Personalized health monitoring and early warning system for elderly people living alone and fusing multi-dimensional data

Abstract

The invention relates to the technical field of intelligent medical monitoring, in particular to a personalized health monitoring and early warning system for elderly people living alone fused with multi-dimensional data, and the system comprises a data collection and preprocessing module which collects multi-dimensional physiological parameter signals of a user and converts the multi-dimensional physiological parameter signals into multi-dimensional medical features; the health prediction module performs health category prediction on the multi-dimensional medical features of the current time node based on a pre-trained deep learning model, generates a data graph under each time node, and stores the data graph and a corresponding health category prediction result in a specified directory; the early warning module analyzes whether the multi-dimensional physiological parameter signals under each time node are abnormal or not in a mode of combining a hard threshold value and a soft threshold value, and if yes, an alarm is given out; and the abnormity output module calls and outputs the data maps of the preset duration before and after the abnormity alarm time node and the corresponding health categories. According to the invention, personalized monitoring and risk early warning of the health state of the elderly can be realized more accurately from multiple dimensions.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical monitoring technology, and more specifically to a personalized health monitoring and early warning system for elderly people living alone that integrates multi-dimensional data. Background Technology

[0002] In recent years, with the increasing aging of the population, the health monitoring of elderly people living alone has received much attention. Currently, common health monitoring devices on the market are mainly divided into contact and non-contact types. However, these commonly available devices all have some shortcomings. For example, contact devices are limited in terms of battery life, wearing comfort, and burden. Non-contact devices are affected by light, obstructions, and personal privacy protection factors, thus failing to achieve comprehensive, real-time, and accurate monitoring. Furthermore, traditional non-contact vital sign monitoring methods have limitations, only monitoring basic vital signs such as heartbeat and respiration, without transforming them into more valuable medical indicators, and thus failing to provide comprehensive and accurate predictions of potential health risks. At the same time, health monitoring often relies excessively on clinical data, without fully considering the specific physical condition and historical data of the monitored individual, making it difficult to conduct precise personalized analysis during monitoring and prediction. Summary of the Invention

[0003] In view of this, the present invention provides a personalized health monitoring and early warning system for elderly people living alone that integrates multi-dimensional data, which can more accurately monitor the health status and provide risk warnings for the elderly from multiple dimensions.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A personalized health monitoring and early warning system for elderly people living alone that integrates multi-dimensional data includes: a data acquisition and preprocessing module, a health prediction module, an early warning module, and an anomaly output module;

[0006] The data acquisition and preprocessing module is used to acquire the user's multidimensional physiological parameter signals, and perform preprocessing and medical feature conversion to obtain multidimensional medical features;

[0007] The health prediction module is used to predict the health category of the multidimensional medical features at the current time node based on the pre-trained deep learning model, and generate data maps at each time node. The data maps at each time node and the corresponding health category prediction results are saved together to the specified directory.

[0008] The early warning module is used to analyze whether the multidimensional physiological parameter signals at each time point are abnormal by combining hard threshold and soft threshold. If abnormal, an alarm is issued.

[0009] The anomaly output module is used to retrieve and output the data map and corresponding health category for a preset duration before and after the alarm issuance time.

[0010] Furthermore, the data acquisition and preprocessing module includes: a signal acquisition unit, a signal preprocessing unit, and a medical feature conversion unit;

[0011] The signal acquisition unit is used to acquire multidimensional physiological parameter signals of the user using millimeter-wave radar;

[0012] The signal preprocessing unit is used to separate heartbeat and respiratory signals from multidimensional physiological parameter signals;

[0013] The medical feature conversion unit is used to extract time-domain features, frequency-domain features, and nonlinear features from the separated heartbeat and respiratory signals, respectively; extract and calculate the peak value and heart rate variability (SDNN) of the signal from the time-domain features; extract and calculate the respiratory / heart rate energy and LF / HF power ratio from the frequency-domain features; extract and calculate the sample entropy and Hurst exponent from the nonlinear features; and integrate the extracted features into multidimensional medical features as a common indicator for predicting health categories.

[0014] Furthermore, the data acquisition and preprocessing module also includes a buffer unit;

[0015] The buffer unit is used to employ a fixed-length buffer with a capacity equal to the size of the sliding window. The sliding step size is calculated based on the proportion of repeated data between two adjacent sliding windows. The preprocessed multidimensional physiological parameter signals are added to the buffer using a first-in-first-out rule. After the buffer is full, medical feature conversion is performed.

[0016] Furthermore, the deep learning model includes a three-layer architecture: a 1D-CNN network, a BiLSTM network, and an attention network;

[0017] The 1D-CNN network includes three branches, each containing a convolutional layer, a layer normalization block, a ReLU activation function layer, and a SAME padding block; multidimensional medical features are processed by these three branches to obtain temporal data containing multi-scale local features.

[0018] The BiLSTM network inputs the temporal data output by the 1D-CNN network into a bidirectional LSTM layer to capture the bidirectional temporal dependencies between layers. After layer normalization, the data is passed to the next layer. Then, another bidirectional LSTM layer is used to mine long-range temporal correlations, retain the temporal output, and perform layer normalization to obtain a sequence containing global temporal features.

[0019] The attention network calculates the importance weights of the features at each time step output by the BiLSTM network, generates a query, key, and value matrix, and then weights and fuses it. This matrix is ​​then concatenated with the original output of the BiLSTM network to focus on key temporal information. Finally, the feature dimensions are compressed through global average pooling, and the classification probability of the health category is output through a fully connected layer activated by Softmax.

[0020] Furthermore, when training the deep learning model, the training data used includes clinical disease data, historical normal data of the monitored users, and historical abnormal data of the monitored users. Each sample consists of multidimensional medical feature data and health category labels.

[0021] Furthermore, the health prediction module also uses visualization tools to dynamically plot the multidimensional medical features acquired in real time, generating a data map. The data map is updated every second and saved every ten seconds, and named with a timestamp. The horizontal axis of the data map is the time series, the vertical axis is the medical feature value, and different medical feature curves are marked with different colors.

[0022] Furthermore, the early warning module includes a threshold initial setting unit, a threshold dynamic adjustment unit, and an alarm unit;

[0023] The threshold initial setting unit is used to establish a dedicated threshold for each physiological parameter signal. The threshold for each physiological parameter signal includes a hard threshold and a soft threshold. The physiological limit value of clinical medical consensus is used as the hard threshold, and 80% of the hard threshold range is used as the initial value of the soft threshold.

[0024] The threshold dynamic adjustment unit is used to analyze the user's physiological parameter signals over a period of time using a Gaussian mixture model, identify two distribution clusters, a resting cluster and an active cluster, select the cluster with a relatively large proportion of data points, and calculate and update the soft threshold based on the mean μ and standard deviation σ of the selected cluster. The upper limit of the updated soft threshold is μ+2σ, and the lower limit is μ-2σ.

[0025] The alarm unit is used to trigger an alarm when the physiological parameter signal at a certain time point exceeds the hard threshold, or exceeds the soft threshold three times in a row.

[0026] Furthermore, the system also includes a scene recognition module;

[0027] The scene recognition module is used to collect infrared images of the area where the user is located and to identify the scene where the user is currently located;

[0028] The threshold dynamic adjustment unit is also used to adjust the soft threshold corresponding to each physiological parameter signal according to the user's current scenario, ensuring that the adjusted soft threshold is less than or equal to the hard threshold, and storing the new soft threshold in the threshold dictionary.

[0029] Furthermore, the health prediction module also includes a model optimization unit; the model optimization unit is used to incrementally train the deep learning model;

[0030] The dataset used for incremental training consists of multidimensional medical feature data and corresponding pseudo-labels for each time point within the past month. The pseudo-labels include normal labels and abnormal labels. If an alarm is issued at a certain time point, the label of that time point is marked as abnormal, and the labels of the other time points are marked as normal.

[0031] Furthermore, the dataset used for incremental training is divided into a training set and a validation set. The model is replaced when the following three conditions are met simultaneously:

[0032] 1) Validation set prediction confidence improved by more than 5%:

[0033]

[0034] in, This represents the variance of the probability sequence from the old model. The variance of the probability sequence of the new model;

[0035] 2) The difference between the proportion of false label anomalies and the historical distribution is ≤2σ:

[0036]

[0037] |r new -μ|≤2σ

[0038] Where, r new To verify the anomaly ratio of pseudo-labels in the training set, μ is the mean of the daily anomaly ratio in the training set, and σ is the standard deviation of the daily anomaly ratio in the training set.

[0039] 3) The consistency rate of predictions between the old and new models is ≥85%:

[0040]

[0041] in, This indicates the predicted category of sample i by the old model. Let I be the predicted class of sample i by the new model, and let I be an indicator function, which is 1 when they are equal and 0 otherwise.

[0042] As can be seen from the above technical solution, compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. This invention addresses the problem of insufficient signal analysis in traditional health monitoring by transforming physiological parameters into medical features. Through the synergistic extraction of time-domain analysis (such as heart rate variability SDNN and peak value calculation), frequency-domain analysis (LF / HF power ratio, respiratory / heart rate energy), and nonlinear analysis (sample entropy, Hurst exponent), the original radar signal is transformed into physiological indicators with clear clinical significance. By fusing multi-dimensional medical features, the deep learning model can make more comprehensive and accurate predictions of the user's health category, breaking through the limitations of single analysis.

[0044] 2. This invention addresses two major pain points of traditional health monitoring systems: high false alarm rates due to static thresholds and poor adaptability to individual differences. It continuously incrementally trains the model and updates the thresholds using historical data from the monitored users. By setting both hard and soft thresholds, with physiological extremes as the hard threshold, the soft threshold is dynamically updated based on the specific historical data of the monitored individuals. The calculated new soft threshold is compared with the physiological extremes to avoid inaccurate thresholds and further improve the accuracy of abnormal alarms.

[0045] 3. This invention also incorporates user-defined scene data to achieve context-aware dynamic adjustment of thresholds. The updated soft thresholds that meet the standards are added to the threshold dictionary for use in the next prediction. If the new soft thresholds have problems, the old thresholds in the threshold dictionary are used for prediction.

[0046] 4. This invention employs a deep learning architecture combining 1D-CNN, BiLSTM, and an Attention mechanism to analyze and predict medical features after vital signs have been converted. The 1D-CNN layer is responsible for capturing local patterns of features, the BiLSTM layer analyzes the temporal dynamic changes of features, and the Attention mechanism automatically learns the importance weights of different time points for prediction. When key events are detected (such as apnea accompanied by heart rate oscillations), the weights of relevant features are automatically enhanced, achieving personalized and accurate prediction of the user's health status. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0048] Figure 1 This invention provides a structural block diagram of a personalized health monitoring and early warning system for elderly people living alone that integrates multi-dimensional data.

[0049] Figure 2This is a schematic diagram illustrating the medical feature conversion of physiological parameter signals by the data acquisition and preprocessing module provided by the present invention.

[0050] Figure 3 An architecture diagram of the deep learning model provided by this invention;

[0051] Figure 4 The flowchart for updating the soft threshold provided by this invention. Detailed Implementation

[0052] The technical solutions of 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.

[0053] like Figure 1 As shown in the figure, this invention discloses a personalized health monitoring and early warning system for elderly people living alone that integrates multi-dimensional data, including: a data acquisition and preprocessing module, a health prediction module, an early warning module, and an anomaly output module;

[0054] The data acquisition and preprocessing module is used to acquire the user's multidimensional physiological parameter signals, and to perform preprocessing and medical feature transformation to obtain multidimensional medical features;

[0055] The health prediction module is used to predict the health category of the multidimensional medical features at the current time point based on the pre-trained deep learning model, and generate data maps at each time point. The data maps at each time point and the corresponding health category prediction results are saved to the specified directory.

[0056] The early warning module is used to analyze whether the multidimensional physiological parameter signals at each time point are abnormal by combining hard thresholds and soft thresholds. If abnormality is found, an alarm is issued.

[0057] The anomaly output module is used to retrieve and output the data graphs and corresponding health categories for a preset duration before and after the alarm issuance time.

[0058] In one embodiment, the data acquisition and preprocessing module includes: a signal acquisition unit, a signal preprocessing unit, and a medical feature conversion unit;

[0059] The signal acquisition unit is used to acquire multidimensional physiological parameter signals of users using millimeter-wave radar;

[0060] The signal preprocessing unit is used to separate heartbeat and respiratory signals from multidimensional physiological parameter signals as vital signs;

[0061] like Figure 2 As shown, the medical feature conversion unit extracts time-domain, frequency-domain, and nonlinear features from the separated heartbeat and respiratory signals, respectively. From the time-domain features, it extracts and calculates the peak value and heart rate variability (SDNN) of the signal; from the frequency-domain features, it extracts and calculates the respiratory / heart rate energy and LF / HF power ratio; from the nonlinear features, it extracts and calculates the sample entropy and Hurst exponent; and it integrates the extracted features into multidimensional medical features as a common indicator for predicting health categories. For example, during the onset of cardiovascular disease, the peak interval decreases, the low-frequency / high-frequency ratio increases, and the sample entropy decreases.

[0062] Because the health risk early warning for elderly people living alone requires extremely high real-time performance, this invention adds a buffer unit to the data acquisition and preprocessing module to achieve real-time analysis of vital signs. The specific working process of the buffer unit is as follows:

[0063] First, create a fixed-length buffer with a capacity equal to the sliding window size. Then, calculate the sliding step size based on the proportion of duplicate data between two adjacent sliding windows (i.e., the overlap rate), expressed as:

[0064]

[0065] Sliding step size = window size × (1 - overlap ratio)

[0066] Next, a first-in, first-out (FIFO) rule is used to add the preprocessed multidimensional physiological parameter signals to the buffer. Once the buffer is full, medical feature conversion is performed. Specifically: real-time multidimensional physiological parameter signal data streams are continuously fed in for data preprocessing. Preprocessed data is continuously added to the buffer. When the data volume exceeds the buffer capacity, the oldest data entering the buffer is automatically evicted, ensuring the buffer always retains the latest N samples. When the buffer is full of N samples, medical feature conversion is performed. After medical feature conversion, the buffer is not cleared; new data continues to be fed in, and the buffer automatically discards old data, allowing the window to slide forward. These steps are repeated continuously, forming a loop of "receiving data → filling the window → extracting features → sliding the window," thereby achieving real-time conversion of human vital signs.

[0067] like Figure 3As shown, the deep learning model comprises a three-layer architecture: a 1D-CNN network, a BiLSTM network, and an attention network. First, the 1D-CNN network receives temporal input data after feature dimension transformation and extracts multi-scale local features through three convolutional layers, combining ReLU activation and layer normalization. Next, two bidirectional LSTM layers in the BiLSTM network capture the sequential dependencies of the temporal data, preserving temporal information and performing layer normalization. Then, the attention network calculates the importance weights for each time step, generating a query, key, and value matrix, which is then weighted and fused. This matrix is ​​then concatenated with the original output of the BiLSTM to focus on key temporal information. Finally, global average pooling compresses the feature dimension, and a fully connected layer activated by Softmax outputs the disease classification probability, realizing a complete disease classification process from local feature extraction to global temporal modeling and then to focusing on key information.

[0068] Specifically, the 1D-CNN network consists of three branches, each containing a convolutional layer, a layer normalization block, a ReLU activation function layer, and a SAME padding block. First, medical features are passed through a 1D convolutional layer with a kernel size of 5 and 64 filters, combined with ReLU activation function and SAME padding to extract local temporal features, and then stabilized by layer normalization. Subsequently, they are passed through 1D convolutional layers with kernel sizes of 3 and 128 filters and convolutional layers with kernel sizes of 3 and 256 filters, and the same layer normalization process is performed, outputting temporal data containing multi-scale local features.

[0069] The BiLSTM network inputs the temporal data output from the 1D-CNN network into a 128-unit bidirectional LSTM layer to capture bidirectional temporal dependencies. After layer normalization, the data is passed to the next layer. Another 64-unit bidirectional LSTM layer then mines long-range temporal correlations, retains the temporal output, and performs layer normalization to obtain a sequence containing global temporal features. The temporal features output by the BiLSTM network are represented as follows:

[0070]

[0071] Where T is the number of time steps (sequence length), dh is the dimension of the BiLSTM hidden layer, and h T ∈R dh Let be the feature vector at the T-th time step.

[0072] The attention network applies the temporal features of the BiLSTM network output through three fully connected layers (weight matrices).

[0073] W Q W K W V Generate a matrix of query, key, and value elements:

[0074] Q = HW Q (Query matrix)

[0075] K = HW K (Key matrix)

[0076] V = HW V (Value matrix)

[0077] Next, the importance weights of each time step in the sequence are calculated using the attention layer, resulting in a weighted attention output:

[0078]

[0079] Where, d k Let Q, K, and V be the projection dimensions.

[0080] The specific steps are as follows:

[0081] 1. Similarity calculation:

[0082]

[0083] Where S is the similarity matrix, and the elements S in matrix S are... ij This represents the similarity between the i-th query and the j-th time step key.

[0084] 2. Weight normalization:

[0085] A = Softmax(S) ∈ R T×T

[0086] Where A represents the attention weight matrix, and the elements A in matrix A are... ij This represents the attention weight of the i-th time step to the j-th time step.

[0087] 3. Weighted output:

[0088]

[0089] Where O is the attention-weighted feature representation, focusing on key temporal features.

[0090] The output of the attention network is concatenated with the original output of the BiLSTM network, fusing key temporal information (the attention-weighted output O) with global features (the original output H of the BiLSTM):

[0091]

[0092] After the above three steps of "1D-CNN+BiLSTM+Attention mechanism", the output data is compressed into a fixed-dimensional vector by global average pooling. Finally, the probability distribution of the disease is output through a fully connected layer. The output is converted into a normalized classification probability by the Softmax activation function. The disease code corresponding to the disease with the highest predicted probability is output. The disease type corresponding to the code is stored on the hard disk together with the subsequent data map.

[0093] When training the deep learning model, the training data used includes clinical disease data, historical normal data of monitored users, and historical abnormal data of monitored users. Each sample consists of multidimensional medical feature data and health category labels. In this embodiment, the training data includes 9,000 sets of clinical disease data, 3,000 sets of historical normal data of monitored users, and 1,500 sets of abnormal data, all stored in CSV file format. These CSV files consist of medical feature data and health category labels, for example: 45 (SDNN), 12 (peak value), 1.2 (respiratory / heart rate energy), 0.8 (LF / HF power comparison), 1.1 (sample entropy), 0.65 (Hurst index), and 0 (health category label).

[0094] Feature data from the training dataset, which consists of clinical disease data and historical data of the monitored individuals, is input into the deep learning model for training. The training output is an encoding that predicts the health category.

[0095] In one embodiment, the system of the present invention can realize dynamic real-time drawing and persistent storage of all data.

[0096] Specifically, the health prediction module uses visualization tools (such as Matlab) to dynamically plot the multidimensional medical features acquired in real time, generating a data map. This data map is updated every second and saved every ten seconds, named with a timestamp (e.g., "SDNN-20250212_120000.png"). The horizontal axis of the data map represents the time series, and the vertical axis represents the medical feature values, with different medical feature curves marked in different colors. The data map and corresponding health category prediction results can be saved in PNG format to a specified directory on the hard drive for persistent storage of all data.

[0097] More advantageously, the present invention can also achieve personalized health prediction for different users, mainly including dynamic threshold updates and model optimization.

[0098] Regarding dynamic threshold updates, this is primarily achieved through an early warning module, specifically, such as... Figure 4 As shown, the early warning module includes a threshold initial setting unit, a threshold dynamic adjustment unit, and an alarm unit;

[0099] The threshold initialization unit is used to establish a dedicated threshold for each physiological parameter signal. Each physiological parameter signal's threshold includes a hard threshold and a soft threshold. The physiological limits agreed upon in clinical medicine are used as the hard thresholds to ensure that human physiological limits are not exceeded; once set, the hard thresholds are not adjusted. 80% of the hard threshold range is used as the initial value for the soft threshold. When the data within the sliding window is less than 3 days, each parameter uses preset medical safety thresholds as the default hard thresholds (e.g., heart rate 35-200 bpm, blood oxygen 70-100%). These hard thresholds are based on clinical guidelines, and an alarm is immediately triggered if the data exceeds the hard threshold.

[0100] Once the data within the sliding window has accumulated for a preset duration, the dynamic threshold adjustment unit is activated. This unit uses a Gaussian mixture model to analyze the user's physiological parameter signals over a period of time (e.g., 3 days), identifying two distribution clusters. The cluster with the larger proportion of data points is selected, and the soft threshold is calculated and updated based on the mean μ and standard deviation σ of the selected cluster. The upper limit of the updated soft threshold is μ+2σ, and the lower limit is μ-2σ, ensuring that the updated soft threshold does not exceed the physiological limit, i.e., the hard threshold.

[0101] Specifically, after accumulating data for 3 days, Gaussian Mixture Model (GMM) analysis is performed independently for each physiological parameter. The specific process is as follows: First, the physiological parameter data from the past 3 days undergoes standardized preprocessing, including outlier removal (deleting data exceeding the clinical data limit ±5%) and feature engineering. Feature engineering refers to constructing state-discriminating features for cluster analysis, such as moving averages and difference sequences in time-domain derived features. This invention's method assumes that the monitoring data is generated from a mixture of resting and active states. The preprocessed data is iteratively optimized using the Expectation-Maximization (EM) algorithm to identify two distribution clusters (resting cluster and active cluster). The cluster with the larger proportion of data points (where the monitored person spends most of their time, resulting in more stable data that better reflects the true health level) is selected, and its mean μ and standard deviation σ are calculated. This generates soft thresholds (upper limit = μ + 2σ, lower limit = μ - 2σ). The generated soft thresholds are then compared with hard thresholds to avoid exceeding physiological thresholds. This dual-cluster design effectively distinguishes the normal physiological range under different scenarios, avoiding misjudgments caused by a single threshold.

[0102] The alarm unit is used to trigger an alarm when the physiological parameter signal at a certain time point exceeds the hard threshold, or exceeds the soft threshold three times in a row.

[0103] To dynamically adapt to different scenarios, the system is also equipped with a scene recognition module;

[0104] The scene recognition module uses an infrared sensor to collect infrared images of the area where the user is located and identifies the scene in which the user is currently located.

[0105] The threshold dynamic adjustment unit adopts a context compensation mechanism. By identifying the user's current scenario, it adjusts the soft threshold corresponding to each physiological parameter signal (the current soft threshold is multiplied by the compensation coefficient) to ensure that the adjusted soft threshold is less than or equal to the hard threshold. For example, when strenuous exercise is detected, the upper limit of the heart rate soft threshold is increased by 30%, and the upper limit of the respiratory rate is increased by 20%.

[0106] After calculation, the new soft threshold is stored in the threshold dictionary, and threshold detection is performed. If the hard threshold is exceeded, an alarm is triggered directly; if the hard threshold is not exceeded, a soft threshold detection is performed. To reduce the module's false alarm rate, an alarm is triggered after three consecutive exceedances of the soft threshold. After detection, the system returns to monitoring mode, continuously receiving new data, calculating the new soft threshold, and using the new soft threshold for judgment.

[0107] In terms of model optimization, this invention includes a model optimization unit within the health prediction module; the model optimization unit is used for incremental training of the deep learning model.

[0108] First, retrieve the user's physiological parameter signal data from the past month and perform standardized preprocessing, mainly to remove outliers (such as data points that exceed physiological limits). Then, perform medical feature transformation on these preprocessed data. Finally, the dataset used for incremental training consists of the user's multidimensional medical feature data and corresponding pseudo-labels for each time point in the past month. The pseudo-labels include normal labels and abnormal labels. If an alarm is issued at a certain time point, the label of that time point is marked as abnormal, and the labels of other time points are marked as normal.

[0109] The dataset used for incremental training is divided into a training set and a validation set. For example, the first 80% (the first 24 days) is used as the training set and the last 20% (the last 6 days) is used as the validation set. This data is then input into the current deep learning model for incremental training. The model is replaced when all three of the following conditions are met:

[0110] 1) Validation set prediction confidence improved by more than 5%:

[0111]

[0112] in, This represents the variance of the probability sequence from the old model. The variance of the probability sequence of the new model;

[0113] 2) The difference between the proportion of false label anomalies and the historical distribution is ≤2σ:

[0114]

[0115] |r new -μ|≤2σ

[0116] Where, rnew To verify the anomaly ratio of pseudo-labels in the training set, μ is the mean of the daily anomaly ratio in the training set, and σ is the standard deviation of the daily anomaly ratio in the training set.

[0117] 3) The consistency rate of predictions between the old and new models is ≥85%:

[0118]

[0119] in, This indicates the predicted category of sample i by the old model. Let I be the predicted class of sample i by the new model, and let I be an indicator function, which is 1 when they are equal and 0 otherwise.

[0120] In one embodiment, the abnormal output module can be connected to a printer. When an abnormal alarm is issued, it retrieves data graphs and corresponding health categories for a preset duration (e.g., 15 minutes) before and after the abnormal time point and prints them out for timely analysis during medical visits.

[0121] Through its various modules, this invention can create personalized and specific datasets based on clinical disease data and the historical data of the monitored individuals. It transforms collected vital signs into medical diagnostic features and integrates multiple features for comprehensive prediction. It can generate data maps in real time, permanently storing them on the hard drive in PNG format. When an anomaly is detected, it promptly outputs the predicted health category. Employing a dual-threshold + soft-threshold dynamic adjustment mechanism, it automatically adjusts the anomaly detection threshold to achieve personalized health risk warnings.

[0122] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0123] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A personalized health monitoring and early warning system for elderly people living alone that integrates multi-dimensional data, characterized in that, include: The system includes a data acquisition and preprocessing module, a health prediction module, an early warning module, and an anomaly output module. The data acquisition and preprocessing module is used to acquire the user's multidimensional physiological parameter signals, and perform preprocessing and medical feature conversion to obtain multidimensional medical features; The health prediction module is used to predict the health category of the multidimensional medical features at the current time node based on the pre-trained deep learning model, and generate data maps at each time node. The data maps at each time node and the corresponding health category prediction results are saved together to the specified directory. The early warning module is used to analyze whether the multidimensional physiological parameter signals at each time point are abnormal by combining hard threshold and soft threshold. If abnormal, an alarm is issued. The anomaly output module is used to retrieve and output the data map and corresponding health category for a preset duration before and after the alarm issuance time.

2. The personalized health monitoring and early warning system for elderly people living alone that integrates multi-dimensional data according to claim 1, characterized in that, The data acquisition and preprocessing module includes: a signal acquisition unit, a signal preprocessing unit, and a medical feature conversion unit; The signal acquisition unit is used to acquire multidimensional physiological parameter signals of the user using millimeter-wave radar; The signal preprocessing unit is used to separate heartbeat and respiratory signals from multidimensional physiological parameter signals; The medical feature conversion unit is used to extract time-domain features, frequency-domain features, and nonlinear features from the separated heartbeat and respiratory signals, respectively; extract and calculate the peak value and heart rate variability (SDNN) of the signal from the time-domain features; extract and calculate the respiratory / heart rate energy and LF / HF power ratio from the frequency-domain features; extract and calculate the sample entropy and Hurst exponent from the nonlinear features; and integrate the extracted features into multidimensional medical features as a common indicator for predicting health categories.

3. The personalized health monitoring and early warning system for elderly people living alone that integrates multi-dimensional data according to claim 2, characterized in that, The data acquisition and preprocessing module also includes a buffer unit; The buffer unit is used to employ a fixed-length buffer with a capacity equal to the size of the sliding window. The sliding step size is calculated based on the proportion of repeated data between two adjacent sliding windows. The preprocessed multidimensional physiological parameter signals are added to the buffer using a first-in-first-out rule. After the buffer is full, medical feature conversion is performed.

4. The personalized health monitoring and early warning system for elderly people living alone that integrates multi-dimensional data according to claim 1, characterized in that, The deep learning model comprises a three-layer architecture: a 1D-CNN network, a BiLSTM network, and an attention network; The 1D-CNN network includes three branches, each containing a convolutional layer, a layer normalization block, a ReLU activation function layer, and a SAME padding block; multidimensional medical features are processed by these three branches to obtain temporal data containing multi-scale local features. The BiLSTM network inputs the temporal data output by the 1D-CNN network into a bidirectional LSTM layer to capture the bidirectional temporal dependencies between layers. After layer normalization, the data is passed to the next layer. Then, another bidirectional LSTM layer is used to mine long-range temporal correlations, retain the temporal output, and perform layer normalization to obtain a sequence containing global temporal features. The attention network calculates the importance weights of the features at each time step output by the BiLSTM network, generates a query, key, and value matrix, and then weights and fuses it. This matrix is ​​then concatenated with the original output of the BiLSTM network to focus on key temporal information. Finally, the feature dimensions are compressed through global average pooling, and the classification probability of the health category is output through a fully connected layer activated by Softmax.

5. The personalized health monitoring and early warning system for elderly people living alone that integrates multi-dimensional data according to claim 1, characterized in that, When training the deep learning model, the training data used includes clinical disease data, historical normal data of monitored users, and historical abnormal data of monitored users. Each sample consists of multidimensional medical feature data and health category labels.

6. The personalized health monitoring and early warning system for elderly people living alone that integrates multi-dimensional data according to claim 1, characterized in that, The health prediction module also uses visualization tools to dynamically plot the multidimensional medical features acquired in real time, generating a data map. The data map is updated every second and saved every ten seconds, and named with a timestamp. The horizontal axis of the data map is the time series, the vertical axis is the medical feature value, and different medical feature curves are marked with different colors.

7. The personalized health monitoring and early warning system for elderly people living alone that integrates multi-dimensional data according to claim 1, characterized in that, The early warning module includes a threshold initial setting unit, a threshold dynamic adjustment unit, and an alarm unit; The threshold initial setting unit is used to establish a dedicated threshold for each physiological parameter signal. The threshold for each physiological parameter signal includes a hard threshold and a soft threshold. The physiological limit value of clinical medical consensus is used as the hard threshold, and 80% of the hard threshold range is used as the initial value of the soft threshold. The threshold dynamic adjustment unit is used to analyze the user's physiological parameter signals over a period of time using a Gaussian mixture model, identify two distribution clusters, a resting cluster and an active cluster, select the cluster with a relatively large proportion of data points, and calculate and update the soft threshold based on the mean μ and standard deviation σ of the selected cluster. The upper limit of the updated soft threshold is μ+2σ, and the lower limit is μ-2σ. The alarm unit is used to trigger an alarm when the physiological parameter signal at a certain time point exceeds the hard threshold, or exceeds the soft threshold three times in a row.

8. The personalized health monitoring and early warning system for elderly people living alone that integrates multi-dimensional data according to claim 7, characterized in that, The system also includes a scene recognition module; The scene recognition module is used to collect infrared images of the area where the user is located and to identify the scene where the user is currently located; The threshold dynamic adjustment unit is also used to adjust the soft threshold corresponding to each physiological parameter signal according to the user's current scenario, ensuring that the adjusted soft threshold is less than or equal to the hard threshold, and storing the new soft threshold in the threshold dictionary.

9. The personalized health monitoring and early warning system for elderly people living alone that integrates multi-dimensional data according to claim 7, characterized in that, The health prediction module further includes a model optimization unit; the model optimization unit is used to incrementally train the deep learning model. The dataset used for incremental training consists of multidimensional medical feature data and corresponding pseudo-labels for each time point within the past month. The pseudo-labels include normal labels and abnormal labels. If an alarm is issued at a certain time point, the label of that time point is marked as abnormal, and the labels of the other time points are marked as normal.

10. The personalized health monitoring and early warning system for elderly people living alone that integrates multi-dimensional data according to claim 9, characterized in that, The dataset used for incremental training is divided into a training set and a validation set. The model is replaced when the following three conditions are met simultaneously: 1) Validation set prediction confidence improved by more than 5%: in, This represents the variance of the probability sequence from the old model. The variance of the probability sequence of the new model; 2) The difference between the proportion of false label anomalies and the historical distribution is ≤2σ: |r new -μ|≤2σ Where, r new To verify the anomaly ratio of pseudo-labels in the training set, μ is the mean of the daily anomaly ratio in the training set, and σ is the standard deviation of the daily anomaly ratio in the training set. 3) The consistency rate of predictions between the old and new models is ≥85%: in, This indicates the predicted category of sample i by the old model. Let I be the predicted class of sample i by the new model, and let I be an indicator function, which is 1 when they are equal and 0 otherwise.

Images