Prediction of the Hypertension Risk of the Elderly in Built Environments Based on the LSTM Deep Learning and Bayesian Fitting Method

Abstract

Hypertension has become the greatest risk factor for death in elderly populations. As factors influencing cardiovascular disease, indoor environmental parameters pose potential risks for older adults. In this study, elderly residents in Dalian (Liaoning Province, China) urban dwellings were selected as the research subjects, and the environmental parameters of the dwellings’ main activity rooms and the blood pressure parameters of the older adults were measured. Based on the Long Short-Term Memory (LSTM) deep learning algorithm and Bayesian fitting method, a hypertension disease model was established using the long-term environmental parameters to predict the hypertension risk of older adults in their building’s environment. The results showed that temperature, humidity, and some air quality parameters had an impact on blood pressure under single environmental factor, and the comprehensive environmental risks of high systolic blood pressure, high diastolic blood pressure, and high blood pressure were 16.44%, 0%, and 16.44% for the male elderly and 14.11%, 7.14%, and 17.55% for the female elderly, respectively. By comparing the results for the blood pressure measurement and prediction, it can be observed that the risk error of hypertension obtained by the algorithm maintains the variables’ relationship, and the result of the algorithm is reliable in this period. This technology can provide a basis for measuring environmental parameters and will be conducive to the development of an ecological smart building environment.

1. Introduction

With the development of urbanization, developed countries experienced a population demographic transformation, and the percentage of older adults increased. It is estimated that by 2035, the number of elderly people and the level of aging in the world will be 392 million and 30.5%, respectively, thus making the world a severely aging society [1]. According to data released by the National Bureau of Statistics of China, by the end of 2019, the population over 60 years old accounted for 18.1% of the total, and the population over 65 years old accounted for 12.6% [2]. The American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE) standard mentions that the indoor environment will affect human health, and the environment is related to physical health conditions [3]. It is mentioned in some papers that, due to harsh indoor environments or environmental pollution, people may become susceptible to related diseases and building syndromes [4,5]. Compared to other people, the elderly have longer indoor activity times, so studying the correlation between the health of the elderly and the indoor environment is of great significance [6]. By 2015, it was the case that nearly half of the world’s elderly people eventually died due to cardiovascular diseases [7]. Compared with other seasons, the incidence of cardiovascular diseases reaches a peak in winter [8]. Studies have shown that temperature difference will affect the morbidity and mortality of cardiovascular diseases [9,10], and cardiovascular diseases are related to air pollution [11,12]. As one of the key determinants of cardiovascular disease, blood pressure has a high predictive value [13] for the elderly; it is more necessary to control systolic blood pressure than diastolic blood pressure [14]. Controlling systolic blood pressure can prevent the risk of cerebrovascular disease and stroke; thus, it is the main target in the effort to improve the prognosis of elderly patients [15]. It is unreasonable to control indoor parameters only to align with building environmental standards. On the one hand, standards are formulated based on the needs of young people; on the other hand, differences in personal physique cannot be ignored [16]. In order to deal with the most relevant problems in the area of environment and health, the three core concepts of the “Healthy China” strategy—”Great Health”, “All-round”, and “Great Environment”—put forward new development goals for intelligent and healthy old-age care [17].

Some scholars proposed Ambient Assisted Living (AAL) technology very early, and some studies have shown that the application of a deep learning model to an AAL service can greatly improve the accuracy and reliability of life records [18,19]. The deep learning algorithm extracts data from Internet of Things (IoT) device sensors in the deployment environment. The algorithm uses advanced methods to greatly improve the performance of existing machine learning technologies and develops rapidly in the fields of biological data and health information [20,21]. Deep learning algorithms can extract feature information from human physiological parameters for pattern recognition and health assessment [22,23]. Some scholars have systematically studied the application of deep learning algorithms in the automatic diagnosis of a series of diabetic diseases. The model used neural networks based on data expansion and data correction technology, covering a variety of physiological parameters of the human body and proving its ability to predict diseases [24,25,26]. The following methods are used to measure and estimate blood pressure: Combining the classical pulse width evaluation model and neural network model to estimate blood pressure [27], blood pressure can be reconstructed using noninvasive ambulatory blood pressure estimation based on photoplethysmography (PPG) signals, and blood pressure training and prediction can be carried out by LSTM [28]. A hybrid deep learning neural network framework based on convolutional neural network (CNN) and LSTM can be used to predict the time series data in electronic health records [29]. In the field of clinical cardiology, one study has shown that deep learning algorithms clearly outperform clinicians in predicting prognosis and future events in patients with pulmonary hypertension [30]. With the development of IoT technology, IoT will gradually become the basis of home appliances [31]. To sum up, many artificial intelligence diagnostic algorithms have been developed to detect diseases, but their prediction accuracy for hypertension is lower. In addition, the prediction of blood pressure parameters ignores the interference of external factors, and indoor environment IoT products do not pay attention to dynamic changes in human health.

At present, the research on indoor air quality is mainly carried out concerning two aspects: evaluation and management [32]. Elderly people are very interested in using visualization technology to monitor their long-term health trends [33]. In this study, we carried out continuous environmental measurements and blood pressure parameter tests in the morning and evening for urban residential buildings in Dalian and prepared to build an Internet of Things data platform and mobile app in the future. The research objectives of this paper are: (1) Monitor indoor temperature, humidity, formaldehyde, carbon dioxide, Total Volatile Organic Compounds (TVOC), and PM_2.5 using the standard for indoor environment evaluation; (2) Analyze the impact of single environmental factor on the blood pressure risk of the elderly; (3) Establish a multi-environmental parameter to characterize the risk of hypertension in the elderly, using the LSTM deep learning algorithm and Bayesian fitting, and evaluate the accuracy of the model by comparing it with a previous probability distribution of blood pressure.

2. Materials and Methods

2.1. Measurement Objects

In the early stage, we conducted a survey using a questionnaire. The contents of the questionnaire included questions about the residential environment, thermal insulation form of envelope, type of floor decorations present, personal thermal comfort of residents, living and eating habits, etc. We sent out 50 questionnaires in the urban area of Dalian (Liaoning Province, China) in order to understand the habits and housing conditions of elderly residents. As IoT devices require good network conditions, this study conducted an actual measurement on a household in the urban area of Dalian. The dates were from 18 December 2018 to 30 January 2019 (the heating season). There were two elderly people in the household: a 69-year-old male with no cardiovascular disease and a 66-year-old female with arrhythmia and high cholesterol but no high blood pressure. According to the preliminary questionnaire survey, the living room, study room, and master bedroom were the main activity spaces. The residential area for the elderly in China is mostly between 60 and 120 m², with one bedroom or more [34], so the area and room type of this house has met the requirements.

2.2. Measurement Method

According to the requirements of the Code for Indoor Environmental Pollution Control of Civil Building Engineering (GB 50325-2010) [35], the area of each room was less than 50 m², and the number of detection points in each room was 1. The distance between the field test point and the inner wall was more than 0.5 m, the height from the floor was 0.8 m, and the vents were avoided. In this study, the environmental parameter acquisition instrument was the Hike-b6a IOT (produced by Beijing Hike Zhidong Technology Development Co., Ltd. Beijing, China), which was placed in three rooms for 24 h of operation. In order to measure the microenvironment around the human body, the instrument was placed in positions where the elderly residents would typically stay for a long time, such as the bedside table in the bedroom, the tea table in the living room, and the study desk in the study room, while avoiding direct sunlight. At the same time, we properly handled the power cord and plug-in of the instrument to ensure that the elderly residents could walk around easily and that their indoor activities would be safe. The acquisition parameters included indoor temperature, humidity, CO₂, TVOC, formaldehyde, PM_2.5 (PM_2.5 accuracy is ±15%, formaldehyde accuracy is ±0.03 mg/m³, CO₂ accuracy is ±40 ppm, TVOC accuracy is ±0.03 mg/ m³, temperature accuracy is 2 °C, humidity accuracy is 5% RH). Data can be uploaded to the cloud platform through WiFi. In order to record the indoor activities of the elderly subjected, we asked them to fill in a record form whenever they left and entered the house. Although it was not convenient for the two elderly people to go downstairs due to their physical mobility issues, leading to them engaging in few outdoor activities, we still excluded the parameters of indoor environment when they were outdoors. The physiological parameters were measured by the Life Sense i5 IOT blood pressure instrument (produced by Guangdong Lexin Medical Electronics Co., Ltd. Zhongshan, China), which had an accuracy of ±3 mmHg. The instrument could measure blood pressure and heart rate. According to the Chinese Elderly Hypertension Management Guidelines 2019 [36], blood pressure monitoring is mainly carried out before going to bed and in the early morning after waking up. In this study, we measured it twice a day and took the average twice every time to eliminate error. Daily morning measurements were taken within an hour of waking up, after urination and before breakfast. Bedtime measurements were taken more than 30 min after the bath. When the men measured, they pressed the blue button; when women measured, they pressed the red button. Two elderly people helped each other to put on the equipment and perform the tests. According to the cloud data and the previous questionnaire survey, we were able to obtain the daily rest time of the two elderly people. We called half an hour in advance to remind the two elderly people to take the blood pressure measurement. The device could be connected to the Internet, and the data could be uploaded to the Lexin data platform. The layout of the room and the schematic diagram of the room where the equipment was placed are shown in Figure 1.

After obtaining the parameters, a single-factor environmental factor analysis was conducted to explore the relationship between potential independent variables and dependent variables. The core idea of the model is the combination of Bayesian fitting and LSTM deep learning to estimate the blood pressure of elderly people under the influence of their environment. A variable relationship between indoor environment and blood pressure was established, and a model for estimating the influence of the indoor environment on blood pressure on a long time scale was developed. Bayesian fitting was used to explore the impact of the environment on blood pressure risk. Data collection and future application ideas are shown in Figure 2.

2.3. Deep Learning Algorithm and Parameter Setting

LSTM is a type of RNN (recurrent neural network) with a time memory. The core idea is to save important information and delete irrelevant information The basic structure of the LSTM is shown in Figure 3. We analyzed the internal structure of the sequence index position of the LSTM model at time t: the multi-dimensional independent variable $X_t$ enters the unit cell through the input gate at time t, denoted as $C_t$ (cell state). The forgetting gate controls whether the upper layer of the fine hidden cell structure is forgotten and with what probability, as well as whether the memory data are retained. Finally, the output gate signal controls whether the output presents $f^{\left(t\right)},W_f$; $U_f$ and $b_f$ are the coefficient and bias of linear relationship; $f^{\left(t\right)}$ represents the probability of forgetting the previous hidden cell state, as shown in Formula (1):

$$f^{\left(t\right)}=\sigma \left(W_f{h}^{\left(t-1\right)}+U_f{x}^{\left(t\right)}+b_f\right)$$

(1)

The first part of the input gate uses sigmoid to activate the function, and the output is the combination of $i^{\left(t\right)}$ and the tanh-activated function $a^{\left(t\right)}$. The cell state $C^{\left(t\right)}$ consists of two parts: the first part is the product of $C^{\left(t-1\right)}$ and the output $f^{\left(t\right)}$ of the forgetting gate, and the second part is the product of $i^{\left(t\right)}$ and $a^{\left(t\right)}$ of the input gate. The mathematical expressions are shown in (2) to (4).

$$i^{\left(t\right)}=\sigma \left(W_i{h}^{\left(t-1\right)}+U_i{x}^{\left(t\right)}+b_i\right)$$

(2)

$$a^{\left(t\right)}=tanh\left(W_a{h}^{\left(t-1\right)}+U_a{x}^{\left(t\right)}+b_a\right)$$

(3)

$$C^{\left(t\right)}=C^{\left(t-1\right)}\times f^{\left(t\right)}+i^{\left(t\right)}\times a^{\left(t\right) }$$

(4)

The output gate consists of two parts: the hidden state $h^{\left(t-1\right)}$ of the previous sequence and $x^{\left(t\right)}$ of the previous sequence generate $o^{\left(t\right)}$ under the action of the σ function. The other part is composed of $C^{\left(t\right)}$ and the activation function tanh, as shown in mathematical Formulas (5) and (6):

$$o^{\left(t\right)}=\sigma \left(W_o{h}^{\left(t-1\right)}+U_o{x}^{\left(t\right)}+b_o\right)$$

(5)

$$h^{\left(t\right)}=o^{\left(t\right)}\times tanh\left(C^{\left(t\right)}\right)$$

(6)

The weight matrix W and offset vector b variables are the hyperparameters of LSTM; the samples are brought into the model and the gradient descent method is used to update the W and b values to minimize the loss function. At this time, the result value can be predicted based on the multidimensional sequence variables.

Due to the time series characteristics of the indoor environmental parameters, time is a key variable that cannot be ignored when studying the long-term correlation between the indoor environment and blood pressure. Therefore, the LSTM model was used to establish the correlation estimation model, and the environmental and blood pressure parameters were used as input variables in the LSTM network structure. At the same time, considering the influence of multiple behavioral factors such as diet and medication behavior on blood pressure [37], the morning measurement was carried out within one hour after waking up, and the bedtime measurement was carried out more than 30 min after bathing and before going to bed. According to the daily routine of the two elderly people, the whole day was divided into two periods: 9 a.m. to 8 p.m., and 9 p.m. to 8 a.m. the next day. Taking the 12 h environmental data as the input sequence, the model corresponding to blood pressure was established. The indoor environmental parameters of the different units were dimensionless, and some or all of the units of the equations involving physical quantities were removed to simplify the calculation. Please refer to “Building Ventilation Effect Evaluation and Evaluation Standard” (JGJ/T 309-2013) [38] and “Indoor Air Quality Standard” (GB/T18883-2002) [39]. The standard value and characteristic value are shown in

Table 1. Characteristic values of the indoor environmental parameters.

Standard	Parameter	Unit	Standard Value	Characteristic Value
Indoor Air Quality Standard	T	°C	16~24	20
	RH	%	30~60	45
	CO2	ppm	1000	1000
	HCOH	μg·m−3	100	100
	TVOC	μg·m−3	600	600
Test and evaluation standard of building ventilation effect	PM2.5	μg·m−3	75	75

.

According to Formula (7), indoor environmental parameters are dimensionless. $X_i$ is the average value of six indoor environmental parameters in three rooms, with a total of 18 independent variable dimensions:

$$X_i^{\prime }=X_i/X_0,i=1,2,3,\dots ,18$$

(7)

The adjustment of the hyperparameters is very important for training neural networks. The setting of the learning rate will affect the fluctuation of the curve and the duration of the training process. We repeatedly adjusted the learning rate during the training process. In most studies, there are no consistent choices of optimizer. The Adam optimizer is widely used, so we used this optimizer. We judged the training result of this model by looking at the training curve, and the loss value was steadily decreasing. Considering that the fitting effect of the training set would affect the generalization ability of the test set, once the test error stopped falling or the error started to increase, we chose to stop training. The hyperparameter settings are shown in

Table 2. Hyperparameters of the LSTM model.

Hyperparameter	Value	Remark
Learning rate	0.01	Iteration steps
Batch size	1	Gradient descent method update weight w and deviation b
Training iterations	85	A total of 85 sets of data
Batch iterations	500	The total number of iterations is 85 × 500
LSTM unit numbers	12	12 h environmental data
LSTM unit dimensions	18	Independent variable dimension

.

2.4. Blood Pressure Probability Model Combined with Bayesian Algorithm

According to the error data of the regression prediction model, the Bayesian fitting of blood pressure prediction can be carried out [40]. Parameter estimation plays an important role in the intelligent analysis of big data, and it is also an important basic guarantee for a machine learning algorithm [41]. In addition, parameter estimation also plays an important role in other fields, such as cell metabolism parameters and biochemical reaction kinetics parameters estimation in the field of biology [42,43,44], the field of energy [45,46], the field of signal processing [47,48,49], the field of computational mathematics [50], etc. The reliability and accuracy of the parameter estimation were shown to be good. Based on the calculation of the linear deviation of normal distribution parameters and the determination of the maximum similarity value of the parameters, the Bayesian function was used to realize the parameter estimation. The characteristics of the error curve and the error between the predicted value and true value were taken as the standard deviation of the normal distribution of the blood pressure probability density curve.

$$\mathrm{F}\left({\mathrm{Y}}_1{,\mathrm{Y}}_2\right)={\left({2\pi \sigma }_1{\sigma }_2\sqrt{{1-\rho }_j{}^2}\right)}^{-1}\exp \left[-\frac{1}{2\left({1-\rho }_j{}^2\right)}\left(\frac{{\left(Y_1{-y}_1\right)}^2}{{\sigma }_1{}^2}-\frac{{2\rho }_{\mathrm{j}}\left(Y_1{-y}_1\right)\left({\mathrm{Y}}_2{-y}_2\right)}{{\sigma }_1{\sigma }_2}+\frac{{\left({\mathrm{Y}}_2{-y}_2\right)}^2}{{\sigma }_2{}^2}\right)\right]$$

(8)

For ${\sigma }_j$, j = 1 for male, and j = 2 for female; for $y_i$, i = 1 represents systolic blood pressure, and i = 2 represents diastolic blood pressure.

According to the classification and definition of the blood pressure level, a systolic blood pressure over 140 mmHg or a diastolic blood pressure over 90 mmHg can be defined as hypertension [51]. Therefore, the probability of elderly people suffering from hypertension is shown in Formula (9).

$$\mathrm{P}\left({\mathrm{Y}}_1\ge 140\cup {\mathrm{Y}}_2\ge 90\right)={{\displaystyle \int }}_{140}^{+\infty }{\mathrm{f}}_1\left({\mathrm{Y}}_1\right){\mathrm{dY}}_1+{{\displaystyle \int }}_{90}^{+\infty }{\mathrm{f}}_2\left({\mathrm{Y}}_2\right){\mathrm{dY}}_2-{{\displaystyle \int }}_{140}^{+\infty }{\mathrm{dY}}_1{{\displaystyle \int }}_{90}^{+\infty }\mathrm{f}\left({\mathrm{Y}}_1{,\mathrm{Y}}_2\right){\mathrm{dY}}_2$$

(9)

3. Results and Discussion

3.1. Indoor Environmental Quality Evaluation

During the actual measurement period, the lowest temperature in the living room was 21 °C, the highest temperature was 26 °C, and the average temperature was 23.2 °C. According to the “Indoor Air Quality Standard” (GB/T 18883-2002), the winter indoor heating temperature standard is 16–24 °C (each picture uses a red line to indicate the standard range); thus, the living room exceeded the standard required temperature 11.7% of the time. The lowest temperature recorded in the study was 16 °C, the highest temperature was 27 °C, the average temperature was 23.0 °C, and the temperature exceeded the standard requirement 7.7% of the time. The lowest temperature recorded in the bedroom was 19 °C, the highest temperature was 30 °C, the average temperature was 26.0 °C, and the temperature exceeded the standard requirement 70.1% of the time. The parameter image of each room is shown in Figure 4.

A list of the indoor environment parameters of different rooms, including the highest value, lowest value, average value, and noncompliance ratio, is shown in

Table 3. Indoor environment of different rooms.

Environmental Parameters	Room Name	Standard Value	Highest Value	Lowest Value	Average Value	Noncompliance Rate
Temperature	Living room	16–24 °C	26	21	23.2	11.70%
	Study room		27	16	23	7.70%
	Master bedroom		30	19	26	70.10%
Humidity	Living room	30–60%	37	18	26.2	87.80%
	Study room		42	19	27	87.10%
	Master bedroom		33	10	22	95.40%
Carbon dioxide	Living room	<1000 ppm	1669	381	733	10%
	Study room		1437	374	748	10.10%
	Master bedroom		1929	382	980	42.30%
Formaldehyde	Living room	<100 μg/m3	940	10	57	6.95%
	Study room		550	10	50.6	3.72%
	Master bedroom		560	10	69.2	6.46%
TVOC	Living room	<600 μg/m3	2110	230	334	2.70%
	Study room		1340	220	317	1.40%
	Master bedroom		1360	210	360	2.10%
PM2.5	Living room	<75 μg/m3	351	1	47.3	19.70%
	Study room		195	1	35.8	10.10%
	Master bedroom		205	1	34	9.50%

. The noncompliance rate is listed in the last column to represent the proportion that does not meet the standard range for indoor environmental parameters.

The results show that many aspects have a certain impact on indoor air quality. For example, indoor air environmental parameters will be affected by residents’ activities. For example, through the previous questionnaire survey, we found that residents would turn on the ventilator when they use the kitchen and bathroom for ventilation, and they would open the window every day for regular ventilation when the weather is sunny. In addition, the room temperature and humidity will be affected by cooking behavior and the internal factors of residential central heating, as well as external sunshine and other factors. The maximum temperature of all rooms was at about 2 p.m., and the minimum temperature almost appeared in the early morning, which was consistent with the trend of outdoor temperature. The average relative humidity could not reach the lower limit of 30%, which may be related to the low air humidity in winter. Because the doors and windows were closed to keep the indoor air warm in winter, CO₂ was significantly affected by people’s behavior and activities, especially in bedrooms, where it was concentrated at 0–8 a.m., when people were asleep; increased CO₂ concentrations make people indoors drowsy [52]. As the living room is the main space for leisure and entertainment and connected with the kitchen, it is easy for it to be affected by cooking activities, so the peak concentrations of formaldehyde, TVOC, and PM_2.5 all appeared in the living room, which was consistent with the peak concentration mentioned in other literature and related to specific family activities (such as cooking or leisure) [53]. According to The Standard of the Measurement and Evaluation for Efficiency of Building Ventilation (JGJ/T 309-2013) of the People’s Republic of China. The concentration limit of PM_2.5 was 75 μg/m³, which fluctuated violently and exceeded 75 μg/m³ frequently. The concentration of PM_2.5 in the daytime was higher than that in the evening. The concentration of PM_2.5 increased significantly at 7:00–9:00, 13:00–15:00, and 22:00, which may have been caused by the activities of the elderly, but the interval mean value was lower than the limit of 75 μg/m³. The TVOC and indoor formaldehyde pollution were not serious.

3.2. Correlation between Single Indoor Environmental Parameter and Blood Pressure of the Elderly

In order to explore the correlation between the indoor environment and the blood pressure of the elderly, we used SPSS 22.0 to calculate the nonparametric correlation coefficients, as shown in

Table 4. Correlation analysis of indoor environment and blood pressure.

Environmental Parameters	Correlation Coefficient
	Male Systolic Pressure	Male Diastolic Pressure	Female Systolic Pressure	Female Diastolic Pressure
Living room temperature	−0.59 ***	−0.20	−0.50 **	−0.41 *
Study room temperature	−0.57 ***	−0.20	−0.52 **	−0.50 **
Bedroom temperature	−0.60 ***	−0.08	−0.53 **	−0.36
Living room relative humidity	−0.57 ***	−0.13	−0.54 ***	−0.36
Study room relative humidity	−0.39 *	−0.01	−0.42 *	−0.29
Bedroom relative humidity	−0.30	0.23	−0.37	−0.30
Living room carbon dioxide concentration	−0.01	0.02	−0.11	0.12
Study room carbon dioxide concentration	0.09	0.10	−0.07	0.10
Bedroom carbon dioxide concentration	0.28	0.36	0.18	0.19
Living room formaldehyde concentration	−0.24	−0.07	−0.47 *	−0.20
Study room formaldehyde concentration	−0.34	−0.25	−0.37	−0.23
Bedroom formaldehyde concentration	−0.31	−0.02	−0.36 *	−0.42 *
Living room TVOC concentration	−0.24	−0.04	−0.38	−0.15
Study room TVOC concentration	−0.29	−0.13	−0.48 *	−0.29
Bedroom TVOC concentration	−0.32	0.06	−0.33 *	−0.34
Living room PM2.5 concentration	−0.22	−0.07	−0.24	−0.09
Study room PM2.5 concentration	−0.25	−0.09	−0.27	−0.11
Bedroom PM2.5 concentration	−0.41 *	−0.23	−0.35	−0.22

Note: * Represents p < 0.05; ** Represents p < 0.01; *** Represents p < 0.001.

. The marks are determined by the p values threshold, and the values in the table are the Spearman correlation coefficients. Since blood pressure data were collected twice a day, respectively before going to bed at night and after getting up in the morning, the indoor environment data within 1 h of blood pressure measurement were used in the correlation analysis.

It can be seen from the results that the temperature of each room had an extremely significant correlation with the systolic blood pressure (p < 0.01), and had a correlation with part of the diastolic blood pressure. Between 16.0 and 32.0 °C, temperature and blood pressure showed a negative correlation. The relative humidity and systolic blood pressure of the living room and the study room were also correlated, and in the relative humidity range: 10–42%, the relative humidity was negatively correlated with blood pressure. Among the air pollutants measured, formaldehyde, TVOC and PM_2.5 were found to be related to changes in blood pressure.

3.3. Blood Pressure Error Curve and Probability Prediction Model

The proportion of each element of the input sequence in the model was calculated by LSTM input layer parameters. The calculation results are shown in

Table 5. Weight ratios at different time points.

Time Point	Male Systolic Blood Pressure	Male Diastolic Blood Pressure	Female Systolic Blood Pressure	Female Diastolic Blood Pressure
1	4.60%	6.40%	3.20%	3.40%
2	4.60%	6.10%	4.30%	5.00%
3	4.50%	5.60%	8.60%	4.90%
4	4.40%	4.40%	10.70%	6.00%
5	4.70%	4.40%	10.80%	5.30%
6	6.20%	3.50%	6.50%	4.90%
7	7.50%	3.90%	6.30%	5.90%
8	8.10%	3.40%	6.30%	5.70%
9	9.10%	3.10%	6.60%	6.50%
10	8.00%	2.90%	6.80%	7.80%
11	7.90%	2.80%	13.30%	8.10%
12	30.30%	53.50%	16.60%	36.50%

, reflecting the weight distribution of data at different times in the input.

It was mentioned in the deep learning algorithm and parameter settings that according to the daily lives of the two elderly people, the whole day was divided into two periods: 9 a.m. to 8 p.m. and 9 p.m. to 8 a.m. the next day. For each segment of 12 h, using 12 h environmental data as the input sequence, a model corresponding to blood pressure was established. Time point 12 was the latest time point for estimating blood pressure. According to Table 5, the weight ratio of systolic and diastolic blood pressure in males was 30.3% and 53.5%, while that in females was 16.6% and 36.5%. When measuring blood pressure, environmental data at time point 12 had the highest weight for blood pressure prediction, so the environmental time series data 1 h before measuring blood pressure was used to predict blood pressure. The least square errors of the four models of blood pressure and environment for the elderly reflect the magnitude of the prediction and actual errors. The results of the iteration are shown in Figure 5, Figure 6, Figure 7 and Figure 8 below. The horizontal axis is the number of iterations, the vertical axis is the root mean square error, and the gradient descent method shows the weights and deviations.

The average value of 50~500 steps of the least square error was selected as the Bayesian standard deviation. According to the bivariate correlation analysis, the Pearson correlation coefficient of diastolic blood pressure and systolic blood pressure in the elderly man was ${\rho }_1$ = 0.634, and it was ${\rho }_2$ = 0.768 for the elderly woman. According to Formula (8), a two-dimensional normal distribution can obtain the blood pressure probability density distribution function; according to Formula (9), to obtain the probability of suffering from hypertension, the LSTM model and blood pressure prediction and the risk of suffering from hypertension are shown in

Table 6. LSTM model and blood pressure prediction and risk of hypertension.

Bayesian Fitting Parameters and Results	Male		Female
	Systolic Pressure	Diastolic Pressure	Systolic Pressure	Diastolic Pressure
LSTM predicted value $y_i$	135.24	72.55	131.1.	78.84
LSTM prediction error	21.8	3.36	23.67	29.92
Single-factor judgment of risk of hypertension	16.44%	0.00%	14.11%	7.14%
Systolic pressure and correlation coefficient of systolic pressure	0.634		0.768
Risk of hypertension	16.44%		17.55%

Note: Single-factor judgment refers to analyzing only systolic or diastolic blood pressure.

. Here, the single-factor judgement refers to the degree of risk described by only a single parameter, either systolic or diastolic blood pressure.

The prediction error of systolic and diastolic blood pressure of the female elderly individual is large, the blood pressure data distribution is scattered, the curve is flat, and both diastolic and systolic blood pressure are at risk of exceeding the standard blood pressure. The prediction error of the diastolic blood pressure of the male elderly individual is very small, and the curve is flat and concentrated around the mean value. The risk of hypertension of the diastolic blood pressure single factor is almost zero, the prediction error of the systolic blood pressure is large, the curve is flat, and the mean value is 135.24 mmHg, which is close to the high pressure limit of 140 mmHg, so the risk of hypertension is high. The above parameters were used to draw the normal distribution curve to predict the probability distribution of blood pressure. This is shown in Figure 9.

3.4. Comparison of Predicted Blood Pressure and Pre-Blood Pressure Distribution

A deep learning algorithm is used to predict parameters through training data, provided that the most accurate ratio between training time and its reproducibility is the best value of the number of iterations [54]. Training samples are used to adjust the parameters in the training factors. When the number of training samples in the network structure is limited and the number of calculation factors increases, its generalization ability will become very poor [55]. Parameter settings produce different results to a certain extent; thus, we need to evaluate the results. In order to further test the accuracy of the probability distribution of blood pressure predicted by the deep learning algorithm, the frequency distribution statistics of the measured blood pressure values were analyzed. Although the time series were not comparable in the strict sense, considering the outdoor meteorological parameters and the subjects’ habitual activities, the residential parameters were in line with the objective variation law and the variation range was small; thus, they can be used to compare the accuracy of the model. The parameters of the Gaussian model and high-pressure risk are shown in

Table 7. Simulated Gaussian value of blood pressure curve of male elderly people.

	Systolic Pressure	Diastolic Pressure
Equation	$y=y_0+\left(\frac{A}{\sqrt{2\pi }w}exp\left(-\frac{{\left(x-x_c\right)}^2}{2w^2}\right)\right)$
$y_0$	4.74262 × 10−4 $\pm$ 0.00131	−3.4417 × 10−4 $\pm$ 0.00105
$x_c$	135.79726 $\pm$ 0.51247	73.18374 $\pm$ 0.14837
w	10.35555 $\pm$ 1.05984	5.56568 $\pm$ 0.30174
$A$	0.92412 $\pm$ 0.08705	1.05515 $\pm$ 0.05114
Reduced Chi-Sqr	2.27889 × 10−4	1.604005 × 10−4
Adjusted R square	0.52783	0.8052
Single-factor high pressure risk	24.25%	0.00%
Multi-factor high pressure risk	24.25%

Note: Single factor refers to analyzing only systolic or diastolic blood pressure.

and

Table 8. Simulated Gaussian value of blood pressure curve of female elderly people.

	Systolic Pressure	Diastolic Pressure
Equation	$y=y_0+\left(\frac{A}{\sqrt{2\pi }w}exp\left(-\frac{{\left(x-x_c\right)}^2}{2w^2}\right)\right)$
$y_0$	4.91559 × 10−4 $\pm$ 9.94549 × 10−4	2.5933 × 10−5 $\pm$ 0.00101
$x_c$	132.55752 $\pm$ 0.45283	78.4307 $\pm$ 0.57203
w	11.53666 $\pm$ 0.94079	14.36104 $\pm$ 1.20235
$A$	0.92135 $\pm$ 0.06967	0.99585 $\pm$ 0.07874
Reduced Chi-Sqr	1.27922 × 10−4	1.23631 × 10−4
Adjusted R square	0.63786	0.62233
Single-factor high pressure risk	18.38%	7.09%
Multi-factor high pressure risk	21.77%

Note: Single factor refers to analyzing only systolic or diastolic blood pressure.

.

Previous blood pressure measurements showed that the risk of single-factor hypertension of systolic blood pressure was 24.25%, the risk of diastolic blood pressure was 0%, and the risk of multi-factor hypertension was 24.25% in the male elderly; the risk of single-factor hypertension of systolic blood pressure was 18.38%, the risk of diastolic blood pressure was 7.09%, and the risk of multi-factor hypertension was 21.77% in the female elderly. For visualization, the distribution frequency of diastolic and systolic blood pressure of male and female elderly people was plotted according to the Gaussian value parameter of the nonlinear fitting curve, as shown in Figure 10.

It can be seen from the above two figures that the trend is consistent with the previous LSTM estimation model curve, indicating that the early blood pressure level has reference value for the later environmental effect on the human body. The previous characteristics also showed that the distribution of the systolic and diastolic blood pressure data of female elderly people was scattered, and that they were at risk of exceeding the standard blood pressure. The diastolic blood pressure of male elderly people was concentrated around the mean value, and the diastolic blood pressure almost did not exceed 90 mmHg, while the systolic blood pressure changed greatly around the mean value of 135.24 mmHg and the risk of hypertension was high.

4. Conclusions

In this paper, the indoor environmental parameters and blood pressure data collected from elderly people in a residence in Dalian city are used as training models, mathematical statistics analysis and deep learning prediction are carried out, and a probability density model of the impact of environmental parameters on blood pressure in the growth time scale is established. The following conclusions can be drawn from the prediction of blood pressure based on the average environmental data during the measurement period.

(1)

Many aspects can have certain impacts on indoor air quality, including internal factors; for example, cooking behavior, smoking, and central heating in the room, as well as external ones, such as weather conditions outside the building. Regular household activities and weather changes will result in similar daily changes in indoor environmental parameters and peak times.

(2)

Indoor environmental parameters may have potential effects on health. It is concluded that the influence of temperature and humidity on systolic blood pressure is significantly higher than that of diastolic blood pressure, and there was a correlation between some air quality parameters and blood pressure, such as formaldehyde, TVOC, and PM_2.5. The effect of air quality on blood pressure risk is small, which may be related to the overall good air quality of the family.

(3)

The risk of systolic hypertension was higher than that of diastolic hypertension for both the two elderly, which is consistent with the fact mentioned in the introduction that controlling systolic blood pressure is more important. By comparing the blood pressure value predicted by the deep learning algorithm with previous blood pressure measurement results, it was found that for the male resident, the risk difference for systolic blood pressure was 7.81%, and the risk difference for diastolic blood pressure was the same. For the female one, the risk difference for systolic blood pressure was 4.27%, the risk difference for diastolic blood pressure was 0.05%, and the total risk difference was 4.22%. The algorithms used in this study follow the same trend in the relationship between variable sizes in the prediction. Although the time series used in these two methods are different, they are still reliable since the risk changes remain the same over a short period of time.

This study is a prediction of the hypertension risk of the elderly in the built environments by using deep learning algorithm. It was found that the environment may have potential influence for the elderly. In the future, we plan to test the reliability of the algorithm in clinical practice, continue our research idea, read relevant papers, and study the mechanisms of other diseases. If it has a good enough reliability, we will extend the algorithm to other disease diagnosis in the future.