Frailty assessment method and system based on multi-source data fusion for the elderly

Abstract

The application discloses a senile frailty evaluation method and system based on multi-source data fusion, and establishes data space nodes at a hospital end, a community end and / or a home end, gathers multi-source data about senile frailty evaluation through access to standardized interfaces, realizes unified semantics of data flow through field mapping, unit conversion and time alignment, and realizes cross-institution data association through privacy protection identity alignment; then, data quality scores of each index are calculated, and evidence credibility weights are generated; then, a state space model is used to perform weighted Bayesian updating on frailty hidden variables, output frailty state posterior distribution and frailty evolution curve; the application further constructs a time-lag causal Bayesian network to predict future frailty prognosis probability and identify key influence paths; individualized intervention suggestions are generated under safety constraints in combination with an intervention knowledge base; the application can realize continuous frailty evaluation, risk prediction and explainable intervention cooperation, and improve the robustness and accuracy of cross-scene management.

Description

Technical Field

[0001] This invention relates to medical or health data processing technology and elderly healthcare information technology, and in particular to a method and system for assessing frailty in the elderly based on multi-source data fusion. Background Technology

[0002] Frailty, a common geriatric syndrome, is characterized by a decline in bodily reserves and reduced tolerance to internal and external stresses. It is significantly associated with the risk of falls, hospitalization, disability, and death. Therefore, establishing a closed loop for early screening, assessment, follow-up, and intervention of frailty in communities and medical institutions is of great significance for improving the quality of elderly health management.

[0003] Among the currently disclosed technologies, one type of approach centers on scales or frailty indices (FI), collecting information on health status, daily living abilities, psychological state, and cognitive function to score and output reports. Some even incorporate voice interaction to lower the barrier to entry. This method is simple to implement, but typically relies on a data collection, scoring, and reporting workflow. This approach struggles to fully utilize continuous data from hospitals, communities, homes, and wearable devices, and rarely depicts the long-term evolution and uncertainty of frailty. Another type of approach targets community health checkup scenarios, calculating health status, activity level, and nutritional status values ​​from checkup data, exercise data, and nutritional intake data respectively. These values ​​are then weighted and summed to obtain an assessment value, which is then mapped to the degree of frailty through preset intervals. This type of approach can cover multidimensional factors to some extent, but its fusion method often uses fixed weights or weakly adaptive weights, making it difficult to handle cross-institutional data conflicts, missing data, and noise. Furthermore, the output is often discrete and hierarchical, making it difficult to form a continuous and updatable frailty trajectory.

[0004] In addition to the two types of approaches mentioned above, there are also approaches that take a single-modal approach. For example, 3D keypoint reconstruction based on multi-view video gait and LSTM temporal action classification can extract motion parameters and map frailty levels; or indices can be calculated based on specific biomarkers (such as peripheral blood mononuclear cell mitochondrial metabolic indicators) for risk assessment. These methods are relatively objective, but often focus on a single dimension, making it difficult to integrate with factors such as chronic disease management, medication burden, nutrition, psychological and environmental safety, which may lead to assessment bias or insufficient interpretability.

[0005] Furthermore, there are existing health risk assessment frameworks that integrate evaluation decision-making, risk prediction, and intervention recommendations. These frameworks select scales / scenarios / direct assessment methods through rules and statistical learning models, and then use graph neural networks to predict future risks and provide intervention options. However, while these frameworks are broad in scope, they are not typically designed specifically around the continuous evolution and regression mechanisms of frailty. They lack reliable modeling of the quality of multi-source data and interpretable inferences about critical paths to frailty regression, making it difficult to establish verifiable risk-benefit evidence for intervention recommendations.

[0006] In summary, existing technologies for assessing frailty in older adults generally suffer from the following shortcomings:

[0007] (1) Data silos exist across multiple institutions and scenarios, making continuous management across institutions difficult;

[0008] (2) The quality of multi-source heterogeneous data varies and there are conflicts, missing data and noise. The lack of credibility modeling leads to the instability of the fusion results.

[0009] (3) Most methods focus on single assessment or discrete grading, lacking characterization and real-time updates of the "continuous trajectory" of the weakened state.

[0010] (4) The lack of mechanistic explanations for the impact pathways and key nodes in prediction and intervention makes it difficult to support clinical decision-making and hierarchical management of public health.

[0011] To address the aforementioned issues, this invention proposes a frailty assessment method and system for the elderly based on multi-source data fusion. This approach introduces data quality scoring and evidence credibility weights on the basis of cross-institutional data interconnection, uses a continuous latent variable state-space model to generate frailty evolution curves, and combines a time-delayed causal Bayesian network to achieve outcome probability inference and critical path identification. Under safety constraints, it outputs individualized intervention recommendations, thereby improving the continuity, robustness, and interpretability of frailty assessment. Summary of the Invention

[0012] In view of this, the purpose of this invention is to propose a frailty assessment method and system for the elderly based on multi-source data fusion that can improve the continuity, robustness and interpretability of frailty assessment.

[0013] To achieve the above-mentioned technical objectives, the technical solution adopted by this invention is as follows:

[0014] A frailty assessment method for the elderly based on multi-source data fusion, comprising:

[0015] S01. Establish data space nodes at the hospital, community, and / or home ends respectively, and then access multi-source health data about the target object through a standardized interface, and preprocess the multi-source health data to obtain a data stream with unified semantics;

[0016] S02. Perform cross-node privacy-preserving identity alignment on the multi-source health data to associate cross-node data belonging to the same target object into the same object data set;

[0017] S03. Calculate the data quality score for each data indicator in the object data set, and determine the evidence credibility weight of the corresponding data based on the data quality score;

[0018] S04. Construct a decay assessment feature vector based on the object data set, wherein the decay assessment feature vector includes multiple sub-features characterizing the health status of the target object;

[0019] S05. Based on the state-space model, the weakened state of the target object is represented as a continuous latent variable that evolves over time. The weakened assessment feature vector is then updated using a weighted Bayesian method according to the evidence credibility weights to obtain the posterior distribution of the weakened state of the target object at each time point and the corresponding weakened evolution curve. Then, based on the posterior distribution of the weakened state and the corresponding weakened evolution curve, a weakened assessment result containing the weakened state and weakened evolution trend of the target object is generated and output.

[0020] As a possible implementation, further, in this solution S01, the data sources of the multi-source health data include hospital information system data, community health information system data, laboratory information system data and / or home monitoring system data, which respectively include one or more of the following: medication record data, physical examination data, chronic disease follow-up data, wearable device data, and home environment and security IoT data.

[0021] As a possible implementation, further, in this solution S01, the preprocessing involves mapping fields, converting units, and aligning time for the multi-source health data to obtain a data stream with unified semantics.

[0022] As a possible implementation, further, in this solution S02, the privacy-protected identity alignment is to perform irreversible desensitization encoding on the identity identifier of the target object, and to complete the association and / or aggregation of the same target object data between different data space nodes based on the desensitization encoding.

[0023] As a preferred implementation option, preferably, in this scheme S03, the data quality score is calculated from at least two or more of the following: integrity score, timeliness score, consistency score, and equipment reliability score, and the data quality score is mapped to the evidence credibility weight of the corresponding data.

[0024] In S01, the multi-source health data, after preprocessing, yields a data stream with unified semantics, defined as follows: The target object In time At that time, from the data source The data obtained was acquired after preprocessing.

[0025] As a preferred implementation option, in this scheme S02, the function that associates cross-node data belonging to the same target object into the same object data set is defined as follows:

[0026]

[0027] in, For target object In time A collection of objects at time. As a data source, For data source collection, For target object In time At that time, from the data source The data stream generated after preprocessing the obtained data, i.e., the preprocessed data subset, has more than one data feature, and each data feature has its corresponding associated data sequence, i.e., observation features. It is represented as the target object. In time The first time Data features, i.e., data characteristics Observational characteristics.

[0028] As a preferred implementation option, preferably, in this scheme S03, the function definition of the integrity score is as follows:

[0029]

[0030] in, For target object In time Time-related data characteristics Integrity score, For target object In time Time-related data characteristics The percentage of missing values.

[0031] The function for the timeliness score is defined as follows:

[0032]

[0033] in, For target object In time Time-related data characteristics Timeliness score For target object The data feature collection time corresponding to the most recent timeliness rating; This is the time constant for age decay.

[0034] The function for the consistency score is defined as follows:

[0035]

[0036] in, For target object In time Time-related data characteristics Consistency score, For data features The corresponding associated data sequence, i.e., the observed features, , The target objects are respectively Data characteristics The mean and standard deviation over a preset historical period It is a constant value, used to prevent the denominator from being 0.

[0037] In this solution, when the data characteristics in the object data set... When data includes wearable device data and / or home environment and security IoT data, the data quality score includes a device reliability score, the function of which is defined as follows:

[0038]

[0039] in, For target object In time Time-related data characteristics The corresponding equipment reliability score, For target object In time At that time, its data characteristics The percentage of data containing anomalies.

[0040] In this scheme, the function for data quality scoring is defined as follows:

[0041]

[0042] in, For target object In time Time-related data characteristics The data quality score, whose value is truncated by the function. Truncate to [0, 1], , , , These are the weighting coefficients, and their sum is 1.

[0043] Based on the above, the function that maps the data quality score to the corresponding data's evidence credibility weight is defined as follows:

[0044]

[0045] in, For target object In time At that time, its included data characteristics The weight of the credibility of the evidence; , These are the lower and upper limits of the evidence credibility weight, respectively, which satisfy... , This is the nonlinear strengthening coefficient, and its value is greater than 1.

[0046] As a preferred implementation option, in S04 of this scheme, a decay assessment feature vector is constructed based on the object data set by concatenating data features, and its definition is... , that is, the observation vector.

[0047] Wherein, the decay assessment feature vector Data characteristics Corresponding observation features The numerical features were also normalized to obtain normalized observation features. .

[0048] As a preferred implementation option, S04 of this scheme further includes: pooling the evidence credibility weights of the data features contained in the attenuation assessment feature vector to form a composite data structure with the attenuation assessment feature vector. Associate the corresponding data weight vector .

[0049] As a preferred implementation option, preferably, in this scheme S04, the sub-features are one or more of the following: demographic and socioeconomic characteristics, lifestyle and nutritional characteristics, chronic disease management characteristics, previous medication burden characteristics, laboratory test characteristics, physical examination characteristics, functional assessment characteristics, wearable continuous monitoring of physiological parameters characteristics, and home environment and safety characteristics.

[0050] The aforementioned characteristics of prior medication burden include one or more of the following: multiple medication quantity indicators, potential inappropriate medication indicators, and drug interaction risk indicators.

[0051] The wearable continuous monitoring physiological parameter characteristics include one or more of the following: heart rate variability, sleep efficiency, daily activity level and activity rhythm, gait stability, or gait speed variability.

[0052] As a preferred implementation option, preferably, in this scheme S05, the weighted Bayes update includes: exponentially weighting the likelihood term of each observation indicator according to its evidence credibility weight, thereby reducing the impact of low-credibility observations on the posterior distribution of the weakened state.

[0053] As a preferred implementation option, this solution S05 preferably includes:

[0054] Based on the state-space model, the decay state of the target object is represented as a continuous latent variable that evolves over time, and the corresponding state transition model function is defined as follows:

[0055]

[0056] in, , The target objects are respectively In time , The continuous decaying hidden state at time; For the target object In time The driving factor vector corresponds to the sub-features in the decay assessment features; For the driving coefficient vector, In time The process noise, used to characterize unobservable disturbances, follows a normal distribution. , This represents the process variance.

[0057] The observation model function is defined as follows:

[0058]

[0059] in, For target object In time At that time, its included data characteristics The normalized observed features, i.e., the decay assessment feature vector. Normalized feature subset; For the target object In time The continuous decaying hidden state at time; For the first A mapping function between an observation feature and a continuous weakened hidden state, which is used to map the continuous weakened hidden state to each observation index. For data features In time The process noise follows a normal distribution. , For observation features The noise variance.

[0060] Based on the above techniques, a weighting of evidence credibility can be introduced. The function for weighted Bayesian update of the decay assessment feature vector based on the evidence credibility weights is defined as follows:

[0061]

[0062] in, , The target objects are respectively In time , The continuous decaying hidden state at time; For the target object From time 1 to time The observation vector at time, i.e., the decay assessment feature vector, has a dimension of . Data vectors; , The target objects are respectively In time At that time, it contains the first Evidence credibility weights for each data feature, and normalized observation features; For the target object From time 1 to time Given all observed vectors, its time Continuous decaying hidden state The posterior distribution of; The state transition probability, i.e., the probability of a continuously decaying hidden state from time... arrive The evolutionary pattern, i.e., prior information; For the first The likelihood of each data feature represents the likelihood in the true state. Under these conditions, observed The probability of.

[0063] Where the likelihood follows a normal distribution It satisfies the following definition:

[0064]

[0065] Then its probability density is:

[0066]

[0067] That is, it is equivalent to adding the evidence confidence weight to the observation variance. When the confidence is lower, the noise of the equivalent observation variance is greater, and the contribution of the observation feature is smaller when updating.

[0068] in, For the first Observation features The mapping function with continuously decaying hidden states, For observation features The noise variance.

[0069] The posterior distribution of the weakened state of the target object at each time point is defined as follows:

[0070]

[0071] in, For the target object In time At that time, it corresponds to a continuously decaying hidden state. The point estimate is the mean of the posterior distribution; For conditional expectation, Represented as in a given observation vector Under the condition of continuous decay hidden state Expected estimate.

[0072] For uncertainty, i.e., the posterior variance, the function is defined as follows:

[0073]

[0074] For the target object In time The uncertainty of the state estimation at time is given by the variance of the posterior distribution; For conditional variance, Represented as in a given observation vector Under the condition of continuous decay hidden state The degree of fluctuation in the estimated value.

[0075] Among them, in time 1 to time period The definition of the decay evolution curve is: That is, time period target object Point estimate The time series.

[0076] As a preferred implementation option, S05 of this scheme further includes: performing change point detection on the weakening evolution curve to identify sudden deterioration events in the weakening state of the target object and generate early warning information.

[0077] The function for detecting change points is defined as follows:

[0078]

[0079] in, , The target objects are respectively In time , The mean of the posterior distribution at time; This represents the posterior distribution change value.

[0080] Based on the above, the posterior distribution change value is compared with the preset deterioration threshold. In comparison, if When this occurs, a sudden deterioration event is triggered, generating and outputting a warning message.

[0081] As a preferred implementation option, this solution also includes:

[0082] S06. Based on the decay evolution curve and the decay assessment feature vector, construct a time-delay causal Bayesian network, perform probability inference to output the probability of decay and deterioration of the target object within a preset time range in the future and its key influence path, generate prediction results, and then aggregate them into the decay assessment results for output.

[0083] The topology learning of the time-delayed causal Bayesian network is combined with domain constraint rules to prohibit candidate connection edges that violate the temporal causal direction.

[0084] As a preferred implementation option, preferably, in this scheme S06, the key influence path is obtained by performing sensitivity analysis on the nodes in the time-delay causal Bayesian network. The sensitivity analysis is used to quantify the contribution of changes in each node to the probability of weakening and deterioration.

[0085] As a preferred implementation option, solution S06 preferably includes:

[0086] Extract at least one sub-feature related to the decay assessment from the decay assessment feature vector as a key influencing factor, denoted as _____. Then, extract the point estimates of the continuous decay hidden states corresponding to the decay assessment feature vector from the decay evolution curve, and set them as follows: The time-delay causal Bayesian network is a dynamic network, which is a dynamic network that is time-delayed. node combination Defined as follows:

[0087]

[0088] in, This is the key influencing factor number, i.e., the node number.

[0089] A time direction constraint is introduced into the edges of the time-delay causal Bayesian network to ensure that intra-piece and inter-piece edges satisfy the following definition:

[0090] The film satisfies the following:

[0091] or

[0092] The space between the pieces satisfies:

[0093] , ,

[0094] in, , Number the nodes. , They are nodes In time , Key influencing factors; For nodes In time Key influencing factors.

[0095] The constraints of the time-delay causal Bayesian network also include prohibiting obvious anti-causal edges, that is, not allowing directed edges whose time direction points to the past.

[0096] Construct a scoring function, defined as follows:

[0097]

[0098] in, For training data; Candidate graph structure; In candidate graph structure The following network parameters, The number of parameters; The number of samples; This is the scoring function.

[0099] Based on the scoring function, the candidate graph structure with the highest score that satisfies the constraints is defined as follows: , Represented as candidate graph structure The set of nodes, Represented as candidate graph structure The set of directed edges;

[0100] For discrete nodes, we use conditional probability tables (CPT) for learning, which are defined as follows:

[0101]

[0102] For continuous nodes, conditional Gaussian or piecewise linear methods are used to transform the problem into a discrete one, which is then processed using CPT. As a network node, it is or ; The set of parent nodes; Represents network nodes The conditional probability distribution; Represents network nodes The conditional probability table.

[0103] Based on the point estimate of the continuously decaying hidden state, the deterioration event in the future time H is defined as... ,in, To worsen the threshold, For time The point estimate of the continuously weakened hidden state is then defined as follows:

[0104]

[0105] in, For the target object From time Starting with predicting the step size in the future. The probability of a worsening event occurring within the organization; For the target object In time Point estimates of the continuously weakening hidden states; For the target object In time The evidence set includes point estimates of key influencing factors and continuously weakening latent states.

[0106] Sensitivity analysis was performed on each key influencing factor, and the formula is defined as follows:

[0107]

[0108] in, For nodes In time Key influencing factors; This indicates the key impact factors. Artificial intervention to a preset reference health level Later, in the evidence set Under the condition that remains unchanged, the posterior probability of the risk of deterioration; For nodes in a dynamic network The node contribution, if The larger the value, the more likely it is to be a node. Key influencing factors The greater the contribution to the risk of deterioration, the more critical the influencing factor will be when it exceeds the preset value. arrive The in-slice edges are defined as high-contribution paths, which are then used as key impact paths and aggregated into the decay assessment results for output.

[0109] As a preferred implementation option, this solution also includes:

[0110] S07. Based on the posterior distribution of the current frailty state, the probability of frailty worsening, the key influencing paths, and the pre-built intervention knowledge base, generate individualized intervention suggestions, and then aggregate them into the frailty assessment results for output.

[0111] The individualized intervention recommendations are generated in accordance with safety constraints, which include one or more of the following: the target subject's chronic disease status, recent history of falls, risk of malnutrition, and risk of drug interactions.

[0112] Based on the above, this solution also proposes an elderly frailty assessment system based on multi-source data fusion, which applies the aforementioned elderly frailty assessment method based on multi-source data fusion, including:

[0113] The data access module is used to establish data space nodes at the hospital, community and / or home end, and then access multi-source health data about the target object through a standardized interface.

[0114] The data semantic standardization module is used to preprocess the multi-source health data accessed by the data access module to obtain a data stream with unified semantics; the preprocessing includes one or more of the following: field mapping, unit conversion, and time alignment.

[0115] The data identity alignment module is used to perform cross-node privacy-preserving identity alignment on the preprocessed multi-source health data, so as to associate cross-node data belonging to the same target object into the same object data set;

[0116] The quality assessment module is used to calculate data quality scores for each data indicator in the object dataset, and to determine the evidence credibility weight of the corresponding data based on the data quality scores.

[0117] The feature construction module is used to construct a decay assessment feature vector based on the object dataset, wherein the decay assessment feature vector includes multiple sub-features characterizing the health status of the target object;

[0118] The decay state estimation module represents the decay state of the target object as a continuous latent variable that evolves over time based on the state space model. It then performs a weighted Bayes update on the decay assessment feature vector according to the evidence credibility weight to obtain the posterior distribution of the decay state of the target object at each time point and the corresponding decay evolution curve. Finally, it generates and outputs a decay assessment result containing the decay state and decay evolution trend of the target object based on the posterior distribution of the decay state and the corresponding decay evolution curve.

[0119] As a preferred implementation option, the system described in this solution preferably also includes:

[0120] The trajectory prediction and interpretation module is used to construct a time-delay causal Bayesian network based on the decay evolution curve and the decay assessment feature vector, and to perform probability inference to output the decay and deterioration probability of the target object within a preset time range and its key influence path, generate prediction results, and then aggregate them into the decay assessment results for output; wherein, the topology learning of the time-delay causal Bayesian network is combined with domain constraint rules to prohibit candidate connection edges that violate the temporal causal direction;

[0121] The intervention recommendation module is used to generate individualized intervention suggestions based on the posterior distribution of the current frailty state, the probability of frailty deterioration, the key impact paths, and the pre-built intervention knowledge base, and then aggregate them into the frailty assessment results for output.

[0122] The multi-terminal interaction module is used to output frailty assessment results, prediction results and / or individualized intervention suggestions to the doctor's terminal, the caregiver's terminal and / or the elderly's terminal, respectively.

[0123] The data access module includes an interface adaptation unit, which is used to interconnect with hospital information systems, community health information systems, laboratory information systems, and IoT platforms.

[0124] The decay state estimation module also includes a change point detection unit, which is used to identify abrupt deterioration events in the decay evolution curve and trigger an early warning.

[0125] The trajectory prediction and interpretation module includes a structure learning unit and a probability inference unit. The structure learning unit determines the topology of the time-delay causal Bayesian network by combining domain constraint rules, and the probability inference unit outputs the probability of decay and deterioration within a future preset time range.

[0126] The intervention recommendation module includes a safety constraint unit, which is used to limit the content of the intervention plan based on factors such as chronic disease status, fall risk, and medication interaction risk.

[0127] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art: This solution establishes data space nodes at the hospital, community, and / or home levels, and then accesses multi-source health data about the target object through standardized interfaces. This enables continuous, robust, and interpretable assessment of frailty in the elderly across multiple scenarios in hospitals, communities, and homes, significantly improving the scientific rigor and feasibility of frailty screening and management. Through distributed data space nodes and standardized interfaces, interconnection and interoperability with hospital information systems, community health information systems, laboratory information systems, and IoT device platforms can be achieved. This allows for the continuous and flexible unified aggregation and standardized processing of data related to frailty assessment in the elderly, fundamentally alleviating the difficulties in continuous management caused by cross-institutional data silos, inconsistent fields, and inconsistent time granularity.

[0128] This solution also introduces a data quality scoring and evidence credibility weighting mechanism. For real-world data issues such as missing data, delays, noise, device drift, and cross-source conflicts, the effectiveness of each indicator is quantified and the interference of low-quality observations is adaptively suppressed in subsequent inferences, making the fusion results more robust and more in line with clinical practice.

[0129] This scheme also models the weakening state as a continuous latent variable that evolves over time, and uses weighted Bayesian updates to output the posterior distribution of the weakening state and the weakening evolution curve. It can not only provide the current weakening level, but also provide information on assessment uncertainty and trend changes at the same time, thereby realizing dynamic monitoring and early warning of the weakening process of deterioration, stabilization or improvement, avoiding the problem that traditional single scoring or fixed threshold grading is difficult to reflect longitudinal changes.

[0130] Furthermore, this solution constructs a time-delayed causal Bayesian network to make probabilistic inferences about the decay outcome based on time-period data, identifies key nodes and critical paths affecting the decay process, and provides a mechanism-level explanatory basis for the technical objective of "why it deteriorates and where to intervene".

[0131] In addition to the above, this solution can also generate individualized intervention recommendations based on the prediction and retrospective results, combined with a pre-built intervention knowledge base and safety constraints. This enables the assessment, prediction and intervention to form a continuous and precise management process, which can effectively improve the pertinence, feasibility and verifiability of interventions, reduce the risk of frailty-related adverse events and improve the efficiency of collaborative management between communities and medical institutions. Attached Figure Description

[0132] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0133] Figure 1 This is a simplified implementation flowchart of one of the methods for assessing the weakness of this scheme.

[0134] Figure 2 This is a simplified implementation flowchart of the second implementation method of the weakness assessment method in this scheme;

[0135] Figure 3 This is a simplified implementation flowchart of the third implementation method of the weakness assessment method in this scheme;

[0136] Figure 4 This is a schematic diagram of the unit module connection of one of the implementation methods of the decay assessment system of this scheme;

[0137] Figure 5 This is a schematic diagram of the unit module connection for the second implementation method of the decay assessment system of this scheme. Detailed Implementation

[0138] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0139] Combination Figure 1 As shown in the figure, this embodiment of the scheme is a frailty assessment method for the elderly based on multi-source data fusion, which includes:

[0140] S01. Establish data space nodes at the hospital, community, and / or home ends respectively, and then access multi-source health data about the target object through a standardized interface, and preprocess the multi-source health data to obtain a data stream with unified semantics;

[0141] S02. Perform cross-node privacy-preserving identity alignment on the multi-source health data to associate cross-node data belonging to the same target object into the same object data set;

[0142] S03. Calculate the data quality score for each data indicator in the object data set, and determine the evidence credibility weight of the corresponding data based on the data quality score;

[0143] S04. Construct a decay assessment feature vector based on the object data set, wherein the decay assessment feature vector includes multiple sub-features characterizing the health status of the target object;

[0144] S05. Based on the state-space model, the weakened state of the target object is represented as a continuous latent variable that evolves over time. The weakened assessment feature vector is then updated using a weighted Bayesian method according to the evidence credibility weights to obtain the posterior distribution of the weakened state of the target object at each time point and the corresponding weakened evolution curve. Then, based on the posterior distribution of the weakened state and the corresponding weakened evolution curve, a weakened assessment result containing the weakened state and weakened evolution trend of the target object is generated and output.

[0145] In this solution, data spatial nodes are established at the hospital, community, and / or home levels, which enables the data source for frailty assessment of the elderly to have a wider coverage and greater continuity, thereby improving the reliability of the assessment results. Correspondingly, in S01 of this solution, the data sources for the multi-source health data of the target object accessed through standardized interfaces include hospital information system data, community health information system data, laboratory information system data, and / or home monitoring system data, which respectively include one or more of the following: medication record data, physical examination data, chronic disease follow-up data, wearable device data, and home environment and safety IoT data.

[0146] As an example, in hospitals, data sources can include hospital information system data and laboratory information system data, while the data types included in the relevant systems can be medication record data, physical examination data, etc.

[0147] At the community level, data sources can include data from community health information systems and testing information systems, and the data types included in these systems can be medication records, physical examination data, chronic disease follow-up data, etc.

[0148] At home, the data sources are mostly IoT devices deployed in homes and wearable devices (such as wristbands, watches or other instruments with health information monitoring). The data types include wearable device data, home environment and security IoT data, as well as self-retained and backed-up medication records, physical examination data, chronic disease follow-up data, etc.

[0149] Since the data stored in the data sources of different data space nodes may have some similarities, deduplication or merging can be achieved through matching. However, the obstacles to data sharing between different data sources are more due to differences in the semantic form of the records. As a possible implementation method, in this solution S01, the preprocessing is to perform field mapping, unit conversion and time alignment on the multi-source health data to obtain a data stream with unified semantics.

[0150] By forming a data stream using a unified semantic approach, the multi-source health data exhibits greater consistency and reliability in data sharing.

[0151] As an example, in this solution, for each data source You can define a field mapping function to map the original data ( The information is mapped to a unified feature set. In this context, the information transformation is completed, and the field mapping function is expressed as follows:

[0152]

[0153] Field mapping function After processing, the data source The original field names, codes, scale entries, and other information of the raw data are mapped to the decay assessment element dictionary, i.e., standard fields; feature set ( () is a unified set of semantic fields, which may include grip strength, walking speed, BMI, HbA1c, ADL, number of medications taken, number of times one gets up at night, fall events, etc.

[0154] For unit conversion, an affine transformation can be applied to the convertible index, and its function is defined as follows:

[0155]

[0156] in, For data source The original data, i.e., the original records. Data in standard units, i.e., standard data after unifying units. , Data sources In data metrics The unit conversion factor can be configured by the unit dictionary, which is existing technology and will not be elaborated here.

[0157] In addition, this solution also performs time alignment processing on the multi-source health data to form a unified time grid, so that the multi-source health data have a high degree of consistency. This is existing technology, and this solution will not be described in detail here.

[0158] In S01, the multi-source health data, after preprocessing, yields a data stream with unified semantics, defined as follows: The target object In time At that time, from the data source The data obtained was acquired after preprocessing.

[0159] In this solution, since the multi-source health data comes from different data space nodes and data sources, there may be cases where the same elderly person (or target object) has multiple sets of IDs, and directly aligning with ID card / mobile phone number would trigger privacy compliance risks. Therefore, this solution completes same-person matching without exposing plaintext identity. As a possible implementation method, further, in S02 of this solution, the privacy-protected identity alignment involves performing irreversible de-identification encoding on the target object's identity identifier, and completing the association and / or aggregation of the same target object's data between different data space nodes based on the de-identification encoding.

[0160] As an example, this scheme performs irreversible encoding on the identity identifier of each object (such as ID number, mobile phone number, name + birthday combination), and its function definition is as follows:

[0161]

[0162] in, It is a one-way hash function (such as the SHA family). Use a random string (to prevent rainbow table attacks). For data concatenation operations; A unified identifier across organizations with privacy protection.

[0163] After performing irreversible desensitization encoding on the target object's identity identifier, it is possible to base the desensitization encoding (i.e., using the desensitization encoding) on ​​the target object's identity identifier. (as a key) to complete the association and / or aggregation of the same target object data between different data space nodes.

[0164] As a preferred implementation option, in this scheme S02, the function that associates cross-node data belonging to the same target object into a single object data set is defined as follows:

[0165]

[0166] in, For the target object In time A collection of objects at time. As a data source, For data source collection, For the target object In time At that time, from the data source The data stream generated after preprocessing the obtained data, i.e., the preprocessed data subset, has one or more data characteristics. Each data feature Each has its corresponding associated data sequence, i.e., observation features. It is represented as the target object. In time The first time Data features, i.e., data characteristics Observational characteristics.

[0167] Because multi-source data may contain missing data, noise, equipment drift, cross-institutional conflicts, expired data, etc., if the reliability of the data is evaluated, the fusion model will be interfered with by low-quality data, which may lead to abnormal results such as misjudgment and false warning.

[0168] Therefore, as a preferred implementation option, preferably, in this scheme S03, the data quality score is calculated from at least two or more of the following: integrity score, timeliness score, consistency score, and equipment reliability score, and the data quality score is mapped to the evidence credibility weight of the corresponding data.

[0169] As a preferred implementation option, preferably, in this scheme S03, the function definition of the integrity score is as follows:

[0170]

[0171] in, For the target object In time Data characteristics of time Integrity score, For the target object In time Time-related data characteristics The percentage of missing values.

[0172] The function for the timeliness score is defined as follows:

[0173]

[0174] in, For the target object In time Time-related data characteristics Timeliness score For the target object The data feature collection time corresponding to the most recent timeliness rating; The aging decay time constant can be 15 days or 30 days, for example.

[0175] The function for the consistency score is defined as follows:

[0176]

[0177] in, For the target object In time Time-related data characteristics Consistency score, For data features The corresponding associated data sequence, i.e., the observed features, , The target objects are respectively Data characteristics In this scheme, the mean and standard deviation over a preset historical period can be obtained using robust estimation methods. It is a constant value, used to prevent the denominator from being 0.

[0178] In this solution, when the data characteristics in the object data set... When data includes wearable device data and / or home environment and security IoT data, the data quality score includes a device reliability score, the function of which is defined as follows:

[0179]

[0180] in, For the target object In time Time-related data characteristics The corresponding equipment reliability score, For the target object In time At that time, its data characteristics The percentage of data containing anomalies.

[0181] In this scheme, the function for data quality scoring is defined as follows:

[0182]

[0183] in, For the target object In time Time-related data characteristics The data quality score, whose value is truncated by the function. Truncate to [0, 1], , , , These are the weighting coefficients, and their sum is 1.

[0184] Based on the above, the function that maps the data quality score to the corresponding data's evidence credibility weight is defined as follows:

[0185]

[0186] in, For the target object In time At that time, its data characteristics The weight of the credibility of the evidence; , These are the lower and upper limits of the evidence credibility weight, respectively, which satisfy... , The nonlinear strengthening coefficient has a value greater than 1. In this scheme, the nonlinear strengthening coefficient... A value greater than 1 can reduce the weight of low- and medium-quality data and make high-quality data stand out, thereby improving the quality of data participating in the fusion evaluation process.

[0187] After processing in steps S01 and S02, this scheme can obtain a data stream with unified semantics. However, its aggregation method is still relatively messy. Therefore, as a better implementation option, preferably, in S04 of this scheme, based on the object data set, a decay assessment feature vector is constructed by concatenating data features, which is defined as follows: , that is, the observation vector.

[0188] Wherein, the decay assessment feature vector Data characteristics Corresponding observation features The numerical features were also normalized to obtain normalized observation features. .

[0189] As a preferred implementation option, S04 of this scheme further includes: pooling the evidence credibility weights of the data features contained in the attenuation assessment feature vector to form a composite data structure with the attenuation assessment feature vector. Associate the corresponding data weight vector .

[0190] Regarding the data structure of the frailty assessment feature vector, in this scheme S04, the sub-features are one or more of the following: demographic and socioeconomic characteristics, lifestyle and nutritional characteristics, chronic disease management characteristics, past medication burden characteristics, laboratory test characteristics, physical examination characteristics, functional assessment characteristics, wearable continuous monitoring of physiological parameters characteristics, and home environment and safety characteristics.

[0191] The aforementioned characteristics of prior medication burden include one or more of the following: multiple medication quantity indicators, potential inappropriate medication indicators, and drug interaction risk indicators.

[0192] The wearable continuous monitoring physiological parameter characteristics include one or more of the following: heart rate variability, sleep efficiency, daily activity level and activity rhythm, gait stability, or gait speed variability.

[0193] As an example, the attenuation assessment feature vector constructed by the scheme in this embodiment can be defined as follows:

[0194]

[0195] Among them, the subscripts of each feature , Represents the target object In time Data, for example: Represented as target object In time The observation vector, i.e., the weakening assessment feature vector, has the following meanings for its sub-features when the subscript is ignored:

[0196] These are demographic and socioeconomic characteristics, which may include information such as age, gender, living alone, education, and income.

[0197] l represents lifestyle and nutritional characteristics, which may include information and data such as intake structure, alcohol and smoking consumption, protein intake, and MNA-SF.

[0198] c represents chronic disease management characteristics, which may include information such as the number of diseases, control status, and history of acute exacerbations.

[0199] m represents the characteristics of past medication burden, which may include information such as the number of medications used, PIM, and drug interaction risks.

[0200] b refers to laboratory test characteristics, which may include information such as Hb (hemoglobin), albumin, CRP, and renal function.

[0201] p represents physical examination characteristics, which may include information such as BMI, blood pressure, grip strength, and walking speed;

[0202] f represents functional assessment characteristics, which may include information and data such as ADL / IADL, TUG, and cognitive / emotional scales.

[0203] w represents the wearable continuous monitoring of physiological parameters, which may include information such as HRV, sleep efficiency, activity rhythm, and gait variability.

[0204] h represents the home environment and safety characteristics, which may include information such as fall risk and abnormal temperature and humidity.

[0205] Correspondingly, this scheme corresponds to the aforementioned attenuation assessment feature vector. The data dimension aggregates the evidence credibility weights of the data features contained in the decay assessment feature vector to form a data dimension that is consistent with the decay assessment feature vector. Associate the corresponding data weight vector Its function is defined as follows:

[0206]

[0207] in, For the total dimensions of the data, Represents the eigenvector of the weakening assessment The Middle Weighting of the credibility of evidence in the data dimension.

[0208] Traditional attenuation assessment schemes mostly use single scoring or fixed threshold grading for attenuation assessment. This approach cannot represent the continuous trend and uncertainty of attenuation, nor can it be stably updated when data quality varies.

[0209] Therefore, in terms of decay assessment, this scheme uses continuous latent variables, namely the target object. In time Continuous decaying hidden state The attenuation intensity of the target object is represented; based on the state-space model, the attenuation state of the target object is represented as a continuous latent variable evolving over time, and a weighted Bayesian update is performed on the attenuation assessment feature vector according to the evidence credibility weights to obtain the posterior distribution of the attenuation state of the target object at each time point (let it be denoted as ). The target object From time 1 to time Given all observed vectors, its time Continuous decaying hidden state The posterior distribution of the weakened state and the corresponding weakening evolution curve are used to generate and output a weakening assessment result containing the weakening state and weakening evolution trend of the target object.

[0210] In addition to the above, as a preferred implementation option, preferably, in this scheme S05, the weighted Bayes update includes: exponentially weighting the likelihood term of each observation indicator according to its evidence credibility weight, thereby reducing the impact of low-credibility observations on the posterior distribution of the weakened state.

[0211] As an example of a preferred implementation option, preferably, solution S05 includes:

[0212] Based on the state-space model, the decay state of the target object is represented as a continuous latent variable that evolves over time, and the corresponding state transition model function is defined as follows:

[0213]

[0214] in, , The target objects are respectively In time , The continuous decaying hidden state at time; For the target object In time The driving factor vector corresponds to the sub-features in the decay assessment features; For the driving coefficient vector, In time The process noise, used to characterize unobservable disturbances, follows a normal distribution. , This represents the process variance.

[0215] The observation model function is defined as follows:

[0216]

[0217] in, For the target object In time At that time, its included data characteristics The normalized observed features, i.e., the decay assessment feature vector. Normalized feature subset; For the target object In time The continuous decaying hidden state at time; For the first A mapping function between an observation feature and a continuous weakened hidden state, which is used to map the continuous weakened hidden state to each observation index. For data features In time The process noise follows a normal distribution. , For observation features The noise variance.

[0218] Based on the above techniques, a weighting of evidence credibility can be introduced. The function for weighted Bayesian update of the decay assessment feature vector based on the evidence credibility weights is defined as follows:

[0219]

[0220] in, , The target objects are respectively In time , The continuous decaying hidden state at time; For the target object From time 1 to time The observation vector at time, i.e., the decay assessment feature vector, has a dimension of . Data vectors; , The target objects are respectively In time At that time, it contains the first Evidence credibility weights for each data feature, and normalized observation features; For the target object From time 1 to time Given all observed vectors, its time Continuous decaying hidden state The posterior distribution of; The state transition probability, i.e., the probability of a continuously decaying hidden state from time... arrive The evolutionary pattern, i.e., prior information; For the first The likelihood of each data feature represents the likelihood in the true state. Under these conditions, observed The probability of.

[0221] Where the likelihood follows a normal distribution It satisfies the following definition:

[0222]

[0223] Then its probability density is:

[0224]

[0225] That is, it is equivalent to adding the evidence confidence weight to the observation variance. When the confidence is lower, the noise of the equivalent observation variance is greater, and the contribution of the observation feature is smaller when updating.

[0226] in, For the first Observation features The mapping function with continuously decaying hidden states, For observation features The noise variance.

[0227] The posterior distribution of the weakened state of the target object at each time point is defined as follows:

[0228]

[0229] in, For the target object In time At that time, it corresponds to a continuously decaying hidden state. The point estimate is the mean of the posterior distribution; For conditional expectation, Represented as in a given observation vector Under the condition of continuous decay hidden state Expected estimate.

[0230] For uncertainty, i.e., the posterior variance, the function is defined as follows:

[0231]

[0232] For the target object In time The uncertainty of the state estimation at time is given by the variance of the posterior distribution; For conditional variance, Represented as in a given observation vector Under the condition of continuous decay hidden state The degree of fluctuation in the estimated value.

[0233] Among them, in time 1 to time period The definition of the decay evolution curve is: That is, time period target object Point estimate The time series.

[0234] This solution introduces a data quality scoring and evidence credibility weighting mechanism to address real-world data issues such as missing data, latency, noise, device drift, and cross-source conflicts. It quantifies the effectiveness of each indicator and adaptively suppresses the interference of low-quality observations in subsequent inferences, making the fusion results more robust and more in line with clinical practice.

[0235] As a preferred implementation option, S05 of this scheme further includes: performing change point detection on the weakening evolution curve to identify sudden deterioration events in the weakening state of the target object and generate early warning information.

[0236] The function for detecting change points is defined as follows:

[0237]

[0238] in, , The target objects are respectively In time , The mean of the posterior distribution at time; This represents the posterior distribution change value.

[0239] Based on the above, the posterior distribution change value is compared with the preset deterioration threshold. In comparison, if When this occurs, a sudden deterioration event is triggered, generating and outputting a warning message.

[0240] This scheme models the weakened state as a continuous latent variable that evolves over time, and uses weighted Bayesian updates to output the posterior distribution of the weakened state and the weakened evolution curve. It can not only provide the current weakened level, but also provide information on assessment uncertainty and trend changes at the same time, thereby realizing dynamic monitoring and early warning of the weakened state deterioration, stabilization or improvement process, avoiding the problem that traditional single scoring or fixed threshold grading cannot reflect longitudinal changes.

[0241] Because there is only the target object In time posterior distribution mean at time The decay evolution curve can only reflect the current status and trend of decay assessment, but it is difficult to directly predict future steps based on the decay evolution curve. Therefore, in addition to the decay assessment methods mentioned above, it is necessary to combine... Figure 2 As shown, based on the aforementioned attenuation assessment method, and as a preferred implementation option, this scheme preferably also includes:

[0242] S06. Based on the decay evolution curve and the decay assessment feature vector, construct a time-delay causal Bayesian network, perform probability inference to output the probability of decay and deterioration of the target object within a preset time range in the future and its key influence path, generate prediction results, and then aggregate them into the decay assessment results for output.

[0243] The topology learning of the time-delayed causal Bayesian network is combined with domain constraint rules to prohibit candidate connection edges that violate the temporal causal direction.

[0244] As a preferred implementation option, preferably, in this scheme S06, the key influence path is obtained by performing sensitivity analysis on the nodes in the time-delay causal Bayesian network. The sensitivity analysis is used to quantify the contribution of changes in each node to the probability of weakening and deterioration.

[0245] As an example of a preferred implementation option, solution S06 includes:

[0246] Extract at least one sub-feature related to the decay assessment from the decay assessment feature vector as a key influencing factor, denoted as _____. Among them, the sub-features can be one or more of the following: lifestyle and nutritional characteristics, chronic disease management characteristics, previous medication burden characteristics, physical examination characteristics, functional assessment characteristics, and wearable continuous monitoring of physiological parameters. For example, it can be a sub-feature involving factors such as nutritional risk, lack of activity, sleep disorders, coexistence of multiple diseases, medication burden, depression and anxiety, insufficient social support, and environmental fall risk.

[0247] In identifying key influencing factors Then, the point estimates of the continuous latent states corresponding to the decay assessment feature vector are extracted from the decay evolution curve, and denoted as . The time-delay causal Bayesian network is a dynamic network, which is a dynamic network that is time-delayed. node combination Defined as follows:

[0248]

[0249] in, This is the key influencing factor number, i.e., the node number.

[0250] For the structure learning of the time-delay causal Bayesian network, the goal is to learn the set of directed edges. This scheme introduces time direction constraints into the edges of the time-delay causal Bayesian network to ensure that intra-piece and inter-piece edges satisfy the following definitions:

[0251] The film satisfies the following:

[0252] or

[0253] The space between the pieces satisfies:

[0254] , ,

[0255] in, , Number the nodes. , They are nodes In time , Key influencing factors; For nodes In time Key influencing factors.

[0256] The constraints of the time-delay causal Bayesian network also include prohibiting obvious anti-causal edges, that is, not allowing directed edges whose time direction points to the past.

[0257] Based on the above, a scoring function is constructed, defined as follows:

[0258]

[0259] in, For training data; Candidate graph structure; In candidate graph structure The following network parameters, The number of parameters; The number of samples; This is the scoring function.

[0260] Based on the scoring function, the candidate graph structure with the highest score that satisfies the constraints is defined as follows: , Represented as candidate graph structure The set of nodes, Represented as candidate graph structure The set of directed edges;

[0261] For discrete nodes, we use conditional probability tables (CPT) for learning, which are defined as follows:

[0262]

[0263] For continuous nodes, conditional Gaussian or piecewise linear methods are used to transform the problem into a discrete one, which is then processed using CPT. As a network node, it is or ; The set of parent nodes; Represents network nodes The conditional probability distribution; Represents network nodes The conditional probability table.

[0264] Based on the point estimate of the continuously decaying hidden state, the deterioration event in the future time H is defined as... ,in, To worsen the threshold, For time The point estimate of the continuously weakened hidden state is then defined as follows:

[0265]

[0266] in, For the target object From time Starting with predicting the step size in the future. The probability of a worsening event occurring within the organization; For the target object In time Point estimates of the continuously weakening hidden states; For the target object In time The evidence set includes point estimates of key influencing factors and continuously weakening latent states.

[0267] Sensitivity analysis was performed on each key influencing factor, and the formula is defined as follows:

[0268]

[0269] in, For nodes In time Key influencing factors; This indicates the key impact factors. Artificial intervention to a preset reference health level Later, in the evidence set Under the condition that remains unchanged, the posterior probability of the risk of deterioration; For nodes in a dynamic network The node contribution, if The larger the value, the more likely it is to be a node. Key influencing factors The greater the contribution to the risk of deterioration, the more critical the influencing factor will be when it exceeds the preset value. arrive The in-slice edges are defined as high-contribution paths, which are then used as key impact paths and aggregated into the decay assessment results for output.

[0270] This solution constructs a time-delayed causal Bayesian network to make probabilistic inferences about the outcome of decay based on time-period data, and identifies key nodes and critical paths that affect the progression of decay. It can provide a mechanism-level explanatory basis for the technical goal of "why it deteriorates and where to intervene".

[0271] The aforementioned solutions focus on frailty assessment and risk prediction, aiming to provide relevant intervention recommendations for different risks or frailty conditions, combined with... Figure 3 As shown, as a preferred implementation option, this solution also includes:

[0272] S07. Based on the posterior distribution of the current frailty state, the probability of frailty worsening, the key influencing paths, and the pre-built intervention knowledge base, generate individualized intervention suggestions, and then aggregate them into the frailty assessment results for output.

[0273] The individualized intervention recommendations are generated in accordance with safety constraints, which include one or more of the following: the target subject's chronic disease status, recent history of falls, risk of malnutrition, and risk of drug interactions.

[0274] As an example, in step S07 of this solution, the pre-built intervention knowledge base can be an intervention set. Its definition is: ; among which, each intervention item It can include the following information:

[0275] 1. Intervention information: A list of measures to be implemented;

[0276] 2. Indications: Corresponding to key nodes, i.e. factors corresponding to key influencing factors, such as nutrition, activity, sleep, depression, medication review, environmental modification, etc.

[0277] 3. Contraindications / risks: such as heart failure, bone and joint problems, recent falls, high-risk drug interactions, etc.

[0278] 4. Expected impact point: Key influencing factors that can be altered by intervention;

[0279] In this scheme, the posterior distribution of the current weakened state, the probability of weakening and deterioration, and the key influencing path can be used as selection conditions to match them with intervention items in a pre-built intervention knowledge base, thereby matching individualized intervention suggestions.

[0280] To improve the reliability and flexibility of individualized intervention recommendations and frailty assessment results, this scheme S07 allows the recipients of relevant information to be doctors, caregivers, and the elderly, respectively, corresponding to the medical service provider, the guardian, and the target individual.

[0281] Combination Figure 4 As shown, based on the above, this solution also proposes an elderly frailty assessment system based on multi-source data fusion, which applies the aforementioned elderly frailty assessment method based on multi-source data fusion, including:

[0282] The data access module is used to establish data space nodes at the hospital, community and / or home end, and then access multi-source health data about the target object through a standardized interface.

[0283] The data semantic standardization module is used to preprocess the multi-source health data accessed by the data access module to obtain a data stream with unified semantics; the preprocessing includes one or more of the following: field mapping, unit conversion, and time alignment.

[0284] The data identity alignment module is used to perform cross-node privacy-preserving identity alignment on the preprocessed multi-source health data, so as to associate cross-node data belonging to the same target object into the same object data set;

[0285] The quality assessment module is used to calculate data quality scores for each data indicator in the object dataset, and to determine the evidence credibility weight of the corresponding data based on the data quality scores.

[0286] The feature construction module is used to construct a decay assessment feature vector based on the object dataset, wherein the decay assessment feature vector includes multiple sub-features characterizing the health status of the target object;

[0287] The decay state estimation module represents the decay state of the target object as a continuous latent variable that evolves over time based on the state space model. It then performs a weighted Bayes update on the decay assessment feature vector according to the evidence credibility weight to obtain the posterior distribution of the decay state of the target object at each time point and the corresponding decay evolution curve. Finally, it generates and outputs a decay assessment result containing the decay state and decay evolution trend of the target object based on the posterior distribution of the decay state and the corresponding decay evolution curve.

[0288] The data access module includes an interface adaptation unit, which is used to access the hospital, community, and home terminals, and to interconnect with the hospital information system, community health information system, laboratory information system, and Internet of Things platform.

[0289] The decay state estimation module also includes a change point detection unit, which is used to identify abrupt deterioration events in the decay evolution curve and trigger an early warning.

[0290] remove Figure 4 In addition to the system implementation shown, combined with Figure 5 As shown, in Figure 4 Based on the system shown, as a preferred implementation option, the system described in this solution preferably further includes:

[0291] The trajectory prediction and interpretation module is used to construct a time-delayed causal Bayesian network based on the decay evolution curve and the decay assessment feature vector, and to perform probability inference to output the decay and deterioration probability of the target object within a preset time range and its key influence path, generate prediction results, and then aggregate them into the decay assessment results for output; wherein, the topology learning of the time-delayed causal Bayesian network is combined with domain constraint rules to prohibit candidate connection edges that violate the temporal causal direction;

[0292] The intervention recommendation module is used to generate individualized intervention suggestions based on the posterior distribution of the current frailty state, the probability of frailty deterioration, the key impact paths, and the pre-built intervention knowledge base, and then aggregate them into the frailty assessment results for output.

[0293] The multi-terminal interaction module is used to output frailty assessment results, prediction results and / or individualized intervention suggestions to the doctor's terminal, the caregiver's terminal and / or the elderly's terminal, respectively.

[0294] The trajectory prediction and interpretation module includes a structure learning unit and a probability inference unit. The structure learning unit determines the topology of the time-delay causal Bayesian network by combining domain constraint rules, and the probability inference unit outputs the probability of decay and deterioration within a preset future time range.

[0295] The intervention recommendation module includes a safety constraint unit, which is used to limit the content of the intervention plan based on factors such as chronic disease status, fall risk, and medication interaction risk.

[0296] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0297] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0298] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for assessing frailty in the elderly based on multi-source data fusion, characterized in that, It includes: S01. Establish data space nodes at the hospital, community, and / or home ends respectively, and then access multi-source health data about the target object through a standardized interface, and preprocess the multi-source health data to obtain a data stream with unified semantics; S02. Perform cross-node privacy-preserving identity alignment on the multi-source health data to associate cross-node data belonging to the same target object into the same object data set; S03. Calculate the data quality score for each data indicator in the object data set, and determine the evidence credibility weight of the corresponding data based on the data quality score; S04. Construct a decay assessment feature vector based on the object data set, wherein the decay assessment feature vector includes multiple sub-features characterizing the health status of the target object; S05. Based on the state-space model, the weakened state of the target object is represented as a continuous latent variable that evolves over time. The weakened assessment feature vector is then updated using a weighted Bayesian method according to the evidence credibility weights to obtain the posterior distribution of the weakened state of the target object at each time point and the corresponding weakened evolution curve. Then, based on the posterior distribution of the weakened state and the corresponding weakened evolution curve, a weakened assessment result containing the weakened state and weakened evolution trend of the target object is generated and output.

2. The frailty assessment method for the elderly based on multi-source data fusion as described in claim 1, characterized in that, In S01, the data sources of the multi-source health data include hospital information system data, community health information system data, laboratory information system data and / or home monitoring system data, which respectively include one or more of the following: medication record data, physical examination data, chronic disease follow-up data, wearable device data, and home environment and safety IoT data. In S01, the preprocessing involves mapping fields, converting units, and aligning time for the multi-source health data to obtain a data stream with unified semantics. In S02, the privacy-preserving identity alignment involves performing irreversible desensitization encoding on the identity identifier of the target object, and completing the association and / or aggregation of the same target object data between different data space nodes based on the desensitization encoding.

3. The frailty assessment method for the elderly based on multi-source data fusion as described in claim 2, characterized in that, In S03, the data quality score is calculated from at least two or more of the following: integrity score, timeliness score, consistency score, and equipment reliability score, and the data quality score is mapped to the evidence credibility weight of the corresponding data. In S01, the multi-source health data, after preprocessing, yields a data stream with unified semantics, defined as follows: The target object In time At that time, from the data source The data obtained was acquired after preprocessing. In S02, the function that associates cross-node data belonging to the same target object into a single object data set is defined as follows: in, For target object In time A collection of objects at time. As a data source, For data source collection, For target object In time At that time, from the data source The data stream generated after preprocessing the obtained data, i.e., the preprocessed data subset, has more than one data feature, and each data feature has its corresponding associated data sequence, i.e., observation features. It is represented as the target object. In time The first time Data features, i.e., data characteristics Observational characteristics; In S03, the function for the integrity score is defined as follows: in, For target object In time Time-related data characteristics Integrity score, For target object In time Time-related data characteristics The percentage of missing values; The function for the timeliness score is defined as follows: in, For target object In time Time-related data characteristics Timeliness score For target object The data feature collection time corresponding to the most recent timeliness rating; The time constant for age decay; The function for the consistency score is defined as follows: in, For the target object In time Time-related data characteristics Consistency score, For data features The corresponding associated data sequence, i.e., the observed features, , The target objects are respectively Data characteristics The mean and standard deviation over a preset historical period It is a constant value, used to prevent the denominator from being 0; When the data characteristics in the object data set When data includes wearable device data and / or home environment and security IoT data, the data quality score includes a device reliability score, the function of which is defined as follows: in, For the target object In time Time-related data characteristics The corresponding equipment reliability score, For the target object In time At that time, its data characteristics The percentage of data containing anomalies; The function for the data quality score is defined as follows: in, For the target object In time Time-related data characteristics The data quality score, whose value is truncated by the function. Truncate to [0, 1], , , , These are weighting coefficients, and their cumulative sum is 1; The function that maps the data quality score to the corresponding evidence credibility weights is defined as follows: in, For the target object In time At that time, its data characteristics The weight of the credibility of the evidence; , These are the lower and upper limits of the evidence credibility weight, respectively, which satisfy... , This is the nonlinear strengthening coefficient, and its value is greater than 1.

4. The frailty assessment method for the elderly based on multi-source data fusion as described in claim 3, characterized in that, In S04, based on the object data set, a decay assessment feature vector is constructed by concatenating data features, and it is defined as follows: , i.e., the observation vector; Wherein, the decay assessment feature vector Data characteristics Corresponding observation features The numerical features were also normalized to obtain normalized observation features. ; S04 further includes: aggregating the evidence credibility weights of the data features contained in the decay assessment feature vector to form a composite data structure with the decay assessment feature vector. Associate the corresponding data weight vector ; In S04, the sub-features are one or more of the following: demographic and socioeconomic characteristics, lifestyle and nutritional characteristics, chronic disease management characteristics, past medication burden characteristics, laboratory test characteristics, physical examination characteristics, functional assessment characteristics, wearable continuous monitoring of physiological parameters characteristics, and home environment and safety characteristics. The aforementioned characteristics of prior medication burden include one or more of the following: multiple medication quantity indicators, potential inappropriate medication indicators, and drug interaction risk indicators; The wearable continuous monitoring physiological parameter characteristics include one or more of the following: heart rate variability, sleep efficiency, daily activity level and activity rhythm, gait stability, or gait speed variability.

5. The frailty assessment method for the elderly based on multi-source data fusion as described in claim 4, characterized in that, In S05, the weighted Bayes update includes: exponentially weighting the likelihood term of each observation indicator according to its evidence confidence weight, thereby reducing the impact of low confidence observations on the posterior distribution of the weakened state. S05 includes: Based on the state-space model, the decay state of the target object is represented as a continuous latent variable that evolves over time, and the corresponding state transition model function is defined as follows: in, , The target objects are respectively In time , The continuous decaying hidden state at time; For the target object In time The driving factor vector corresponds to the sub-features in the decay assessment features; For the driving coefficient vector, In time The process noise, used to characterize unobservable disturbances, follows a normal distribution. , For process variance; The observation model function is defined as follows: in, For the target object In time At that time, its data characteristics The normalized observed features, i.e., the decay assessment feature vector. Normalized feature subset; For the target object In time The continuous decaying hidden state at time; For the first A mapping function between an observation feature and a continuous weakened hidden state, which is used to map the continuous weakened hidden state to each observation index. For data features In time The process noise follows a normal distribution. , For observation features The noise variance; Introducing evidence credibility weights The function for weighted Bayesian update of the decay assessment feature vector based on the evidence credibility weights is defined as follows: in, , The target objects are respectively In time , The continuous decaying hidden state at time; For the target object From time 1 to time The observation vector at time, i.e., the decay assessment feature vector, has a dimension of . Data vectors; , The target objects are respectively In time At that time, it contains the first Evidence credibility weights for each data feature, and normalized observation features; For the target object From time 1 to time Given all observed vectors, its time Continuous decaying hidden state The posterior distribution of; The state transition probability, i.e., the probability of a continuously decaying hidden state from time... arrive The evolutionary pattern, i.e., prior information; For the first The likelihood of each data feature represents the likelihood in the true state. Under these conditions, observed The probability of; When the likelihood follows a normal distribution It satisfies the following definition: Then its probability density is: That is, it is equivalent to adding the evidence confidence weight to the observation variance. When the confidence is lower, the noise of the equivalent observation variance is greater, and the contribution of the observation feature is smaller when updating. in, For the first Observation features The mapping function with continuously decaying hidden states, For observation features The noise variance; The posterior distribution of the weakened state of the target object at each time point is defined as follows: in, For target object In time At that time, it corresponds to a continuously decaying hidden state. The point estimate is the mean of the posterior distribution; For conditional expectation, Represented as in a given observation vector Under the condition of continuous decay hidden state Expected estimate; Uncertainty, i.e., the function of posterior variance, is defined as follows: For target object In time The uncertainty of the state estimation at time is given by the variance of the posterior distribution; For conditional variance, Represented as in a given observation vector Under the condition of continuous decay hidden state The degree of fluctuation in the estimated value; At time 1 to time period The definition of the decay evolution curve is: time period target object Point estimate The time series.

6. The frailty assessment method for the elderly based on multi-source data fusion as described in claim 5, characterized in that, S05 further includes: performing change point detection on the weakening evolution curve to identify sudden deterioration events in the weakening state of the target object and generating early warning information; The function for detecting change points is defined as follows: in, , The target objects are respectively In time , The mean of the posterior distribution at time; This represents the change in the posterior distribution; The posterior distribution change value is compared with the preset deterioration threshold. In comparison, if When this occurs, a mutation and deterioration event is triggered, generating and outputting a warning message.

7. The frailty assessment method for the elderly based on multi-source data fusion as described in claim 6, characterized in that, It also includes: S06. Based on the decay evolution curve and the decay assessment feature vector, construct a time-delay causal Bayesian network, perform probability inference to output the decay deterioration probability of the target object within a preset time range and its key influence path, generate prediction results, and then aggregate them into the decay assessment results for output. The topology learning of the time-delayed causal Bayesian network is combined with domain constraint rules to prohibit candidate connection edges that violate the temporal causal direction.

8. The frailty assessment method for the elderly based on multi-source data fusion as described in claim 7, characterized in that, In S06, the key influence path is obtained by performing sensitivity analysis on the nodes in the time-delay causal Bayesian network. The sensitivity analysis is used to quantify the contribution of changes in each node to the probability of weakening and deterioration. S06 includes: Extract at least one sub-feature related to the decay assessment from the decay assessment feature vector as a key influencing factor, denoted as _____. Then, extract the point estimates of the continuous decay hidden states corresponding to the decay assessment feature vector from the decay evolution curve, and set them as follows: The time-delay causal Bayesian network is a dynamic network, which is a dynamic network that is time-delayed. node combination Defined as follows: in, Key influencing factors are numbered, i.e., node numbers; A time direction constraint is introduced into the edges of the time-delay causal Bayesian network to ensure that intra-piece and inter-piece edges satisfy the following definition: The film satisfies the following: or The space between the pieces satisfies: 、 、 in, , Number the nodes. , They are nodes In time , Key influencing factors; For nodes In time Key influencing factors; The constraints of the time-delay causal Bayesian network also include prohibiting obvious anti-causal edges, that is, not allowing directed edges whose time direction points to the past time. Construct a scoring function, defined as follows: in, For training data; Candidate graph structure; In candidate graph structure The following network parameters, The number of parameters; The number of samples; For scoring functions; Based on the scoring function, the candidate graph structure with the highest score that satisfies the constraints is defined as follows: , Represented as candidate graph structure The set of nodes, Represented as candidate graph structure The set of directed edges; For discrete nodes, we use conditional probability tables (CPT) for learning, which are defined as follows: For continuous nodes, use conditional Gaussian or piecewise linear methods to transform them into discrete problems, and then use CPT to process them. in, As a network node, it is or ; The set of parent nodes; Represents network nodes The conditional probability distribution; Represents network nodes Conditional probability table; Based on the point estimate of the continuously decaying hidden state, the deterioration event in the future time H is defined as... ,in, To worsen the threshold, For time The point estimate of the continuously weakened hidden state is then defined as follows: in, For the target object From time Starting with predicting the step size in the future. The probability of a worsening event occurring within the organization; For the target object In time Point estimates of the continuously weakening hidden states; For the target object In time The evidence set, which includes point estimates of key influencing factors and continuously weakening latent states; Sensitivity analysis was performed on each key influencing factor, and the formula is defined as follows: in, For nodes In time Key influencing factors; This indicates the key impact factors. Artificial intervention to a preset reference health level Later, in the evidence set Under the condition that remains unchanged, the posterior probability of the risk of deterioration; For nodes in a dynamic network The node contribution, if The larger the value, the more likely it is to be a node. Key influencing factors The greater the contribution to the risk of deterioration, the more critical the influencing factor will be when it exceeds the preset value. arrive The in-slice edges are defined as high-contribution paths, which are then used as key impact paths and aggregated into the decay assessment results for output.

9. The frailty assessment method for the elderly based on multi-source data fusion as described in claim 8, characterized in that, It also includes: S07. Based on the posterior distribution of the current weakened state, the probability of weakened deterioration, the key impact paths, and the pre-built intervention knowledge base, generate individualized intervention suggestions, and then aggregate them into the weakened assessment results for output. The individualized intervention recommendations are generated in accordance with safety constraints, which include one or more of the following: the target subject's chronic disease status, recent history of falls, risk of malnutrition, and risk of drug interactions.

10. A frailty assessment system for the elderly based on multi-source data fusion, which applies the frailty assessment method for the elderly based on multi-source data fusion as described in any one of claims 1 to 9, characterized in that, It includes: The data access module is used to establish data space nodes at the hospital, community and / or home end, and then access multi-source health data about the target object through a standardized interface. The data semantic standardization module is used to preprocess the multi-source health data accessed by the data access module to obtain a data stream with unified semantics; the preprocessing includes one or more of the following: field mapping, unit conversion, and time alignment. The data identity alignment module is used to perform cross-node privacy-preserving identity alignment on the preprocessed multi-source health data, so as to associate cross-node data belonging to the same target object into the same object data set; The quality assessment module is used to calculate data quality scores for each data indicator in the object dataset, and to determine the evidence credibility weight of the corresponding data based on the data quality scores. The feature construction module is used to construct a decay assessment feature vector based on the object dataset, wherein the decay assessment feature vector includes multiple sub-features characterizing the health status of the target object; The decay state estimation module represents the decay state of the target object as a continuous latent variable that evolves over time based on the state space model. It then performs a weighted Bayes update on the decay assessment feature vector according to the evidence credibility weight to obtain the posterior distribution of the decay state of the target object at each time point and the corresponding decay evolution curve. Finally, it generates and outputs a decay assessment result containing the decay state and decay evolution trend of the target object based on the posterior distribution of the decay state and the corresponding decay evolution curve.

Images