Health risk assessment method and system based on big data analysis

Abstract

The invention discloses a health risk assessment method and system based on big data analysis, and the method comprises the following steps: obtaining multi-source health data, completing the data standardization, and obtaining a multi-modal time sequence sample; generating feature vectors and quality vectors of each mode, and mapping the feature vectors into basic trust distribution; dividing the basic trust distribution into a hard evidence layer and a soft evidence layer to obtain hard evidence layer discount evidence and soft evidence layer discount evidence; calculating the conflict quality of the hard evidence layer discount evidence and performing routing to obtain a hard layer fusion evidence and a hard layer conflict degree; executing conflict routing on the soft evidence layer discount evidence and redistributing by adopting a PCR6 rule to obtain soft layer fusion evidence and a soft layer conflict degree; calculating inter-layer conflict quality and redistributing by adopting a PCR6 rule to obtain an inter-layer fusion result and an inter-layer conflict degree; and outputting the risk level and the risk confidence coefficient. According to the method, a PCR6 conflict redistribution mechanism is fused, and the stability of a health risk assessment result is improved.

Description

Technical Field

[0001] This invention relates to the field of medical and health big data technology, and in particular to a health risk assessment method and system based on big data analysis. Background Technology

[0002] With the rapid development of wearable devices, electronic medical record systems, health management platforms, and medical testing technologies, personal health-related data is characterized by diverse sources, large scale, and frequent updates. Existing health risk assessment technologies are usually based on physical examination indicators, historical medical records, or data from a single monitoring device, and assess individual health status through statistical analysis or machine learning models. They have been applied to some extent in chronic disease management, risk warning, and decision support. However, in practical applications, multi-source health data vary significantly in terms of collection frequency, data quality, timeliness, and reliability. Different data sources often give inconsistent or even contradictory judgments on the same health risk, posing a significant challenge to health risk assessment.

[0003] To address the issue of multi-source health data fusion, existing technologies typically employ feature concatenation, weighted averaging, or traditional probability models for unified modeling. Some methods introduce evidence theory to fuse multi-source information in an effort to address uncertainty. However, existing evidence fusion methods often use fixed fusion rules, which fail to adequately differentiate the credibility of each data source. In high-conflict scenarios, this can easily lead to information amplification or result distortion. Furthermore, existing technologies often treat all health data as equal evidence for fusion, lacking a tiered processing mechanism for key, highly reliable data versus auxiliary, less reliable data. This results in assessment results that are highly sensitive to noisy and outlier data, leading to insufficient stability.

[0004] Therefore, how to provide a health risk assessment method and system based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a health risk assessment method and system based on big data analysis. This invention introduces a layered fusion mechanism of hard evidence layer and soft evidence layer based on reliability gating, and combines the PCR6 conflict redistribution rule to adaptively process conflict information in multi-source health data, effectively reducing the interference of high-conflict data on assessment results, and improving the stability, credibility and robustness of health risk assessment in complex, multi-source and inconsistent data scenarios.

[0006] A health risk assessment method based on big data analysis according to an embodiment of the present invention includes the following steps:

[0007] Acquire multi-source health data of target users and complete data standardization to obtain multimodal time-series samples with sliding time windows;

[0008] For multimodal time series samples, feature vectors and quality vectors for each modality are generated. Reliability coefficients are calculated from the quality vectors, and feature vectors are mapped to basic trust assignments.

[0009] The basic trust allocation is divided into a hard evidence layer and a soft evidence layer, and a reliability discounting process is performed to obtain hard evidence layer discounted evidence and soft evidence layer discounted evidence.

[0010] Calculate and route the conflict quality of the hard evidence layer discount evidence, transfer the conflict quality with a reliability coefficient below the gating threshold to the unknown set, and input the remaining conflict quality into the PCR6 rule for redistribution to obtain the hard layer fused evidence and hard layer conflict degree.

[0011] Conflict routing is performed on the discounted evidence in the soft evidence layer and the PCR6 rule is used for redistribution to obtain the soft layer fused evidence and the soft layer conflict degree.

[0012] Hard-layer fusion evidence and soft-layer fusion evidence are fused together, the interlayer conflict quality is calculated and redistributed using the PCR6 rule. If the hard evidence validity condition is met, the hard layer allocation weight is increased; otherwise, the interlayer conflict quality is transferred to the unknown set and the uncertainty is increased, thus obtaining the interlayer fusion result and the interlayer conflict degree.

[0013] The degree of inter-layer conflict is used to determine the conflict trend criterion, and the risk level and risk confidence level are output by combining the inter-layer fusion results.

[0014] Optionally, obtaining the multimodal time series samples specifically includes:

[0015] Access and aggregate multi-source health data of target users, establish a list of data sources and data access rules, read raw records from each data source according to the data access rules, write them into a unified raw data area, and output the raw data set;

[0016] Perform standardization on the original dataset, and write each standardized record and its corresponding standardization tag into the standardized data area to output the standardized dataset;

[0017] A sliding time window is constructed based on a standardized dataset. The start time of the evaluation period, the width of the time window, and the sliding step size are determined. Continuous time window intervals are generated sequentially according to the sliding step size. Standardized records in the standardized dataset whose recording time falls into each time window interval are assigned to the corresponding time window. For each time window, the data are merged according to the data source to form a multimodal window subset set. Multimodal time series samples organized by time window number are output.

[0018] Optionally, the determination of the basic trust allocation and reliability coefficient specifically includes:

[0019] Based on multimodal time series samples, the set of records within the window is extracted according to modality, and the timestamp parsing, record deduplication and sorting are completed. The number of records within the window is counted and the modal record sequence and the corresponding number of records are generated.

[0020] For each modal record sequence, feature vector generation processing is performed. Records within the window are aggregated according to preset feature generation rules to obtain statistical features, trend features, and fluctuation features. These features are then concatenated into modal feature vectors in a preset dimension order. Modal identifiers and time window identifiers are written in the same way, and a set of modal feature vectors is output.

[0021] For each modal record sequence, quality vector generation and evidence mapping are performed. Missing data is generated based on the expected number of records and the actual number of records. Timeliness is generated based on the window end time and the most recent record time. Consistency is generated based on the proportion of abnormal records. Reliability is generated by reading the trusted configuration of the data source. The modal quality vector is assembled and the reliability coefficient is calculated. The modal feature vector is input into the modal scoring function to obtain the support for each health risk level. The support is scaled according to the reliability coefficient. The remaining quality is allocated to the set of ignorance to form a basic trust allocation. The basic trust allocation includes the trust quality allocated to each preset health risk level and the trust quality allocated to the set of ignorance.

[0022] Optionally, obtaining the hard evidence layer discount evidence and the soft evidence layer discount evidence specifically includes:

[0023] Read and verify the modality identifier and time window identifier of the reliability coefficient and basic trust allocation. Perform a removal process on records with missing modality identifier or time window identifier. Merge the records that pass the verification according to the modality identifier and output the reliability coefficient and basic trust allocation corresponding to the modality.

[0024] Read the preset evidence layer division rules and generate evidence layer identifiers for each modality. Based on the evidence layer identifiers, classify the basic trust allocations corresponding to each modality into hard evidence layer basic trust allocations and soft evidence layer basic trust allocations respectively. Perform integrity verification on the classification results and generate hard evidence layer basic trust allocation lists and soft evidence layer basic trust allocation lists.

[0025] Reliability discounting is performed on the basic trust allocation of the hard evidence layer and the soft evidence layer respectively. Based on the trust quality of each health risk level of each modality, the scaling is performed according to the reliability coefficient corresponding to the modality. The remaining unallocated trust quality after scaling is merged into the trust quality of the ignorance set of the current modality to form discounted evidence. The discounted evidence of each modality in the hard evidence layer and the soft evidence layer is aggregated to obtain the discounted evidence of the hard evidence layer and the discounted evidence of the soft evidence layer respectively.

[0026] Optionally, obtaining the hard-layer fusion evidence and the hard-layer conflict degree specifically includes:

[0027] The discount evidence and reliability coefficient are matched one by one according to the modality identifier and time window identifier. For the records that are successfully matched, a hard evidence layer processing list is established and the hard evidence layer discount evidence set and hard evidence layer reliability coefficient set are output.

[0028] The hard evidence layer discount evidence set is traversed pairwise, and the trust quality corresponding to the basic trust assignment is read respectively. The health risk level combination pointed to by the two discount evidences is enumerated item by item. When the health risk levels pointed to by the two discount evidences are not the same, the product of the trust quality corresponding to the combination item is included in the conflict quality of the evidence pair and the accumulation is completed to obtain the hard evidence layer conflict quality set. The hard evidence layer conflict quality set is summarized to obtain the hard evidence layer conflict degree.

[0029] A preset gate threshold is set, and gate labels are generated for each item in the hard evidence layer reliability coefficient set. Discounted evidence with a reliability coefficient lower than the gate threshold is marked as low-reliability discounted evidence. The conflict quality of low-reliability discounted evidence that has been included in the formation of the hard evidence layer conflict quality set is assigned to the transfer conflict quality set; otherwise, it is assigned to the conflict quality set to be reassigned. The transfer conflict quality set is summarized by evidence pair index, and the summary result is added to the trust quality corresponding to the ignorance set to form the ignorance set increment.

[0030] The PCR6 conflict redistribution process is invoked to perform a fusion operation on the discounted evidence corresponding to the set of conflict quality to be redistributed. Two pieces of discounted evidence to be fused are selected sequentially according to a preset combination order, and a temporary fusion result is generated. For each piece of discounted evidence to be fused, the trust quality assigned to each preset health risk level is enumerated item by item, and a level-to-level product term is formed. When the health risk levels pointed to by the two pieces of discounted evidence to be fused are not the same, the product term is taken as a conflict term. The conflict term is split into two parts according to the proportion of the trust quality of each piece of discounted evidence to be fused in the conflict term, and then added back to the trust quality of the corresponding health risk level. This completes one PCR6 redistribution. After all conflict terms are split and added back, the fusion result is obtained and used as the input for the next piece of discounted evidence to continue iterative fusion. At the same time, the increment of the ignorance set is accumulated to the trust quality corresponding to the ignorance set in the iterative fusion result. When the remaining pieces of evidence to be fused in the hard evidence layer discounted evidence list are empty, the iterative fusion ends and the hard evidence layer fused evidence is output.

[0031] Optionally, obtaining the soft layer fusion evidence and the soft layer conflict degree specifically includes:

[0032] Read the discounted soft evidence and its corresponding reliability coefficient within the current sliding time window, match them one by one according to the modality identifier and time window identifier, and remove inconsistent records to form a soft evidence processing list;

[0033] Perform pairwise traversal on the soft evidence layer discount evidence set, read the trust quality corresponding to the basic trust assignment of each piece of evidence, enumerate the level combinations and multiply and sum the trust quality of different level combinations to obtain the conflict quality, write it into the soft evidence layer conflict quality set according to the evidence pair index, and summarize the soft evidence layer conflict quality set to obtain the soft evidence layer conflict degree.

[0034] Read the gating threshold and generate gating tags. Classify the conflict quality containing low-reliability discount evidence into the transfer conflict quality and sum them up to the unknown set trust quality. Input the evidence corresponding to the remaining conflict quality into the PCR6 rule to perform conflict redistribution. For each conflict item, split it proportionally and add it back to the level trust quality that caused the conflict and iterate until the evidence is exhausted. Output the soft evidence layer fused evidence.

[0035] Optionally, obtaining the inter-layer fusion result and the inter-layer conflict degree specifically includes:

[0036] Verify the consistency of time window identifiers between the hard evidence layer fused evidence and the soft evidence layer fused evidence and complete the alignment, and establish inter-layer fused input evidence pairs;

[0037] For inter-layer fusion input evidence pairs, calculate the inter-layer conflict quality. When the health risk levels pointed to by the two are not the same, the corresponding trust quality product is included in the conflict item and accumulated to obtain the inter-layer conflict quality, which is then written into the inter-layer conflict quality record.

[0038] The validity of the fused evidence at the hard evidence layer is determined. When the conditions for the validity of hard evidence are met, a valid mark is generated; when the conditions for the validity of hard evidence are not met, an invalid mark is generated.

[0039] When a valid marker is determined, inter-layer bias processing is performed. The preset inter-layer bias strategy is read and hard evidence layer bias weights are generated. The trust quality in the hard evidence layer fused evidence is amplified according to the hard evidence layer bias weights. The amplified trust quality and the trust quality corresponding to the set of ignorance are normalized. The biased hard evidence layer fused evidence is then output.

[0040] Interlayer PCR6 fusion processing is performed. When a label is determined to be valid, the fused evidence from the biased hard evidence layer and the fused evidence from the soft evidence layer are used as fusion inputs. The conflict items of the two are split and added back according to the PCR6 ratio to obtain the interlayer fusion result. At the same time, the interlayer conflict quality is output as the interlayer conflict degree. When a label is determined to be invalid, the interlayer conflict quality is transferred to the trust quality corresponding to the unknown set and the allocation ratio of the unknown set is increased to obtain the interlayer fusion result.

[0041] Optionally, the output of the risk level and risk confidence level specifically includes:

[0042] Write the inter-layer conflict degree into the conflict degree sequence according to the time window identifier, and retrieve the inter-layer conflict degree of the sliding time window to form a conflict degree history sequence;

[0043] Read the preset conflict degree threshold, preset conflict increment threshold and preset continuous window number threshold, compare the current inter-layer conflict degree with the preset conflict threshold to generate a trend mark, calculate the increment of the inter-layer conflict degree of adjacent layers in the conflict degree history sequence and compare it with the preset conflict increment threshold to generate an increment mark, count the increment marks in the most recent continuous window and output the conflict trend criterion result;

[0044] The trust quality is read from the inter-layer fusion results. The preset health risk level with the highest trust quality is selected as the risk level output. The trust quality of the ignorance set and the inter-layer conflict degree are converted into confidence decay factors and combined to obtain the risk confidence level.

[0045] A health risk assessment system based on big data analysis according to an embodiment of the present invention includes:

[0046] The data access standardization module is used to access multi-source health data of target users and unify fields, codes, time bases and units to output standardized health data.

[0047] The time window sample construction module is used to construct sliding time windows based on standardized health data, divide records by time windows and merge them by data source to form multimodal time series samples;

[0048] The feature quality modeling module is used to generate record sequences of multimodal time series samples according to modality, extract statistics, trend and fluctuation features to form modal feature vectors, and calculate reliability coefficients;

[0049] The basic trust allocation module is used to input modal feature vectors into a scoring function to obtain the support for each health risk level, scale it in combination with a reliability coefficient, and allocate the remaining quality to the set of ignorance.

[0050] The layered discount gating module is used to divide the basic trust allocation into a hard evidence layer and a soft evidence layer and perform reliability discounting, calculate the conflict quality, and transfer low-reliability conflicts to the set of ignorance;

[0051] The PCR6 fusion and risk output module is used to perform PCR6 conflict redistribution and interlayer fusion on hard and soft layers, and output the risk level and risk confidence.

[0052] The beneficial effects of this invention are:

[0053] This invention standardizes multi-source health data and constructs multimodal time-series samples with sliding time windows, enabling health data from different sources, frequencies, and formats to be aligned and compared under the same assessment scale, thereby improving the consistency and time-series availability of subsequent risk assessment inputs.

[0054] This invention generates feature vectors and quality vectors for each modality and calculates reliability coefficients accordingly. It explicitly introduces data quality differences into the basic trust allocation process, enabling the assessment process to adjust for reliability issues such as missing data, reduced timeliness, and abnormal noise, thereby reducing the impact of low-quality data on risk assessment.

[0055] This invention achieves adaptive conflict handling and controllable uncertainty expression in high-conflict scenarios by dividing the basic trust allocation into a hard evidence layer and a soft evidence layer, performing reliability discounting, and then routing conflict quality by combining a gating threshold. Conflict quality caused by low-reliability evidence is transferred to the set of ignorance, and the remaining conflicts are redistributed using the PCR6 rule. This improves the stability and robustness of the fusion results.

[0056] This invention introduces hard evidence validity conditions through interlayer fusion, dynamically increases the weight of hard layer allocation or increases the proportion of the ignorant set when invalid, and combines the degree of interlayer conflict to form a conflict trend criterion and output risk confidence. This enables health risk assessment to not only output risk level, but also simultaneously reflect the credibility of the results and the characteristics of conflict evolution, thereby enhancing the interpretability and application value of the assessment results in real health management scenarios. Attached Figure Description

[0057] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0058] Figure 1 This is a flowchart of a health risk assessment method based on big data analysis proposed in this invention;

[0059] Figure 2 This is a schematic diagram of a health risk assessment system based on big data analysis proposed in this invention;

[0060] Figure 3 This diagram illustrates the hard and soft stratified evidence fusion and PCR6 conflict redistribution in a health risk assessment method based on big data analysis proposed in this invention. Detailed Implementation

[0061] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0062] refer to Figures 1-3 A health risk assessment method based on big data analysis includes the following steps:

[0063] Acquire multi-source health data of target users and complete data standardization to obtain multimodal time-series samples with sliding time windows;

[0064] For multimodal time series samples, feature vectors and quality vectors for each modality are generated. Reliability coefficients are calculated from the quality vectors, and feature vectors are mapped to basic trust assignments.

[0065] The basic trust allocation is divided into a hard evidence layer and a soft evidence layer, and a reliability discounting process is performed to obtain hard evidence layer discounted evidence and soft evidence layer discounted evidence.

[0066] Calculate and route the conflict quality of the hard evidence layer discount evidence, transfer the conflict quality with a reliability coefficient below the gating threshold to the unknown set, and input the remaining conflict quality into the PCR6 rule for redistribution to obtain the hard layer fused evidence and hard layer conflict degree.

[0067] Conflict routing is performed on the discounted evidence in the soft evidence layer and the PCR6 rule is used for redistribution to obtain the soft layer fused evidence and the soft layer conflict degree.

[0068] Hard-layer fusion evidence and soft-layer fusion evidence are fused together, the interlayer conflict quality is calculated and redistributed using the PCR6 rule. If the hard evidence validity condition is met, the hard layer allocation weight is increased; otherwise, the interlayer conflict quality is transferred to the unknown set and the uncertainty is increased, thus obtaining the interlayer fusion result and the interlayer conflict degree.

[0069] The degree of inter-layer conflict is used to determine the conflict trend criterion, and the risk level and risk confidence level are output by combining the inter-layer fusion results.

[0070] In this embodiment, obtaining the multimodal time series samples specifically includes:

[0071] Access and aggregate multi-source health data of target users, establish a list of data sources and data access rules, read raw records from each data source according to the data access rules, the raw records contain record content and record time information, and write them into a unified raw data area, output the raw data set, the multi-source health data includes at least wearable device physiological monitoring data, physical examination data, electronic medical record summary data, medication record data, and user self-reported questionnaire data;

[0072] The original dataset is subjected to standardization processing, which includes unifying the format and encoding of record fields, unifying the time base of record time information, removing duplicate records, identifying abnormal records and removing or correcting them according to preset processing rules, performing missing record processing and generating missing markers for missing records, performing noise suppression processing on noisy records, and performing unit conversion processing on records with inconsistent units. Each standardized record and its corresponding standardization marker are written into the standardized data area, and a standardized dataset is output.

[0073] A sliding time window is constructed based on a standardized dataset. The start time of the evaluation period, the width of the time window, and the sliding step size are determined. Continuous time window intervals are generated sequentially according to the sliding step size. Standardized records in the standardized dataset whose recording times fall into each time window interval are assigned to the corresponding time window. For each time window, the data are merged according to the data source to form a multimodal window subset set. Multimodal time series samples organized by time window number are output. The multimodal time series samples contain window subsets of data from each data source within the same time window and the corresponding time window identifier.

[0074] In this embodiment, the determination of the basic trust allocation and reliability coefficient specifically includes:

[0075] Based on multimodal time series samples, the set of records within the window is extracted according to modality, and the timestamp parsing, record deduplication and sorting are completed. The number of records within the window is counted and the modal record sequence and the corresponding number of records are generated.

[0076] For each modal record sequence, feature vector generation processing is performed. Records within the window are aggregated according to preset feature generation rules to obtain statistical features, trend features, and fluctuation features. These features are then concatenated into modal feature vectors in a preset dimension order. Modal identifiers and time window identifiers are written in the same way, and a set of modal feature vectors is output.

[0077] For each modal record sequence, quality vector generation and evidence mapping are performed. Missing data is generated based on the expected number of records and the actual number of records. Timeliness is generated based on the window end time and the most recent record time. Consistency is generated based on the proportion of abnormal records. Reliability is generated by reading the trusted configuration of the data source. The modal quality vector is assembled and the reliability coefficient is calculated. The modal feature vector is input into the modal scoring function to obtain the support of each health risk level. The support is scaled according to the reliability coefficient, and the remaining quality is allocated to the unknown set to form a basic trust allocation. The modal scoring function is a logistic regression-based scoring function. The parameters of the scoring function are obtained and fixed by maximum likelihood training using the modal feature vectors of historical health samples and the corresponding health risk level labels. The unknown set is a set containing all preset health risk levels. The basic trust allocation includes the trust quality allocated to each preset health risk level and the trust quality allocated to the unknown set. The trust quality is a normalized weight value allocated to each preset health risk level and a weight value allocated to the unknown set, which are scaled according to the reliability coefficient, and the sum of the two is one.

[0078] This invention utilizes the joint modeling of feature vectors and quality vectors from multimodal time-series samples. It forms a reliability coefficient based on missing data, timeliness, consistency, and data source reliability. This reliability scaling is applied to the support of the health risk level output by logistic regression, and the remaining quality is assigned to the unknown set, resulting in a normalizable basic trust allocation. This suppresses noise misleading under conditions of missing, lagging, and conflicting multi-source data, enhances the credibility and uncertainty expression of the fused input, and improves the stability and interpretability of health risk assessment results.

[0079] In this embodiment, obtaining the hard evidence layer discount evidence and the soft evidence layer discount evidence specifically includes:

[0080] Read and verify the modality identifier and time window identifier of the reliability coefficient and basic trust allocation. Perform a removal process on records with missing modality identifier or time window identifier. Merge the records that pass the verification according to the modality identifier and output the reliability coefficient and basic trust allocation corresponding to the modality.

[0081] Read the preset evidence layer division rules and generate evidence layer identifiers for each modality. Based on the evidence layer identifiers, classify the basic trust allocations corresponding to each modality into hard evidence layer basic trust allocations and soft evidence layer basic trust allocations respectively. Perform integrity verification on the classification results and generate hard evidence layer basic trust allocation lists and soft evidence layer basic trust allocation lists.

[0082] Reliability discounting is performed on the basic trust allocation of the hard evidence layer and the soft evidence layer respectively. Based on the trust quality of each health risk level of each modality, the scaling is performed according to the reliability coefficient corresponding to the modality. The remaining unallocated trust quality after scaling is merged into the trust quality of the ignorance set of the current modality to form discounted evidence. The discounted evidence of each modality in the hard evidence layer and the soft evidence layer is aggregated to obtain the discounted evidence of the hard evidence layer and the discounted evidence of the soft evidence layer respectively.

[0083] In this embodiment, obtaining the hard-layer fusion evidence and the hard-layer conflict degree specifically includes:

[0084] The discount evidence and reliability coefficient are matched one by one according to the modality identifier and time window identifier. For the records that are successfully matched, a hard evidence layer processing list is established and the hard evidence layer discount evidence set and hard evidence layer reliability coefficient set are output.

[0085] The hard evidence layer discount evidence set is traversed pairwise, and the trust quality corresponding to the basic trust assignment is read respectively. The health risk level combination pointed to by the two discount evidences is enumerated item by item. When the health risk levels pointed to by the two discount evidences are not the same, the product of the trust quality corresponding to the combination item is included in the conflict quality of the evidence pair and the accumulation is completed to obtain the hard evidence layer conflict quality set. The hard evidence layer conflict quality set is summarized to obtain the hard evidence layer conflict degree.

[0086] A preset gating threshold is set, and gating labels are generated for each item in the hard evidence layer reliability coefficient set. Discounted evidence with reliability coefficients lower than the gating threshold is marked as low-reliability discounted evidence. The conflict quality of low-reliability discounted evidence that has been included in the formation and written into the hard evidence layer conflict quality set is assigned to the transfer conflict quality set; otherwise, it is assigned to the conflict quality set to be reassigned. The transfer conflict quality set is summarized by evidence pair index, and the summary result is added to the trust quality corresponding to the ignorance set to form the ignorance set increment. The conflict quality is the sum of the cross-product terms of trust quality assigned to different health risk levels by different discounted evidence.

[0087] The PCR6 conflict redistribution process is invoked to perform a fusion operation on the discounted evidence corresponding to the set of conflict quality to be redistributed. Two pieces of discounted evidence to be fused are selected sequentially according to a preset combination order, and a temporary fusion result is generated. For each piece of discounted evidence to be fused, the trust quality assigned to each preset health risk level is enumerated item by item, and a level-to-level product term is formed. When the health risk levels pointed to by the two pieces of discounted evidence to be fused are not the same, the product term is taken as a conflict term. The conflict term is split into two parts according to the proportion of the trust quality of each piece of discounted evidence to be fused in the conflict term, and then added back to the trust quality of the corresponding health risk level. This completes one PCR6 redistribution. After all conflict terms are split and added back, the fusion result is obtained and used as the input for the next piece of discounted evidence to continue iterative fusion. At the same time, the increment of the ignorance set is accumulated to the trust quality corresponding to the ignorance set in the iterative fusion result. When the remaining pieces of evidence to be fused in the hard evidence layer discounted evidence list are empty, the iterative fusion ends and the hard evidence layer fused evidence is output.

[0088] This invention performs consistency matching between discounted evidence and reliability coefficients within the hard evidence layer to form a processing list. It determines the conflict quality based on the cross-product of trust quality between different health risk levels, summarizes the conflict degree, and transfers the conflict quality caused by low-reliability discounted evidence to the ignorance set through a gating threshold to form an uncertainty increment. The remaining conflicts are input into the PCR6 rule, proportionally split and added back to the corresponding risk level trust quality, and iteratively fused until evidence is exhausted. The hard-layer fused evidence and hard-layer conflict degree are output, improving the stability and credibility of the assessment results in high-conflict scenarios.

[0089] In this embodiment, obtaining the soft layer fusion evidence and the soft layer conflict degree specifically includes:

[0090] Read the discounted soft evidence and its corresponding reliability coefficient within the current sliding time window, match them one by one according to the modality identifier and time window identifier, and remove inconsistent records to form a soft evidence processing list;

[0091] Perform pairwise traversal on the soft evidence layer discount evidence set, read the trust quality corresponding to the basic trust assignment of each piece of evidence, enumerate the level combinations and multiply and sum the trust quality of different level combinations to obtain the conflict quality, write it into the soft evidence layer conflict quality set according to the evidence pair index, and summarize the soft evidence layer conflict quality set to obtain the soft evidence layer conflict degree.

[0092] Read the gating threshold and generate gating tags. Classify the conflict quality containing low-reliability discount evidence into the transfer conflict quality and sum them up to the unknown set trust quality. Input the evidence corresponding to the remaining conflict quality into the PCR6 rule to perform conflict redistribution. For each conflict item, split it proportionally and add it back to the level trust quality that caused the conflict and iterate until the evidence is exhausted. Output the soft evidence layer fused evidence.

[0093] In this embodiment, obtaining the inter-layer fusion result and the inter-layer conflict degree specifically includes:

[0094] Verify the consistency of time window identifiers between the hard evidence layer fused evidence and the soft evidence layer fused evidence and complete the alignment, and establish inter-layer fused input evidence pairs;

[0095] The inter-layer conflict quality is calculated for the inter-layer fusion input evidence pair. The inter-layer conflict quality is obtained by enumerating the combination of the health risk level pointed to by the hard evidence fusion evidence and the health risk level pointed to by the soft evidence fusion evidence. When the health risk levels pointed to by the two are not the same, the corresponding trust quality product is included in the conflict item and accumulated to obtain the inter-layer conflict quality, and written into the inter-layer conflict quality record.

[0096] The validity of the fused evidence at the hard evidence layer is determined. When the conditions for the validity of hard evidence are met, a valid mark is generated; when the conditions for the validity of hard evidence are not met, an invalid mark is generated.

[0097] When a valid marker is determined, inter-layer bias processing is performed. The preset inter-layer bias strategy is read and hard evidence layer bias weights are generated. The trust quality in the hard evidence layer fused evidence is amplified according to the hard evidence layer bias weights. The amplified trust quality and the trust quality corresponding to the set of ignorance are normalized. The biased hard evidence layer fused evidence is then output.

[0098] Interlayer PCR6 fusion processing is performed. When a label is determined to be valid, the fused evidence from the biased hard evidence layer and the fused evidence from the soft evidence layer are used as fusion inputs. The conflict items of the two are split and added back according to the PCR6 ratio to obtain the interlayer fusion result. At the same time, the interlayer conflict quality is output as the interlayer conflict degree. When a label is determined to be invalid, the interlayer conflict quality is transferred to the trust quality corresponding to the unknown set and the allocation ratio of the unknown set is increased to obtain the interlayer fusion result.

[0099] This invention aligns hard evidence and soft evidence in interlayer fusion within a time window to establish a fusion input. It uses the trust quality multiplication and accumulation of different health risk level combinations to characterize the quality of interlayer conflict and outputs the degree of interlayer conflict. It introduces a hard evidence validity judgment and interlayer bias weighting mechanism. When the hard evidence is deemed valid, the contribution of the hard layer is increased and normalized to maintain consistent allocation. When the hard evidence is deemed invalid, the conflict quality is transferred to the set of ignorance and the allocation ratio of the set of ignorance is increased. Combined with PCR6 conflict redistribution splitting and back-addition, a more robust interlayer fusion result and risk characterization are obtained.

[0100] In this embodiment, the output of the risk level and risk confidence level specifically includes:

[0101] Write the inter-layer conflict degree into the conflict degree sequence according to the time window identifier, and retrieve the inter-layer conflict degree of the sliding time window to form a conflict degree history sequence;

[0102] Read the preset conflict degree threshold, preset conflict increment threshold and preset continuous window number threshold, compare the current inter-layer conflict degree with the preset conflict threshold to generate a trend mark, calculate the increment of the inter-layer conflict degree of adjacent layers in the conflict degree history sequence and compare it with the preset conflict increment threshold to generate an increment mark, count the increment marks in the most recent continuous window and output the conflict trend criterion result;

[0103] The trust quality is read from the inter-layer fusion results. The preset health risk level with the highest trust quality is selected as the risk level output. The trust quality of the ignorance set and the inter-layer conflict degree are converted into confidence decay factors and combined to obtain the risk confidence level.

[0104] A health risk assessment system based on big data analytics includes:

[0105] The data access standardization module is used to access multi-source health data of target users and unify fields, codes, time bases and units to output standardized health data.

[0106] The time window sample construction module is used to construct sliding time windows based on standardized health data, divide records by time windows and merge them by data source to form multimodal time series samples;

[0107] The feature quality modeling module is used to generate record sequences of multimodal time series samples according to modality, extract statistics, trend and fluctuation features to form modal feature vectors, and calculate reliability coefficients;

[0108] The basic trust allocation module is used to input modal feature vectors into a scoring function to obtain the support for each health risk level, scale it in combination with a reliability coefficient, and allocate the remaining quality to the set of ignorance.

[0109] The layered discount gating module is used to divide the basic trust allocation into a hard evidence layer and a soft evidence layer and perform reliability discounting, calculate the conflict quality, and transfer low-reliability conflicts to the set of ignorance;

[0110] The PCR6 fusion and risk output module is used to perform PCR6 conflict redistribution and interlayer fusion on hard and soft layers, and output the risk level and risk confidence.

[0111] Example 1: To verify the feasibility of this invention in practice, it was applied to a continuous health management risk assessment scenario for people with chronic diseases and sub-health conditions. The target users are those who receive daily health follow-ups and remote monitoring. Data sources cover continuous physiological monitoring from wearable devices, physical examination indicators, electronic medical record summaries, medication records, and user self-reported questionnaires. The core problem that has long existed in this scenario is that the sampling frequency and quality of multi-source data vary significantly. Wearable data is missing and subject to noise drift. Physical examination data is sparsely and delayed. Questionnaires have subjective biases. Medication and medical record information are encoded inconsistently across different systems. This leads to strong conflicts in the support for risk levels from different sources within the same sliding time window. As a result, traditional weighted fusion or simple voting models produce output jitter and confidence distortion. In some cases, risk levels may even reverse frequently in the short term, making it difficult to support stable early warning and subsequent intervention.

[0112] In this scenario, the present invention first standardizes multi-source health data to ensure consistency in field format, coding system, time base, and unit caliber. Under the unified data caliber, a sliding time window is constructed to form multimodal time series samples. Subsequently, statistical features, trend features, and fluctuation features are aggregated within the window for each modality, and a quality vector is generated simultaneously to characterize missingness, timeliness, consistency, and data source reliability. The quality vector is further used to calculate the reliability coefficient. The modal feature vector is input into a modal scoring function based on logistic regression to obtain the support for each health risk level. The support is then scaled by the reliability coefficient, and the remaining quality is allocated to the set of ignorance, thereby forming a basic trust allocation that includes the trust quality of each risk level and the trust quality of the set of ignorance.

[0113] The basic trust allocation is divided into hard evidence and soft evidence layers according to the evidence layer division rules, and a reliability discount is applied. Conflict quality is calculated for each hard and soft evidence layer, and gating routing is performed. Conflict quality caused by low-reliability evidence is transferred to the ignorance set to explicitly increase uncertainty. The remaining conflict quality is entered into the PCR6 rule for conflict redistribution. Conflict items are proportionally split and added back to the risk level trust quality that caused the conflict to obtain the fused evidence and conflict degree of each layer. Finally, when performing inter-layer fusion, a hard evidence validity judgment is introduced. When hard evidence is valid, the hard layer allocation weight is increased and normalized to stabilize the contribution of the main evidence. When hard evidence is invalid, the inter-layer conflict quality is transferred to the ignorance set and the allocation ratio of the ignorance set is increased to avoid misjudgment. At the same time, PCR6 is used to complete the inter-layer conflict redistribution, outputting the inter-layer fusion result, inter-layer conflict degree, risk level, and risk confidence. A conflict trend criterion is formed based on the historical sequence of inter-layer conflict degree to support risk interpretation and follow-up strategy adjustment.

[0114] To quantify and verify the beneficial effects, a comparative method was set up on the same business process and the same batch of target user data. Traditional weighted average fusion, traditional evidence fusion rules, and the method of this invention were compared and evaluated. The evaluation period used multiple consecutive sliding time windows and covered typical situations of daily monitoring and inconsistencies in multi-source data. The specific comparison data is shown in Table 1:

[0115] Table 1. Comparison of different multi-source fusion strategies in health risk assessment tasks.

[0116] index Traditional weighted average fusion Traditional evidence fusion rules This invention (gated routing + PCR6 + inter-layer bias) High conflict window ratio 0.31 0.31 0.31 Robustness of high-conflict windows (proportion of level consistency) 0.62 0.68 0.81 Number of times adjacent window levels change (times / user / cycle) 3.4±1.9 2.8±1.6 1.6±1.1 Confidence level calibration error (absolute deviation) 0.091 0.078 0.043 False alarm rate (percentage) under low-quality modal interference 0.14 0.12 0.07 Review trigger rate (percentage) 0.18 0.16 0.11 Consistency (proportion) between risk level and clinical outcome 0.73 0.75 0.82

[0117] As shown in Table 1, under the condition of the same proportion of high-conflict windows (0.31), the present invention achieves a robustness score of 0.81 in high-conflict window performance, which is significantly better than the 0.62 of traditional weighted average fusion and the 0.68 of traditional evidence fusion rules. Regarding output stability, the number of adjacent window level changes in the present invention is 1.6 ± 1.1 times / user / cycle, which is significantly lower than the 3.4 ± 1.9 of weighted average fusion and the 2.8 ± 1.6 of traditional evidence fusion. In terms of confidence reliability, the confidence calibration error of the present invention is 0.043. The lower confidence level (0.091 for weighted average fusion and 0.078 for traditional evidence fusion) indicates that the invention can maintain a more reasonable confidence level even when conflict and uncertainty are enhanced. Furthermore, the invention exhibits a false positive rate of 0.07 and a review trigger rate of 0.11 under low-quality modal interference, both lower than the control method (false positive rate 0.14 / 0.12, review trigger rate 0.18 / 0.16). It also achieves a higher consistency (0.82) between risk level and clinical outcome, demonstrating a comprehensive performance improvement compared to the control method's 0.73 and 0.75.

[0118] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A health risk assessment method based on big data analysis, characterized in that, Includes the following steps: Acquire multi-source health data of target users and complete data standardization to obtain multimodal time-series samples with sliding time windows; For multimodal time series samples, feature vectors and quality vectors for each modality are generated. Reliability coefficients are calculated from the quality vectors, and feature vectors are mapped to basic trust assignments. The basic trust allocation is divided into a hard evidence layer and a soft evidence layer, and a reliability discounting process is performed to obtain hard evidence layer discounted evidence and soft evidence layer discounted evidence. Calculate and route the conflict quality of the hard evidence layer discount evidence, transfer the conflict quality with a reliability coefficient below the gating threshold to the unknown set, and input the remaining conflict quality into the PCR6 rule for redistribution to obtain the hard layer fused evidence and hard layer conflict degree. Conflict routing is performed on the discounted evidence in the soft evidence layer and the PCR6 rule is used for redistribution to obtain the soft layer fused evidence and the soft layer conflict degree. Hard-layer fusion evidence and soft-layer fusion evidence are fused together, the interlayer conflict quality is calculated and redistributed using the PCR6 rule. If the hard evidence validity condition is met, the hard layer allocation weight is increased; otherwise, the interlayer conflict quality is transferred to the unknown set and the uncertainty is increased, thus obtaining the interlayer fusion result and the interlayer conflict degree. The degree of inter-layer conflict is used to determine the conflict trend criterion, and the risk level and risk confidence level are output by combining the inter-layer fusion results.

2. The health risk assessment method based on big data analysis according to claim 1, characterized in that, The acquisition of the multimodal time series samples specifically includes: Access and aggregate multi-source health data of target users, establish a list of data sources and data access rules, read raw records from each data source according to the data access rules, write them into a unified raw data area, and output the raw data set; Perform standardization on the original dataset, and write each standardized record and its corresponding standardization tag into the standardized data area to output the standardized dataset; A sliding time window is constructed based on a standardized dataset. The start time of the evaluation period, the width of the time window, and the sliding step size are determined. Continuous time window intervals are generated sequentially according to the sliding step size. Standardized records in the standardized dataset whose recording time falls into each time window interval are assigned to the corresponding time window. For each time window, the data are merged according to the data source to form a multimodal window subset set. Multimodal time series samples organized by time window number are output.

3. The health risk assessment method based on big data analysis according to claim 1, characterized in that, The determination of the basic trust allocation and reliability coefficient specifically includes: Based on multimodal time series samples, the set of records within the window is extracted according to modality, and the timestamp parsing, record deduplication and sorting are completed. The number of records within the window is counted and the modal record sequence and the corresponding number of records are generated. For each modal record sequence, feature vector generation processing is performed. Records within the window are aggregated according to preset feature generation rules to obtain statistical features, trend features, and fluctuation features. These features are then concatenated into modal feature vectors in a preset dimension order. Modal identifiers and time window identifiers are written in the same way, and a set of modal feature vectors is output. For each modal record sequence, quality vector generation and evidence mapping are performed. Missing data is generated based on the expected number of records and the actual number of records. Timeliness is generated based on the window end time and the most recent record time. Consistency is generated based on the proportion of abnormal records. Reliability is generated by reading the trusted configuration of the data source. The modal quality vector is assembled and the reliability coefficient is calculated. The modal feature vector is input into the modal scoring function to obtain the support for each health risk level. The support is scaled according to the reliability coefficient. The remaining quality is allocated to the set of ignorance to form a basic trust allocation. The basic trust allocation includes the trust quality allocated to each preset health risk level and the trust quality allocated to the set of ignorance.

4. The health risk assessment method based on big data analysis according to claim 1, characterized in that, The acquisition of the hard evidence layer discount evidence and the soft evidence layer discount evidence specifically includes: Read and verify the modality identifier and time window identifier of the reliability coefficient and basic trust allocation. Perform a removal process on records with missing modality identifier or time window identifier. Merge the records that pass the verification according to the modality identifier and output the reliability coefficient and basic trust allocation corresponding to the modality. Read the preset evidence layer division rules and generate evidence layer identifiers for each modality. Based on the evidence layer identifiers, classify the basic trust allocations corresponding to each modality into hard evidence layer basic trust allocations and soft evidence layer basic trust allocations respectively. Perform integrity verification on the classification results and generate hard evidence layer basic trust allocation lists and soft evidence layer basic trust allocation lists. Reliability discounting is performed on the basic trust allocation of the hard evidence layer and the soft evidence layer respectively. Based on the trust quality of each health risk level of each modality, the scaling is performed according to the reliability coefficient corresponding to the modality. The remaining unallocated trust quality after scaling is merged into the trust quality of the ignorance set of the current modality to form discounted evidence. The discounted evidence of each modality in the hard evidence layer and the soft evidence layer is aggregated to obtain the discounted evidence of the hard evidence layer and the discounted evidence of the soft evidence layer respectively.

5. A health risk assessment method based on big data analysis according to claim 1, characterized in that, The determination of the hard-layer fusion evidence and the hard-layer conflict degree specifically includes: The discount evidence and reliability coefficient are matched one by one according to the modality identifier and time window identifier. For the records that are successfully matched, a hard evidence layer processing list is established and the hard evidence layer discount evidence set and hard evidence layer reliability coefficient set are output. The hard evidence layer discount evidence set is traversed pairwise, and the trust quality corresponding to the basic trust assignment is read respectively. The health risk level combination pointed to by the two discount evidences is enumerated item by item. When the health risk levels pointed to by the two discount evidences are not the same, the product of the trust quality corresponding to the combination item is included in the conflict quality of the evidence pair and the accumulation is completed to obtain the hard evidence layer conflict quality set. The hard evidence layer conflict quality set is summarized to obtain the hard evidence layer conflict degree. A preset gate threshold is set, and gate labels are generated for each item in the hard evidence layer reliability coefficient set. Discounted evidence with a reliability coefficient lower than the gate threshold is marked as low-reliability discounted evidence. The conflict quality of low-reliability discounted evidence that has been included in the formation of the hard evidence layer conflict quality set is assigned to the transfer conflict quality set; otherwise, it is assigned to the conflict quality set to be reassigned. The transfer conflict quality set is summarized by evidence pair index, and the summary result is added to the trust quality corresponding to the ignorance set to form the ignorance set increment. The PCR6 conflict redistribution process is invoked to perform a fusion operation on the discounted evidence corresponding to the set of conflict quality to be redistributed. Two pieces of discounted evidence to be fused are selected sequentially according to a preset combination order, and a temporary fusion result is generated. For each piece of discounted evidence to be fused, the trust quality assigned to each preset health risk level is enumerated item by item, and a level-to-level product term is formed. When the health risk levels pointed to by the two pieces of discounted evidence to be fused are not the same, the product term is taken as a conflict term. The conflict term is split into two parts according to the proportion of the trust quality of each piece of discounted evidence to be fused in the conflict term, and then added back to the trust quality of the corresponding health risk level. This completes one PCR6 redistribution. After all conflict terms are split and added back, the fusion result is obtained and used as the input for the next piece of discounted evidence to continue iterative fusion. At the same time, the increment of the ignorance set is accumulated to the trust quality corresponding to the ignorance set in the iterative fusion result. When the remaining pieces of evidence to be fused in the hard evidence layer discounted evidence list are empty, the iterative fusion ends and the hard evidence layer fused evidence is output.

6. The health risk assessment method based on big data analysis according to claim 1, characterized in that, The specific methods for obtaining the soft layer fusion evidence and the soft layer conflict degree include: Read the discounted soft evidence and its corresponding reliability coefficient within the current sliding time window, match them one by one according to the modality identifier and time window identifier, and remove inconsistent records to form a soft evidence processing list; Perform pairwise traversal on the soft evidence layer discount evidence set, read the trust quality corresponding to the basic trust assignment of each piece of evidence, enumerate the level combinations and multiply and sum the trust quality of different level combinations to obtain the conflict quality, write it into the soft evidence layer conflict quality set according to the evidence pair index, and summarize the soft evidence layer conflict quality set to obtain the soft evidence layer conflict degree. Read the gating threshold and generate gating tags. Classify the conflict quality containing low-reliability discount evidence into the transfer conflict quality and sum them up to the unknown set trust quality. Input the evidence corresponding to the remaining conflict quality into the PCR6 rule to perform conflict redistribution. For each conflict item, split it proportionally and add it back to the level trust quality that caused the conflict and iterate until the evidence is exhausted. Output the soft evidence layer fused evidence.

7. A health risk assessment method based on big data analysis according to claim 1, characterized in that, The specific methods for obtaining the inter-layer fusion result and the inter-layer conflict degree include: Verify the consistency of time window identifiers between the hard evidence layer fused evidence and the soft evidence layer fused evidence and complete the alignment, and establish inter-layer fused input evidence pairs; For inter-layer fusion input evidence pairs, calculate the inter-layer conflict quality. When the health risk levels pointed to by the two are not the same, the corresponding trust quality product is included in the conflict item and accumulated to obtain the inter-layer conflict quality, which is then written into the inter-layer conflict quality record. The validity of the fused evidence at the hard evidence layer is determined. When the conditions for the validity of hard evidence are met, a valid mark is generated; when the conditions for the validity of hard evidence are not met, an invalid mark is generated. When a valid marker is determined, inter-layer bias processing is performed. The preset inter-layer bias strategy is read and hard evidence layer bias weights are generated. The trust quality in the hard evidence layer fused evidence is amplified according to the hard evidence layer bias weights. The amplified trust quality and the trust quality corresponding to the set of ignorance are normalized. The biased hard evidence layer fused evidence is then output. Interlayer PCR6 fusion processing is performed. When a label is determined to be valid, the fused evidence from the biased hard evidence layer and the fused evidence from the soft evidence layer are used as fusion inputs. The conflict items of the two are split and added back according to the PCR6 ratio to obtain the interlayer fusion result. At the same time, the interlayer conflict quality is output as the interlayer conflict degree. When a label is determined to be invalid, the interlayer conflict quality is transferred to the trust quality corresponding to the unknown set and the allocation ratio of the unknown set is increased to obtain the interlayer fusion result.

8. The health risk assessment method based on big data analysis according to claim 1, characterized in that, The output of the risk level and risk confidence level specifically includes: Write the inter-layer conflict degree into the conflict degree sequence according to the time window identifier, and retrieve the inter-layer conflict degree of the sliding time window to form a conflict degree history sequence; Read the preset conflict degree threshold, preset conflict increment threshold and preset continuous window number threshold, compare the current inter-layer conflict degree with the preset conflict threshold to generate a trend mark, calculate the increment of the inter-layer conflict degree of adjacent layers in the conflict degree history sequence and compare it with the preset conflict increment threshold to generate an increment mark, count the increment marks in the most recent continuous window and output the conflict trend criterion result; The trust quality is read from the inter-layer fusion results. The preset health risk level with the highest trust quality is selected as the risk level output. The trust quality of the ignorance set and the inter-layer conflict degree are converted into confidence decay factors and combined to obtain the risk confidence level.

9. A health risk assessment system based on big data analysis, comprising executing the health risk assessment method based on big data analysis as described in any one of claims 1 to 8, characterized in that, include: The data access standardization module is used to access multi-source health data of target users and unify fields, codes, time bases and units to output standardized health data. The time window sample construction module is used to construct sliding time windows based on standardized health data, divide records by time windows and merge them by data source to form multimodal time series samples; The feature quality modeling module is used to generate record sequences of multimodal time series samples according to modality, extract statistics, trend and fluctuation features to form modal feature vectors, and calculate reliability coefficients; The basic trust allocation module is used to input modal feature vectors into a scoring function to obtain the support for each health risk level, scale it in combination with a reliability coefficient, and allocate the remaining quality to the set of ignorance. The layered discount gating module is used to divide the basic trust allocation into a hard evidence layer and a soft evidence layer and perform reliability discounting, calculate the conflict quality, and transfer low-reliability conflicts to the set of ignorance; The PCR6 fusion and risk output module is used to perform PCR6 conflict redistribution and interlayer fusion on hard and soft layers, and output the risk level and risk confidence.

Images