Data real-time analysis method and system for smart old-age care community

Abstract

The invention discloses a data real-time analysis method and system for an intelligent old-age care community, and the method comprises the following steps: collecting the heart rate, blood pressure, body temperature and other sign data of old people, generating a multi-channel time sequence data stream, carrying out the data preprocessing and standardization, inputting the standardized data into a TabPFN model, and carrying out the health state prediction. An improved TinyML algorithm is combined to realize real-time anomaly detection and risk assessment of the health data; through analysis of health state deviation and change trend, a health risk assessment result can be generated, and health early warning information can be pushed in time. According to the invention, the precision, real-time performance and adaptability of health monitoring can be significantly improved, and more intelligent health monitoring service is provided for the elderly.

Description

Technical Field

[0001] This invention relates to the field of data processing and analysis technology, and in particular to a real-time data analysis method and system for smart elderly care communities. Background Technology

[0002] Current health monitoring technologies for the elderly in smart senior living communities face some shortcomings. Existing health monitoring systems mainly rely on a single data source, such as assessing the health status of the elderly solely through vital signs data. However, a single data source cannot comprehensively reflect the multidimensional changes in the health status of the elderly, especially in the face of sudden health events, where the monitoring system lacks sufficient real-time feedback capabilities and early warning mechanisms.

[0003] Current health monitoring methods are mainly based on traditional manual examinations or simple data analysis, which suffer from low monitoring accuracy and poor real-time performance, failing to promptly identify subtle changes in the health status of the elderly. When faced with real-time acquisition of multi-channel data, the computing efficiency and data processing capabilities of existing systems cannot meet the needs of low-power devices, resulting in inefficient real-time analysis and feedback of health data.

[0004] In addition, most existing health monitoring systems lack adaptive adjustment mechanisms, meaning they cannot automatically adjust the model's analysis strategy and early warning thresholds when health data fluctuates, resulting in low accuracy in predicting health risks and an inability to operate efficiently on low-power devices.

[0005] Therefore, how to provide a real-time data analysis method and system for smart elderly care communities is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a real-time data analysis method and system for smart elderly care communities. This invention fully integrates multi-channel real-time health data monitoring with adaptive learning algorithms, constructing a closed-loop optimized process including data acquisition, preprocessing, feature extraction, health status prediction, anomaly identification, and real-time feedback. By introducing regression analysis based on the TabPFN model and an improved TinyML algorithm, this invention achieves accurate prediction and health risk assessment of multi-dimensional health data in elderly health monitoring. This invention possesses efficient data processing capabilities, accurate health risk identification, real-time health status updates, and adaptive adjustment capabilities, significantly improving the accuracy and response speed of health monitoring. This effectively solves the problems of low data processing efficiency, delayed risk assessment, and lack of real-time feedback and adjustment in existing health monitoring methods.

[0007] A real-time data analysis method for smart elderly care communities according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect vital sign data of the elderly to form a time-series data stream of vital signs. Based on the differences in the value range and collection frequency of different vital sign data, perform preprocessing to generate multi-channel standard data.

[0009] S2. Input the multi-channel standardized data into the TabPFN model for regression analysis to predict future changes in the health status of the elderly and output the predicted data.

[0010] S3. Input the predicted data generated by the TabPFN model and the current real-time health data in the multi-channel standard data into the improved TinyML algorithm for analysis, detect abnormal signs and mark the abnormal time and the data type of the signs.

[0011] S4. Compare the predicted data generated by TabPFN with real-time health data, detect abnormal signs and mark abnormal time periods, analyze the deviation of health status and the trend of health changes, and output the risk level based on the health risk score.

[0012] S5. Integrate abnormal time periods, abnormal signs, deviations in health status, trends in health changes, and risk levels to generate a structured output report, which is then sent to senior living community staff and seniors' families via user-end display and notification push functions.

[0013] Optionally, S1 specifically includes:

[0014] S11. Collect vital sign data of elderly people, including heart rate, blood pressure, body temperature and posture changes, and form four sets of time-series data streams of vital signs.

[0015] S12. Based on the differences in the numerical ranges of the four types of vital signs data (heart rate, blood pressure, body temperature, and posture), the units are standardized for each data type, and all data are retained to two decimal places.

[0016] S13. Use the sliding window averaging method to smooth each set of vital signs time series data streams. Take 5 consecutive data points as a window, calculate the average value of the data in the window, and assign the average value to the center point of the window. Move the window one time point to the right. Repeat the process until the complete vital signs time series data stream is processed.

[0017] S14. Perform time alignment on the smoothed vital sign time series data stream to determine a unified sampling time series. If a vital sign channel is missing data at a certain time point, find the two closest known data points before and after the time point, calculate the time difference between the previous and subsequent values, as well as the time difference between the missing point and the previous value, and use linear interpolation to estimate the missing data by weighting the previous value, fill in all the missing data, and form multi-channel standard data.

[0018] Optionally, S2 specifically includes:

[0019] S21. Sort the multi-channel standardized data according to the time series and reconstruct it into a tabular structure for input into the TabPFN model. Each row corresponds to a time point and each column corresponds to a vital sign data. All vital sign data at each time point will be used as an instance of the TabPFN model input.

[0020] S22. Within the TabPFN model, for each vital sign, the data is mapped to a high-dimensional continuous vector space by searching the pre-trained embedding matrix. The weights of various vital sign data are dynamically preset according to different needs and weighted, and the output is a sequence of vital sign data vectors after feature embedding and weighting.

[0021] The S23 and TabPFN models use the Transformer architecture to encode the sequence of vital sign data vectors. Each sequence of vital sign data vectors is processed through a self-attention mechanism to obtain the weighted feature representation of the vital signs at each time point.

[0022] The S24 and TabPFN models input the weighted feature representation into a multilayer perceptron or fully connected layer for further nonlinear mapping. The weighted feature representation is then entered into an MLP or fully connected layer. The weighted feature representation is multiplied by the preset weight matrix of each layer, a bias term is added, and a nonlinear transformation is performed through the ReLU activation function to generate activation values, which are then used as inputs to the next layer. Through the multilayer structure, deep feature vectors are gradually extracted from the weighted feature representation.

[0023] S25. Based on the feedforward neural network, the features are combined. The feedforward neural network structure is formed by multiple fully connected layers. The input deep feature vector is multiplied by the preset weight matrix of each layer and a bias is added. The result is passed to the Sigmoid activation function for nonlinear transformation and the abstract feature vector is extracted step by step.

[0024] S26. Multiply the abstract feature vector by the output layer weight matrix and add a bias term to map the abstract feature vector to a single numerical output, which serves as the prediction data for the regression model.

[0025] Optionally, the self-attention mechanism specifically includes the following processing:

[0026] Add a positional code with time as the data location to the vital sign data vector in each vital sign data vector sequence to form a new input vector with time information, and then arrange it in time sequence and input it into the Transformer model.

[0027] The Transformer model uses a self-attention mechanism to calculate the query, key, and value vectors at each time point of the input vector. It calculates the dot product similarity between the query vector and all key vectors, normalizes the dot product results using the Softmax function, and then sums the corresponding value vectors with the normalized weights to obtain the output of the current query.

[0028] The outputs of the current query for all input vectors will be concatenated and combined through a linear transformation to obtain a weighted feature representation of the vital signs at each time point.

[0029] Optionally, S3 specifically includes:

[0030] S31. Obtain the predicted data and real-time health data generated by the TabPFN model. The predicted data and real-time data contain the values ​​of each vital sign at different time points. Ensure that the timestamp of each vital sign data is consistent with the timestamp of the corresponding predicted data. By aligning the timestamps, combine the predicted data at each time point with the corresponding real-time health data into a data instance to form a multi-channel time series data stream.

[0031] S32. Input the multi-channel time series data stream into the improved TinyML algorithm for analysis. Normalize the multi-channel time series data stream according to the sampling frequency and data amplitude of each vital sign. Subtract the mean from the data of each vital sign and divide by the standard deviation to generate vital sign data of the same scale.

[0032] S33. Use the sliding window method to perform time series analysis on multi-channel time series data streams. Set a window and process the time series data of each vital sign in segments. Each time the sliding window moves, the data in the window is smoothed. In each sliding window, the data is sorted and the rank difference and the sum of squared rank differences of each pair of vital sign data in the sorting are calculated to obtain the nonlinear correlation score of each group of data.

[0033] S34. Use the nonlinear correlation score to measure the nonlinear dependence of multi-channel time series data stream within a specific time window, that is, the strength of the relationship between vital signs data. If the change of a certain vital sign exceeds the normal range of the predicted data, the normal range is the average value of the nonlinear correlation score plus or minus the standard deviation, then it is considered that an abnormal fluctuation in health status has occurred, and the abnormal time and vital sign data type are marked.

[0034] Optionally, the improved TinyML algorithm specifically includes:

[0035] Introducing the RAdam optimizer, when using RAdam, the initial learning rate is set to a constant value, and the gradient momentum and the mean of the weighted squared gradient are calculated for each parameter;

[0036] During training, the mean squared error loss is calculated based on each batch of data, and the gradients of all trainable parameters are calculated. For each parameter, the RAdam algorithm calculates the first moment estimate, which is the weighted average of the gradients, and the second moment estimate, which is the weighted average of the squares of the gradients.

[0037] RAdam calculates the corrected learning rate for each parameter based on the weighted average of the gradients and the weighted average of the squares of the gradients. In the first two rounds of gradient updates, RAdam uses bias correction to adjust the learning rate, which gradually makes the estimates of the first and second moments closer to the true values.

[0038] Based on the corrected first-order moment estimates and second-order moment estimates, the RAdam algorithm calculates the adaptive learning rate for each parameter. It adds the corrected mean gradient to the corrected mean squared gradient and divides by the square root of the corrected squared gradient to obtain the learning rate for each parameter. The global learning rate is multiplied by the corrected mean gradient and divided by the square root of the corrected squared gradient, then a constant value is added to obtain the adaptive learning rate for each parameter.

[0039] The calculated adaptive learning rate is used to update each parameter. The value of each parameter at the previous time step is subtracted from the calculated adaptive learning rate, and the parameter value at the current time step is updated.

[0040] Optionally, S4 specifically includes:

[0041] S41. Align the predicted data generated by the TabPFN model with the real-time health data by time point to detect abnormal signs. Compare the predicted data with the corresponding real-time health data. Use the absolute difference method to measure the difference between the real-time health data and the predicted data for each sign. If the difference of a certain sign exceeds the set abnormal sign threshold, it is marked as an abnormal sign.

[0042] S42. Analyze multiple time points that indicate abnormal vital signs. If multiple consecutive time points are marked as abnormal, define them as abnormal segments and record the start and end times of the abnormal segments. Output the time when the abnormality occurred.

[0043] S43. Obtain real-time health status data and historical health status data. Extract the data of each vital sign at the current time point and the historical time point to calculate the difference value. Squar the difference value at each time point, sum the squared differences, divide by the total number of data points to obtain the mean square error. If the mean square error exceeds the set health status deviation threshold, it is considered that there is a significant deviation in the health status.

[0044] S44. Extract the values ​​of the same vital signs at each time point from real-time health data and historical data respectively. Subtract the value of the previous time point from the current time point value and divide by the time difference to obtain the rate of change of each vital sign. Perform a weighted average of the rate of change of all vital signs with different preset feature weights to obtain the weighted average value. Output the health change trend according to the health status change trend rules.

[0045] S45. The nonlinear correlation score of the signs marked as abnormal is weighted and summed with the artificially preset feature weights of the signs. The minimum value is subtracted from each weighted summation result, and the result is divided by the difference between the maximum and minimum values ​​to obtain the normalized health risk score. Low risk, medium risk and high risk are output according to the preset intervals.

[0046] Optionally, the rules governing the trend of health status changes specifically include:

[0047] Acquire each vital sign from real-time health data, calculate the change range of each vital sign between the current time point and the previous time point, and obtain the change range of the same vital sign in historical health data.

[0048] The real-time changes in vital signs are compared with the average of historical changes. If the real-time changes exceed the preset range of historical changes, it is considered an abnormal escalation. If the changes gradually decrease from above the preset range to within the preset range, it is considered an abnormal relief. If the current changes are within the preset range of the historical average changes, the health status is considered to be stable.

[0049] Optionally, S5 specifically includes: combining the time period of abnormal occurrence, abnormal signs, deviation of health status, trend of health change, and risk level in a structured format to form a health status report. Each record is constructed according to time-sign-status-trend-level. Each structured data is sent to the terminal devices of the bound elderly care community staff and the elderly's family members through a message push interface to push health monitoring results in real time, ensuring that changes in health status can be transmitted in a timely manner and remotely warned and intervened.

[0050] A real-time data analysis system for smart elderly care communities according to an embodiment of the present invention includes:

[0051] The data acquisition and preprocessing module is used to collect vital sign data of the elderly, perform noise reduction, normalization and time series standardization processing, and generate multi-channel standard data in a unified format.

[0052] The health status prediction module is used to input standardized data into the TabPFN model, perform regression analysis, and predict future changes in the health status and potential risks of the elderly.

[0053] The anomaly detection module is used to analyze and predict data and real-time health data using the TinyML algorithm, detect abnormal signs, and mark abnormal fluctuations and sign types.

[0054] The health risk assessment module is used to compare predicted data and real-time health data, detect abnormal signs, analyze deviations and trends in health status, and output the health risk level.

[0055] The report generation and push module is used to integrate information on abnormal time periods, vital signs, status, and risk levels, generate structured reports, and push them to staff and their families through user terminals.

[0056] The beneficial effects of this invention are:

[0057] This invention proposes a real-time data analysis method and system for smart elderly care communities. Addressing the problems of low data processing efficiency, delayed health risk assessment, and lack of real-time feedback and adjustment in existing elderly health monitoring methods, this invention improves the accuracy of health status prediction and effectively achieves timely assessment and prediction of health risks by combining multi-channel health data acquisition with adaptive learning algorithms. Using the TabPFN model for regression analysis, combined with an improved TinyML algorithm, this invention enhances the processing capacity of health data and can automatically adjust the analysis model to ensure more efficient health management. This method operates efficiently on low-power devices, can monitor and provide feedback on health changes in real time, and offers personalized health warnings and interventions. It effectively solves the problems of data lag, low accuracy, and slow feedback in traditional health monitoring methods, thereby improving the health management level of the elderly in smart elderly care communities. Attached Figure Description

[0058] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0059] Figure 1 This is a flowchart of a real-time data analysis method for smart elderly care communities proposed in this invention;

[0060] Figure 2 This is a structural diagram of a real-time data analysis system for smart elderly care communities proposed in this invention;

[0061] Figure 3 This is a flowchart of the operation of health status prediction and risk assessment based on the combination of the TabPFN model and the improved TinyML algorithm proposed in this invention.

[0062] Figure 4 The flowchart illustrates the introduction of an adaptive learning rate adjustment mechanism in the improved TinyML algorithm proposed in this invention. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0064] refer to Figure 1-4 A real-time data analysis method for smart elderly care communities includes the following steps:

[0065] S1. Collect vital sign data of elderly people, including heart rate, blood pressure, body temperature and posture changes, to form a set of time-series data streams containing multiple types of indicators. Based on the differences in the value range and collection frequency of different vital sign data, preprocess them separately to generate multi-channel standard data.

[0066] S2. Input the multi-channel standardized data into the TabPFN model for regression analysis to predict future changes in the health status of the elderly and output predicted data including potential health risks and health change trends.

[0067] S3. Input the predicted data generated by the TabPFN model and the current real-time health data in the multi-channel standard data into the improved TinyML algorithm for analysis, detect abnormal signs and mark the abnormal time and the data type of the signs.

[0068] S4. Compare the predicted data generated by TabPFN with real-time health data, detect abnormal signs and mark abnormal time periods, analyze the deviation of health status and the trend of health changes, and output the risk level based on the health risk score.

[0069] S5 integrates abnormal time periods, abnormal signs, deviations in health status, trends in health changes, and risk levels to generate a structured output report. This report is then sent to senior living community staff and seniors' families via user-end display and notification push functions to achieve real-time early warning and health status monitoring.

[0070] This implementation method collects real-time data on multiple vital signs of elderly individuals, including heart rate, blood pressure, body temperature, and posture changes. This data is then standardized to ensure consistency and accuracy. Regression analysis of health status using the TabPFN model predicts future health trends and potential health risks. An improved TinyML algorithm further analyzes this data, detecting abnormal vital signs and marking health fluctuations in real time, ensuring timely identification of any health risks. The system analyzes deviations and trends in health status to output risk levels, helping elderly care community staff and families stay informed about the health status of the elderly. Combining real-time health data with health risk assessments, structured reports are generated and timely warnings are issued via push notifications. This not only improves the accuracy of health monitoring but also significantly enhances response speed, reducing delays and errors in traditional detection methods, ensuring the efficiency and accuracy of elderly health management, and truly achieving rapid assessment and intervention of health risks.

[0071] In this embodiment, S1 specifically includes:

[0072] S11. Collect vital sign data of elderly people, including heart rate, blood pressure, body temperature and posture changes. Heart rate data is recorded at a frequency of 10 times per second, blood pressure data is recorded at a frequency of 5 times per second, body temperature data is recorded at a frequency of 1 time per second, and posture data is set to be sampled at 15 times per second according to the output frequency of different sensors, forming four sets of time-series data streams of vital signs, each set of data having a different sampling frequency.

[0073] S12. Based on the differences in the numerical range of the four types of vital signs data, namely heart rate, blood pressure, body temperature and posture, the units are standardized. The heart rate unit is standardized to bpm, the body temperature unit is standardized to degrees Celsius, the blood pressure unit is standardized to mmHg, and the posture data is converted into angle or acceleration values ​​according to the sensor type. All data are retained to two decimal places.

[0074] S13. The sliding window averaging method is used to smooth each set of vital signs time series data streams. Every 5 consecutive data points are used as a window. The average value of the data in the window is calculated and the average value is assigned to the center point of the window. The window is moved one time point to the right. The process is repeated until the complete vital signs time series data stream is processed, thereby removing short-term fluctuations caused by factors such as respiratory tremors and unstable equipment contact.

[0075] S14. Perform time alignment on the smoothed vital sign time series data stream to determine a unified sampling time series. If a vital sign channel is missing data at a certain time point, find the two closest known data points before and after the time point, calculate the time difference between the previous and subsequent values, as well as the time difference between the missing point and the previous value, and use linear interpolation to estimate the missing data by weighting the previous value, fill in all the missing data, and form multi-channel standard data.

[0076] In this embodiment, S2 specifically includes:

[0077] S21. Sort the multi-channel standardized data according to the time series to ensure that the timestamps corresponding to each vital sign data are consistent, and reconstruct it into a tabular structure to input into the TabPFN model. Each row corresponds to a time point, and each column corresponds to a vital sign data. All vital sign data at each time point will be used as an instance of the TabPFN model input.

[0078] S22. Within the TabPFN model, for each vital sign, the data is mapped to a high-dimensional continuous vector space by searching the pre-trained embedding matrix. The weights of various vital sign data are dynamically preset according to different needs and weighted, and the output is a sequence of vital sign data vectors after feature embedding and weighting, so as to better understand the relationship between various vital signs.

[0079] The S23 and TabPFN models use the Transformer architecture to encode the sequence of vital sign data vectors. Each sequence of vital sign data vectors is processed through a self-attention mechanism to obtain the weighted feature representation of the vital signs at each time point.

[0080] The S24 and TabPFN models input the weighted feature representation into a multilayer perceptron or fully connected layer for further nonlinear mapping. The weighted feature representation is then entered into an MLP or fully connected layer. The weighted feature representation is multiplied by the preset weight matrix of each layer, a bias term is added, and a nonlinear transformation is performed through the ReLU activation function to generate activation values, which are then used as inputs to the next layer. Through the multilayer structure, deep feature vectors are gradually extracted from the weighted feature representation.

[0081] S25. Based on the feedforward neural network, the features are combined and the structure of the feedforward neural network is formed by multiple fully connected layers. The input deep feature vector is multiplied by the preset weight matrix of each layer and a bias is added. It is then passed to the Sigmoid activation function for nonlinear transformation, gradually extracting higher-order abstract feature vectors and capturing complex patterns that affect changes in health status.

[0082] S26. Multiply the abstract feature vector by the output layer weight matrix and add a bias term to map the abstract feature vector to a single numerical output, which serves as the prediction data for the regression model, representing the change in health status at a future point in time or over a period of time. This prediction data can be a prediction of health risk, a change in vital signs, or a predicted value of a certain indicator of health status.

[0083] This implementation sorts multi-channel standardized data according to time series, ensuring consistency in the timestamps corresponding to each vital sign data point, and reconstructs it into a tabular structure as input to the TabPFN model. Vital sign data at each time point is mapped to a high-dimensional continuous vector space and weighted to better understand the relationships between different vital signs. The TabPFN model, based on the Transformer architecture, encodes this vital sign data and obtains weighted feature representations for each time point through a self-attention mechanism, ensuring fine-grained capture of health data. The weighted feature representations are input to a multilayer perceptron or fully connected layer for nonlinear mapping, further extracting deep features to enhance the predictive ability of health status changes. A feedforward neural network is used to combine deep features, capturing complex patterns of health status changes and ensuring accurate extraction of key features influencing health changes. The abstract feature vector is multiplied by the weight matrix of the output layer and mapped to a single numerical value through a nonlinear activation function, serving as the predictive output of the regression model and accurately representing future changes in health status. This invention, by comprehensively considering multidimensional vital sign data, significantly improves the accuracy and real-time performance of health status prediction, effectively enhancing the precision and response speed of health risk warnings.

[0084] In this embodiment, the self-attention mechanism specifically includes the following processing:

[0085] Add a positional code with time as the data location to the vital sign data vector in each vital sign data vector sequence to form a new input vector with time information, and then arrange it in time sequence and input it into the Transformer model.

[0086] The Transformer model uses a self-attention mechanism to calculate the query, key, and value vectors at each time point of the input vector. It calculates the dot product similarity between the query vector and all key vectors, normalizes the dot product results using the Softmax function, and then sums the corresponding value vectors with the normalized weights to obtain the output of the current query.

[0087] The outputs of the current query for all input vectors will be concatenated and combined through a linear transformation to obtain a weighted feature representation of the vital signs at each time point.

[0088] In this embodiment, S3 specifically includes:

[0089] S31. Obtain the predicted data and real-time health data generated by the TabPFN model. The predicted data and real-time data contain the values ​​of each vital sign at different time points. Ensure that the timestamp of each vital sign data is consistent with the timestamp of the corresponding predicted data. By aligning the timestamps, combine the predicted data at each time point with the corresponding real-time health data into a data instance to form a multi-channel time series data stream.

[0090] S32. Input the multi-channel time series data stream into the improved TinyML algorithm for analysis. Normalize the multi-channel time series data stream according to the sampling frequency and data amplitude of each vital sign so that the model can process the input data of different vital signs. Subtract the mean and divide the data of each vital sign by the standard deviation to generate vital sign data of the same scale, ensuring that the numerical range of each vital sign is on the same scale.

[0091] S33. Use the sliding window method to perform time series analysis on multi-channel time series data streams. Set a window with a size of 16 and a step size of 4. Segment the time series data of each vital sign. When the sliding window moves, smooth the data in the window. Remove short-term fluctuations caused by unstable equipment contact, breathing and other factors by average filtering. In each sliding window, sort the data and calculate the rank difference and the sum of squared rank differences for each pair of vital sign data in the sorting to obtain the nonlinear correlation score of each group of data.

[0092] S34. Use the nonlinear correlation score to measure the nonlinear dependence of multi-channel time series data stream within a specific time window, that is, the strength of the relationship between vital signs data. If the change of a certain vital sign exceeds the normal range of the predicted data, the normal range is the average value of the nonlinear correlation score plus or minus the standard deviation, then it is considered that an abnormal fluctuation in health status has occurred, and the abnormal time and vital sign data type are marked.

[0093] This implementation aligns the predicted data generated by the TabPFN model with real-time health data using timestamps, forming a multi-channel time-series data stream for synchronous analysis. This data is then input into an improved TinyML algorithm for analysis. Data normalization ensures that different vital signs have the same scale, resolving differences in numerical ranges and guaranteeing accurate model processing. A sliding window method is used for time-series analysis. Data within each window is smoothed to remove short-term fluctuations, and the sum of squared rank differences is calculated to derive a nonlinear correlation score, further assessing the strength of relationships between various vital signs. Based on this, the nonlinear correlation score is used to identify abnormal fluctuations in health status. If a vital sign exceeds the normal range, it is marked as an abnormal fluctuation, and a timely warning is issued. This method significantly improves the accuracy and sensitivity of health data analysis, enabling rapid detection of subtle changes in health status and effective response to abnormal fluctuations, ensuring more precise and real-time health management for the elderly.

[0094] In this embodiment, the improved TinyML algorithm specifically includes:

[0095] Introducing the RAdam optimizer, when using RAdam, the initial learning rate is set to a constant value, and the gradient momentum and the mean of the weighted squared gradient are calculated for each parameter;

[0096] During training, the mean squared error loss is calculated based on each batch of data, and the gradients of all trainable parameters are calculated. For each parameter, the RAdam algorithm calculates the first moment estimate, which is the weighted average of the gradients, and the second moment estimate, which is the weighted average of the squares of the gradients.

[0097] RAdam calculates the corrected learning rate for each parameter based on the weighted average of the gradients and the weighted average of the squares of the gradients. In the first two rounds of gradient updates, since the estimates of the first and second moments are inaccurate, RAdam uses bias correction to adjust the learning rate. Bias correction gradually makes the estimates of the first and second moments closer to the true values, thereby stabilizing the learning rate and ensuring that the learning rate adjustment is more accurate even in the early stages.

[0098] Based on the corrected first-order moment estimates and second-order moment estimates, the RAdam algorithm calculates the adaptive learning rate for each parameter. It adds the corrected mean gradient to the corrected mean squared gradient and divides by the square root of the corrected squared gradient to obtain the learning rate for each parameter. The global learning rate is multiplied by the corrected mean gradient and divided by the square root of the corrected squared gradient, then a constant value is added to obtain the adaptive learning rate for each parameter.

[0099] The calculated adaptive learning rate is used to update each parameter. The value of each parameter at the previous time step is subtracted from the calculated adaptive learning rate, and the parameter value at the current time step is updated. This process ensures that the model's parameters are adjusted according to the adaptive learning rate in each iteration, thereby improving the model's training efficiency and optimizing the parameter update process.

[0100] In this embodiment, S4 specifically includes:

[0101] S41. Align the predicted data generated by the TabPFN model with the real-time health data by time point to detect abnormal signs. Compare the predicted data with the corresponding real-time health data. Use the absolute difference method to measure the difference between the real-time health data and the predicted data for each sign. If the difference of a certain sign exceeds the set abnormal sign threshold, it is marked as an abnormal sign.

[0102] S42. Analyze multiple time points that indicate abnormal vital signs. If multiple consecutive time points are marked as abnormal, define them as abnormal segments and record the start and end times of the abnormal segments. Output the time when the abnormality occurred.

[0103] S43. Obtain real-time health status data and historical health status data. Extract the data of each vital sign at the current time point and the historical time point to calculate the difference value. Squar the difference value at each time point, sum the squared differences, divide by the total number of data points to obtain the mean square error. If the mean square error exceeds the set health status deviation threshold, it is considered that there is a significant deviation in the health status.

[0104] S44. Extract the values ​​of the same vital signs at each time point from real-time health data and historical data respectively. Subtract the value of the previous time point from the current time point value and divide by the time difference to obtain the rate of change of each vital sign. Perform a weighted average of the rate of change of all vital signs with different preset feature weights to obtain the weighted average value. Output the health change trend according to the health status change trend rules.

[0105] S45. The nonlinear correlation score of the signs marked as abnormal is weighted and summed with the artificially preset feature weights of the signs. The minimum value is subtracted from each weighted summation result, and the result is divided by the difference between the maximum and minimum values ​​to obtain the normalized health risk score. Low risk, medium risk and high risk are output according to the preset intervals.

[0106] In this embodiment, the health status change trend rule specifically includes:

[0107] Acquire each vital sign from real-time health data, calculate the change range of each vital sign between the current time point and the previous time point, and obtain the change range of the same vital sign in historical health data.

[0108] The real-time changes in vital signs are compared with the average of historical changes. If the real-time changes are more than 30% greater than the historical changes, it is considered an abnormal escalation, indicating a significant deterioration in health. If the real-time changes are less than 30% of the historical changes and gradually decrease to less than 30%, it is considered an abnormal remission, indicating some improvement in health. If the current changes are within 30% of the historical average changes, the health is considered stable, indicating no significant fluctuations in health status.

[0109] In this embodiment, S5 specifically includes: combining the time period of abnormal occurrence, abnormal signs, deviation of health status, trend of health change, and risk level in a structured format to form a health status report. Each record is constructed according to time-sign-status-trend-level to ensure the clarity and completeness of the information. Each piece of structured data is sent to the terminal devices of the bound elderly care community staff and the elderly's family members through a message push interface to push health monitoring results in real time, ensuring that changes in health status can be transmitted in a timely manner and remotely warned and intervened.

[0110] A real-time data analysis system for smart elderly care communities includes:

[0111] The data acquisition and preprocessing module is used to collect vital sign data of the elderly, perform noise reduction, normalization and time series standardization processing, and generate multi-channel standard data in a unified format.

[0112] The health status prediction module is used to input standardized data into the TabPFN model, perform regression analysis, and predict future changes in the health status and potential risks of the elderly.

[0113] The anomaly detection module is used to analyze and predict data and real-time health data using the TinyML algorithm, detect abnormal signs, and mark abnormal fluctuations and sign types.

[0114] The health risk assessment module is used to compare predicted data and real-time health data, detect abnormal signs, analyze deviations and trends in health status, and output the health risk level.

[0115] The report generation and push module is used to integrate information on abnormal time periods, vital signs, status, and risk levels, generate structured reports, and push them to staff and their families through user terminals.

[0116] Example 1:

[0117] To verify the feasibility of this invention in practice, it was applied to a real-time health data analysis and monitoring system in a large smart elderly care community in a certain province. Within this community, the amount of health data collected from the elderly population is large, and their health status changes rapidly. Traditional health monitoring methods cannot meet the needs of real-time monitoring and prediction. The community collects over 5,000 health data entries daily, and each elderly resident requires real-time tracking of multiple vital signs such as heart rate, blood pressure, and body temperature. Furthermore, with the increasing number of elderly residents, the pressure on real-time health data analysis, health risk assessment, and early warning response is also growing.

[0118] Currently, the health monitoring systems used in senior living communities mainly rely on traditional manual methods for health checks and management. This method suffers from problems such as delayed data processing, lack of real-time feedback, and difficulty in responding quickly to health risks. In particular, when faced with sudden health emergencies, it is impossible to assess health risks in a timely and accurate manner, resulting in the failure to address the health problems of the elderly in a timely manner, which in turn affects the quality and effectiveness of health management.

[0119] To address the aforementioned issues, this invention proposes a real-time data analysis method and system based on the TabPFN model and an improved TinyML algorithm. In this approach, we acquire multi-channel data, such as heart rate, blood pressure, and body temperature, and combine this with the TabPFN regression model to predict the health status of the elderly. Simultaneously, we use the improved TinyML algorithm for real-time health risk assessment and anomaly detection. Through comprehensive analysis of multi-dimensional vital sign data, this system can acquire health data in real time and predict health trends, achieving rapid feedback and early warning of health data. We compared the effectiveness of this invention with traditional health monitoring methods in terms of health data monitoring, risk assessment, and early warning response. The experiment was conducted in a senior living community for 12 months, and a comparative test was performed between traditional manual detection and the method of this invention. The results are shown in the table below.

[0120] Table 1. Comparison of the effects of the method of the present invention and manual detection.

[0121] Detection method Monitoring Projects Total number of tests (items) Detection rate (%) False positive rate (%) False negative rate (%) Average inspection time per item (seconds) Timeliness of health risk assessment response (%) Manual inspection Heart rate monitoring 12800 92.5 5.4 7.8 4.5 70.1 This invention Heart rate monitoring 12800 98.3 2.2 1.6 1.2 96.5 Manual inspection Blood pressure monitoring 15500 89.8 6.1 10.5 5.0 72.3 This invention Blood pressure monitoring 15500 97.7 3.3 2.5 1.0 95.7 Manual inspection Body temperature monitoring 14200 86.6 8.0 12.2 5.3 68.9 This invention Body temperature monitoring 14200 99.1 1.3 0.7 1.4 98.3

[0122] As shown in Table 1, the method of this invention improves the defect detection accuracy by approximately 8.8 percentage points compared to manual inspection methods, and also performs better in terms of false positive and false negative rates. For example, in heart rate monitoring, the false positive rate of the method of this invention decreased from 5.4% to 2.2%, and the false negative rate decreased from 7.8% to 1.6%, with similar improvements observed in other monitoring items. The method of this invention also demonstrates significant advantages in terms of health data monitoring efficiency. For example, in heart rate monitoring, the average detection time per item of this invention is reduced from 4.5 seconds for manual inspection to 1.2 seconds, saving approximately 70% of the time. This greatly improves the efficiency of health monitoring, reduces waiting time for elderly individuals during health monitoring, and maintains high accuracy.

[0123] Furthermore, the method of this invention outperforms manual detection methods in terms of the timeliness of health risk assessment. In heart rate monitoring, the timeliness of the health risk assessment of this invention is 96.5%, while that of manual detection methods is only 70.1%. This indicates that this invention can promptly capture health fluctuations and provide real-time feedback, helping staff and families in elderly care communities to make rapid interventions. This invention is particularly suitable for monitoring complex health data. Taking blood pressure monitoring as an example, the false negative rate of the method of this invention is reduced from 10.5% for manual detection to 2.5%. This is due to the use of multi-channel data fusion and real-time dynamic adjustment mechanisms in this invention, which can more accurately identify subtle health changes, making it particularly suitable for health monitoring and intervention in high-risk elderly populations.

[0124] In summary, this invention not only improves the accuracy of health data monitoring, but also significantly enhances the efficiency of health monitoring and the timeliness of early warning response through adaptive learning algorithms and health risk assessment feedback mechanisms. It solves the problems of lag and inefficiency in traditional monitoring methods and has broad application value, especially in improving the health management of the elderly in smart elderly care communities.

[0125] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for real-time data analysis for a smart community for the elderly, characterized in that, The method comprises the following steps: S1, collecting the physical sign data of the elderly to form a physical sign time series data stream, and respectively preprocessing according to the differences in the value range and the acquisition frequency of different physical sign data to generate multi-channel standard data; S2, inputting the multi-channel standardized data into the TabPFN model for regression analysis to predict the future health status change of the elderly and output prediction data; S3, inputting the prediction data generated by the TabPFN model and the current real-time health data in the multi-channel standard data into the improved TinyML algorithm for analysis to detect abnormal signs and mark the abnormal time and sign data type; S4, comparing the prediction data generated by the TabPFN with the real-time health data to detect abnormal signs and mark the abnormal time period, analyze the health status deviation and health change trend, output the risk level according to the health risk score; S5, integrating the abnormal time period, abnormal signs, health status deviation, health change trend and risk level to generate a structured output report and send it to the staff of the elderly community and the family members of the elderly through the user end display and notification push function. 2.The data real-time analysis method for smart community for the elderly according to claim 1, wherein, The S1 specifically comprises: S11, collecting the physical sign data of the elderly, such as heart rate, blood pressure, body temperature and posture change, to form four groups of physical sign time series data streams; S12, according to the numerical range differences of the four types of physical sign data of heart rate, blood pressure, body temperature and posture, respectively performing unit uniform processing, and retaining the decimal places of all data to two digits; S13, using the sliding window average method to smooth each group of physical sign time series data streams, taking every 5 consecutive data points as a window, calculating the average value of the data in the window, and assigning the average value to the window center point, sliding the window to the next time point, and repeating the process until the complete physical sign time series data stream is processed; S14, time aligning the smoothed physical sign time series data stream to determine a unified sampling time sequence, if a physical sign channel is missing data at a certain time point, finding the nearest two known data points before and after the time point, calculating the time difference between the previous value and the next value, and the time difference between the missing point and the previous value, estimating the missing data by weighting the previous value using linear interpolation method, filling all missing data to form multi-channel standard data. 3.The data real-time analysis method for smart community for the elderly according to claim 1, wherein, The S2 specifically comprises: S21, sorting the multi-channel standardized data according to the time sequence and reconstructing it into a tabular structure to input the TabPFN model, each row corresponds to a time point, and each column corresponds to a physical sign data, wherein all physical sign data of each time point will be an instance input into the TabPFN model; S22, in the TabPFN model, for each physical sign, mapping each physical sign data to a high-dimensional continuous vector space by searching the pre-trained embedding matrix, artificially dynamically presetting the weight of various physical sign data according to different needs and performing weighted processing, and outputting the physical sign data vector sequence after feature embedding and weighting processing; S23, the TabPFN model encodes the sequence of sign data vectors using the architecture of the Transformer, each sequence of sign data vectors is processed by a self-attention mechanism to obtain a weighted feature representation of the sign at each time point; S24, the TabPFN model inputs the weighted feature representation into a multi-layer perceptron or a fully connected layer for further nonlinear mapping. After entering the MLP or fully connected layer, the weighted feature representation is multiplied by the preset weight matrix of each layer, a bias term is added, and a ReLU activation function is used for nonlinear transformation to generate an activation value and serve as the input of the next layer. Through the multi-layer structure, deep feature vectors are gradually extracted from the weighted feature representation; S25, the features are combined based on the feedforward neural network. The structure of the feedforward neural network is composed of multiple fully connected layers. The deep feature vectors are multiplied by the preset weight matrix of each layer and added with a bias, and then transmitted to a Sigmoid activation function for nonlinear transformation to gradually extract abstract feature vectors; S26, multiply the abstract feature vectors by the output layer weight matrix and add a bias term to map the abstract feature vectors to a single numerical output as the prediction data of the regression model. 4.The data real-time analysis method for smart community for the elderly according to claim 3, wherein, The self-attention mechanism processing specifically includes: Adding position encoding with time as the data position to the sign data vector in each sign data vector sequence to form a new input vector with time information and input into the Transformer model in time sequence order; The Transformer model uses the self-attention mechanism to calculate the query, key and value vectors at each time point. By calculating the dot product similarity between the query vector and all key vectors, using the Softmax function to normalize the dot product result, and weighting and summing the corresponding value vectors after normalization, the output of the current query is obtained; The outputs of the current query of all input vectors are spliced and combined through linear transformation to obtain the weighted feature representation of the sign at each time point. 5.The data real-time analysis method for smart community for the elderly according to claim 1, wherein, S3 specifically includes: S31, obtain the prediction data generated by the TabPFN model and the real-time health data. The prediction data and the real-time data respectively contain the values of each sign at different time points. The timestamp of each sign data ensures consistency with the timestamp of the corresponding prediction data. Through timestamp alignment, the prediction data at each time point and the corresponding real-time health data form a data instance to form a multi-channel time series data stream; S32, input the multi-channel time series data stream into the improved TinyML algorithm for analysis. Normalize the data of each sign according to the sampling frequency and data amplitude of each sign, subtract the mean value of each sign data and divide by the standard deviation to generate sign data of the same scale; S33, using a sliding window method to perform time series analysis on the multi-channel time series data stream, setting a window, segmenting the time series data of each sign, smoothing the data in the window each time the window moves, sorting the data in each sliding window, calculating the rank difference and rank difference square sum of each pair of sign data in the sorting, and obtaining the nonlinear correlation score of each group of data; S34, using the nonlinear correlation score to measure the nonlinear dependence of the multi-channel time series data stream in a specific time window, i.e. the mutual relationship strength between the sign data, if the change of a sign exceeds the normal range of the predicted data, the normal range being the average value of the nonlinear correlation score plus or minus the standard deviation, it is considered that an abnormal fluctuation of the health state has occurred, and the abnormal time and sign data type are marked. 6.The data real-time analysis method for smart community for the elderly according to claim 5, wherein, The improved TinyML algorithm specifically includes: Introducing the RAdam optimizer, setting the initial learning rate to a constant value when using RAdam, and calculating the momentum of the gradient and the weighted square gradient mean for each parameter; During the training process, the mean square error loss is calculated according to each batch of data, and the gradient of all trainable parameters is calculated. For each parameter, the RAdam algorithm calculates the first moment estimate, i.e. the weighted average of the gradient, and the second moment estimate, i.e. the weighted average of the square of the gradient; RAdam will calculate the modified learning rate of each parameter according to the weighted average of the gradient and the weighted average of the square of the gradient. In the first two rounds of gradient update, RAdam will use bias correction to adjust the learning rate. The bias correction will gradually make the estimates of the first and second moments closer to the true values. According to the modified first moment estimate and the second moment estimate, the RAdam algorithm calculates the adaptive learning rate of each parameter. The modified gradient mean is added to the modified gradient square mean, divided by the square root of the modified square gradient, and the learning rate of each parameter is obtained. The global learning rate is multiplied by the modified gradient mean, divided by the modified square root of the square gradient plus a constant value, and the adaptive learning rate of each parameter is obtained.

7. Using the calculated adaptive learning rate to update each parameter, subtracting the calculated adaptive learning rate from the value of each parameter at the previous time, and updating the parameter value at the current time. 8.The data real-time analysis method for smart community for the elderly according to claim 1, wherein, The S4 specifically includes: S41, aligning the predicted data generated by the TabPFN model with the real-time health data by time point for abnormal sign detection, comparing the predicted data and the corresponding real-time health data, using the absolute difference method to measure the difference between the real-time health data and the predicted data of each sign, and if the difference of a sign exceeds the set sign abnormal threshold, it is marked as an abnormal sign; S42, analyzing multiple time points marked as abnormal signs, defining an abnormal segment if multiple consecutive time points are marked as abnormal, and recording the start time and end time of the abnormal segment, and outputting the abnormal occurrence time; S43, obtain real-time health state data and historical health state data, extract the data of each sign at the current time point and the historical time point, calculate the difference value, square the difference value at each time point, sum the squared difference value, divide by the total number of data points, and obtain the mean square error, if the mean square error exceeds the set health state deviation threshold, it is considered that the health state deviates significantly; S44, extract the values of the same sign at each time point from the real-time health data and the historical data, subtract the value at the previous time point from the value at the current time point, divide by the time difference to obtain the change rate of each sign, and perform weighted average on the change rates of all signs according to the preset different feature weights to obtain the weighted average value, and output the health change trend according to the health state change trend rule; S45, the nonlinear correlation score of the sign marked as an abnormal sign is weighted and summed with the preset feature weight of the sign, each weighted sum result is subtracted by the minimum value, and divided by the difference between the maximum value and the minimum value to obtain the normalized health risk score, and low risk, medium risk and high risk are output according to the preset interval. 9.The data real-time analysis method for smart community for the elderly according to claim 7, wherein, The health state change trend rule specifically includes: Obtain each sign in real-time health data, calculate the difference value between each sign at the current time point and the previous time point to obtain the change amplitude, and obtain the change amplitude of the same sign in historical health data; Compare the real-time sign change amplitude with the average value of the historical sign change amplitude, if the real-time sign change amplitude is greater than the preset range of the historical sign change amplitude, it is determined as abnormal escalation, if it gradually decreases from the preset range to within the preset range, it is determined as abnormal alleviation, and if the current change amplitude and the historical average change amplitude differ within the preset range, it is considered that the health state is stable and maintained. 10.The data real-time analysis method for smart community for the elderly according to claim 1, wherein, The S5 specifically includes: combining the abnormal occurrence time period, abnormal sign, health state deviation, health change trend, risk level according to the structured format to form a health state report, each record is constructed according to time-sign-state-trend-level, each structured data is sent to the terminal equipment of the bound pension community staff and the old people's family members through the message push interface, the health monitoring result is pushed in real time, and the health state change can be transmitted in time and remote early warning and intervention is ensured.

11. A data real-time analysis system for a smart community for the aged, which executes the data real-time analysis method for a smart community for the aged according to any one of claims 1 to 9, characterized in that, It includes: A data acquisition and preprocessing module for acquiring elderly sign data, denoising, normalizing and time series standardization processing, and generating uniform format multi-channel standard data; A health state prediction module for inputting standardized data into a TabPFN model for regression analysis to predict future health state changes and potential risks of the elderly; An anomaly detection module for analyzing predicted data and real-time health data using TinyML algorithm to detect abnormal signs and mark abnormal fluctuations and sign types; A health risk assessment module for comparing predicted data and real-time health data, detecting abnormal signs and analyzing health state deviation and change trend to output health risk level; The report generation and pushing module is configured to integrate information of the abnormal time period, the sign, the state, and the risk level, generate a structured report, and display and notify the staff and the family members through the user end.

Images