A real-time fall risk assessment system and method based on big data

Abstract

The present invention discloses a real-time fall risk assessment system and method based on big data, which relates to the field of risk assessment. The system includes a data acquisition and preprocessing module, a big data processing and integration module, a feature extraction and selection module, a fall risk assessment module, and a real-time monitoring and warning module that are connected in sequence and work collaboratively. Among them: The data acquisition and preprocessing module is responsible for collecting data from multiple sources and converting it into digital signals; the big data processing and integration module is responsible for the secure transmission, aggregation, and preprocessing of data; the feature extraction and selection module is responsible for extracting features related to fall risk from the preprocessed data; the fall risk assessment module is responsible for the real-time assessment of fall risk based on the extracted features; the real-time monitoring and warning module is responsible for triggering warning signals according to the fall risk assessment results. This application realizes the accurate and real-time assessment of fall risk, the accurate and timely output of assessment results, and the rapid response to warning signals.

Description

Technical Field

[0001] The present invention relates to the technical field of risk assessment, and particularly to a real-time fall risk assessment system and method based on big data. Background Art

[0002] Falling is a serious health problem faced by people of all ages, especially the elderly and those with physical disabilities. It may lead to serious consequences such as fractures and traumatic brain injuries, which not only affect the quality of life of individuals but also increase medical costs and social burdens. Therefore, accurately and real-time assessing the fall risk is crucial for preventing falls.

[0003] However, the existing fall risk assessment technologies have many limitations. In terms of data collection, traditional methods are not comprehensive enough. The collection of skin microbial metabolic signals by biosensors may be inaccurate, and the sensitivity to environmental factors is insufficient, unable to fully reflect the impact of the environment on the fall risk. In addition, with the advent of the big data era, data security and privacy protection have become new challenges. Existing systems may lack effective data encryption technologies and cannot guarantee the integrity and security of data. In terms of data processing and feature extraction, the existing systems have limited ability to fuse and process multi-source data, making it difficult to integrate data from different data sources, resulting in low data quality. At the same time, the feature extraction algorithm may not be able to accurately extract features related to the fall risk, or the extracted feature dimensions are too high, affecting the calculation efficiency of the evaluation model. In terms of the evaluation model, the accuracy and real-time performance of existing models are insufficient. They fail to fully combine various factors such as individual biological characteristics, environmental factors, and historical data, and cannot update the evaluation results and issue warnings in a timely manner.

[0004] In summary, the existing fall risk assessment technologies have deficiencies in data collection, data security, data processing, feature extraction, evaluation models, as well as monitoring and warning. There is an urgent need for new technologies to improve the accuracy and real-time performance of fall risk assessment. Summary of the Invention

[0005] In view of this, the present invention proposes a real-time fall risk assessment system and method based on big data to solve the problems of poor assessment accuracy and real-time performance in the existing technology.

[0006] The specific technical solution of the present invention is as follows: A real-time fall risk assessment system based on big data, including the following modules connected in sequence and working in cooperation:

[0007] A data collection and conversion module, used to collect data from multiple sources and convert it into digital signals suitable for subsequent processing; this module includes a sensing unit that uses sensors to respectively obtain bio-related electronic signals and environmental analog signals, and converts them into digital signals through circuits and converters;

[0008] The big data processing and integration module is used for the secure transmission, aggregation, and preprocessing of data;

[0009] The feature extraction and selection module is used to extract features related to risk assessment from the preprocessed data; this module includes a federated learning and generative adversarial network sub-module that constructs a central node to coordinate the federated learning process at different data source ends and uses the generative adversarial network to extract risk-related features; it also includes an information bottleneck theory application sub-module that, based on the information bottleneck theory framework, selects a key feature set that is useful for risk assessment and mutually independent, and performs refined screening in combination with domain knowledge;

[0010] The fall risk assessment module is used for the real-time assessment of risk based on the extracted features to obtain a fall risk assessment result; this module contains a hybrid model sub-module that fuses two learning models to construct a hybrid model, interacts with the environment through an agent to learn the optimal strategy, constructs the model structure based on prior knowledge and expert experience, trains and optimizes the model using a large amount of data, and introduces transfer learning technology to improve the generalization ability of the model;

[0011] The real-time monitoring and warning module is used to trigger a warning signal according to the fall risk assessment result.

[0012] Specifically, for the data acquisition and conversion module, a biosensor array based on the principle of microbial fuel cells is used to closely adhere to the skin to capture the electronic signals generated by microbial metabolism in real time. Let the output signal of the biosensor array be , where t represents time;

[0013] An environmental sensor made of composite materials is used to collect the analog signals of the environment in real time. Let the output signal of the environmental sensor be ;

[0014] The biological signal and the environmental signal are converted into digital signals through a circuit and a converter. The digital signal of the biological signal is represented as , and the digital signal of the environmental signal is represented as ; The conversion process of the biological signal can be expressed as , where represents the biological signal conversion function; The conversion process of the environmental signal can be expressed as , where represents the environmental signal conversion function;

[0015] During the conversion process, noise suppression technology and bioelectrocatalytic amplification analog-to-digital conversion technology are adopted. Let the biological signal after noise suppression be , then ; Let the output of the environmental signal after signal conditioning be , then ; The final digital signal and It is transmitted to the big data processing and integration module for subsequent processing.

[0016] Specifically, let the feature vector obtained through feature extraction and selection be , which respectively represent the first, second, and nth eigenvalue in the key feature set that is useful for fall risk assessment and mutually independent after passing through the feature extraction and selection module; the weight vector of deep reinforcement learning is , which respectively represent the first, second, and nth weight values learned in the deep reinforcement learning part; the conditional probability matrix of the Bayesian network is , where represents the conditional probability value in the Bayesian network, the prior knowledge influence factor is , the expert experience influence factor is , and the construction of the Bayesian network structure is affected by prior knowledge and expert experience as , where is the activation function, is the number of relevant probabilities, represents the probability of a certain event related to the th feature occurring under specific conditions. These probability values together constitute the conditional probability matrix of the Bayesian network, which is used to assist in determining the final fall risk value based on probability reasoning in fall risk assessment. represents the

[0017] Specifically, the federated learning and generative adversarial network sub-module of the feature extraction and selection module constructs a central node to coordinate and manage the federated learning process of different data source ends, ensuring that the privacy and security of data are strictly protected. Each data source end deploys the TensorFlow federated learning framework to establish a local feature learning model and communicates with the central node through a secure communication protocol to share the learned features or model parameters. On the basis of federated learning, a generative adversarial network is constructed, and the generator and discriminator cooperate to extract features related to fall risk.

[0018] Specifically, let there be a total of v data source ends, which are respectively denoted as , the central node is C, the generator of the generative adversarial network is G, the discriminator is D, the features or model parameters learned from the data source ends are F, and the feature set related to fall risk finally extracted is P;

[0019] In the federated learning stage, each data source end Deploy the federated learning framework and establish a local feature learning model , where is a function for establishing a local model;

[0020] Each data source end communicates with the central node C through a secure communication protocol and shares the learned features or model parameters with the central node: , is a communication function;

[0021] The central node C collects the information shared by all data source ends to form an information set , respectively represent the 1st, 2nd, and vth key features related to the fall risk obtained after being processed by the feature extraction and selection module.

[0022] Specifically, based on the information set F obtained by federated learning, a generative adversarial network is constructed. The generator G generates latent features related to the fall risk according to the input: ; The discriminator D discriminates the features generated by the generator G, judges its authenticity and the relevance of the fall risk, and through adversarial training, both continuously optimize the network parameters: , is a training function, and are the generator and discriminator after training optimization respectively; Finally, the feature set related to the fall risk is extracted by the generative adversarial network after training optimization , is a function for extracting the final features.

[0023] Specifically, the big data processing and integration module includes a secure transmission sub-module, which uses the homomorphic encryption algorithm to encrypt sensitive data to ensure security during the transmission process, and uses the cyclic redundancy check and message authentication code technology at the receiving end to verify the integrity of the data; It also includes a data aggregation and preprocessing sub-module. The data aggregation and preprocessing sub-module is designed with flexible data interfaces, which can receive data from the data collection module and external data sources, and perform format unification, duplicate removal, and merging processing.

[0024] Specifically, the warning signal generation sub-module of the real-time monitoring and warning module uses the bio-electronic-photonic integration technology to design a signal generation circuit inside the wearable device, which is precisely managed by the computer system. According to the fall risk result, the fall risk level is divided, and then corresponding warning signals are generated in real time according to the change of the fall risk level.

[0025] Specifically, let Let \(R\) be the fall risk level, \(S\) be the warning signal, \(H\) represent the high risk level, and \(L\) represent the low risk level. A function that generates a signal when the risk is high. A function that generates a signal when the risk is low. The execution formula is: .

[0026] This application also proposes a real-time fall risk assessment method based on big data, including the following steps:

[0027] Step 1, data collection and conversion: Collect data from multiple sources, and use a bio-sensor array based on the principle of microbial fuel cells to closely adhere to the skin to capture the electronic signals generated by microbial metabolism in real time. At the same time, use an environmental sensor made of composite materials to collect the analog signals of the environment in real time; Subsequently, convert these signals into digital signals that can be used for subsequent processing through circuits and converters.

[0028] Step 2, big data processing and integration: Securely transmit the collected digital signals, use the homomorphic encryption algorithm to encrypt sensitive data to ensure data security during transmission; At the receiving end, use cyclic redundancy check and message authentication code technology to verify the integrity of the data; Then, design a flexible data interface to receive data from the data collection step and external data sources, and perform format unification, duplicate removal, and merging processing.

[0029] Step 3, feature extraction and selection: Construct a central node to coordinate and manage the federated learning process of different data source ends, and use a generative adversarial network to extract features related to fall risk; At the same time, use the information bottleneck theory as the theoretical framework for feature selection and compression, select a key feature set that is useful for fall risk assessment and mutually independent, and refine the selection of features in combination with domain knowledge.

[0030] Step 4, fall risk assessment: Based on the extracted features, fuse deep reinforcement learning and Bayesian networks to construct a hybrid model for real-time assessment of fall risk; Let the agent interact with the environment to learn the optimal strategy, and construct the Bayesian network structure according to prior knowledge and expert experience; Use a large amount of data for model training and optimization, and at the same time introduce transfer learning technology to improve the generalization ability of the model.

[0031] Step 5, real-time monitoring and warning: Continuously obtain the fall risk status of an individual. When the fall risk level changes, use an integration technology combined with personalized threshold settings to generate a warning signal to accurately prompt the individual's current fall risk level; And ensure the timely transmission of warning information through a multi-channel push mechanism.

[0032] The beneficial effects of the present invention are as follows:

[0033] The present invention uses a microbial fuel cell biosensor and an environmental sensor, combined with a circuit and a converter, to achieve precise and real-time data acquisition and conversion, providing a reliable data basis for fall risk assessment.

[0034] The present invention introduces a homomorphic encryption algorithm and a cyclic redundancy check technology to comprehensively ensure the integrity and security of data during transmission, enhancing users' confidence in privacy protection.

[0035] The present invention efficiently extracts and screens key features related to fall risk by constructing a federated learning network and utilizing a generative adversarial network and the information bottleneck theory, improving the computational efficiency.

[0036] The present invention integrates deep reinforcement learning and a Bayesian network to construct a hybrid model, and through transfer learning technology, ensures the accuracy and real-time nature of fall risk assessment, providing support for timely intervention.

[0037] The system of the present invention monitors the fall risk status all day long, quickly generates warning signals by using integration technology and personalized threshold settings, and uses bio-electronic-photonic integration technology to achieve rapid transmission of warning signals, winning valuable time for preventive measures. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0039] Figure 1 is a schematic structural diagram of the system of the present invention;

[0040] Figure 2 is a schematic flow diagram of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention clearer, the following further details the present invention with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0042] The present invention proposes a real-time fall risk assessment system and method based on big data.

[0043] Please refer to Figure 1 , a real-time fall risk assessment system based on big data, including the following modules that are connected in sequence and work together:

[0044] The data acquisition and conversion module is responsible for collecting data from multiple sources and converting it into digital signals for subsequent processing. It includes: a biological and environmental sensing unit, which uses a bio-sensor array based on the principle of microbial fuel cells and an environmental sensor made of composite materials. These are respectively closely attached to the skin to capture the electronic signals generated by microbial metabolism in real time and collect the analog signals of the environment in real time, and convert these signals into digital signals through circuits and converters.

[0045] The big data processing and integration module is responsible for the secure transmission, aggregation, and preprocessing of data. It includes: a secure transmission sub-module, which encrypts sensitive data using a homomorphic encryption algorithm to ensure security during the transmission process, and uses cyclic redundancy check and message authentication code technologies at the receiving end to verify the integrity of the data.

[0046] The data aggregation and preprocessing sub-module is designed with flexible data interfaces, capable of receiving data from the data acquisition module and external data sources, and performing format unification, duplicate removal, and merging processing.

[0047] The feature extraction and selection module is responsible for extracting features related to fall risk from the processed data. It includes: a federated learning and generative adversarial network sub-module, which constructs a central node to coordinate and manage the federated learning process at different data source ends, and uses a generative adversarial network to extract features related to fall risk.

[0048] The information bottleneck theory application sub-module uses the information bottleneck theory as the theoretical framework for feature selection and compression, selects a key feature set that is useful for fall risk assessment and mutually independent, and combines domain knowledge to refine the selection of features.

[0049] The fall risk assessment module is responsible for the real-time assessment of fall risk based on the extracted features. It includes: a deep reinforcement learning and Bayesian network hybrid model sub-module, which fuses deep reinforcement learning and Bayesian networks to construct a hybrid model, learns the optimal strategy through the interaction between the agent and the environment, constructs the Bayesian network structure based on prior knowledge and expert experience, trains and optimizes the model using a large amount of data, and introduces transfer learning technology to improve the generalization ability of the model.

[0050] The real-time monitoring and warning module is responsible for triggering warning signals according to the fall risk assessment results. It includes:

[0051] The warning signal generation sub-module is connected to the fall risk assessment module, continuously obtains the fall risk status of an individual, and when the fall risk level changes, uses an integration technology combined with personalized threshold settings to generate warning signals, accurately indicating the current fall risk level of the individual, and ensuring the timely transmission of warning information through a multi-channel push mechanism.

[0052] For the data acquisition and conversion module, this application designs a biosensor array strictly in accordance with the principle of microbial fuel cells. By deeply studying the influence of different materials and structures on the electron transfer efficiency, the optimal combination is finally determined. This application uses a metal oxide electrode material with high catalytic activity. This material not only has excellent electron conduction performance but also exhibits outstanding chemical stability, thus significantly improving the efficiency of electron transfer. This application installs the sensor array inside a smart bracelet or belt, closely fitting the skin to ensure that it can efficiently capture the electronic signals generated by skin microbial metabolism.

[0053] When microorganisms metabolize on the skin surface, weak bioelectric signals and various biological metabolites are generated. To comprehensively capture these signals, the biosensor array is equipped with a precise current detection circuit with a resolution up to the nanoampere level, enabling even weak signal changes to be accurately captured. At the same time, this application incorporates noise suppression technology to ensure the accuracy of the collected signals. In addition, microfluidic chip technology is used to detect the concentration changes of biological metabolites. Based on the enzyme-linked immunosorbent assay principle, this technology can simultaneously detect at least 20 key metabolites with a detection accuracy up to the picomole level. Finally, the weak biological signals collected need to be converted into digital signals for subsequent processing and analysis. This application uses a bioelectrocatalytic amplification analog-to-digital conversion technology, which performs signal conversion under the control of a computer system. During the conversion process, the computer precisely adjusts the bioelectrocatalytic reaction conditions (such as temperature, pH value, substrate concentration, etc.) according to a preset program, carefully designs the analog-to-digital conversion circuit, and uses a high-precision analog-to-digital converter, ultimately ensuring that the error rate of the digital signal is less than 0.1%, thus guaranteeing the accuracy and reliability of the data.

[0054] In the design process of the environmental sensor unit, this application selects graphene and nano metal oxides with special adsorption properties as composite materials, and determines the best composite method of these two materials through computer simulation technology. This composite method combines the excellent conductivity and mechanical strength of graphene with the sensitive adsorption properties of nano metal oxides to environmental factors (such as gas concentration, humidity, etc.), thus significantly enhancing the sensitivity of the sensor to environmental factors. During the manufacturing process, this application uses micro-nano processing technologies such as photolithography and electron beam evaporation processes. The manufactured sensor has structures such as microelectrodes and microchannels, which enable the sensor to more sensitively detect environmental factors closely related to the risk of falling, such as tiny electric field changes, tiny ground deformations, and changes in air ion concentration.

[0055] When the environmental sensor unit starts to work, it collects analog signals from the environment in real time. These signals may include changes in physical quantities such as electric field strength, deformation degree, ion concentration, etc., which are closely related to factors such as fall risk. In order to convert these analog signals into digital signals that can be processed by a computer, this application uses a high-performance signal conditioning circuit and an analog-to-digital converter. The signal conditioning circuit includes components such as a low-noise amplifier and a filter, whose function is to amplify the signal and remove noise interference. The analog-to-digital converter converts the continuous analog signal into a discrete digital signal for subsequent processing by the computer. In order to improve the quality of the digital signal (i.e., the signal-to-noise ratio), this application optimizes the circuit. By adjusting the parameters of the low-noise amplifier and the filter, the noise suppression ratio is increased to at least 30 dB or higher. This means that the sensor can more effectively suppress noise interference, thereby improving the accuracy and reliability of the digital signal.

[0056] Specifically, a biosensor array based on the principle of microbial fuel cells is closely attached to the skin to capture the electronic signals generated by microbial metabolism in real time. Let the output signal of the biosensor array be , where t represents time;

[0057] An environmental sensor made of composite materials is used to collect analog signals of the environment in real time. Let the output signal of the environmental sensor be ;

[0058] The biological signal and the environmental signal are converted into digital signals through a circuit and a converter. The digital signal of the biological signal is represented as , and the digital signal of the environmental signal is represented as ; The conversion process of the biological signal can be expressed as , where represents the biological signal conversion function; the conversion process of the environmental signal can be expressed as , where represents the environmental signal conversion function;

[0059] During the conversion process, noise suppression technology and an analog-to-digital conversion technology with bioelectrocatalytic amplification are adopted. Let the biological signal after noise suppression be , then ; Let the output of the environmental signal after signal conditioning be , then ; The final digital signals and are transmitted to the big data processing and integration module for subsequent processing.

[0060] For the big data processing and integration module, in the big data processing process, the secure transmission and efficient preprocessing of data are two crucial steps. To ensure the security and integrity of data, this application introduces a secure transmission protocol based on homomorphic encryption. This protocol uses advanced homomorphic encryption algorithms (such as lattice-based homomorphic encryption algorithms) to encrypt sensitive data, ensuring a high level of security during data transmission. At the same time, after receiving the encrypted data, the data aggregation end will comprehensively use two technologies, cyclic redundancy check (CRC) and message authentication code (MAC), to strictly check whether the data has been tampered with or damaged during transmission, so as to ensure that the integrity of the data reaches an extremely high level of over 99%.

[0061] To meet the requirements of multi-source data fusion in big data processing, the big data processing and integration module also designs multi-source data aggregation interfaces. These interfaces can simultaneously receive data from the data acquisition module and data transmitted from external data sources after homomorphic encryption. After receiving this data, the computer will start a series of complex and delicate programs to process them, including key steps such as converting the data into a unified format, removing duplicate data items, and merging relevant data records, for subsequent unified processing and analysis.

[0062] In addition, to improve the quality and usability of data, this application introduces a data preprocessing sub-module based on self-supervised learning. For the aggregated big data, this sub-module will build a series of self-supervised tasks in the computer system. For example, for biological signal data (such as electrocardiograms, electroencephalograms, etc.), the sub-module will build data reconstruction tasks, requiring the computer system to use partial data to predict the overall data, so as to learn the overall structure and characteristics of the data. For environmental data (such as temperature, humidity, air quality, etc.), the sub-module will build contrastive learning tasks, requiring the computer system to compare the data characteristics under different environments and discover the differences and similarities between them. The construction of these self-supervised tasks is completely realized by computer algorithms without manual intervention, thus greatly improving the efficiency and accuracy of data preprocessing.

[0063] After constructing the self-supervised tasks, the sub-module will use a self-supervised learning algorithm framework based on the deep learning autoencoder structure to perform operations such as data cleaning, denoising, and normalization. In the data cleaning stage, the computer algorithm will automatically identify and remove abnormal data, which may be caused by equipment failures, data transmission errors, etc. By adjusting the algorithm parameters according to experiments and data analysis, it can ensure that the removal rate of abnormal data reaches over 95%. In the denoising stage, the computer algorithm will learn the internal structure of the data and remove the noise in it, thereby improving the signal-to-noise ratio and usability of the data. In the normalization stage, the computer algorithm will perform linear or non-linear transformations on the data so that the deviation after data normalization is within ±2%.

[0064] For the feature extraction and selection module, this application constructs a robust computer device as the central node of federated learning. This device not only has efficient data processing capabilities but also has a rigorous secure communication mechanism built-in. This device is responsible for coordinating and managing the federated learning process from different data source ends, ensuring that data privacy and security are strictly protected throughout the process. This application integrates multiple data sources such as medical institutions and health monitoring device manufacturers into this federated learning network. A TensorFlow federated learning framework or other federated learning frameworks are deployed at each data source end to establish local feature learning models. These models operate independently on their respective data sources, without uploading the original data to the central node, but communicate with the central node through a secure communication protocol to share the learned features or model parameters. In this way, multiple data sources can jointly learn and extract features without revealing their respective data privacy.

[0065] Based on federated learning, this application further constructs a generative adversarial network (GAN). The GAN consists of a generator and a discriminator, which cooperate together to extract features related to fall risk. The generator adopts a deep convolutional neural network structure and can generate potential features related to fall risk according to the input data (preprocessed big data). These features may cover multiple aspects such as the dynamic balance features of bio-environment coordination and the fall tendency features of individual physiological rhythms. The discriminator uses a multi-layer perceptron structure to discriminate the features generated by the generator, judging their authenticity and correlation with fall risk. Through adversarial training, the generator and the discriminator compete with and learn from each other, continuously optimizing their respective network parameters. Eventually, the computer can extract more accurate features related to fall risk from the preprocessed big data. These features are represented in the form of high-dimensional vectors of 200 - 1000 dimensions, providing effective input for the subsequent risk assessment model.

[0066] For the extracted high-dimensional feature vectors, this application uses the information bottleneck theory as the theoretical framework for feature selection and compression. By using the Shannon entropy calculation method, the mutual information between the features and fall risk and the redundant information between the features are accurately calculated. The mutual information reflects the strength of the correlation between the features and fall risk, while the redundant information reveals the degree of similarity between the features. These information provide strong support for deeply understanding the information relationship between the features and fall risk and the redundancy degree between the features.

[0067] Based on the calculated mutual information and redundancy information, the present application sets a reasonable information threshold for screening features useful for fall risk assessment. At the same time, an efficient screening algorithm is designed, which can significantly reduce the number of features (at least reduce by 40%) while retaining more than 90% of the evaluation information. The screening algorithm carefully selects a set of key features that are useful for fall risk assessment and independent of each other according to mutual information and redundancy information. Through feature selection, the input dimension of the subsequent evaluation model is successfully reduced, thereby significantly improving the computational efficiency of the model. This not only speeds up the training and prediction speed of the model but also reduces the demand for computing resources.

[0068] Specifically, let there be v data source ends, denoted as respectively, the central node is C, the generator of the generative adversarial network is G, the discriminator is D, the features or model parameters learned from the data source ends are F, and the set of features related to fall risk finally extracted is P;

[0069] In the federated learning stage, each data source end deploys a federated learning framework and establishes a local feature learning model where is the function for establishing the local model;

[0070] Each data source end communicates with the central node C through a secure communication protocol and shares the learned features or model parameters with the central node: , is the communication function;

[0071] The central node C collects the information shared by all data source ends to form an information set , respectively represent the first, second, and vth key features related to fall risk obtained after being processed by the feature extraction and selection module.

[0072] Based on the information set F obtained from federated learning, a generative adversarial network is constructed. The generator G generates latent features related to fall risk according to the input: ; the discriminator D discriminates the features generated by the generator G, judges their authenticity and the relevance to fall risk, and through adversarial training, both continuously optimize the network parameters: , is the training function, and are the generator and discriminator after training optimization respectively; finally, the set of features related to fall risk is extracted by the generative adversarial network after training optimization, is the function for extracting the final features.

[0073] For the fall risk assessment module, a hybrid model that integrates deep reinforcement learning and Bayesian network aims to combine the advantages of both to improve the accuracy and real-time performance of fall risk assessment. In the deep reinforcement learning part, an agent is defined. It adopts an advanced neural network structure, the deep Q-network (DQN). The core task of the agent is to learn how to select the optimal action according to the current environmental state (including individual biometric features, environmental factors, etc., and these environmental states contain the key feature set obtained from the feature extraction and selection module) to maximize the long-term cumulative reward. At the same time, this application designs a reward function, which is the key to evaluating the quality of the agent's actions. It motivates the agent to learn better evaluation strategies based on prediction accuracy and fall event avoidance. In the Bayesian network part, this application constructs the network structure according to prior knowledge and expert experience, determines the probability relationships between various nodes (such as biometric features, environmental factors, etc.), and these relationships reflect how various factors jointly act on the fall risk. Through expert knowledge or data learning, this application determines the conditional probability table for the nodes in the Bayesian network, providing a solid foundation for probability inference.

[0074] Specifically, let the feature vector obtained through feature extraction and selection be , respectively represent the first, second, and nth eigenvalue in the key feature set that is useful for fall risk assessment and mutually independent after passing through the feature extraction and selection module; the weight vector of deep reinforcement learning is , respectively represent the first, second, and nth weight values learned in the deep reinforcement learning part; the conditional probability matrix of the Bayesian network is , represents the conditional probability value in the Bayesian network, the prior knowledge influence factor is , the expert experience influence factor is , and the construction of the Bayesian network structure is affected by prior knowledge and expert experience as , where K and E are quantities abstractly representing prior knowledge and expert experience; the fall risk assessment result , is the activation function, is the number of relevant probabilities, represents the probability of a certain event related to the rd feature occurring under specific conditions. These probability values jointly form the conditional probability matrix of the Bayesian network, which is used to assist in determining the final fall risk value based on probability inference in fall risk assessment, represents the i-th eigenvalue, represents the i-th weight value.

[0075] Specifically, let be the fall risk level, S be the warning signal, H represent the high-risk level, and L represent the low-risk level. is a function that generates a signal when it is a high risk. is a function that generates a signal when it is a low risk. The execution formula is: .

[0076] To train this hybrid model, this application collected a large amount of clinical data and actual application feedback data. These data cover the biometric characteristics, environmental factors, and fall event records of different individuals, providing rich samples for the training of the model. During the deep reinforcement learning training process, the agent interacts with the environment, selects actions according to the current environmental state, and obtains feedback according to the reward function. The computer continuously adjusts the agent's strategy based on these feedbacks, enabling it to gradually learn the optimal fall risk assessment strategy. The training of the Bayesian network may involve adjusting the conditional probability table to better reflect the dependencies in the data. This application adopts a method that combines data learning and expert adjustment to ensure the accuracy and reliability of the network structure. To further improve the performance of the model, this application adopts an optimization method based on the simulated annealing algorithm. By controlling the temperature parameter to search for the optimal solution in the solution space, the parameters of the hybrid model are continuously adjusted until the predetermined performance index is reached. The accuracy of the optimized model for fall risk assessment reaches more than 85%, significantly improving the accuracy of the assessment.

[0077] When the feature extraction and selection module outputs the key feature set, these feature sets are input into the hybrid model in real time. The hybrid model calculates the fall risk probability value (ranging from 0 to 1) of an individual in real time in the computer system. This probability value reflects the likelihood of the individual having a fall event in the future, providing strong support for fall prevention. To more intuitively understand the fall risk status of an individual, this application divides the fall risk into three levels: low risk, medium risk, and high risk according to a preset threshold. These thresholds can be adjusted according to clinical experiment data and actual application feedback to ensure the accuracy and practicality of the division. For example, low risk may correspond to a probability value less than 0.3, medium risk corresponds to a probability value between 0.3 and 0.7, and high risk corresponds to a probability value greater than 0.7. The computer can automatically divide the fall risk level according to these thresholds and output corresponding prevention measures and suggestions, providing targeted support for individuals with different risk levels.

[0078] For the real-time monitoring and early warning module, the real-time monitoring and early warning module is closely connected to the fall risk assessment module to form a closed-loop and highly integrated monitoring system. The fall risk assessment module comprehensively considers various physiological parameters of an individual (including but not limited to heart rate, blood pressure, gait stability, etc.), historical fall records, current activity conditions (such as walking speed, whether engaging in strenuous exercise, environmental light conditions, etc.), and possible external factors (such as whether the ground is slippery, whether there are obstacles or steps, etc.) to comprehensively evaluate the fall risk level of the individual. When the fall risk level of an individual changes, whether it rises or falls to a certain critical threshold, the system will immediately respond and automatically trigger the warning signal generation mechanism according to the preset rules and algorithms to ensure the timeliness and effectiveness of the warning, providing valuable reaction time for the individual to take preventive measures.

[0079] To achieve the rapid and accurate transmission of warning signals, this application adopts bio-electronic-photonic integration technology. This technology ingeniously combines the principles of biology, electronics, and photonics and can realize the generation and efficient transmission of complex signals in wearable devices (such as smart bracelets, smart watches, etc.). Inside the wearable device, a signal generation circuit is designed, which is precisely managed by a computer system and can generate corresponding warning signals in real time according to the change of the fall risk level. When the fall risk level of an individual is relatively high, the circuit will, under the instruction of the computer system, generate high-intensity and high-frequency optical signals, electrical stimulation signals, or make the wearable device generate specific vibration signals, which are designed to quickly attract the attention of the individual and remind them of the current fall risk. On the contrary, when the fall risk level is relatively low, the intensity and frequency of the generated signals will be correspondingly reduced to maintain continuous attention to the fall risk while avoiding unnecessary interference or panic to the individual. In addition, this application also realizes the conversion function from digital signals to visual signals, and by displaying the icon or text prompt of the fall risk level on the screen of devices such as smart watches in real time, the individual can intuitively understand their own fall risk status, thereby more actively taking corresponding preventive measures and reducing the probability of fall events.

[0080] Please refer to Figure 2 , this embodiment also discloses a real-time fall risk assessment method based on big data, including the following steps:

[0081] Step 1, data collection and conversion. Collect data from multiple sources, and use a bio-sensor array based on the principle of microbial fuel cells to closely adhere to the skin to capture the electronic signals generated by microbial metabolism in real time. At the same time, use an environmental sensor made of composite materials to collect the analog signals of the environment in real time; then, convert these signals into digital signals that can be used for subsequent processing through circuits and converters;

[0082] Step 2, Big Data Processing and Integration: Securely transmit the collected digital signals, and use the homomorphic encryption algorithm to encrypt sensitive data to ensure data security during transmission. At the receiving end, use cyclic redundancy check and message authentication code technologies to verify data integrity. Then, design a flexible data interface to receive data from the data collection step and external data sources, and perform format unification, duplicate removal, and merging processing.

[0083] Step 3, Feature Extraction and Selection: Construct a central node to coordinate and manage the federated learning process at different data source ends, and use a generative adversarial network to extract features related to fall risk. At the same time, adopt the information bottleneck theory as the theoretical framework for feature selection and compression, select a key feature set that is useful for fall risk assessment and mutually independent, and refine the features in combination with domain knowledge.

[0084] Step 4, Fall Risk Assessment: Based on the extracted features, fuse deep reinforcement learning and Bayesian network to construct a hybrid model for real-time fall risk assessment. The agent interacts with the environment to learn the optimal strategy, and constructs the Bayesian network structure according to prior knowledge and expert experience. Use a large amount of data for model training and optimization, and introduce transfer learning technology to improve the generalization ability of the model.

[0085] Step 5, Real-time Monitoring and Warning: Continuously obtain the fall risk status of an individual. When the fall risk level changes, use the integration technology combined with personalized threshold settings to generate warning signals to accurately prompt the individual's current fall risk level. And ensure the timely transmission of warning information through a multi-channel push mechanism.

[0086] The beneficial effects of the present invention are as follows:

[0087] High efficiency and accuracy of data collection and conversion: By adopting a biosensor array based on the principle of microbial fuel cells and an environmental sensor made of composite materials, the present invention can closely adhere to the skin and capture the electronic signals generated by microbial metabolism and environmental simulation signals in real time. Combined with circuits and converters, the error rate after signal conversion into digital signals is accurately controlled below 0.1%, significantly improving the accuracy and reliability of data, and providing a solid foundation for subsequent data analysis and fall risk assessment.

[0088] Security of big data processing and integration: The present invention innovatively introduces the homomorphic encryption algorithm to encrypt sensitive data, ensuring data security during transmission. At the same time, using cyclic redundancy check and message authentication code technologies effectively verifies data integrity, prevents data from being tampered with or damaged during transmission, comprehensively guarantees data integrity and security, and enhances users' confidence in privacy protection.

[0089] High precision and innovation in feature extraction and selection: By constructing a federated learning network, the present invention successfully integrates the feature learning models of multiple data sources, protecting data privacy while extracting features related to fall risk. Using generative adversarial networks and the information bottleneck theory, a key feature set that is useful and mutually independent for fall risk assessment is further screened, reducing the input dimension of the subsequent assessment model and significantly improving computational efficiency.

[0090] Accuracy and real-time performance of fall risk assessment: By deeply integrating deep reinforcement learning and Bayesian networks, the present invention constructs a hybrid model with both prior knowledge and expert experience, greatly improving the accuracy of fall risk assessment. Through a large amount of data for model training and optimization, and introducing transfer learning technology, the generalization ability of the model is further enhanced, ensuring the real-time performance and accuracy of the assessment results, providing strong support for timely taking effective intervention measures.

[0091] Timeliness and effectiveness of real-time monitoring and early warning: The system of the present invention can monitor the fall risk status of an individual in real time all day long. Once the risk level changes, it quickly generates a warning signal by using integration technology combined with personalized threshold setting. By adopting advanced bio-electronic-photonic integration technology, rapid and accurate transmission of the warning signal is achieved, winning valuable response time for the individual to take preventive measures in a timely manner, further demonstrating the practical value of the present invention.

[0092] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A real-time fall risk assessment system based on big data, characterized in that: Includes the following modules that are connected in sequence and work together: A data acquisition and conversion module is used to collect data from multiple sources and convert it into digital signals suitable for subsequent processing; the module includes a sensor unit, which uses sensors to obtain biological related electronic signals and environmental analog signals respectively, and converts them into digital signals through circuits and converters; Big data processing and integration module, used for secure transmission, aggregation and pre-processing of data; Feature extraction and selection module, used to extract risk assessment-related features from preprocessed data; This module includes the Federated Learning and Generative Adversarial Network submodules, which build a central node to coordinate the federated learning process of different data sources and use the Generative Adversarial Network to extract risk-related features. It also includes the Information Bottleneck Theory Application submodule, which uses the Information Bottleneck Theory as a framework to select key feature sets that are useful and independent of each other for risk assessment, and conducts refined screening in combination with domain knowledge. The fall risk assessment module is used to conduct real-time risk assessment based on the extracted features and obtain the fall risk assessment results. This module contains a hybrid model submodule, which integrates two learning models to build a hybrid model, learns the optimal strategy through the interaction between the agent and the environment, builds the model structure based on prior knowledge and expert experience, uses a large amount of data to train and optimize the model, and introduces transfer learning technology to improve the generalization ability of the model. Real-time monitoring and early warning module, used to trigger early warning signals based on fall risk assessment results; Suppose the feature vector obtained after feature extraction and selection is , They represent the first, second, and nth eigenvalues ​​in the key feature set that is useful and independent of fall risk assessment after feature extraction and selection modules; the weight vector of deep reinforcement learning is , They represent the first, second, and nth weight values ​​learned in the deep reinforcement learning part respectively; the conditional probability matrix of the Bayesian network is , Represents the conditional probability value in the Bayesian network, and the prior knowledge influence factor is , the expert experience impact factor is , the Bayesian network structure construction is affected by prior knowledge and expert experience. , where K and E are quantities that abstractly represent prior knowledge and expert experience; fall risk assessment results , is the activation function, is the relevant probability quantity, Indicates that under certain conditions, The probability of a certain event occurring related to a feature. These probability values ​​together constitute the conditional probability matrix of the Bayesian network. , used to assist in determining the final fall risk value based on probabilistic reasoning in fall risk assessment, represents the i-th eigenvalue, Represents the i-th weight value.

2. The real-time fall risk assessment system based on big data as claimed in claim 1, characterized in that: For the data acquisition and conversion module, a biosensor array based on the principle of microbial fuel cells is used to fit closely to the skin to capture the electronic signals generated by microbial metabolism in real time. The output signal of the biosensor array is set to , where t represents time; The environmental sensor made of composite materials is used to collect the analog signal of the environment in real time. The output signal of the environmental sensor is ; The biological signal and the environmental signal are converted into digital signals through circuits and converters. The digital signal of the biological signal is expressed as , the digital signal of the environmental signal is expressed as ; The biological signal conversion process is expressed as ,in represents the biological signal conversion function; the environmental signal conversion process is expressed as ,in represents the environmental signal conversion function; In the conversion process, noise suppression technology and bio-electrocatalytic amplification analog-to-digital conversion technology are used. The biological signal after noise suppression is assumed to be ,but ; Assume that the output of the ambient signal after signal conditioning is ,but ; The final digital signal and It is transmitted to the big data processing and integration module for subsequent processing.

3. The real-time fall risk assessment system based on big data as claimed in claim 1, characterized in that: The federated learning and generative adversarial network submodules of the feature extraction and selection module construct a central node to coordinate and manage the federated learning process of different data sources to ensure that the privacy and security of the data are strictly protected. Each data source deploys the TensorFlow federated learning framework to establish a local feature learning model, and communicates with the central node through a secure communication protocol to share the learned features or model parameters. A generative adversarial network is constructed based on federated learning, and the generator and discriminator work together to extract features related to the risk of falling.

4. The real-time fall risk assessment system based on big data as claimed in claim 3 is characterized in that: There are v data source terminals, which are respectively , the central node is C, the generator of the generative adversarial network is G, the discriminator is D, the features or model parameters learned from the data source are F, and the final feature set related to fall risk is P; In the federated learning phase, each data source Deploy the federated learning framework and build a local feature learning model ,in A function for building a local model; Each data source Communicate with the central node C through a secure communication protocol and transmit the learned features or model parameters Shared to the central node: , is the communication function; The central node C collects the information shared by all data sources to form an information collection , They respectively represent the first, second, and vth key features related to the risk of falling obtained after being processed by the feature extraction and selection module.

5. The real-time fall risk assessment system based on big data as claimed in claim 4, characterized in that: Based on the information set F obtained by federated learning, a generative adversarial network is constructed, and the generator G generates potential features related to the risk of falling according to the input: ; The discriminator D generates features from the generator G To identify the authenticity and relevance of the fall risk, the two continuously optimize the network parameters through adversarial training: , is the training function, and They are the generator and discriminator after training and optimization respectively; finally, the feature set related to the risk of falling is extracted by the trained and optimized generative adversarial network , is the function for extracting the final features.

6. The real-time fall risk assessment system based on big data as claimed in claim 1, characterized in that: The big data processing and integration module includes a secure transmission submodule, which uses a homomorphic encryption algorithm to encrypt sensitive data to ensure security during transmission, and uses cyclic redundancy check and message authentication code technology at the receiving end to verify the integrity of the data; it also includes a data aggregation and preprocessing submodule, which is designed with a flexible data interface and can receive data from the data acquisition module and external data sources, and perform format unification, deduplication, and merging processing.

7. The real-time fall risk assessment system based on big data as claimed in claim 1, characterized in that: The warning signal generating submodule of the real-time monitoring and warning module adopts bio-electronic-photonic integration technology, designs a signal generating circuit inside the wearable device, is accurately managed by a computer system, divides the fall risk level according to the fall risk results, and then generates a corresponding warning signal in real time according to the change of the fall risk level.

8. The real-time fall risk assessment system based on big data as claimed in claim 1, characterized in that: set up is the fall risk level, S is the warning signal, H indicates a high risk level, and L indicates a low risk level. is the function that generates signals when the risk is high, It is a function that generates signals when the risk is low. The execution formula is: .

9. A real-time fall risk assessment method based on big data, applied to the real-time fall risk assessment system based on big data as claimed in claim 1, characterized in that: The following steps are involved: Step 1, data collection and conversion, collects data from multiple sources, and uses a biosensor array based on the principle of microbial fuel cells to fit closely to the skin to capture electronic signals generated by microbial metabolism in real time. At the same time, environmental sensors made of composite materials are used to collect analog signals from the environment in real time. Subsequently, these signals are converted into digital signals that can be used for subsequent processing through circuits and converters; Step 2, big data processing and integration, securely transmit the collected digital signals, use homomorphic encryption algorithm to encrypt sensitive data, and ensure data security during transmission; at the receiving end, use cyclic redundancy check and message authentication code technology to verify the integrity of the data; then, design a flexible data interface to receive data from the data collection step and external data sources, and unify the format, remove duplicates, and merge them; Step 3: Feature extraction and selection. A central node is built to coordinate and manage the federated learning process of different data sources. Generative adversarial networks are used to extract features related to fall risk. At the same time, the information bottleneck theory is used as the theoretical framework for feature selection and compression to select key feature sets that are useful and independent of each other for fall risk assessment, and the features are refined by combining domain knowledge. Step 4: Fall risk assessment: Based on the extracted features, a hybrid model is constructed by integrating deep reinforcement learning and Bayesian network to conduct real-time assessment of fall risk; the optimal strategy is learned through the interaction between the agent and the environment, and the Bayesian network structure is constructed based on prior knowledge and expert experience; a large amount of data is used for model training and optimization, and transfer learning technology is introduced to improve the generalization ability of the model; Step 5: Real-time monitoring and early warning, continuously obtaining the individual's fall risk status. When the fall risk level changes, the integrated technology combined with personalized threshold settings will generate early warning signals to accurately indicate the individual's current fall risk level; And through a multi-channel push mechanism, ensure the timely communication of early warning information.

Images