A method for constructing a health impact prediction model based on coupling of multi-source air data

Abstract

The application discloses a kind of based on the construction method of health influence prediction model of multi-source air data coupling, it is related to air quality health monitoring technical field, including obtaining the multi-source air environment data in selected space as environmental index, simultaneously obtaining the multi-source physiological signal of person in selected space as physiological index, based on the feature vector of environmental index and physiological index, the coupling relationship between air environment feature and physiological feature is constructed, the nonlinear coupling relationship between air feature and physiological feature is modeled using time series deep learning model, the action law of air pollution cumulative exposure to physiological index dynamic change is captured.The application can not only reveal the comprehensive influence law of air pollution on human blood pressure, blood oxygen, body temperature, heart rate, heart rate variability and blood glucose physiological index by feature extraction and cross-modal coupling, but also realize dynamic prediction of health risk using deep learning model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of air quality health monitoring, and particularly relates to a method for constructing a health influence prediction model based on coupling of multi-source air data. BACKGROUND

[0002] In current health building and indoor environment research, although there are many monitoring and standard limit value provisions about air quality (such as CO2, PM2.5, TVOC and indoor temperature and humidity), the common practice at present is to detect whether the environmental indicators meet the IAQ (indoor air quality) limit value standard, but in practice, even if the standard is met for a short time, long-term cumulative exposure, for example, the influence of PM2.5 and TVOC on the daily average or weekly average level, may still have chronic effects on the cardiovascular, metabolic and respiratory systems, and the prior art lacks effective evaluation of the dynamic relationship between long-term exposure of air quality and human health.

[0003] In addition, traditional research is often based on population data to statistically analyze the correlation between air pollution and disease incidence, but it is difficult to fall into real-time health monitoring at the individual level.

[0004] At the individual application level, air monitoring and physiological monitoring are often fragmented, that is, air sensors only deal with environmental data, wearable devices only deal with physiological signals, and the two are not coupled, so it is impossible to further analyze problems such as "how long does a certain air factor affect the individual's blood pressure / heart rate". SUMMARY

[0005] I) Technical problem to be solved

[0006] The present application provides a method for constructing a health influence prediction model based on coupling of multi-source air data, which can analyze the influence of multi-source air data and human physiological factors in the time sequence dimension.

[0007] II) Technical scheme

[0008] To achieve the above object, the present application provides the following technical scheme: a method for constructing a health influence prediction model based on coupling of multi-source air data, comprising the following steps:

[0009] Continuously acquire multi-source air environment data in a selected space as environmental indicators, and acquire multi-source physiological signals of a person in the selected space as physiological indicators; wherein the environmental indicators include carbon dioxide, PM2.5, TVOC, temperature and humidity, and the physiological indicators include blood pressure, blood oxygen, body temperature, heart rate, heart rate variability and blood glucose;

[0010] The collected environmental indicators and physiological indicators are normalized and preprocessed, and are aligned in the time dimension to form time-synchronized sequence data;

[0011] Based on the pre-processed sequence data, the time sequence features of the environmental indicators and the physiological indicators are extracted within a set time window, the time sequence features of the environmental indicators include the change amplitude, the local time fluctuation and the cumulative exposure, and the time sequence features of the physiological indicators include the trend change, the local time fluctuation, the time domain feature and the frequency domain feature; each type of time sequence feature is converted into a corresponding feature vector sequence;

[0012] A dual-channel deep learning model is constructed, and the environmental feature vector sequence and the physiological feature vector sequence are respectively input into the corresponding channels of the deep learning model in time steps, for capturing the cumulative effect and the local fluctuation feature of the environmental indicators in the time dimension, and simultaneously capturing the change pattern of the physiological indicators over time; the deep learning model is trained to fuse and process the capture results output by each channel, so as to establish the coupling relationship between the cumulative exposure of the environmental indicators and the response of the physiological indicators;

[0013] The constructed and trained deep learning model takes the air environmental feature sequence and the individual feature information as input, generates a prediction of the physiological indicator time sequence and its dynamic change, and obtains the health impact prediction model based on the coupling of multiple sources of air data.

[0014] In an implementable embodiment, in the selected space, the carbon dioxide, PM2.5, TVOC, temperature and humidity are continuously monitored and sampled by a multi-source environmental monitoring device, and various air environmental data are continuously obtained at a set time interval to form the time sequence of the environmental indicators;

[0015] Meanwhile, through a wearable human monitoring device, the blood pressure, blood oxygen, body temperature, heart rate, heart rate variability and blood glucose of the human body in the selected space are obtained, and various physiological signals are continuously obtained at the same time interval to form the time sequence of the physiological indicators corresponding to the environmental indicators.

[0016] In an implementable embodiment, the continuously collected and obtained environmental indicators and physiological indicators are uniformly pre-processed, and the pre-processing operation includes detecting and filling missing values, identifying and rejecting abnormal values, smoothing noise signals, and normalizing the environmental indicators and physiological indicators;

[0017] After data preprocessing, the environmental indicators and the physiological indicators are aligned according to the time stamp through time stamp matching to form time-synchronized sequence data.

[0018] In an implementable embodiment, after data preprocessing and time alignment, based on a set time window, the time sequence features of the environmental indicators are extracted within each time window; correspondingly:

[0019] For each air environment data in the environmental indicators, the increment between adjacent sampling points is calculated to quantify the change amplitude in the time dimension to reflect the instantaneous fluctuation trend;

[0020] In each time window, the dispersion of each air environment data value in the corresponding time period is counted as the local fluctuation feature of the air environment data;

[0021] In each time window, the values of each air environment data in the time period are accumulated to calculate the cumulative exposure of the air environment data.

[0022] In a feasible embodiment, after data preprocessing and time alignment, the time sequence features of the physiological indicators are extracted in each time window based on the set time window. Correspondingly:

[0023] For each physiological signal in the physiological indicators, the overall change trend in the time window is calculated, including the direction and its change amplitude;

[0024] In each time window, the dispersion of each physiological signal in the corresponding time period is counted as the local fluctuation feature of the physiological signal;

[0025] In each time window, the mean and extreme values of each physiological signal in the corresponding time period are calculated as the time domain features;

[0026] In each time window, the oscillation intensity of each physiological signal in the frequency spectrum is extracted through frequency spectrum analysis to represent the physiological oscillation pattern.

[0027] In a feasible embodiment, the time sequence features of the environmental indicators extracted in each time window are combined to form an environmental feature vector, and the time sequence features of the physiological indicators extracted in each time window are combined to form a physiological feature vector;

[0028] The feature vectors of each time window are arranged in time sequence to form an environmental feature vector sequence and a physiological feature vector sequence.

[0029] In a feasible embodiment, the deep learning model is a dual-channel architecture. The time-sequenced environmental feature vector sequence and physiological feature vector sequence are respectively input into the corresponding channels of the deep learning model. Each channel is provided with an independent encoder to dynamically extract the feature vectors of the respective input sequences. The encoder outputs the hidden state sequence of the corresponding channel, and each hidden state sequence retains the time sequence information. The deep learning model includes an environmental feature channel and a physiological feature channel. The environmental feature channel is used to capture the cumulative exposure effect and local fluctuation features of the environmental indicators in the time dimension. The physiological feature channel is used to capture the change pattern and local change features of the physiological indicators over time.

[0030] The hidden state sequence of the two channel outputs at each time step is input to the intermediate fusion layer of the deep learning model, the two hidden state sequences at the corresponding time step are paired, the environmental feature vector is calculated using the attention mechanism, and the contribution weight of the physiological feature vector at the time step is generated to generate a weighted fusion vector.

[0031] In an available embodiment, for the weighted fusion vector generated by the intermediate fusion layer at each time step, the prediction layer of the deep learning model is input, the prediction layer outputs the predicted values of various physiological signals of the physiological indicators by regression to obtain the predicted values of the physiological indicators in a future set time interval; if the set time interval is a single time step, the output is the predicted values of various physiological signals of the next time step;

[0032] If the set time interval is a step of multiple consecutive time steps, the prediction results are recursively output on the future multiple consecutive time steps in a sequence decoding manner to form a set of predicted values of various physiological signals and their change trend;

[0033] The prediction layer combines the cumulative environmental exposure in the set prediction time interval to predict and output the risk level of various physiological signals.

[0034] In an available embodiment, in the training phase of the deep learning model, the fusion sequence of the historical environmental feature vector sequence and the individual feature information is input, and the real physiological indicators at the corresponding time step are used as labels, wherein the individual feature information includes the age, gender, BMI and disease history of the individual; the error between the predicted physiological indicators and the various physiological signals of the real physiological indicators is defined as a loss function, the gradient is calculated in the time dimension using the back propagation algorithm, and the parameters of the encoder corresponding to the environmental feature channel and the physiological feature channel, the attention mechanism of the intermediate fusion layer and the prediction layer are updated.

[0035] In an available embodiment, according to the prediction results of various physiological signals in the physiological indicators predicted and output by the deep learning model, the corresponding risk threshold of various physiological signals in the physiological indicators is set, and when the prediction result of any physiological signal exceeds the corresponding preset risk threshold, health prompt information is generated and visualized;

[0036] The selected space is provided with an air conditioning device, the prediction result of the physiological indicators is combined with the prediction result of the environmental indicators at the corresponding time point to generate a control reference signal, and after the control reference signal is mapped to the corresponding control amount of the air conditioning device, the air conditioning device is adjusted and driven.

[0037] III) Beneficial effects:

[0038] Compared with the prior art, the present application has the following beneficial effects:

[0039] The present application can make up for the limitations of relying only on a single air pollution index or a single health index in the prior art by jointly analyzing environmental indicators and human physiological indicators in the long time dimension.

[0040] By feature extraction and cross-modal coupling, multi-source environmental indicators (such as carbon dioxide, PM2.5, TVOC, temperature and humidity), human physiological indicators (blood pressure, blood oxygen, body temperature, heart rate, heart rate variability, blood glucose) and individual characteristic information (age, gender, BMI and disease history) are jointly analyzed, which not only reveals the comprehensive influence law of air pollution and environmental conditions on different individual core physiological indicators, but also utilizes the advantages of deep learning model in time sequence characteristics to realize dynamic prediction of individualized health risk. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 A flowchart of a construction method of a health impact prediction model based on multi-source air data coupling provided by an embodiment of the present application;

[0042] Figure 2 In the construction method of the health impact prediction model based on multi-source air data coupling provided by the embodiment of the present application, a schematic diagram of each operation process is shown.

[0043] Figure 3 In the construction method of the health impact prediction model based on multi-source air data coupling provided by the embodiment of the present application, a processing schematic diagram of the input data of the constructed deep learning module is shown.

[0044] Figure 4 In the construction method of the health impact prediction model based on multi-source air data coupling provided by the embodiment of the present application, a schematic diagram of a user wearing a wearable device to measure various physiological signals is shown.

[0045] Figure 5 In the construction method of the health impact prediction model based on multi-source air data coupling provided by the embodiment of the present application, according to the constructed health impact prediction model, a plurality of environmental data in a plurality of selected spaces and a plurality of physiological data of human body corresponding to the selected spaces are obtained within a certain time period, and a visual interface diagram of warning and regulating indoor air equipment after coupling prediction is shown. DETAILED DESCRIPTION

[0046] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be clearly and completely described below, obviously, the described embodiments are only a part of the embodiments of the present application, and not all the embodiments of the present application, based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the protection scope of the present application.

[0047] In the description of the present application, it should be understood that the orientation or positional relationship indicated by the terms "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0048] In addition, if the terms "first", "second" and the like are used, they are only used to distinguish description, and cannot be understood as indicating or implying relative importance.

[0049] It should be noted that the features in the embodiments of the present application can be combined with each other without conflict.

[0050] In combination Figures 1 to 5 The method for constructing a health impact prediction model based on multi-source air data coupling shown can not only reveal the comprehensive influence law of air pollution on the core physiological indicators of blood pressure, blood oxygen, body temperature, heart rate, heart rate variability and blood glucose of human body, but also can utilize the advantage of deep learning model on time sequence feature to realize dynamic prediction of health risk, not only improve the accuracy and sensitivity of prediction, but also provide more timely and personalized reference basis for individual health management and public health intervention.

[0051] Specifically, first refer to Figure 1 S10: continuously acquiring multi-source air environment data in the selected space as environmental indicators, and acquiring multi-source physiological signals of people in the selected space as physiological indicators; wherein the environmental indicators include carbon dioxide, PM2.5, TVOC, temperature and humidity, and the physiological indicators include blood pressure, blood oxygen, body temperature, heart rate, heart rate variability and blood glucose. The step is mainly to obtain time sequence data of the environmental indicators and the physiological indicators synchronously, and to provide a basis for subsequent pretreatment and feature extraction.

[0052] Regarding the selected space, which is the collection object of the environment, the key indicators in the air are mainly monitored in the selected indoor or enclosed space, including carbon dioxide (CO2), PM2.5 (fine particulate matter), and volatile organic compounds (TVOC), as well as the temperature and humidity in the room.

[0053] For the detection of carbon dioxide, in some embodiments of the present application, the CO2 concentration in the air is continuously measured in the selected space by using infrared non-dispersive absorption technology (NDIR) or setting corresponding electrochemical sensors. Such sensors output corresponding concentration signals by detecting the degree of absorption of specific wavelength infrared light by CO2 molecules.

[0054] For the detection of PM2.5 (fine particulate matter), laser scattering or optical particle counting sensor principles can be used. When the air flows through the measurement area of the sensor, the particulate matter scatters the laser, and the sensor detects the scattered light intensity to calculate the particulate matter concentration.

[0055] For volatile organic compounds (TVOC), metal oxide semiconductor (MOS) sensors or photoionization detection (PID) technology can be used to continuously detect the total volatile organic compound concentration in the air.

[0056] It needs to be further understood that temperature and humidity are also considered as part of the environmental indicators because temperature and humidity can affect the release rate of volatile organic compounds (such as TVOC) in the air. High temperature environment may accelerate the volatilization of TVOC, and high humidity environment may affect the distribution of particulate matter concentration through water vapor adsorption or dilution. In addition, the suspension characteristics of PM2.5 and other particulate matter are also affected by air density, humidity changes, and air flow, thereby affecting the amount of particulate matter inhaled by the human body.

[0057] Overall, the temperature and humidity in the selected space are considered as environmental background variables because they can change the sensitivity of the human body to the same concentration of pollutants. For example, under high temperature and high humidity, the blood pressure and heart rate may be more significantly affected by air pollution exposure; while in a low temperature and dry environment, respiratory indicators may be more susceptible to influence. The subsequent prediction model can learn the interaction between physical conditions such as temperature and humidity in the environment and air particulate matter exposure and physiological response, further capturing the dynamic action rules of cumulative exposure on physiological indicators.

[0058] The temperature and humidity in the room can be detected by environmental monitoring devices (such as temperature and humidity sensors) to obtain the temperature and relative humidity values in the air in real time. The collected data has a timestamp and can be synchronized with the existing carbon dioxide, PM2.5, and TVOC environmental indicator data to form a multi-source air environment time series.

[0059] In summary, by arranging multi-source environmental monitoring devices in the space, continuous monitoring of CO2, PM2.5, TVOC and temperature and humidity can be achieved. The sampling frequency of each device is not specifically limited, as long as the sampling frequency of each device is consistent and consistent with the sampling frequency of the subsequent physiological detection.

[0060] For the collection of various physiological signals of the person in the selected space, the person in the selected space can be monitored in real time by a wearable bracelet, which refers to Figure 4 The person monitors multiple physiological indicators in real time through a wearable bracelet, including blood pressure, blood oxygen, heart rate, heart rate variability (HRV), body temperature and blood glucose.

[0061] More specifically, for the detection of body temperature, blood pressure and blood oxygen, a wearable smart bracelet can be used. For blood oxygen and blood pressure detection, a photoplethysmography (PPG) sensor is built into the wearable smart bracelet, which detects blood flow changes through light signals and indirectly calculates blood pressure fluctuations; blood oxygen saturation (SpO2) is calculated by the PPG light signal infrared absorption ratio.

[0062] For heart rate and heart rate variability, the PPG or integrated electrocardiogram (ECG) sensor in the smart bracelet is used to obtain the heart rate interval sequence, calculate short-term and long-term heart rate variability indicators, and reflect autonomic nervous regulation.

[0063] For blood glucose in the physiological signal, if a wearable continuous blood glucose monitoring device can be synchronized with the bracelet data, a continuous blood glucose time series is generated; if the bracelet does not directly measure blood glucose, the CGM device data can be synchronized through the interface.

[0064] By wearing a bracelet for each participant, ensuring stable wearing and good sensor contact, starting continuous data collection, it should be noted that the sampling frequency of the bracelet for various physiological signals needs to be synchronized with the air environment data collection time to ensure time series alignment. Finally, the collected various physiological signals are uploaded in real time to the data management platform through Bluetooth or wireless network to form the physiological indicators corresponding to each time step, and finally form the multi-source time series data of the physiological indicators collected by the wearable bracelet and the environmental indicators.

[0065] In addition, on the basis of the above, in some embodiments of the present application, on the basis of collecting multi-source air environment data and physiological signals in the selected space, in order to construct a health impact prediction model that can reflect individual differences, individual characteristic information of each participant needs to be collected, including but not limited to:

[0066] Gender and age: reflecting the basic physiological differences of individuals;

[0067] BMI (Body Mass Index): used to reflect the body size differences between individuals, and different body size individuals may have an impact on the changes of physiological indicators;

[0068] Disease history: such as chronic diseases such as hypertension, diabetes, which will affect the sensitivity of physiological response.

[0069] In combination with the deep learning model built later, the multi-source physiological signal sequence data (blood pressure, blood oxygen, body temperature, heart rate, heart rate variability and blood glucose) of each individual during the experiment or monitoring, together with the air environment time series data constitute an input pair of a sample, where the individual characteristic information is used as a covariate, and the multi-source air environment data (environmental indicator feature vector) are input into the deep learning model built later, so as to form individual modeling capability. It can also be understood that a plurality of individual sample sets are combined to form a training set for fitting the time series deep learning model to learn the dynamic relationship between environmental indicators, individual characteristics and physiological indicators.

[0070] In summary, it can be understood that the individual characteristics are used as covariates to enable the model to capture the different physiological indicator response differences of individuals of different genders, ages, BMIs or disease histories under the same air environment conditions, and to improve the model generalization ability, so that the trained model is not only suitable for a specific individual, but also can be extended to other people with similar characteristics, and is more suitable for health risk assessment.

[0071] After data collection, data preprocessing is performed, specifically S20: normalizing the collected environmental indicators and physiological indicators, and aligning them in the time dimension to form time-synchronized sequence data. This step mainly aims to uniformly preprocess the data after the multi-source data collection of environmental indicators and physiological indicators, to ensure the reliability and effectiveness of subsequent feature extraction and coupled modeling.

[0072] First of all, it needs to be understood that the main purpose of preprocessing is to ensure that the environmental indicators and physiological indicators are aligned in the time dimension to form synchronized multi-source time series data, eliminate possible abnormalities, noise or missing data in the data collection process, and ensure the stability of model training and prediction.

[0073] Considering that during the data collection process, due to sensor failure, wireless transmission delay or non-standard wearing, there may be missing data. Therefore, in some embodiments of the present application, an interpolation method based on sliding window mean can be used to fill in the missing points and maintain the continuity of the time series. Other methods include but are not limited to linear interpolation, forward filling and backward filling, etc. The specific missing value detection and filling method is not limited here.

[0074] Considering that during the data collection process, abnormal peaks may be generated due to sensor abnormalities, external interference or collection errors. Therefore, identification and corresponding rejection of these abnormal values are needed. In some embodiments of the present application, a statistical method can be used, specifically using Z-score method, for example, when a certain value is detected to be more than 3 times the standard deviation of all values in a set time period, it is determined that the monitoring data is an abnormal value, and it is rejected or replaced by the adjacent valid value.

[0075] In addition, the time series collected by the sensor may contain high-frequency jitter or random noise, which can be removed by using a smoothing filter to remove short-term random fluctuations, making the data trend clearer.

[0076] Finally, it should be noted that due to the different dimensions of environmental indicators and physiological indicators, such as CO2 measured in , PM2.5 measured in , and blood pressure measured in , directly inputting the subsequently constructed model will cause training bias. Therefore, after all data is preprocessed, linear normalization or standardization of the data of each indicator is needed to convert the data of each indicator to a unified scale range, for example, all mapped to a value in the interval range of 0-1, to improve the stability of subsequent model training.

[0077] After preprocessing and normalizing the data, considering the collection frequency of environmental indicators and physiological indicators, in order to establish a feature coupling relationship, it is necessary to align at each time step.

[0078] Specifically, in some embodiments of the present application, by taking the environmental indicator collection frequency as the reference time step, or selecting a suitable unified sampling frequency as mentioned above, the physiological signal is then sampled or aggregated, for example, taking the average value or the last value in the time step, to ensure that each time step contains an environmental indicator vector and a physiological indicator vector.

[0079] It can be further understood that the aligned data forms a time-synchronized multi-source sequence, ensuring that each time step corresponds to a complete input vector, which facilitates subsequent feature extraction and coupling modeling.

[0080] In summary, the preprocessing and synchronization operations ensure the complete correspondence of environmental indicators and physiological indicators at each time step, eliminate collection errors and noise, and improve data quality, thereby providing a reliable foundation for subsequent feature extraction and coupling modeling.

[0081] Then, S30 is performed: based on the preprocessed sequence data, time series features of the environmental indicators and the physiological indicators are respectively extracted within a set time window, the time series features of the environmental indicators include a change amplitude, a local time fluctuation, and a cumulative exposure, the time series features of the physiological indicators include a trend change, a local time fluctuation, a time domain feature, and a frequency domain feature, and then each type of time series feature is converted into a corresponding feature vector sequence.

[0082] For this step, a summary understanding is first performed, that is, for each type of original time series (each type of air composition data, each type of physiological signal), a group of features describing time series dynamics (trend, fluctuation, cumulative, frequency domain) are respectively calculated on several time scales, and then these features of each source are structured into a feature vector sequence in a fixed format on each unified time step. Finally, the air feature vector sequence and the physiological feature vector sequence are aligned and combined according to the time stamp to obtain a fusion input vector sequence at each time step, which is used as an input for subsequent coupling modeling.

[0083] For carbon dioxide, PM2.5, TVOC, and temperature and humidity, etc., the air environment data forms a time series (such as a PM2.5 concentration sequence, a carbon dioxide concentration sequence, a TVOC concentration sequence, and a temperature and humidity sequence) after preprocessing, and the following features are extracted on this basis.

[0084] One is a trend feature (overall direction and amplitude), which is obtained by linear fitting or sliding window regression to obtain the slope of the change of the indicator over time, wherein a positive slope represents an upward trend, a negative slope represents a downward trend, and a value close to zero represents stability. Then, the average growth amplitude in a certain time range can be calculated, such as an increase of 10 .

[0085] One is a fluctuation feature (including stability and dispersion), which can be calculated by calculating the variance, standard deviation, and coefficient of variation within a set window to quantify the fluctuation strength of the indicator. In addition, in some embodiments, extreme value difference (maximum and minimum) can also be added to measure the fluctuation amplitude in a short time.

[0086] One is a cumulative exposure feature (long-term load), which can be obtained by integrating or weightedly accumulating the indicator in the time dimension to obtain the total exposure in the observation period, such as the cumulative PM2.5 exposure for several hours. In addition, in some embodiments, the proportion of exposure time exceeding the set value can also be calculated, such as the proportion of PM2.5 exceeding the set value of 75 The cumulative time.

[0087] Finally, considering that these particulate matter particles may exhibit periodicity or sudden occurrences during indoor decoration, some embodiments additionally consider extracting frequency domain features. This can be achieved by performing a Fast Fourier Transform (FFT) on the environmental indicator sequence to extract the spectral energy distribution, reflecting the periodic fluctuation characteristics of pollutants (such as diurnal variations and short-term peaks). Furthermore, the dominant frequency (the frequency corresponding to the maximum energy) and the energy concentration range can be marked to represent the dominant period of pollution fluctuations.

[0088] After preprocessing physiological signals to obtain smooth time series, the following types of features are extracted.

[0089] One is the trend characteristic (the trend of changes in health status), which is obtained by linear fitting or sliding window averaging of each physiological signal to obtain the direction and magnitude of change. For example, a sustained increase in heart rate may indicate that air quality is causing stress on the cardiovascular system.

[0090] One is the fluctuation characteristic (physiological stability), which can be reflected by calculating the standard deviation or coefficient of variation within each time window to show whether the physiological signal is stable. For example, short-term high fluctuations, such as fluctuating respiratory rate, may indicate that the person is responding to stressful environmental stimuli.

[0091] One is time-domain features (direct statistical features), including but not limited to statistical averages, maximums, minimums, medians, etc., which characterize the overall level. It can also include the frequency of peak occurrences, such as the number of heart rate spikes, and mutation point detection, which is used to identify time points of rapid changes in physiological responses.

[0092] The last one is frequency domain characteristics (physiological oscillation characteristics). This mainly involves performing spectral analysis on heart rate and blood pressure to obtain the power spectrum distribution, where the main peak frequency of the heart rate spectrum can correspond to the actual heart rate in a relaxed state.

[0093] After extracting a set of feature values ​​from the above air environment data at each time window (time step), an air feature vector is formed. Similarly, within the same time window, corresponding trend, fluctuation, time domain, and frequency domain features of physiological signals are extracted to form a physiological feature vector.

[0094] Then, align the air feature vectors and physiological feature vectors according to the timestamps, such as... Figure 2 As shown, at each time step (time window), the vectors are combined into a fusion vector. Finally, the fusion vectors from all time steps are arranged sequentially to form a fusion input sequence, which reflects the temporal coupling relationship between the air environment and physiological responses. In this way, the deep learning model subsequently constructed can directly learn the dynamic mapping relationship between air and physiology.

[0095] Afterwards, S40 is performed: a two-channel deep learning model is constructed, and the environment feature vector sequence and the physiological feature vector sequence are respectively input into the corresponding channels of the deep learning model in time steps, so as to capture the cumulative effect and local fluctuation characteristics of the environment index in the time dimension, and capture the change pattern of the physiological index over time; after the deep learning model fuses and processes the capture results output by each channel, a coupling relationship between the cumulative exposure of the environment index and the response of the physiological index is established.

[0096] First, in the preparation of input data during the training of the above-constructed deep learning, the following data need to be prepared.

[0097] One is an air environment feature sequence, that is, a time sequence feature vector formed after preprocessing and feature extraction of carbon dioxide, PM2.5, TVOC, and temperature and humidity multi-source indexes. Each time step corresponds to a multi-dimensional feature, and the whole is a time sequence.

[0098] One is a physiological feature sequence, that is, a time sequence feature vector formed after feature extraction of physiological signal indexes such as blood pressure, blood oxygen, body temperature, heart rate, heart rate variability, and blood glucose.

[0099] It should be noted here that the air environment features and the physiological features are corresponded at the same time step through a unified time reference, so as to ensure that the two types of sequences are strictly aligned.

[0100] The last one is an individual static covariate, that is, the above-mentioned gender, age, BMI, disease history, and the like. Such data does not change with time, but will affect the relationship between air exposure and physiological response, so it needs to be introduced into the model as background information at the individual level.

[0101] In this way, it can be understood that a sample is composed of three parts:

[0102] An air feature sequence (changes with time);

[0103] A physiological feature sequence (changes with time);

[0104] A static feature vector (at the individual level).

[0105] Regarding the deep learning model, in the embodiments of the present application, a two-channel architecture is adopted, and reference is made to Figure 3 , and it is further understood that the first sub-channel is an air channel encoder, which can also be represented as an environment feature channel; the second sub-channel is a physiological channel encoder, which can also be represented as a physiological feature channel.

[0106] Specifically, the first sub-channel, representing the environmental feature channel, takes an air feature sequence as input, encodes it using a temporal neural network, and outputs a hidden state sequence for the air channel, representing the dynamic characteristics of air over time. The second sub-channel, representing the physiological channel encoder, takes a physiological feature sequence as input, encodes it using the same structure, and outputs a hidden state sequence for the physiological channel, representing the dynamic characteristics of physiological indicators over time.

[0107] It's crucial to understand why feature vectors for indicators are constructed based on time-series features, rather than directly using detected data. While directly using detected data is the most straightforward approach—simply concatenating all values ​​collected by the sensor within a specific time window—it has drawbacks. This data is rigid, sensitive to noise, and lacks an understanding of change patterns. Deep learning models would then need to learn trends and fluctuations during training, often increasing training difficulty and reducing predictive stability.

[0108] Constructing feature vectors based on time-series features requires understanding that these features are not simply a single data point. Rather, as mentioned above, they are time-series features extracted by analyzing the sequence of indicators within a time window, describing their evolutionary patterns. Combining these features forms the feature vector corresponding to that window, and arranging them in window order yields a time-series sequence of feature vectors.

[0109] Therefore, based on the above, it can be understood that feature extraction makes the input data more stable. Compared to directly inputting the original time-point sequence, the extracted features are often more condensed, reducing redundancy while retaining key information, making it easier for the learning model to converge. Essentially, deep learning models extract and learn the trend of indicators within a time window, rather than single-point values. This allows them to learn evolution patterns across time windows more naturally, making them more suitable for capturing "chronic cumulative effects" or "sudden fluctuations."

[0110] The environmental and physiological characteristic channels are processed by LSTM and Transformer encoders, respectively. The advantage of LSTM lies in its ability to memorize short- to medium-term temporal dependencies and capture the lag in physiological responses to air pollution. Transformer, on the other hand, relies on attention mechanisms to identify "which periods of air exposure have the greatest impact on the results" over a long time span.

[0111] Specifically, the environmental feature channel (LSTM encoder) receives a sequence of environmental feature vectors, and the input of each step is the environmental feature vector corresponding to a certain time window. During the training and learning of the deep learning model, the memory unit (cell state) of the LSTM can retain the cumulative information of the previous window, and at the same time, the fusion of the "current input" and the "historical memory data" is adjusted through the input gate, the forgetting gate and the output gate. In this way, the corresponding hidden state sequence of the LSTM can reflect the cumulative exposure effect and local fluctuation characteristics of the environmental indicators in the time dimension, such as the cumulative effect of high PM2.5 in a continuous period of time or a sudden short-term peak.

[0112] The physiological feature channel (Transformer encoder) receives a sequence of physiological feature vectors. In terms of programming language, the features of the physiological indicators of each time window are regarded as a "token", and through the self-attention mechanism, the model calculates the correlation weight between different time points in the sequence, so as to find out which time point of the physiological features is more critical to the current prediction. After the processing of multiple attention and feedforward networks, the output hidden state sequence condenses the information of "global time sequence dependence" and "key period".

[0113] At each time step, the environmental hidden state output by the LSTM and the physiological hidden state output by the Transformer are paired, and the attention mechanism of the fusion layer calculates the "contribution weight of the environmental hidden state to the physiological hidden state". For example, at a certain time step, the "cumulative high exposure of PM2.5" in the environmental hidden state may have a higher weight on the "heart rate fluctuation"; while at another time step, the "sudden rise in temperature" in the environmental hidden state may contribute more to the "blood oxygen fluctuation".

[0114] After weighting, the weighted fusion vector is generated, which not only contains the time sequence information of the environmental exposure, but also embeds the action strength of the environmental exposure on the physiological response. These weighted fusion vectors are input into the prediction layer in chronological order as the basis of prediction.

[0115] Regarding the prediction layer in the architecture of the deep learning model, the architecture design needs to consider the span of the prediction time, and different settings correspond to different output methods. For example, in some embodiments of the present application, single-step prediction is performed, that is, the time interval is set to be equal to the step length of one time step (for example, 30 minutes or 60 minutes), and the prediction layer directly outputs the predicted values of each physiological indicator in the next time step.

[0116] In some embodiments of the present application, the model is configured to perform a single-step prediction, i.e., the time interval covers only one time step. In this case, the prediction layer outputs a single value of the physiological indicator at the future time step. In other embodiments of the present application, the model is configured to perform multi-step prediction, i.e., the time interval covers multiple time steps. In this case, the prediction layer outputs a sequence of values of the physiological indicator at the future time steps. The decoding process is recursive, i.e., the output of the previous step of prediction is used as one of the inputs of the next step of prediction, and the process is repeated until the prediction of the physiological indicator at the future time steps is completed.

[0117] In summary, the input of the model is a feature vector composed of time-series features, and the output is a predicted value of the physiological indicator. It is understood that the prediction layer in the model architecture is a regression module, typically a fully connected layer or a sequence decoding layer. The input of the prediction layer is the weighted fusion vector, i.e., the fusion vector at the current time step or during recursive prediction, and the output of the prediction layer is the numerical value of the physiological indicator. This is because the fusion vector contains the information captured by the time-series features, and the prediction layer acts as a mathematical mapping function that maps the abstract vector to the numerical space of the original indicator. Thus, the output of the model is the predicted value that has been mapped to the specific numerical space. Moreover, the loss function of the model mentioned later is also based on the error calculation of the true physiological value, and the optimization is the mathematical mapping function mentioned above.

[0118] In addition, while outputting the predicted value, the prediction layer also introduces reference information of the cumulative environmental exposure, i.e., the time range of the cumulative exposure is aligned with the prediction time interval, for example, predicting the future 30 minutes, and calculating the corresponding cumulative exposure in the past.

[0119] For health risk assessment under the cumulative environmental exposure, the predicted values of each physiological signal in the predicted physiological indicator are mapped to the risk level interval with the cumulative exposure. For example, if the future heart rate prediction shows a continuous rising trend, and the PM2.5 or CO2 cumulative exposure level exceeds the threshold, the prediction layer outputs “medium risk” or “high risk”. Conversely, if the predicted physiological indicator remains stable and the cumulative exposure is below the safety threshold, it outputs “low risk”.

[0120] Regarding the access of static covariates, this is flexible in the embodiments of the present application, and individual characteristics can be added at different levels. For example:

[0121] In the input layer, the static characteristics are mapped and spliced to the input of each time step;

[0122] In the hidden state layer, the hidden state is conditionally modulated (e.g., gating or scaling offset);

[0123] In the prediction output layer, the static characteristics are spliced with the fusion representation into the prediction layer.

[0124] After encoding, the outputs of the two channels are fused with static covariates to obtain a comprehensive time series representation. This representation is equivalent to a joint characterization of the relationship between “air exposure-physiological response-individual difference”.

[0125] In summary, it can be understood that by setting the input in this way, the model can learn the sensitivity differences of different individuals to the same air exposure, and achieve individualized prediction. Then at each time step, the hidden states of the air channel and the physiological channel are spliced or interacted through an attention mechanism to obtain a fusion vector. In fact, the coupling relationship between air features and physiological features is established in the time dimension.

[0126] In the prediction phase, the model outputs two types of results:

[0127] The first is a set of predicted values of physiological indicators, as well as the trend of changes in each physiological signal, such as the magnitude of the increase in heart rate in the future (the next time step or the next time interval).

[0128] The second is the health risk assessment value under the cumulative exposure level, which can be a probability score value or a risk level.

[0129] In the training process of the deep learning model, the samples are constructed in a sliding time window manner, and the continuous air and physiological time series are divided into training samples. Each sample input is an air sequence, a physiological sequence, and static features, and each sample output is the future (future continuous time steps) physiological indicator change and health risk assessment value. These outputs are compared with the real physiological observation data (labels), and the prediction error is measured using methods including but not limited to mean square error, cross-entropy, or multi-task weighted loss.

[0130] Backpropagation is performed according to the conventional process of deep learning, calculating the gradient layer by layer and updating the model parameters. In some embodiments, to avoid gradient explosion or overfitting, methods including but not limited to gradient clipping, dropout, weight regularization, etc. are used during training.

[0131] At the same time, in some embodiments, a validation set is also introduced to monitor the training effect, and once the validation error no longer improves within a certain number of rounds, the training is terminated through the #early stopping# instruction to save the best model.

[0132] It is particularly emphasized here that individual feature information (gender, age, BMI, disease history) plays a role as a conditional variable in training. That is, even if the air pollution level is the same, the prediction results of different individuals may be different.

[0133] In summary, it can be understood that in the training process of the deep learning model, each sample contains two parts of input data.

[0134] One is the environmental feature vector sequence composed of the time series characteristics (change amplitude, local volatility, cumulative exposure) of carbon dioxide, PM2.5, TVOC, temperature and humidity in historical data;

[0135] The other is individual feature information, i.e., age, gender, BMI, disease history, etc., as static features, which are input into the deep learning model as individual condition parameters together with the environmental feature vector sequence.

[0136] In addition, the training sample also includes the real physiological label corresponding to the historical time node mentioned above. The model learns the coupling relationship between air environment features and physiological index sequence, considers the adjustment effect of individual characteristics on physiological response, captures the nonlinear influence of environmental exposure on different individuals, and in the training process, individual characteristics enable the model to distinguish different physiological responses of different people under the same air environment conditions, so that the prediction result has individualized characteristics.

[0137] Therefore, it can be further understood that the deep learning model learns the corresponding relationship between the "individual difference-air exposure-physiological outcome" in a large number of samples to realize personalized prediction of different populations. This makes the trained model not only able to give overall trend judgment, but also to generate customized health risk assessment according to the input individual characteristics.

[0138] After training is completed, the parameters of the time series deep learning model have been fixed, at which time the model has the mapping ability from air environment input to physiological response prediction. Next, S50 is performed: the constructed and trained deep learning model is used to generate a prediction of the physiological index time series and its dynamic change based on the air environment feature sequence and individual feature information as input, to obtain a health impact prediction model based on multi-source air data coupling.

[0139] In this step, the first input is new air environment data, which is also a time series collected from multiple sources, such as continuous monitoring values of CO2, PM2.5, TVOC and temperature and humidity in the target space. Before input, these data also need to be preprocessed in the same way as in the training phase, including missing value filling, outlier removal, normalization, etc., to ensure that the distribution of the input is consistent with that in the training.

[0140] In the time dimension, the new air environment data is divided into windows, and the data in each window is used as the input of the prediction. It should be noted that, unlike the training stage, the new physiological signal is not input at this time, but only the air environment features are relied on, combined with individual characteristics information (gender, age, BMI, disease history), to infer the physiological indicator trend in the future time period through the coupling relationship learned by the model.

[0141] After processing by the dual-channel structure of the deep learning model, the air channel converts the new environmental indicator sequence into a time series feature representation, and the physiological channel can receive the existing physiological state as a conditional input in the inference stage, supplemented by static covariates to supplement individual difference information. The two channels of data are coupled in the fusion layer to capture the dynamic effects of cumulative exposure to air pollution on individual physiological states.

[0142] The final prediction output includes two types of results: one is the predicted value of the physiological indicator and its trend in the predicted time range, and the other is the health risk assessment, which provides a quantitative result of the health impact based on the cumulative exposure level and individual differences.

[0143] In summary, it can be understood that the learned "air environment-physiological response-individual difference" coupling rule in the training stage is used to map the new air environment time series data to the prediction of future physiological changes. This prediction is not only a single value at a time point, but also a trend result, reflecting the dynamic impact of air exposure on the human physiological process.

[0144] Finally, in some embodiments of the present application, considering that the health prediction result has been obtained, the physiological indicator trend and health risk assessment value output by the model can be integrated with real-time air environment data to achieve multi-dimensional closed-loop control and feedback. For details, please refer to Figure 5 .

[0145] First, through the visual dashboard, the various air environment data (such as CO2, PM2.5, TVOC, and temperature and humidity) in the environmental indicators and the corresponding predicted physiological signal data (blood pressure, blood oxygen, body temperature, heart rate, heart rate variability, and blood glucose) are dynamically displayed on the time axis, allowing users or management systems to intuitively view the current air environment state and its potential impact on health.

[0146] At the same time, the management system can compare the predicted health risk value with the preset threshold in real time, that is, when the risk value output by the model exceeds the threshold, the health warning mechanism is triggered, and the user is sent a reminder, such as a wristband vibration or a mobile phone notification, prompting the user to pay attention to air exposure and take protective measures in a timely manner.

[0147] On the other hand, the prediction result can also be used as an input of the air conditioning system. By connecting the model-predicted cumulative exposure of the environment with the control strategy of the indoor air conditioning equipment, such as the fresh air system, air purifier, air conditioner and the like, the system can map the difference between the cumulative exposure of the environment and the safety threshold to the control amount required by the air conditioning equipment to run, and then automatically adjust the equipment operation, thereby reducing the potential impact of air pollution on the human body and achieving active regulation.

[0148] In this way, it can be understood that the prediction result of the constructed deep learning model not only provides health risk assessment and early warning, but also directly feeds back to the environmental control link to form an air environment monitoring, health prediction, risk early warning and automatic regulation closed-loop system. This closed loop ensures the real-time coupling between air environment management and individual health protection, and dynamically optimizes the interaction between indoor environment and physiological response.

[0149] The above is only a preferred embodiment of the present application and is not intended to limit the present application. The patent protection scope of the present application is subject to the claims, and any equivalent structural changes made by applying the contents of the specification and drawings shall also be included in the protection scope of the present application.

Claims

1. A method for constructing a health impact prediction model based on multi-source air data coupling, characterized in that, The method comprises the following steps: obtaining multi-source air environment data in a selected space as an environmental indicator, and simultaneously obtaining multi-source physiological signals of a person in the selected space as a physiological indicator; wherein the environmental indicator comprises carbon dioxide, PM2.5, TVOC, temperature and humidity, and the physiological indicator comprises blood pressure, blood oxygen, body temperature, heart rate, heart rate variability and blood glucose; normalizing and preprocessing the collected environmental indicator and physiological indicator, and aligning them in the time dimension to form time-synchronized sequence data; based on the preprocessed sequence data, extracting time sequence features of the environmental indicator and the physiological indicator in a set time window, wherein the time sequence features of the environmental indicator comprise change amplitude, local time fluctuation and cumulative exposure, and the time sequence features of the physiological indicator comprise trend change, local time fluctuation, time domain feature and frequency domain feature; each type of time sequence feature is converted into a corresponding feature vector sequence; a dual-channel deep learning model is constructed, and the environmental feature vector sequence and the physiological feature vector sequence are input into the corresponding channels of the deep learning model in time steps, for capturing the cumulative effect and local fluctuation features of the environmental indicator in the time dimension, and simultaneously capturing the change pattern of the physiological indicator over time; each channel is provided with an independent encoder for dynamically extracting the input feature vector sequence, and the encoder outputs a hidden state sequence of the corresponding channel, each hidden state sequence retaining time sequence information; the hidden state sequences output by the two channels at each time step are input into the middle fusion layer of the deep learning model, the two hidden state sequences are paired at the corresponding time step, the contribution weight of the environmental feature vector to the physiological feature vector at the time step is calculated by using an attention mechanism, and a weighted fusion vector is generated; the deep learning model is trained to fuse and process the captured results output by each channel, so as to establish a coupling relationship between the cumulative exposure of the environmental indicator and the response of the physiological indicator; the weighted fusion vector is input into a prediction layer of the deep learning model, and the prediction layer outputs predicted values of various physiological signals of the physiological indicator by regression, thereby obtaining predicted values of the physiological indicator in a future set time interval; the constructed and trained deep learning model takes air environment feature sequences and individual feature information as input, generates a prediction of a time sequence of a physiological indicator and its dynamic change, and obtains a health impact prediction model based on multi-source air data coupling. 2.The method of claim 1, wherein, In the selected space, multi-source environmental monitoring devices are used to continuously monitor and sample carbon dioxide, PM2.5, TVOC, temperature and humidity, and various air environment data is continuously obtained at a set time interval, thereby forming a time sequence of the environmental indicator; Meanwhile, wearable human monitoring equipment is used to obtain blood pressure, blood oxygen, body temperature, heart rate, heart rate variability and blood glucose of a person in the selected space, and various physiological signals are continuously obtained at the same time interval, thereby forming a time sequence of the physiological indicator corresponding to the environmental indicator. 3.The method of claim 1, wherein, The environment indicators and physiological indicators continuously collected are uniformly preprocessed, and the preprocessing operations include detecting and filling in missing values, identifying and removing outliers, smoothing noise signals, and normalizing the environment indicators and physiological indicators; After data preprocessing, the environment indicators and physiological indicators are aligned according to timestamps through timestamp matching to form time-synchronized sequence data. 4.The method of claim 1, wherein, After data preprocessing and time alignment, time series features of the environment indicators are extracted in each time window based on a set time window. For each air environment data in the environment indicators, the increment between adjacent sampling points is calculated to quantify the change amplitude in the time dimension and reflect the instantaneous fluctuation trend. In each time window, the dispersion degree of each air environment data value in the corresponding time period is calculated as the local fluctuation feature of the air environment data. In each time window, the cumulative exposure of each air environment data in the time period is calculated by accumulating the numerical values of each air environment data in the time period.

5. The method of claim 4, wherein, After data preprocessing and time alignment, time series features of the physiological indicators are extracted in each time window based on a set time window. Correspondingly, For each physiological signal in the physiological indicators, the overall change trend in the time window is calculated, including the direction and change amplitude. In each time window, the dispersion degree of each physiological signal in the corresponding time period is calculated as the local fluctuation feature of the physiological signal. In each time window, the mean and extreme values of each physiological signal in the corresponding time period are calculated as the time domain features. In each time window, the oscillation intensity of each physiological signal in the frequency spectrum is extracted through frequency spectrum analysis to represent the physiological oscillation pattern.

6. The method of claim 5, wherein the method is characterized by: The time series features of the environment indicators extracted in each time window are combined to form an environment feature vector, and the time series features of the physiological indicators extracted in each time window are combined to form a physiological feature vector. The feature vectors of each time window are arranged in chronological order to form an environment feature vector sequence and a physiological feature vector sequence.

7. The method of claim 1, wherein, The deep learning model is a dual-channel architecture, and the time-sequenced environment feature vector sequence and physiological feature vector sequence are input into the corresponding channels of the deep learning model, respectively. The deep learning model includes an environment feature channel and a physiological feature channel. The environment feature channel is used to capture the cumulative exposure effect and local fluctuation features of the environment indicators in the time dimension. The physiological feature channel is used to capture the change pattern and local change features of the physiological indicators over time. 8.The method of claim 7, wherein, The prediction layer of the deep learning model is used to obtain the predicted values of the physiological indicators in a future set time interval. If the set time interval is a single time step, the output is the predicted values of various physiological signals in the next time step. If the set time interval is a continuous multiple time steps, the prediction results are recursively output on the future multiple continuous time steps through sequence decoding to form a set of predicted values and change trends of various physiological signals. The prediction layer combines to set the environmental cumulative exposure in the prediction time interval, and predicts the risk level of various physiological signals. 9.The method of claim 7, wherein, In the training stage of the deep learning model, a fusion sequence of a historical environmental feature vector sequence and individual feature information is taken as input, and a real physiological index of a corresponding time step in history is taken as a label, wherein the individual feature information includes age, gender, BMI, and disease history of the individual; various physiological signals of the predicted physiological index are defined, and an error between the various physiological signals of the predicted physiological index and the various physiological signals of the real physiological index is taken as a loss function; a gradient is calculated in a time dimension by using a back propagation algorithm, and parameters of an encoder corresponding to the environmental feature channel and the physiological feature channel, an attention mechanism of an intermediate fusion layer, and a prediction layer are updated.

10. The method of claim 1, wherein, According to the prediction result of the various physiological signals in the physiological index predicted and output by the deep learning model, a corresponding risk threshold of the various physiological signals in the physiological index is set, and when the prediction result of any physiological signal exceeds the corresponding preset risk threshold, health prompt information is generated and visually displayed; The selected space is provided with an air conditioning device, the prediction result of the physiological index is combined with the prediction result of the environmental index at the corresponding time point, a control reference signal is generated, and after the control reference signal is mapped into a corresponding control amount of the air conditioning device, the air conditioning device is adjusted and driven.

Images