residual distribution and the preset fractal characteristics is calculated to obtain the verification residual fidelity. The initial encryption quality assessment value is obtained by allocating the weights of each indicator using the analytic hierarchy process (AHP) and then summing the weights, combined with the encryption continuity assessment value.

6. The personalized risk prediction and data encryption method based on health big data according to claim 1, characterized in that, The process for outputting the initial personalized health risk prediction value is as follows: The weighting of four types of health data in the risk assessment of the target disease is pre-set. Combining the actual coverage and time span of each type of data, the initial encryption quality assessment value is used as the credibility correction coefficient to calculate the health feature contribution value of each type of data. After accumulating the total risk correction base, input it into the preset disease risk assessment model and output the initial personalized health risk prediction value; The historical baseline for health risk prediction is dynamically adjusted. The historical baseline is adjusted frame by frame within the sliding window by superimposing the baseline drift correction coefficient onto the baseline parameters of the risk assessment model.

7. The personalized risk prediction and data encryption method based on health big data according to claim 1, characterized in that, The process of generating the risk prediction residual sequence is as follows: Set a fixed-duration sliding time window, extract the initial personalized health risk prediction values ​​within the window according to the preset step size to form a time series risk value sequence, match the historical baseline values ​​of the corresponding window to form a time series baseline sequence, and perform point-by-point subtraction and adaptive denoising and smoothing to obtain a standardized risk prediction residual sequence. The process of generating the chaotic key seed is as follows: The multifractal spectrum of the residual sequence is calculated using the sliding box dimension method to obtain the pointwise fractal dimension sequence. At the same time, the number of local extreme points per unit time is counted to obtain the local extreme value density sequence. The two types of sequences are normalized and then spliced ​​together pointwise over time to form a combined feature sequence, which is then mapped to a preset numerical range to generate a chaotic key seed.

8. The personalized risk prediction and data encryption method based on health big data according to claim 7, characterized in that, The process of generating the dynamic encryption key is as follows: The chaotic key seed is used as a perturbation parameter input to the key expansion function of the basic encryption key, so that the number of iterations of the expansion function is positively correlated with the value of the chaotic key seed. After multiple rounds of iteration to spread the randomness of the key, the dynamic encryption key corresponding to the time window is output. The steps for block re-encryption are as follows: Encryption operations are performed sequentially on each data block. The fractal dimension code corresponding to the current moment is written into the redundant padding bits of each data block to form an implicit verification identifier for fractal features.

9. The personalized risk prediction and data encryption method based on health big data according to claim 1, characterized in that, The generation process of the fractal parameter correction coefficient and the baseline drift correction coefficient is as follows: Integrity verification is performed on each re-encrypted data block to obtain the corresponding verification residual value. The mean, variance, skewness and fractal dimension distribution characteristics of all residuals within the statistical window are then analyzed. The fractal feature implicit verification identifier in the decoded data block is compared with the actual fractal dimension of the verification residual to obtain the fractal feature fidelity. Based on this, fractal parameter correction coefficients are generated to dynamically adjust the window size of the sliding box dimension during the next round of multifractal feature extraction. The baseline drift correction coefficient is obtained by weighting and summing the residual mean, variance, and skewness according to preset weights. This coefficient is used to offset the baseline deviation caused by data acquisition drift and model parameter aging.

10. The personalized risk prediction and data encryption method based on health big data according to claim 1, characterized in that, Based on the obtained corrected personalized health risk prediction values, the risk level range to which the risk level range belongs is determined. The processing steps after determining the risk level range are as follows: For high-risk areas, identify the corresponding high-sensitivity data categories and data volumes, match the corresponding level of encryption granularity and key length, and use dynamic encryption keys to perform secondary encryption on high-sensitivity data such as medical history and genetic susceptibility. During secondary encryption, embed the current risk prediction residual code as an integrity check code, and simultaneously output a health intervention plan that includes risk factor analysis and graded adjustment suggestions. For low-risk ranges, the computational resource utilization rate of the current encryption strategy is calculated. If it is lower than the preset threshold, the current key and period remain unchanged. If it is higher than the preset threshold, the encryption granularity is reduced according to the fluctuation range of the risk prediction residual sequence. The residual feature lightweight encryption mode is adopted to balance security and resource consumption.

Images