A digital-based smart elderly care institution operation and management method and system

Abstract

This invention relates to the field of smart elderly care technology, specifically to a digitally based smart elderly care institution operation and management method and system. It includes an edge computing layer that preprocesses data to generate multi-source fused time series data; an embedded time-series causal prediction model that infers the probability of future risk events and causal paths based on the current state of an individual digital twin; and a care plan generation module that infers personalized intervention strategies through a care knowledge graph. This solution decomposes strategies into atomic tasks and assigns them to execution terminals, while simultaneously feeding back execution data to the digital twin construction module to update the twin's state, forming a closed loop. The time-series causal prediction model is a deep neural network integrating spatiotemporal attention mechanisms and structural causal models. By constructing individual digital twins and introducing causal inference, it achieves the analysis of the intrinsic mechanisms of risk and proactive intervention, significantly improving the intelligence level and care safety of elderly care institution operations.

Description

Technical Field

[0001] This invention relates to the field of smart elderly care technology, and more specifically, to a digital-based smart elderly care institution operation and management method and system. Background Technology

[0002] With the accelerating aging of the population, the demand for intelligent operation and management in elderly care institutions is becoming increasingly urgent. Existing technical solutions mainly focus on single-dimensional sensing and monitoring and passive alarms, which have revealed the following technical shortcomings in practical applications.

[0003] First, at the data acquisition and sensing level, existing solutions mostly rely on independently deployed sensors, such as standalone infrared detectors, emergency call buttons, or single wearable devices. These sensors operate within closed data channels, often collecting indicators of a single, isolated dimension, such as simply monitoring whether someone has passed by or recording only the wearer's heart rate. However, the health risks of the elderly are essentially an emergent phenomenon resulting from the coupled effects of physiological instability, behavioral changes, and environmental stress. For example, an abnormally high bedroom temperature at night may increase cardiovascular load, triggering frequent turning over in the elderly, which can then evolve into orthostatic hypotension and a higher risk of falls. Existing standalone sensor architectures cannot capture such cross-modal, coordinated change patterns, and the acquired information fragments are insufficient to form a complete characterization of the elderly person's overall condition, fundamentally limiting the depth of subsequent analysis.

[0004] Second, at the level of risk identification and early warning, existing technologies generally employ rule engines based on static thresholds or simple pattern matching algorithms. These methods compare a single collected indicator with preset upper and lower limits, triggering an alarm once the limit is exceeded. This approach has three inherent flaws. First, it cannot detect the joint distribution drift of multidimensional indicators. For example, the co-occurrence of abnormalities in relative blood pressure changes and heart rate variability often precedes the exceeding of limits by a single indicator, but rule engines are powerless to detect this. Second, rule thresholds are insensitive to individual differences; uniform and fixed standards are difficult to adapt to elderly populations with different underlying diseases, different medication regimens, and different lifestyles, leading to numerous false alarms or critical missed alarms. Third, and most critically, existing methods lack causal reasoning capabilities; they can only determine whether an abnormality has occurred, but cannot extrapolate potential risks and their evolutionary paths in the future. For example, they can detect "current gait abnormality," but cannot predict which causal chain will lead to a fall in the next thirty minutes, let alone quantify the probability of each causal path.

[0005] Third, at the operational scheduling and care execution level, existing technical solutions typically operate independently of the risk identification system. The generation of care tasks often relies on fixed scheduling plans or manual calls, lacking a real-time and effective connection mechanism with the results of front-end perception and analysis. When the risk assessment module identifies potential problems, it cannot automatically drive the dynamic allocation of subsequent care resources and the orderly execution of intervention measures. Even if some systems attempt to establish linkages, these linkages are mostly unidirectional and linear, lacking continuous tracking and feedback correction of the actual effects of the execution process. Once delays, interruptions, or deviations occur during task execution, the system cannot detect these changes, let alone re-optimize the intervention plan based on the latest status. This results in a significant lag in care response, making it difficult for intervention strategies to adapt to the elderly's real-time changes in physical and behavioral states, thus missing the optimal window for risk intervention.

[0006] Fourth, at the overall system architecture level, existing technical solutions lack a closed-loop mechanism that runs through "perception—modeling—deduction—decision—execution—feedback." Data flow, analysis flow, and action flow are fragmented, failing to form a complete loop: dynamically constructing individualized models from multimodal data, conducting forward-looking risk deductions based on the models, generating personalized care strategies based on the deduction results and precisely scheduling their execution, and then feeding back the execution results to update the model. This architectural fragmentation prevents the technical capabilities of each link from being integrated to form system-level intelligence. The operation and management of elderly care institutions heavily rely on the continuous attention and experience-based judgment of human personnel, and the system itself lacks the ability to learn autonomously, continuously optimize, and iterate adaptively.

[0007] To address the aforementioned issues, there is an urgent need for a digitally-based smart elderly care institution operation and management method and system. Summary of the Invention

[0008] The purpose of this invention is to provide a digital-based smart elderly care institution operation and management method and system to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, one of the objectives of this invention is to provide a digitally-based smart elderly care institution operation and management method, comprising the following steps: S1. Collect multimodal sensing data, including at least physiological, behavioral, and environmental data, through multiple sensor nodes deployed in elderly care facilities; S2. The multimodal sensing data is preprocessed using the configured edge computing layer, including data cleaning, timestamp alignment and missing value processing, to generate a multi-source fusion time series. S3. The edge computing layer uses a lightweight health event early warning model obtained through knowledge distillation to generate local early warning signals in real time based on the multi-source fusion time series. S4. Build a cloud-based digital twin platform, and in conjunction with the multi-source fusion time series, construct and continuously update an individual digital twin for each elderly person. The corresponding individual digital twin includes a physiological state sub-model, a behavioral pattern sub-model, and an environmental interaction sub-model. S5. The risk inference module in the cloud-based digital twin platform inputs the current multidimensional state vector of the individual digital twin into a pre-trained temporal causal prediction model to infer the probability of occurrence of various care risk events within a preset time window and the causal influence path between events, and outputs the risk probability distribution; the temporal causal prediction model is a deep neural network that integrates spatiotemporal attention mechanism and structural causal model. S6. Construct a care plan generation module, input the risk probability distribution into a preset care knowledge graph, and generate a personalized intervention strategy for the elderly through graph reasoning algorithm; S7. Configure the closed-loop scheduling execution layer to decompose the personalized intervention strategy into atomic task queues, and assign and execute tasks based on the status of each execution terminal, and receive execution status feedback. S8. The closed-loop scheduling execution layer feeds back the execution status to the cloud-based digital twin platform to drive the status update of the individual digital twin, forming a closed loop.

[0010] As a further improvement to this technical solution, the various sensors in S1 include at least a piezoelectric thin film mattress sensor, an ultra-wideband radar sensor, a wearable multi-parameter physiological monitoring bracelet, an environmental passive infrared sensor, a temperature and humidity sensor, a light sensor, a noise sensor, and a sensor built into a smart pillbox.

[0011] As a further improvement to this technical solution, the method for preprocessing multimodal sensing data in S2 includes the following steps: S2.1. Perform outlier detection and removal for each sensing data channel. Outlier detection uses the sliding interquartile range method. S2.2 Incorporate multi-source heterogeneous data into a unified time series analysis framework, and perform precise alignment through a multi-scale dynamic time warping algorithm. Each time step after alignment contains valid data from all sensors at that moment or the nearest neighbor moment. S2.3 The missing value processing submodule uses a multiple interpolation algorithm based on chain equations to complete the missing segments.

[0012] As a further improvement to this technical solution, the method for generating local early warning signals in real time in S3 includes the following steps: S3.1 The lightweight health event early warning model is acquired using a teacher-student knowledge distillation framework. The teacher model is a deep temporal hybrid network pre-trained on large-scale historical data in the cloud, and the student model adopts a lightweight two-layer temporal convolutional network structure. S3.2 Utilize the inference engine built into the lightweight health event early warning model to receive the multi-source fusion time series output by the data preprocessing unit in a continuous stream manner; S3.3 The lightweight health event early warning model performs sliding window inference on the multi-source fusion time series. It maintains a sliding window with a width of 300 time steps, corresponding to a 30-second historical data segment. The window slides forward 10 time steps each time. S3.4 After the multidimensional time series segments within the window are normalized to zero mean and unit variance, they are fed into the student model for a forward propagation calculation, outputting the probability estimate of the occurrence of each preset health event at that moment. The corresponding output is a probability vector, with each dimension corresponding to a health event category.

[0013] As a further improvement to this technical solution, the real-time generation of local early warning signals in S3 includes: The alert trigger timestamp is accurate to milliseconds. Health event categories use predefined health event classification codes; The event confidence score is the smoothed probability value rounded to the nearest percentile. The extracted brief context snapshot includes the mean and rate of change of key physiological indicators such as heart rate and respiratory rate in the three seconds prior to the triggering of the warning.

[0014] As a further improvement to this technical solution, the method for building a cloud-based digital twin platform in S4 includes the following steps: S4.1 Retrieve the elderly person's electronic health record from the institutional information system. The record contains structured demographic information, a list of past medical history, current medication regimen, allergy history, surgical history, a sequence of physical examination indicators for the past six months, and nursing assessment results upon admission. Map this to an initial ontology instance. The ontology is coded using standard medical system nomenclature and clinical terminology. S4.2 The cloud-based digital twin platform uses the internal hidden state of an individual digital twin as the variable to be estimated through the unscented particle filter algorithm; The hidden state vector contains the dynamic parameters of the physiological state sub-model, the hidden state sequence of the behavioral pattern sub-model, and the interaction strength parameters of the environmental interaction sub-model. S4.3. Use unscented particle filtering to maintain a set of 500 particles, each particle containing three parts: The first part consists of the latent variables of the physiological state sub-model, specifically the cardiovascular dynamics parameter values ​​and respiratory dynamics parameter values ​​at the current moment; The second part is the hidden state identifier of the behavior pattern sub-model, specifically the current bottom-level behavior atom category code, the middle-level activity segment category code, the remaining dwell time in seconds of the middle-level activity, and the top-level day-night rhythm time period index; The third part is the interaction intensity parameter of the environmental interaction sub-model, specifically the current interaction probability value of each functional area; S4.4. For each particle, the physiological state sub-model is advanced using the fourth-order Runge-Kutta numerical integration method to obtain the predicted physiological state; S4.5. Stepwise solve the lumped parameter ordinary differential equations of cardiovascular and respiratory dynamics to generate the evolution trajectory of physiological indicators such as heart rate, blood pressure, and blood oxygen. S4.6, Promoting the behavioral pattern sub-model, we obtain the predicted mid-level activity categories and new remaining dwell time; S4.7, Promoting the environmental interaction sub-model, we obtained a complete prior prediction of the current time step state; S4.8 Perform model calibration on the constructed model. After calculating the mixture likelihood weights of 500 particles, normalize the weight vector so that the sum of all weights equals one.

[0015] As a further improvement to this technical solution, the method for outputting the risk probability distribution in S5 includes the following steps: S5.1 Input the current multidimensional state vector of the individual digital twin into the pre-trained temporal causal prediction model through the risk inference module; S5.2 The risk inference module embeds a pre-trained temporal causal prediction model, which is a deep neural network that integrates spatiotemporal attention mechanism and structural causal model. The model structure consists of three layers: a graph structure learning layer, a spatiotemporal graph attention encoder, and a causal convolutional decoder. S5.3 Introducing Counterfactual Intervention Loss: Randomly intervene in some causal variables in the training samples to adjust their values, forming counterfactual samples. This forces the model to learn that the change in the probability distribution after intervention is consistent with the true causal effect, and outputs the risk probability distribution.

[0016] As a further improvement to this technical solution, the method for generating a personalized intervention strategy for the elderly in step S6 includes the following steps: S6.1 Configure care knowledge graph construction units and graph reasoning engine; S6.2 The care knowledge graph is constructed using RDF triples format, defining a directed heterogeneous graph. ,in For a set of entity nodes, Let be the set of edges. For a set of relation types, A collection of attributes; S6.3 The graph reasoning engine adopts a link prediction algorithm based on graph attention network, which calculates the link score between the node and all intervention entities through a bilinear graph attention network. S6.4. Use bundle search to generate Top-K intervention paths. Each path consists of multiple ordered intervention actions. Each action is marked with execution priority, suggested execution time window and required resources. The resulting personalized intervention strategy is a structured XML or JSON object. Generate a personalized intervention strategy for the elderly.

[0017] The second objective of this invention is to provide a system for realizing a digital-based smart elderly care institution operation and management method, including a multimodal data acquisition layer, an edge computing layer, a cloud-based digital twin platform, a care plan generation module, and a closed-loop scheduling and execution layer. The multimodal data acquisition layer includes multiple sensor nodes deployed within the elderly care institution for collecting multimodal perception data that includes at least physiological data, behavioral data, and environmental data. The edge computing layer is communicatively connected to the multimodal data acquisition layer and includes a data preprocessing unit and a local inference engine. The data preprocessing unit is used to clean, timestamp align and handle missing values ​​of the multimodal sensing data to generate a multi-source fusion time series. The local inference engine is deployed with a lightweight health event early warning model obtained by knowledge distillation, which is used to generate local early warning signals in real time based on the multi-source fusion time series. The cloud-based digital twin platform is connected to the edge computing layer via a secure communication link. The cloud-based digital twin platform includes a digital twin construction module, a risk simulation module, and a care plan generation module. The digital twin construction module is used to construct and continuously update an individual digital twin for each elderly person based on the multi-source fusion time series. The individual digital twin includes a physiological state sub-model representing cardiovascular and respiratory dynamics, a behavioral pattern sub-model representing daily activity sequences, and an environmental interaction sub-model representing the interaction between the elderly person and environmental facilities. The risk extrapolation module embeds a pre-trained temporal causal prediction model to receive the current multidimensional state vector of the individual digital twin, and extrapolate the probability of occurrence of various care risk events and the causal influence path between events within a preset time window, outputting a risk probability distribution; the temporal causal prediction model is a deep neural network that integrates spatiotemporal attention mechanism and structural causal model. The care plan generation module is used to input the risk probability distribution into a preset care knowledge graph and generate a personalized intervention strategy for the elderly through graph reasoning algorithm; The closed-loop scheduling and execution layer is communicatively connected to the cloud-based digital twin platform and includes a task scheduler and multiple execution terminals. The task scheduler is used to decompose the personalized intervention strategy into atomic task queues and assign tasks based on the status of the execution terminals. The execution terminals are used to execute the assigned tasks and report the execution status. The closed-loop scheduling and execution layer also reports the execution status to the digital twin building module to drive the status update of the individual digital twin.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This digital-based smart elderly care institution operation and management method and system decomposes strategies into atomic tasks and assigns them to execution terminals. At the same time, the execution is fed back to the digital twin construction module to update the state of the twin, forming a closed loop. The temporal causal prediction model is a deep neural network that integrates spatiotemporal attention mechanism and structural causal model. By constructing individual digital twins and introducing causal inference, it realizes the analysis of the internal mechanism of risk and the proactive intervention, which significantly improves the intelligence level of elderly care institution operation and care safety. Attached Figure Description

[0019] Figure 1 This is a diagram illustrating the overall method steps of the present invention; Figure 2 This is an overall system block diagram of the present invention. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 As shown, one of the objectives of this invention is to provide a digitally-based smart elderly care institution operation and management method, comprising the following steps: S1. Collect multimodal sensing data, including at least physiological, behavioral, and environmental data, through multiple sensor nodes deployed in elderly care facilities; S2. The multimodal sensing data is preprocessed at the edge computing layer, including data cleaning, timestamp alignment and missing value processing, to generate a multi-source fusion time series. S3. The edge computing layer uses a lightweight health event early warning model obtained through knowledge distillation to generate local early warning signals in real time based on the multi-source fusion time series. S4. The cloud-based digital twin platform constructs and continuously updates an individual digital twin for each elderly person based on the multi-source fusion time series. The individual digital twin includes a physiological state sub-model, a behavioral pattern sub-model, and an environmental interaction sub-model. S5. The risk inference module in the cloud-based digital twin platform inputs the current multidimensional state vector of the individual digital twin into a pre-trained temporal causal prediction model to infer the probability of occurrence of various care risk events within a preset time window and the causal influence path between events, and outputs the risk probability distribution; the temporal causal prediction model is a deep neural network that integrates spatiotemporal attention mechanism and structural causal model. S6. The care plan generation module inputs the risk probability distribution into a preset care knowledge graph and generates a personalized intervention strategy for the elderly through graph reasoning algorithm. S7. The closed-loop scheduling execution layer decomposes the personalized intervention strategy into atomic task queues, assigns and executes tasks based on the status of each execution terminal, and receives execution status feedback. S8. The closed-loop scheduling execution layer feeds back the execution status to the digital twin construction module to drive the status update of the individual digital twin, forming a closed loop.

[0022] The specific plan is as follows: In the process of smart elderly care operation and management, due to the differences in the physical condition of different individuals, the corresponding health data are diverse. In order to achieve targeted management, it is necessary to collect data on the physical condition of each individual. This solution collects multimodal perception data, including at least physiological data, behavioral data and environmental data, through multiple sensor nodes deployed in elderly care institutions.

[0023] The sensor nodes specifically include: A piezoelectric film mattress sensor laid under the mattress collects cardiac impaction signals at a sampling rate of 100Hz, from which physiological data such as heart rate, respiratory rate, and body movement are extracted; An ultra-wideband radar sensor installed in the corner of the room ceiling acquires human point cloud data at a sampling rate of 10Hz and outputs behavioral data such as indoor position trajectory and posture changes. The wearable multi-parameter physiological monitoring bracelet worn by the elderly collects photoplethysmography (PPG) waves and acceleration signals from the wrist at a sampling rate of 1Hz, and obtains information such as heart rate variability, blood pressure trend, and step count. Passive infrared sensors and temperature, humidity, light, and noise sensors are used to collect environmental data such as room temperature, humidity, light intensity, and ambient noise. The smart pillbox has built-in sensors that record pillbox opening events, generating data on medication use behavior.

[0024] Furthermore, due to the differences in the corresponding data states among the collected data, abnormal data, redundant data, and missing data are easily generated. To ensure the validity of the data, an edge computing layer is used to preprocess the multimodal sensing data. The specific method is as follows: First, outlier detection and removal are performed on each sensing data channel. Outlier detection uses the sliding quartile method, which involves maintaining a sliding window of 60 seconds for each sensor channel, with the window moving forward one data point at a time. Within the window, the data is sorted from smallest to largest, and the first quartile Q1 and the third quartile Q3 are calculated to obtain the interquartile range IQR = Q3 - Q1. Establish statistical standards; Exceeding the window Data points within the data range are marked as outliers; The window is in Data points within the data range are marked as normal values.

[0025] Data points marked as outliers are removed from the original sequence, leaving the corresponding positions blank for subsequent missing value processing steps to complete. For some sensor channels with obviously physically unreachable values, such as body temperature sensor readings below 30 degrees Celsius or above 45 degrees Celsius, or heart rate sensor readings below 20 beats per minute or above 250 beats per minute, a preset physiological feasibility boundary table is used for hard filtering, and out-of-bounds values ​​are directly marked as missing.

[0026] After this process, only normal observations in the statistical and physical sense are retained in the data sequence of each sensor channel, while abnormal locations exist as structured vacancies.

[0027] After removing outlier values, due to differences in sampling frequencies and clock sources among different sensor nodes—for example, a piezoelectric film mattress sensor samples at 100Hz and has its own high-precision crystal oscillator clock, an ultra-wideband radar sensor samples at 10Hz and is synchronized with the network time protocol service cycle, a wearable wristband samples at 1Hz and relies on a Bluetooth gateway timestamp, and an environmental sensor samples at 0.1Hz and is synchronized with the edge gateway clock—to incorporate these multi-source heterogeneous data into a unified time series analysis framework, precise alignment needs to be performed using a multi-scale dynamic time warping algorithm. The specific solution is as follows: The first step is to select a reference time axis, using the 10Hz sampling time series of the ultra-wideband radar sensor as the benchmark reference, because it has good intermediate characteristics in terms of sampling rate, time accuracy and coverage.

[0028] The second step involves multi-scale downsampling and coarse alignment. The 100Hz high-frequency signal is downsampled to 10Hz by a factor of 10 after being filtered by an anti-aliasing low-pass filter. The 1Hz low-frequency signal is upsampled to 10Hz using cubic spline interpolation. After unifying the signals of each channel to an approximate 10Hz time grid, the dynamic time warping distortion path between each channel signal and the reference channel signal is calculated to obtain the preliminary time alignment offset.

[0029] The third step is fine resampling. Based on the distorted path obtained in the previous step, the original high-frequency or low-frequency sequences of each channel are resampled point by point in a non-uniform manner, so that the sampling time of all channels is precisely anchored on a uniform 10Hz time axis, and the timestamp deviation is controlled within 5 milliseconds.

[0030] Each aligned time step contains valid data from all sensors at that moment or the nearest neighbor moment.

[0031] After outlier removal and timestamp alignment, some channels at certain time steps still have missing data. The sources of this missing data include gaps left by outlier removal, data loss due to brief interruptions in sensor communication, and uncertainties at interpolation boundaries during time alignment. The missing value processing submodule uses a chain equation-based multiple interpolation algorithm to complete the missing segments.

[0032] The algorithm is implemented as follows: First, the number of iterations for multiple imputation is set to M = 5, representing the generation of 5 complete imputation datasets. In each iteration, for each channel containing missing values, the complete observation portion of that channel is used as the training sample, and the corresponding time data of all other channels are used as prediction variables to construct a regularized linear regression model. This regression model is used to predict the value at the missing position of that channel, and a normally distributed random perturbation term is added with the predicted value as the mean and the standard deviation of the regression residual as the perturbation amplitude to generate an imputation value. After completing one round of imputation for all channels, the imputed dataset is used as the initial value for the next round of iteration, and the above process is repeated until all missing positions converge. After five rounds of iteration convergence, 5 complete multidimensional time series copies are obtained. Finally, the arithmetic mean of these 5 complete copies at each missing position is taken as the final imputation value. Through this process, a multi-source fusion time series with no missing values ​​and all channels cleaned and aligned is generated on a unified 10Hz time axis.

[0033] Furthermore, in order to perform targeted processing based on the monitored multi-source fusion time series data, the edge computing layer in this scheme utilizes a lightweight health event early warning model obtained through knowledge distillation to generate local early warning signals in real time based on the multi-source fusion time series. The specific scheme is as follows: The lightweight health event early warning model was acquired using a teacher-student knowledge distillation framework.

[0034] The teacher model is a deep temporal hybrid network pre-trained on large-scale historical data in the cloud. Its structure consists of a three-layer dilated causal convolutional network to extract multi-scale temporal features, a two-layer Transformer encoder to capture cross-channel interaction dependencies, and a fully connected classification head, with a total of approximately 3.5 million parameters. During the training phase, supervised training was conducted using de-identified multi-source fused time-series historical data from dozens of partner elderly care institutions, along with corresponding manually labeled health event tags. These health event tags include, but are not limited to, emergency event categories such as arrhythmia, sleep apnea, rapid postural changes, fall warning waves, and sudden drop in blood oxygen levels.

[0035] The corresponding student model employs a lightweight two-layer temporal convolutional network structure. The first layer is a standard temporal convolutional layer containing 32 one-dimensional convolutional kernels of length 7, and the second layer is a depthwise separable temporal convolutional layer containing 64 one-dimensional convolutional kernels of length 5. This is followed by global average pooling and a single-layer fully connected classifier. The total number of parameters is approximately one-third that of the teacher model, or about 110,000. The distillation process uses the posterior probability distribution of the class output from the teacher model as soft labels and the ground truth hard labels as supervision anchors. A weighted sum of the distillation loss function and the classification loss function is jointly constructed to train the student model end-to-end. Specifically, the distillation loss uses KL divergence to measure the difference between the student's output distribution and the teacher's output distribution, with a temperature parameter set to four. The classification loss uses cross-entropy loss. After sufficient iterations, the student model maintains a high degree of discriminative ability consistent with the teacher model on the test set, with a macro-average AUC above 0.93. Simultaneously, the inference speed is improved by approximately 40 times, with a single inference time of less than two milliseconds, significantly adapting to the computing power environment of edge gateways.

[0036] After the lightweight health event early warning model is built, the inference engine built into the lightweight health event early warning model is used to receive the multi-source fusion time series output by the data preprocessing unit in a continuous stream manner. The time series is a uniform time grid of 10Hz, and each time step contains a complete feature vector from all sensor channels. The dimension depends on the type and number of sensors deployed. Under typical configuration, it contains about forty to sixty features, covering heart rate, respiratory rate, heart rate variability low-frequency and high-frequency power, body motion intensity, indoor location coordinates, attitude angle, room temperature, relative humidity, light level, noise decibels, medicine box status, etc.

[0037] Simultaneously, the lightweight health event early warning model performs sliding window inference on the multi-source fusion time series. It maintains a sliding window with a width of 300 time steps, corresponding to 30-second historical data segments. The window slides forward 10 time steps each time, performing one inference per second. After each multi-dimensional time series segment within the window is normalized to zero mean and unit variance, it is fed into the Student model for a forward propagation calculation, outputting the probability estimate of the occurrence of each preset health event at that moment. This output is in the form of a probability vector, with each dimension corresponding to a health event category, and the probability value between zero and one.

[0038] To suppress random noise in single inference iterations and reduce false alarm rates, the inference engine performs exponentially weighted moving average smoothing on the probability vectors of continuous outputs. The choice of smoothing coefficient balances response speed and stability, with a typical value set to 0.6, ensuring that the model's output responds quickly to real-world events while filtering out isolated spikes. When the probability of any type of health event after smoothing exceeds a preset local warning threshold, the inference engine immediately generates a warning signal.

[0039] The generated local warning signal includes the following structured information fields: warning trigger timestamp, accurate to milliseconds; health event category, using a predefined health event classification code; event confidence level, a smoothed probability value rounded to a percentage; and an extracted brief context snapshot, including the mean and rate of change of key physiological indicators such as heart rate and respiratory rate within three seconds before the warning trigger, as well as the elderly person's current behavioral state, such as lying down, sitting, or moving. This warning signal is pushed to the on-duty nurse station terminal via the lightweight message queue telemetry transmission protocol built into the edge gateway, using a reliable delivery mode at Quality of Service Level 1. Upon receiving the warning, the on-duty terminal displays a prominent audio-visual alert on the interactive interface and shows the corresponding context snapshot information for quick confirmation and response by the on-duty nursing staff.

[0040] Furthermore, in order to predict and provide targeted interventions for the health status of individual individuals, this solution constructs a cloud-based digital twin platform to build and continuously update an individual digital twin for each elderly person based on multi-source fusion time series. The individual digital twin includes a physiological state sub-model, a behavioral pattern sub-model, and an environmental interaction sub-model.

[0041] The physiological state sub-model adopts a lumped-parameter differential equation model based on cardiovascular and respiratory dynamics to describe the dynamic behavior of the elderly circulatory and respiratory systems under neurohumoral regulation.

[0042] The core state variables in the cardiovascular component include left ventricular pressure, aortic pressure, right atrial pressure, heart rate, and peripheral vascular resistance. The model uses a set of ordinary differential equations to characterize the interactions between these variables: changes in left ventricular pressure are determined by myocardial contractility, current heart rate, and arterial load; changes in aortic pressure are determined by the mass balance between left ventricular ejection and outflow via peripheral resistance; and autonomic regulation of heart rate is determined by the dynamic balance between sympathetic and parasympathetic tone, which itself is regulated by a baroreceptor feedback loop that uses changes in aortic pressure as input. The model includes approximately twelve identifiable physiological parameters, including baseline myocardial contractility, arterial compliance, baseline peripheral resistance, baroreceptor gain, and sympathetic activation time constant.

[0043] The respiratory model is coupled with the cardiovascular model. Core state variables include alveolar oxygen partial pressure, alveolar carbon dioxide partial pressure, arterial oxygen saturation, tidal volume, and respiratory rate. The model characterizes the exchange dynamics between pulmonary ventilation and blood gas partial pressures, as well as the feedback regulation of the respiratory center by chemoreceptors. The coupling between the cardiovascular and respiratory systems is mainly reflected in two aspects: first, the sinus arrhythmia effect of heart rate and respiratory rate, i.e., the coupled oscillation of increased heart rate during inspiration and decreased heart rate during expiration; second, the supply-demand balance constraint between blood oxygen saturation and myocardial oxygen demand.

[0044] In each particle filter update cycle, the latent variables of the physiological state sub-model, namely the aforementioned physiological parameters and some unobservable state variables, are re-weighted and estimated by the particle ensemble. The model uses heart rate, respiratory rate, heart rate variability frequency domain index, blood oxygen saturation trend, and cardiac impact waveform characteristics from multi-source fusion time series as observed variables. Through likelihood calculation using particle filtering, the observed information is fused into the state estimation, ensuring that the individual digital twin always tracks the elderly person's current true physiological state.

[0045] The behavioral pattern sub-model adopts a hierarchical hidden Markov model to capture the sequence structure, rhythmic patterns, and individual preferences of the elderly's daily activities.

[0046] The model is divided into three time-scale layers. The bottom layer is the short-time state layer, containing typical behavioral atoms such as standing, walking, sitting, lying down, turning, and bending over. The transitions of the bottom-level states occur on a second-level time scale. The modeling adopts a first-order hidden Markov model, whose emission probabilities are connected to human point cloud features and acceleration features extracted from the fusion of ultra-wideband radar and wearable wristband, mapping the original sensor signals to the posterior probability distribution of behavioral atoms.

[0047] The middle layer is the activity fragment layer, which aggregates continuous sequences of behavioral atoms into semantic daily activities. Middle-level states are categorized by activity semantics, including approximately twenty activity categories such as sleeping, eating, washing and toileting, indoor walking, watching television, taking medication, and receiving care. Transitions in middle-level states occur on a minute-level timescale. Modeling employs a hidden semi-Markov model with explicit duration, and the dwell time distribution for each activity is modeled as a gamma distribution with parameters including shape and scale parameters. Personalized dwell time characteristics are obtained through maximum likelihood estimation of the elderly's historical activity data. Middle-level state identification is achieved through Viterbi decoding of the bottom-level behavioral atomic sequences.

[0048] The top layer is the circadian rhythm layer, describing the dominant behavioral patterns of the elderly at different times of the day. A time-varying transition probability Markov model is used, with the transition probability matrix divided into six time periods: early morning, morning, noon, afternoon, evening, and late night, each with different activity transition tendencies. The parameters of the rhythm layer are obtained by fitting multi-day behavioral data of the elderly using Fourier series, capturing their sleep preferences and daytime activity patterns.

[0049] In each digital twin update cycle, the behavior pattern sub-model receives indoor location trajectory, posture changes and movement intensity from multi-source fusion time series as observations, calculates the most likely current behavior state online through forward and backward algorithms, and continuously updates the transition probability parameters and dwell time parameters in each level of the model, so that the model gradually converges to the elderly's real behavior pattern.

[0050] The behavioral pattern sub-model adopts a hierarchical hidden Markov model to capture the sequence structure, rhythmic patterns, and individual preferences of the elderly's daily activities.

[0051] The model is divided into three time-scale layers. The bottom layer is the short-time state layer, containing typical behavioral atoms such as standing, walking, sitting, lying down, turning, and bending over. The transitions of the bottom-level states occur on a second-level time scale. The modeling adopts a first-order hidden Markov model, whose emission probabilities are connected to human point cloud features and acceleration features extracted from the fusion of ultra-wideband radar and wearable wristband, mapping the original sensor signals to the posterior probability distribution of behavioral atoms.

[0052] The middle layer is the activity fragment layer, which aggregates continuous sequences of behavioral atoms into semantic daily activities. Middle-level states are categorized by activity semantics, including approximately twenty activity categories such as sleeping, eating, washing and toileting, indoor walking, watching television, taking medication, and receiving care. Transitions in middle-level states occur on a minute-level timescale. Modeling employs a hidden semi-Markov model with explicit duration, and the dwell time distribution for each activity is modeled as a gamma distribution with parameters including shape and scale parameters. Personalized dwell time characteristics are obtained through maximum likelihood estimation of the elderly's historical activity data. Middle-level state identification is achieved through Viterbi decoding of the bottom-level behavioral atomic sequences.

[0053] The top layer is the circadian rhythm layer, describing the dominant behavioral patterns of the elderly at different times of the day. A time-varying transition probability Markov model is used, with the transition probability matrix divided into six time periods: early morning, morning, noon, afternoon, evening, and late night, each with different activity transition tendencies. The parameters of the rhythm layer are obtained by fitting multi-day behavioral data of the elderly using Fourier series, capturing their sleep preferences and daytime activity patterns.

[0054] In each digital twin update cycle, the behavior pattern sub-model receives indoor location trajectory, posture changes and movement intensity from multi-source fusion time series as observations, calculates the most likely current behavior state online through forward and backward algorithms, and continuously updates the transition probability parameters and dwell time parameters in each level of the model, so that the model gradually converges to the elderly's real behavior pattern.

[0055] The specific methods for building a cloud-based digital twin platform are as follows: First, the elderly person's electronic health record is retrieved from the institutional information system. This record contains structured demographic information, a list of past medical history, current medication regimen, allergy history, surgical history, a sequence of physical examination indicators over the past six months, and the nursing assessment results upon admission. This information is mapped to an initial ontology instance, coded using standard medical system nomenclature and clinical terminology to ensure semantic interoperability. The initial ontology instance contains the elderly person's static attributes and initial physiological parameter estimates, such as baseline blood pressure, resting heart rate, predicted vital capacity, and body mass index, forming the prior knowledge foundation for the individual's digital twin.

[0056] After initialization, the cloud-based digital twin platform enters continuous update mode. Each time the module receives the latest segment of the multi-source fused time series uploaded from the edge computing layer, it triggers a state update cycle. The update employs a Bayesian state estimation framework, specifically implemented using the unscented particle filter algorithm.

[0057] Unscented particle filtering uses the internal hidden states of an individual digital twin as variables to be estimated. The hidden state vector contains the dynamic parameters of the physiological state sub-model, the hidden state sequence of the behavioral pattern sub-model, and the interaction strength parameters of the environmental interaction sub-model.

[0058] The particle set maintained by the unscented particle filter contains five hundred particles. Each particle is a complete data structure containing three parts: the first part is the latent variables of the physiological state sub-model, specifically the current cardiovascular and respiratory dynamic parameters, including but not limited to the scalar values ​​of myocardial contractility, peripheral vascular resistance, arterial compliance, baroreceptor gain coefficient, sympathetic activation time constant, current left ventricular pressure, current aortic pressure, current tidal volume, and current alveolar oxygen partial pressure. The second part is the hidden state identifier of the behavior pattern sub-model, specifically the current bottom-level behavior atom category code, the middle-level activity segment category code, the remaining dwell time in seconds of the middle-level activity, and the top-level day-night rhythm time period index; The third part is the interaction intensity parameter of the environment interaction sub-model, which is the current interaction probability value of each functional area, represented by a floating-point vector with a length equal to the number of functional areas.

[0059] The state structure of each particle is denoted as... superscript Particle numbers range from 1 to 500, with subscripts... This is the current time step. The particle set was resampled at the end of the previous update and is now an equally weighted approximation of the posterior distribution of the previous time step.

[0060] In the specific prediction process, the task of the prediction step is to change the state of each particle from the previous time step. Advance one time step to the current moment to generate the prior predicted state. The driving force for migration comes from the state transition equations of the three sub-models.

[0061] For each particle, the physiological state sub-model is first advanced. Using the physiological latent variables carried by the particle from the previous time step as initial values, and the approximate values ​​of heart rate and respiratory rate observed in the multi-source fusion time series from the previous time step as external driving inputs, a fourth-order Runge-Kutta numerical integration method is employed. The corresponding algorithm formula is as follows: Step length microstep length ,common Microstep.

[0062] right to :

[0063]

[0064]

[0065] ; ; in , , as well as These are the four intermediate slope estimates in the Runge-Kutta method. Let be the physiological state vector of the j-th microstep. Let the function on the right-hand side of the system of differential equations be the function of the final state after integration. Applying process noise , This is the physiological state vector after integration. The diagonal matrix of the noise covariance of the physiological state process is used to obtain the predicted physiological state. After ten micro-step iterations, This is the physiological state vector after a complete time step. Using a step size of 0.1 seconds, the lumped parameter ordinary differential equations of cardiovascular and respiratory dynamics are solved stepwise. The corresponding equations are as follows: Aortic pressure: ; in, Aortic pressure is used to describe the pressure of blood within the arterial system. Left ventricular ejection flow rate is used to describe the amount of blood pumped out by the ventricle per unit time. Peripheral vascular resistance describes the level of resistance to blood flow through peripheral blood vessels. Arterial compliance describes the elastic expansion capacity of the arterial wall in response to pressure; Left ventricular pressure:

[0066] in, Left ventricular pressure, It is a time-varying elastic function. Left ventricular volume Zero pressure volume of the left ventricle; Sympathetic tone: ; in, For sympathetic nerve tension, This represents the maximum activation value of the sympathetic nervous system. The sympathetic activation time constant is This is the activation function for the pressure receptor; Parasympathetic tone: ; in, For parasympathetic tone, This represents the maximum activation of the parasympathetic nervous system. The parasympathetic activation time constant, This is the function by which the baroreceptor regulates the parasympathetic response; Pressure receptor activation function: ; in This refers to the kurtosis parameter of the sigmoid function. Set the setpoint pressure value for the pressure sensor; Instantaneous heart rate: ; in, Instantaneous heart rate, Based on baseline heart rate, This represents the gain coefficient for the effect of sympathetic nervous system on heart rate. This is the gain coefficient for the effect of parasympathetic nervous system on heart rate. alveolar oxygen partial pressure: ; in This refers to the partial pressure of oxygen in the alveoli. This refers to alveolar ventilation. Metabolic oxygen consumption rate, The partial pressure of oxygen in inhaled gas. Effective lung capacity; Arterial oxygen saturation: ; in Where n is arterial oxygen saturation, and n is Hill's coefficient. This is the partial pressure of half-saturated oxygen.

[0067] Ten microsteps of integration yield the evolution result for a complete time step. During integration, the model generates the evolution trajectories of physiological indicators such as heart rate, blood pressure, and blood oxygenation. To simulate the random fluctuations of real physiological systems and the unmodeled dynamics, an additive process noise is applied to the obtained physiological state vector after integration. The noise follows a zero-mean multivariate Gaussian distribution with a diagonal covariance matrix. The variance of each dimension is pre-determined through systematic identification residual analysis of the elderly's historical data.

[0068] Next, the behavior pattern sub-model is advanced. The state evolution of this sub-model is based on probability transitions. For bottom-level behavior atoms, based on the bottom-level behavior atom identifier carried by the current particle, the predicted bottom-level behavior atom category is obtained by sampling probabilistically from the bottom-level transition probability matrix learned from the behavior pattern sub-model. For mid-level activity segments, if the particle's current remaining dwell time is greater than zero, the remaining time is reduced by one while maintaining the activity category; if the remaining time has been reduced to zero, based on the mid-level activity transition probability matrix and the gamma distribution parameters of the corresponding target activity's personalized dwell time, the target activity category is first sampled, and then its dwell time is randomly sampled to obtain the predicted mid-level activity category and the new remaining dwell time. The top-level circadian rhythm index is updated deterministically based on the current system time period.

[0069] Then, the environmental interaction sub-model is advanced. The prior evolution of the interaction probability parameters in this sub-model adopts a zero-order preserved superposition random diffusion strategy. That is, when no new observations arrive directly, the interaction probability value remains unchanged from the previous time step estimate, and a weak additive Gaussian white noise is superimposed to allow for slow parameter drift. The noise variance is taken as one percent of the interaction probability value. When the location information indicating that the elderly person's location has entered a new functional region is received from the multi-source fusion time series, the corresponding functional region interaction probability of the current particle will be updated with a large weight in the correction step.

[0070] Through the parallel evolution of the three sub-models, each particle obtains a complete prior prediction of the current time step state, forming a set of five hundred predicted particles. The predicted particle set preserves particle diversity and reflects the natural evolutionary trend of the system state.

[0071] The model calibration is then performed, and the calibration step involves using the actual multi-source fused time series observation vectors arriving at the current moment. The weight of each particle is adjusted so that particles whose hidden states are closer to the actual observed state receive higher weights, and vice versa. The specific steps are as follows: First, a simulated observation vector is constructed for each predicted particle. The predicted physiological and behavioral states of the particle are then substituted into the forward observation models of each sensor to calculate the expected readings that each sensor should produce under that particle state. For example, the expected waveform characteristics of the photoplethysmogram are calculated based on the aortic pressure and heart rate within the particle, thus obtaining the expected values ​​of blood oxygen saturation and heart rate variability in the frequency domain. The expected outputs for attitude angle and body motion intensity are generated based on the particle's underlying behavioral atom identifiers combined with the emission probability matrix of the behavioral pattern sub-model. The expected distribution mean of the indoor position coordinates is generated based on the particle's current functional region identifier. The dimension of the simulated observation vector is compared with that of the actual multi-source fusion time-series observation vector. Exactly the same.

[0072] Then, the likelihood weight for each particle is calculated. The likelihood function is a multivariate Gaussian distribution with the mean being the simulated observation vector of that particle. The covariance matrix is ​​the observation noise covariance matrix. This matrix is ​​determined based on the actual calibration data of each sensor deployed in the elderly care facility environment; it is a diagonal block matrix, with each block corresponding to the measurement noise variance of the corresponding sensor channel. Likelihood weight Equal to a multivariate Gaussian distribution on the true observation vector The probability density value at that point, i.e. ,in The likelihood weights are the Gaussian distribution components. This represents the current real multi-source fused time series observation vector. The observation noise covariance diagonal block matrix is ​​given, where d is the dimension of the observation vector. Let be the determinant of the covariance matrix; To handle potential anomalous observations that may occur in some channels at a certain time step, a heavy-tailed student is introduced into the likelihood calculation. Mixture models of distributions, i.e. ,in For the likelihood weights of the t-distribution components of the student, Let be the degree of freedom parameter of the student's t-distribution. Calculate the mixed likelihood weights based on the likelihood weights of the Gaussian distribution components and the likelihood weights of the student's t-distribution components, i.e. This enhances the filter's robustness to occasional sensor distortion.

[0073] After calculating the mixture likelihood weights for 500 particles, the weight vector is normalized so that the sum of all weights equals one. The normalized weights are the empirical representation parameters of the posterior distribution of the hidden state at the current time step.

[0074] After weighting is completed, the risk projection module in the corresponding cloud-based digital twin platform inputs the current multidimensional state vector of the individual digital twin into a pre-trained temporal causal prediction model. This model projects the probability of various care risk events occurring within a preset time window and the causal paths between these events, outputting a risk probability distribution. The risk projection module embeds a pre-trained temporal causal prediction model, which is a deep neural network integrating spatiotemporal attention mechanisms and structural causal models. This model has three layers: The first layer, the graph structure learning layer, takes the multivariate state sequence of an individual digital twin as input and uses a continuous optimization algorithm with L1 regularization to adaptively learn a directed acyclic causal graph between variables. Nodes include heart rate, mean blood pressure, body position, activity entropy, medication time, etc. The learned adjacency matrix represents the direct causal strength between variables, and the corresponding algorithm is as follows.

[0075] The second layer, the spatiotemporal graph attention encoder, combines the state vector with the learned causal graph adjacency matrix at each time step, applies multi-head graph attention to calculate the causal attention weights of each node to its neighboring nodes, and introduces temporal position encoding and cross-time step self-attention mechanism to capture multi-step time lag causal effects and output a context feature sequence with causal structure.

[0076] The third layer, the causal convolutional decoder, constructs causal convolution kernels based on the causal graph. Convolution operations are performed only on variables with causal edges, decoding the probability of various care risk events occurring in the next 15, 30, and 60 minutes from the encoded feature sequence. During the training phase, a counterfactual intervention loss is introduced: certain causal variables in the training samples are randomly intervened to adjust their values, forming counterfactual samples. This forces the model to learn that the change in the probability distribution after intervention is consistent with the true causal effect, thereby enhancing the model's ability to identify causal paths. The risk inference module ultimately outputs a risk probability distribution, such as "fall risk probability 0.82, and the causal path is decreased heart rate variability → orthostatic hypotension → gait instability → fall," while also providing the strength of each causal edge.

[0077] After completing the prediction, this solution uses a configured care plan generation module to input the risk probability distribution into a preset care knowledge graph, and then uses a graph reasoning algorithm to generate a personalized intervention strategy for the elderly person. The specific solution is as follows: First, the care plan generation module includes a care knowledge graph construction unit and a graph reasoning engine. The care knowledge graph is constructed using RDF triples. Entities cover nursing interventions such as turning over, back tapping, and toileting assistance; medications such as antihypertensive drugs and sedatives; assistive devices such as walkers and guardrails; emergency plans such as calling emergency services and green channels; and personnel roles such as caregivers, nurses, and doctors. Relationships include "intervention required," "contraindications," "alternatives," and "responsible parties," with the corresponding algorithms as follows: Define a directed heterogeneous graph ,in For a set of entity nodes, Let be the set of edges. For a set of relation types, A collection of attributes; Corresponding node type Include: Nursing interventions: ; Drugs: ; Assistive devices: ; Emergency response plans: ; Personnel Roles: ; Risk events: .

[0078] Corresponding relation type Its domain is as follows: R1: (Intervention required) ; R2: (Recommended medication) ; R3: (Suggested assistive device) ; R4: (Triggered contingency plan) ; R5: (Responsible for execution) ; R6: (Taboo) or ; R1: (Alternative Solution) ; R1: (Time Constraint) (Attribute relationship, indicating the upper limit of execution time); The corresponding resource description framework triple storage format is: Each triplet is ,in , ; Then, the entity attribute embedding initialization is performed: Each entity Initial feature vector ,in The corresponding initialization method is as follows: For entities that can be described by text, a pre-trained clinical BERT model is used to extract the mean pooling vector of its descriptive text; For structural entities without text descriptions, Xavier uniform distribution random initialization is used; The corresponding relation embedding is initialized as follows: Each type of relationship Initialize learnable embedding vectors ,in ,use Uniform distribution initialization.

[0079] Simultaneously, the care plan generation module is equipped with a graph inference engine, which employs a link prediction algorithm based on graph attention networks. It uses risk event entities in the risk probability distribution, such as "fall risk," as query nodes. Through a bilinear graph attention network, it calculates the link score between this node and all intervention entities. Considering time constraints and resource availability, it uses bundle search to generate Top-K intervention paths. Each path consists of multiple ordered intervention actions, with each action labeled with its execution priority, suggested execution time window, and required resources. The resulting personalized intervention strategy is a structured XML or JSON object, for example, containing the action sequence "immediately push medication reminder > arrange for a caregiver to assist with toileting within 10 minutes > adjust room lighting to night mode," generating a personalized intervention strategy for the elderly person.

[0080] Finally, the task scheduler in the coordinated closed-loop scheduling execution layer receives the personalized intervention strategy and decomposes it. The decomposition rules are as follows: If the intervention involves manual services, an atomic task containing "operation type, target elderly ID, target location, and expected execution time" is generated; if it involves equipment, a device control command is generated.

[0081] The task scheduler maintains a real-time status table for all execution terminals, including the caregiver's current location obtained from indoor positioning tags, current task load, service capability type, and online / offline status of smart devices. A constraint-based greedy assignment algorithm is employed to maximize the global task completion rate while satisfying time windows and capability constraints, matching the optimal execution terminal for each atomic task.

[0082] The execution terminals include mobile terminals worn by caregivers, large screens at nurse stations, smart medicine boxes, service robots, and environmental control panels. Upon receiving a task, the execution terminal executes it through interface prompts, voice announcements, or automatic control. Functionally, after task completion, the terminal automatically sends a confirmation message, or sensors detect actions such as opening the medicine box, providing feedback on the execution status.

[0083] The closed-loop scheduling execution layer is also equipped with a task status monitor and an exception rollback module.

[0084] The task status monitor periodically collects the execution status of each atomic task from each execution terminal and sensor, marking it as "pending execution, in execution, completed, timed out", and records the actual execution time.

[0085] The exception rollback module periodically scans the task status table. When it detects that an atomic task is in the "timeout" state or its execution time deviates from the predetermined baseline by more than a preset threshold (e.g., exceeding the predetermined time by 50% or being delayed by more than 5 minutes), it triggers the strategy rollback process: immediately retrieves the latest updated individual digital twin status snapshot of the elderly person from the digital twin construction module, re-invokes the risk simulation module and care plan generation module, generates an updated personalized intervention strategy that takes into account the impact of the current deviation, cancels all unexecuted tasks in the original task queue that were not anchored for rollback, and re-injects the newly generated atomic task queue into the task scheduler for reassignment. In this way, the system can dynamically adapt to unexpected interruptions and delays during execution, ensuring the effective achievement of care intentions.

[0086] Finally, the execution status is fed back to the cloud-based digital twin platform through the closed-loop scheduling execution layer to drive the status update of the individual digital twin, thus forming a closed loop.

[0087] Specific examples are as follows: An elderly person has a history of hypertension and coronary heart disease, and multimodal sensors have been installed in their home.

[0088] At 11:20 PM, the cardiac impact map collected by the piezoelectric film mattress sensor showed a continuous downward trend in heart rate variability for 15 minutes. The ultra-wideband radar showed that the elderly person was in a supine position and frequently turned over, while the room temperature data was normal. Based on this, the lightweight early warning model of the edge computing layer generated a local warning of "abnormal heart rate variability" and sent it to the nurse station.

[0089] Simultaneously, the multi-source fusion time series data was synchronized to the cloud-based digital twin platform. The digital twin construction module updated the individual digital twin of the elderly person. The dynamic identification results of the physiological state sub-model showed that the estimated value of peripheral vascular resistance parameters increased by 18% compared with the daytime baseline. The behavioral pattern sub-model showed that the current state was a transition from deep sleep to light sleep. The environmental interaction sub-model showed continuous interaction in the bed area.

[0090] At 11:25 PM, the risk simulation module used the updated multidimensional state vector as input, and the time-series causal prediction model predicted the risks for the next 30 minutes: the probability of falling due to orthostatic hypotension was 0.76, and the causal path was weakened autonomic nervous regulation → blood pressure fluctuation → change of body position when getting up to urinate at night → insufficient blood supply to the brain → fall.

[0091] The care plan generation module calls the care knowledge graph, and the graph reasoning engine retrieves relevant intervention paths: "increase blood pressure monitoring frequency → assist with getting up earlier → enable bedside rails → adjust nighttime lighting". The priorities are sorted in order, personalized intervention strategies are generated, and then sent to the task scheduler.

[0092] The task scheduler breaks down the strategy into three atomic tasks: Task A, immediately reminding the family member via the smart pillbox and family terminal to adjust the bedtime blood pressure medication time; Task B, assigning the nearest caregiver to check on the patient and measure blood pressure at 11:45 PM; Task C, instructing the bedside rails to automatically rise and the night light to be switched to a soft mode. The caregiver's mobile terminal receives Task B and reports their location and status in real time.

[0093] At 11:48 PM, the task status monitor detected that Task B was in the "Not Started" state, exceeding the preset threshold, triggering the anomaly rollback module. The anomaly rollback module retrieved the latest digital twin status and found that the elderly man, Zhang, had suddenly experienced a rise in heart rate from 68 bpm to 92 bpm, exacerbating the already high blood pressure estimate. The risk simulation was rerun, updating the fall probability to 0.91, and a new enhanced intervention strategy was generated: immediately call two nearby caregivers to assist him in getting up and contact the on-duty doctor. The system canceled the original Task B, reassigned the collaborative task, and ultimately the two caregivers arrived at 11:52 PM and successfully assisted the elderly man, preventing a fall. The execution feedback caused the digital twin status to be updated again, and the risk probability dropped back to 0.12. The entire process formed a continuous adaptive closed loop.

[0094] This solution preprocesses data through an edge computing layer to generate multi-source fused time series data, and a lightweight health event early warning model generates local early warnings. A cloud-based digital twin platform constructs and dynamically updates an individual digital twin for each elderly resident. This twin includes three sub-models: physiological state, behavioral patterns, and environmental interaction. The platform's embedded temporal causal prediction model extrapolates the probability and causal path of future risk events based on the current state of the individual digital twin. The care plan generation module infers personalized intervention strategies through a care knowledge graph. A closed-loop scheduling execution layer decomposes the strategy into atomic tasks and assigns them to execution terminals, while simultaneously feeding back execution data to the digital twin construction module to update the twin's state, forming a closed loop. The temporal causal prediction model is a deep neural network that integrates spatiotemporal attention mechanisms and structural causal models. By constructing individual digital twins and introducing causal inference, it achieves the analysis of the intrinsic mechanisms of risk and proactive intervention, significantly improving the intelligence level and care safety of elderly care institution operations.

[0095] Please see Figure 2 As shown, the second objective of this invention is to provide a system for realizing a digital-based smart elderly care institution operation and management method, including a multimodal data acquisition layer, an edge computing layer, a cloud-based digital twin platform, a care plan generation module, and a closed-loop scheduling and execution layer. The multimodal data acquisition layer includes multiple sensor nodes deployed within the elderly care institution to collect multimodal perception data that includes at least physiological data, behavioral data, and environmental data. The edge computing layer is connected to the multimodal data acquisition layer, including a data preprocessing unit and a local inference engine. The data preprocessing unit is used to clean, timestamp align and handle missing values ​​of multimodal perception data to generate multi-source fusion time series. The local inference engine is equipped with a lightweight health event early warning model obtained by knowledge distillation, which is used to generate local early warning signals in real time based on the multi-source fusion time series. The cloud-based digital twin platform is connected to the edge computing layer via a secure communication link. The cloud-based digital twin platform includes a digital twin construction module, a risk simulation module, and a care plan generation module. The digital twin construction module is used to build and continuously update an individual digital twin for each elderly person based on multi-source fusion time series. The individual digital twin includes a physiological state sub-model representing cardiovascular and respiratory dynamics, a behavioral pattern sub-model representing daily activity sequences, and an environmental interaction sub-model representing the interaction between the elderly person and environmental facilities. The risk extrapolation module embeds a pre-trained temporal causal prediction model to receive the current multidimensional state vector of an individual digital twin and extrapolate the probability of various care risk events occurring within a preset time window and the causal influence path between events, outputting the risk probability distribution; the temporal causal prediction model is a deep neural network that integrates spatiotemporal attention mechanism and structural causal model. The care plan generation module is used to input the risk probability distribution into a preset care knowledge graph and generate personalized intervention strategies for the elderly through graph reasoning algorithms. The closed-loop scheduling execution layer communicates with the cloud-based digital twin platform and includes a task scheduler and multiple execution terminals. The task scheduler decomposes personalized intervention strategies into atomic task queues and assigns tasks based on the status of the execution terminals. The execution terminals execute the assigned tasks and report the execution status. The closed-loop scheduling execution layer also reports the execution status to the digital twin building module to drive the status update of individual digital twins.

[0096] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A digitally-based smart elderly care institution operation and management method, characterized in that: Includes the following steps: S1. Collect multimodal sensing data, including at least physiological, behavioral, and environmental data, through multiple sensor nodes deployed in elderly care facilities; S2. The multimodal sensing data is preprocessed using the configured edge computing layer, including data cleaning, timestamp alignment and missing value processing, to generate a multi-source fusion time series. S3. The edge computing layer uses a lightweight health event early warning model obtained through knowledge distillation to generate local early warning signals in real time based on the multi-source fusion time series. S4. Build a cloud-based digital twin platform, and in conjunction with the multi-source fusion time series, construct and continuously update an individual digital twin for each elderly person. The corresponding individual digital twin includes a physiological state sub-model, a behavioral pattern sub-model, and an environmental interaction sub-model. S5. The risk inference module in the cloud-based digital twin platform inputs the current multidimensional state vector of the individual digital twin into a pre-trained temporal causal prediction model to infer the probability of occurrence of various care risk events within a preset time window and the causal influence path between events, and outputs the risk probability distribution; the temporal causal prediction model is a deep neural network that integrates spatiotemporal attention mechanism and structural causal model. S6. Construct a care plan generation module, input the risk probability distribution into a preset care knowledge graph, and generate a personalized intervention strategy for the elderly through graph reasoning algorithm; S7. Configure the closed-loop scheduling execution layer to decompose the personalized intervention strategy into atomic task queues, and assign and execute tasks based on the status of each execution terminal, and receive execution status feedback. S8. The closed-loop scheduling execution layer feeds back the execution status to the cloud-based digital twin platform to drive the status update of the individual digital twin, forming a closed loop.

2. The digital-based smart elderly care institution operation and management method according to claim 1, characterized in that: The various sensors in S1 include at least a piezoelectric thin film mattress sensor, an ultra-wideband radar sensor, a wearable multi-parameter physiological monitoring bracelet, an environmental passive infrared sensor, a temperature and humidity sensor, a light sensor, a noise sensor, and a sensor built into a smart pillbox.

3. The digital-based smart elderly care institution operation and management method according to claim 1, characterized in that: The method for preprocessing multimodal sensing data in S2 includes the following steps: S2.

1. Perform outlier detection and removal for each sensing data channel. Outlier detection uses the sliding interquartile range method. S2.2 Incorporate multi-source heterogeneous data into a unified time series analysis framework, and perform precise alignment through a multi-scale dynamic time warping algorithm. Each time step after alignment contains valid data from all sensors at that moment or the nearest neighbor moment. S2.3 The missing value processing submodule uses a multiple interpolation algorithm based on chain equations to complete the missing segments.

4. The digital-based smart elderly care institution operation and management method according to claim 1, characterized in that: The method for generating local early warning signals in real time in S3 includes the following steps: S3.1 The lightweight health event early warning model is acquired using a teacher-student knowledge distillation framework. The teacher model is a deep temporal hybrid network pre-trained on large-scale historical data in the cloud, and the student model adopts a lightweight two-layer temporal convolutional network structure. S3.2 Utilize the inference engine built into the lightweight health event early warning model to receive the multi-source fusion time series output by the data preprocessing unit in a continuous stream manner; S3.3 The lightweight health event early warning model performs sliding window inference on the multi-source fusion time series. It maintains a sliding window with a width of 300 time steps, corresponding to a 30-second historical data segment. The window slides forward 10 time steps each time. S3.4 After the multidimensional time series segments within the window are normalized to zero mean and unit variance, they are fed into the student model for a forward propagation calculation, outputting the probability estimate of the occurrence of each preset health event at that moment. The corresponding output is a probability vector, with each dimension corresponding to a health event category.

5. The digital-based smart elderly care institution operation and management method according to claim 4, characterized in that: The real-time generation of local early warning signals in S3 includes: The alert trigger timestamp is accurate to milliseconds. Health event categories use predefined health event classification codes; The event confidence score is the smoothed probability value rounded to the nearest percentile. The extracted brief context snapshot includes the mean and rate of change of key physiological indicators such as heart rate and respiratory rate in the three seconds prior to the triggering of the warning.

6. The digital-based smart elderly care institution operation and management method according to claim 1, characterized in that: The method for building a cloud-based digital twin platform in S4 includes the following steps: S4.1 Retrieve the elderly person's electronic health record from the institutional information system. The record contains structured demographic information, a list of past medical history, current medication regimen, allergy history, surgical history, a sequence of physical examination indicators for the past six months, and nursing assessment results upon admission. Map this to an initial ontology instance. The ontology is coded using standard medical system nomenclature and clinical terminology. S4.2 The cloud-based digital twin platform uses the internal hidden state of an individual digital twin as the variable to be estimated through the unscented particle filter algorithm; The hidden state vector contains the dynamic parameters of the physiological state sub-model, the hidden state sequence of the behavioral pattern sub-model, and the interaction strength parameters of the environmental interaction sub-model. S4.

3. Use unscented particle filtering to maintain a set of 500 particles, each particle containing three parts: The first part consists of the latent variables of the physiological state sub-model, specifically the cardiovascular dynamics parameter values ​​and respiratory dynamics parameter values ​​at the current moment; The second part is the hidden state identifier of the behavior pattern sub-model, specifically the current bottom-level behavior atom category code, the middle-level activity segment category code, the remaining dwell time in seconds of the middle-level activity, and the top-level day-night rhythm time period index; The third part is the interaction intensity parameter of the environmental interaction sub-model, specifically the current interaction probability value of each functional area; S4.

4. For each particle, the physiological state sub-model is advanced using the fourth-order Runge-Kutta numerical integration method to obtain the predicted physiological state; S4.

5. Stepwise solve the lumped parameter ordinary differential equations of cardiovascular and respiratory dynamics to generate the evolution trajectory of physiological indicators such as heart rate, blood pressure, and blood oxygen. S4.6, Promoting the behavioral pattern sub-model, we obtain the predicted mid-level activity categories and new remaining dwell time; S4.7, Promoting the environmental interaction sub-model, we obtained a complete prior prediction of the current time step state; S4.8 Perform model calibration on the constructed model. After calculating the mixture likelihood weights of 500 particles, normalize the weight vector so that the sum of all weights equals one.

7. The digital-based smart elderly care institution operation and management method according to claim 1, characterized in that: The method for outputting the risk probability distribution in S5 includes the following steps: S5.1 Input the current multidimensional state vector of the individual digital twin into the pre-trained temporal causal prediction model through the risk inference module; S5.2 The risk inference module embeds a pre-trained temporal causal prediction model, which is a deep neural network that integrates spatiotemporal attention mechanism and structural causal model. The model structure consists of three layers: a graph structure learning layer, a spatiotemporal graph attention encoder, and a causal convolutional decoder. S5.3 Introducing Counterfactual Intervention Loss: Randomly intervene in some causal variables in the training samples to adjust their values, forming counterfactual samples. This forces the model to learn that the change in the probability distribution after intervention is consistent with the true causal effect, and outputs the risk probability distribution.

8. The digital-based smart elderly care institution operation and management method according to claim 1, characterized in that: The method for generating a personalized intervention strategy for the elderly in step S6 includes the following steps: S6.1 Configure care knowledge graph construction units and graph reasoning engine; S6.2 The care knowledge graph is constructed using RDF triples format, defining a directed heterogeneous graph. ,in For a set of entity nodes, Let be the set of edges. For a set of relation types, A collection of attributes; S6.3 The graph reasoning engine adopts a link prediction algorithm based on graph attention network, which calculates the link score between the node and all intervention entities through a bilinear graph attention network. S6.

4. Use bundle search to generate Top-K intervention paths. Each path consists of multiple ordered intervention actions. Each action is marked with execution priority, suggested execution time window and required resources. The resulting personalized intervention strategy is a structured XML or JSON object. Generate a personalized intervention strategy for the elderly.

9. A system for implementing the digital-based smart elderly care institution operation and management method as described in claim 1, characterized in that: It includes a multimodal data acquisition layer, an edge computing layer, a cloud-based digital twin platform, a care plan generation module, and a closed-loop scheduling and execution layer. The multimodal data acquisition layer includes multiple sensor nodes deployed in elderly care institutions to collect multimodal perception data, which includes at least physiological data, behavioral data, and environmental data. The edge computing layer is communicatively connected to the multimodal data acquisition layer and includes a data preprocessing unit and a local inference engine. The data preprocessing unit is used to clean, timestamp align and handle missing values ​​of the multimodal sensing data to generate a multi-source fusion time series. The local inference engine is deployed with a lightweight health event early warning model obtained by knowledge distillation, which is used to generate local early warning signals in real time based on the multi-source fusion time series. The cloud-based digital twin platform is connected to the edge computing layer via a secure communication link. The cloud-based digital twin platform includes a digital twin construction module, a risk simulation module, and a care plan generation module. The digital twin construction module is used to construct and continuously update an individual digital twin for each elderly person based on the multi-source fusion time series. The individual digital twin includes a physiological state sub-model representing cardiovascular and respiratory dynamics, a behavioral pattern sub-model representing daily activity sequences, and an environmental interaction sub-model representing the interaction between the elderly person and environmental facilities. The risk extrapolation module embeds a pre-trained temporal causal prediction model to receive the current multidimensional state vector of the individual digital twin, and extrapolate the probability of various care risk events occurring within a preset time window and the causal influence path between events, outputting a risk probability distribution; the temporal causal prediction model is a deep neural network that integrates spatiotemporal attention mechanism and structural causal model. The care plan generation module is used to input the risk probability distribution into a preset care knowledge graph and generate a personalized intervention strategy for the elderly through graph reasoning algorithm; The closed-loop scheduling and execution layer is communicatively connected to the cloud-based digital twin platform and includes a task scheduler and multiple execution terminals. The task scheduler is used to decompose the personalized intervention strategy into atomic task queues and assign tasks based on the status of the execution terminals. The execution terminals are used to execute the assigned tasks and report the execution status. The closed-loop scheduling and execution layer also reports the execution status to the digital twin building module to drive the status update of the individual digital twin.

Images