Spatial Modeling of Asthma-Prone Areas Using Remote Sensing and Ensemble Machine Learning Algorithms

Abstract

In this study, asthma-prone area modeling of Tehran, Iran was provided by employing three ensemble machine learning algorithms (Bootstrap aggregating (Bagging), Adaptive Boosting (AdaBoost), and Stacking). First, a spatial database was created with 872 locations of asthma patients and affecting factors (particulate matter (PM₁₀ and PM_2.5), ozone (O₃), sulfur dioxide (SO₂), carbon monoxide (CO), nitrogen dioxide (NO₂), rainfall, wind speed, humidity, temperature, distance to street, traffic volume, and a normalized difference vegetation index (NDVI)). We created four factors using remote sensing (RS) imagery, including air pollution (O₃, SO₂, CO, and NO₂), altitude, and NDVI. All criteria were prepared using a geographic information system (GIS). For modeling and validation, 70% and 30% of the data were used, respectively. The weight of evidence (WOE) model was used to assess the spatial relationship between the dependent and independent data. Finally, three ensemble algorithms were used to perform asthma-prone areas mapping. According to the Gini index, the most influential factors on asthma occurrence were distance to the street, NDVI, and traffic volume. The area under the curve (AUC) of receiver operating characteristic (ROC) values for the AdaBoost, Bagging, and Stacking algorithms was 0.849, 0.82, and 0.785, respectively. According to the findings, the AdaBoost algorithm outperforms the Bagging and Stacking algorithms in spatial modeling of asthma-prone areas.

1. Introduction

Asthma is a chronic and inflammatory condition of the airways that affects more than 300 million people worldwide. According to a report by the Global Initiative for Asthma (GINA), this number is expected to reach 400 million by 2025 [1,2]. The death rate from asthma is so high that it kills 250,000 people annually worldwide [3]. Asthma prevalence has been rising globally in recent decades. It also tends to haunt a patient for the rest of their life [2,4]. There is no definitive cure for asthma, but it can be controlled and managed [5], and in this case, the risk of asthma attacks and resulting mortality is reduced. Asthma is a reversible airway obstruction and bronchospasm condition that affects the lungs [6]. Wheezing, coughing, and shortness of breath are common asthma symptoms caused by a combination of genetic and environmental conditions. In other words, asthma is caused by genetically susceptible individuals being exposed to environmental risk factors [7]. The occurrence of asthma is influenced by genetic predisposition, environmental influences such as climatic parameters, air pollution, allergens, and airborne chemical irritants [4,8]. While genetics play a significant role in asthma growth, the increase observed in the last two decades cannot be explained merely by genetic changes [9]. Understanding the asthma risk factors is crucial to avoiding or reducing the severity of the symptoms of the disease. Investigating environmental factors and their role in the growth of asthma is one of the best approaches to control this disease [10].

The Geographical Information System (GIS) is a helpful tool for assessing the links between disease incidence and environmental quality [11]. GIS is used to process health data, analyze spatial spread, and track disease variation. Furthermore, this technique allows for the spatial localization of the tracked disease and layers combined with knowledge about the environmental quality [11,12]. Asthma maps offer helpful knowledge to epidemiologists and allow them consider asthma risk factors such as air pollution and identify vulnerable areas. These maps provide a graphic representation of disease incidence, used in public health [13]. A primary feature of an early warning system is a spatio-temporal map that tracks the disease’s spread [14]. One of the main components of GIS and spatial modeling is data. Remote Sensing (RS) is a convenient tool for monitoring environmental variables anywhere, anytime. This tool can play an influential role in creating a spatial database in GIS [15]. RS uses satellite imagery to monitor various parameters, including pollution monitoring. Satellite data help provide spatial data owing to their accessibility, high spatial resolution, and coverage of a wide range of study areas [16,17]. The aim of GIS mapping distribution is to gather new knowledge about diseases or health issues [18]. Disease distribution may be predicted using environmental factors gathered from many sources, such as geographic information and remote sensing data. This method has been proven to be effective in the prevention of disease and in the prediction of epidemics, which is critical for health systems’ preparedness to deal with such outbreaks [19,20]. So far, various studies have used spatial analysis to study different diseases. BenBella and Ghosh [21] examined the combination of spatial analysis with HIV care intervention to identify different indicators of HIV/AIDS treatment in Uganda. Pham et al. [22] evaluated and modeled dengue vulnerability in the Mekong Delta of Vietnam using spatial and time-series approaches. Vincent et al. [23] conducted geospatial mapping, epidemiological modeling, statistical correlation, and analysis of COVID-19 with forest cover and population in Tamil Nadu, India. Abdullah et al. [24] investigated the environmental factors associated with the distribution of visceral leishmaniasis in indigenous areas of Bangladesh.

However, few studies on asthma mapping have been conducted, and previous research has been mainly limited to an exploratory visualization of existing asthma prevalence data. Gordian et al. [25] investigated the relationship between traffic exposure and asthma diagnosis in children using GIS. In New York, the USA, Gorai et al. [26] analyzed the spatial association between air pollution parameters and asthma. Samuels-Kalow and Camargo [27] used geographic data to improve asthma care and population health. Using a Bayesian approach, Ouédraogo et al. [28] investigated the spatial patterns and determinants of asthma prevalence and healthcare use in Ontario. Zook et al. [29] integrated spatial analysis into policy formulation and traffic and asthma exposure. Pala et al. [30] examined the spatial potential of major cities to enable the aggregation and study of environmental, geographic, social, and health data related to asthma. Leynaert et al. [31] investigated environmental risk factors for the development of asthma in children. Kinghorn et al. [32] examined socioeconomic and environmental factors for pediatric asthma in an Indian-American community. Ahmed Khan et al. [10] evaluated asthma susceptible areas in Karachi, Pakistan using environmental factors and GIS. Krautenbacher et al. [33] determined asthma in farm children by genetic polymorphism and in non-farm children by environmental factors. Hauptman et al. [34] assessed proximity to major roads and asthma symptoms in an inner-city school asthma study. Rodríguez-Orozco et al. [35] performed a spatial analysis of asthma in Morelia, Mexico, from 2010–2010. Razavi-Termeh et al. [36] investigated six air pollutants affecting Spatio-temporal modeling of asthma.

Numerous studies have been focused on the geographic distribution of asthma and the association between asthma and environmental factors. However, spatial modeling of asthma-prone areas using the integration of GIS, RS, and machine learning algorithms has received less attention in previous research. In spatial modeling using GIS and RS, we are constantly faced with a large amount of data. For spatial analysis and modeling of this volume of data, machine learning is a suitable tool for training, predicting, and extracting spatial patterns [37]. As a result, this study aimed to use ensemble machine learning algorithms to model asthma-prone areas in Tehran using GIS and RS. In modeling with machine learning algorithms, there are always errors that affect the accuracy of the output. Noise, bias, and variance are the three primary sources of learning errors in machine learning algorithms. These errors can be reduced using ensemble machine learning algorithms [38]. Therefore, to improve the modeling accuracy, in this research three ensemble algorithms (Bootstrap aggregating (Bagging), Adaptive Boosting (AdaBoost), and Stacking) were used. Therefore, the regression type of machine learning algorithms was used in this study owing to the nature of the data and the continuous prediction of asthma-prone areas. These three algorithms have been very accurate in different spatial modeling so far [39,40]. This research has three innovations: (1) using RS factors (air pollution factors, normalized difference vegetation index (NDVI), and altitude) in spatial modeling of asthma, (2) spatial mapping of asthma with three ensemble machine learning algorithms, and (3) integration of GIS, RS, and machine learning for asthma-prone area modeling.

2. Materials and Methods

This research was conducted in five main steps (Figure 1). The materials (study area and data) and methods (machine learning methods, statistical methods, and validation methods) used in this research are described below.

2.1. Study Area

Tehran is Iran’s largest city and capital, as well as the center of the Tehran Province. It has a population of 8,244,535 and is the 25th most populous city in the world. The area of this city is 730 km². Tehran is located in northern Iran, in the southern foothills of the Alborz Mountains, at a longitude of 51°2′E to 51°36′E, and a latitude of 35°34′N to 35°50′N. The height of the city in the highest points of the north reaches about 2000 m. Tehran’s climate is affected by mountains in the north and plains in the south. North of Tehran has a temperate and humid climate, and in other parts of the city is hot and dry and slightly cold in winter. The most important source of rainfall in Tehran is the humid Mediterranean and Atlantic winds that blow from the West. July and January are the hottest and coldest months of the year in Tehran, respectively. The maximum relative humidity is about 70% in the cold months and drops to 32% in the warm months. In Tehran, the highest rainfall occurs in winter with 43% of total rainfall and then in spring with 36%. The main land cover of Tehran includes residential (28.8%), streets (18.6%), and green space (11.4%). Among the problems of Tehran are heavy car traffic and air pollution, which affects respiratory disease (Figure 2).

2.2. Asthma Data

The location of asthma patients was used as dependent data in the modeling of asthma disease. The locations of asthmatic patients in Tehran in 2019 were gathered using the information system of Tehran hospital. A total of 872 asthmatic patient locations were obtained, with 70% (611 locations) being used in modeling and 30% (261 locations) being used in the assessment. The holdout method was used to divide the training and test data (Figure 2).

2.3. Effective Criteria

In this study, altitude, meteorology factors (rainfall, temperature, humidity, and wind speed), air pollutants (carbon monoxide (CO), Nitrogen dioxide (NO₂), Ozone (O₃), Sulfur dioxide (SO₂), and particle matter (PM_2.5 and PM₁₀)), distance to the street, traffic volume, and NDVI were considered as factors affecting the occurrence of asthma (Figure 3). The spatial resolution of all effective factors was considered to be 30 × 30 m. Each of these factors is described below.

• Altitude

Altitude can affect asthma by affecting oxygen levels, air pollutants, and climatic parameters [41]. A digital elevation model (DEM) was used to prepare the altitude map in 2019. Advanced space-borne thermal emission and reflection radiometer (ASTER) images with a pixel size of 30 m were used to create the DEM. Altitude layer processing and preparation for modeling were performed in ArcGIS 10.3 software.

• Meteorology data

Owing to the high impact of pollutants on the distribution and density of meteorological parameters, careful study of the relationship between air quality and weather conditions can help improve air pollution models to predict pollution crises, including its impact on human health such as lung diseases and asthma [42]. Therefore, in this study, meteorological parameters of rainfall, wind speed, temperature, and humidity were used from 2009 to 2019, and the annual average of these data were used to construct meteorological parameter maps for 12 meteorological stations in the Tehran province. The kriging interpolation technique in ArcGIS 10.3 software was used to prepare raster maps of meteorological parameters. Interpolation validation was carried by using Equations (1) and (2) (root mean square error (RMSE), and % RMSE). Because RMSE is sensitive to outlier data, % RMSE can be used instead. The lower the value of this index, the higher the accuracy of interpolation. The acceptable limit for % RMSE is <40, while values more than 70% indicate estimate point imprecision [43].

$$\mathrm{RMSE}={(\frac{1}{\mathrm{n}}{\displaystyle \sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{E}}_{\mathrm{i}}-{\mathrm{A}}_{\mathrm{i}}\right)}^2)}^{\frac{1}{2}}$$

(1)

$$\%\mathrm{RMSE}=\frac{\mathrm{RMSE}}{\mathsf{\mu }}$$

(2)

where ${\mathrm{A}}_{\mathrm{i}}$ is the average variable measured at each station, ${\mathrm{E}}_{\mathrm{i}}$ is the estimated predictor via kriging, and n is the total number of stations. Each measurement component average is represented by μ. The accuracy of interpolation of environmental factors is summarized in

Table 1. Result of interpolation accuracy.

Factors	RMSE	% RMSE	Functioning
Rainfall	95.86	30.64	Acceptable
Temperature	3.03	22.85	Acceptable
Humidity	2.9	6.47	Acceptable
Wind speed	2.63	17.13	Acceptable
CO	0.36	24.86	Acceptable
NO2	16.54	32.61	Acceptable
O3	3.78	18.83	Acceptable
SO2	5.45	34.95	Acceptable
PM2.5	5.8	18.14	Acceptable
PM10	13.23	16.6	Acceptable

.

The evaluation results of the kriging method showed acceptable accuracy for all meteorology factors (% RMSE < 40).

• Air pollutants

Toxic particles from air pollution can enter the lungs through the nose and cause variable damage to respiratory health. The prevalence of asthma and chronic pulmonary obstruction is directly related to increased air pollution [44]. Therefore, air pollutants are one of the factors affecting asthma. Sentinel 5P satellite imagery was used to prepare SO₂, NO₂, O₃, and CO pollutants. The measurements’ spatial resolution (3.5 × 7 km² for NO₂, SO₂, and O₃, and 7 × 7 km² for CO) allows air pollution observations. The average maps of these four pollutants were prepared for the time mentioned (July 2018 to December 2019) in the Google Earth Engine (GEE) platform and transferred to ArcGIS 10.3 software for further processing. Owing to the impossibility of direct monitoring of PM₁₀ and PM_2.5 pollutants by satellite images, station data were used to map these two pollutants. From 2009 to 2019, an average of air pollution data (PM₁₀ and PM_2.5) was collected from 23 air pollution monitoring stations in Tehran. In ArcGIS 10.3, the Kriging interpolation technique was employed to map these two pollutant parameters. The results of the interpolation evaluation of air pollution factors (Table 1) revealed that these factors were prepared with acceptable accuracy (% RMSE < 40).

• Distance to street

Streets are effective in creating traffic and air pollution. Street data were obtained from the open street map (OSM) at a scale of 1:100, 1000 in 2019. This criterion was prepared in ArcGIS 10.3 software using the Euclidean distance tool.

• Traffic volume

Vehicle fuel combustion gases are one of the most important air pollutants. Therefore, traffic plays an essential role in air pollution. This research used traffic volume to analyze the traffic impact on asthma [45]. Traffic volume is the number of vehicles that cross a road in one or more lanes at a given time. The traffic volume layer was created using the average annual traffic volume from 2015 to 2019. These data were collected from Tehran Traffic Control Company and processed in ArcGIS 10.3 software.

• Normalized difference vegetation index (NDVI)

By absorbing lead, dust, and soot from the air, green space helps to clean it up [46]. Therefore, green space can play an influential role in reducing air pollutants. The NDVI is a suitable method for calculating vegetation cover from satellite images and measuring vegetation volume [47]. The NDVI map was created using Landsat 7 images and the enhanced thematic mapper plus (ETM+) sensor. To create the NDVI map, the 2009–2019 annual average was used in the GEE platform. The scan line corrector (SLC) and the gap in Landsat images were corrected using a focal median filter. The NDVI index is calculated using Equation (3).

$$\mathrm{NDVI}=\frac{{\rho }_{NIR}-{\rho }_{red}}{{\rho }_{NIR}+{\rho }_{red}}$$

(3)

where${\rho }_{NIR}$ denotes near-infrared reflectance (band 4—Landsat 7), and ${\rho }_{red}$ indicates red reflectance (band 3—Landsat 7). NDVI map with 30 m pixel size was prepared and transferred to ArcGIS 10.3 software for processing.

2.4. Factors Importance Using Gini Index

Breiman presented the Gini index, a divergence-based attribute splitting approach commonly used in the random forest (RF) algorithms [48]. When a variable is randomly selected, the Gini index measures the probability of it being incorrectly labeled [49]. The Gini index is calculated using Equation (4):

$$\mathrm{Gini} \mathrm{index}=1-{\displaystyle \sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{P}}_{\mathrm{i}}\right)}^2$$

(4)

where ${\mathrm{P}}_{\mathrm{i}}$ denotes the probability of an element.

2.5. Multicollinearity Analysis

Multicollinearity occurs when two or more independent factors are highly interdependent. The inflation variance index (VIF) assesses that the parameters affecting asthma are independent of each other and can participate in modeling. Multicollinearity analysis is appropriate, according to previous studies, if the VIF value is less than 5 [50].

2.6. Weight of Evidence (WOE) Model

The weight of evidence is a data-driven method known as one of the methods of Bayesian theory in the form of the linear logarithm. The WOE model is defined based on the positive ($W^{+}$) and negative ($W^{-}$) weights [51]. The weight of each factor of asthma occurrence (A) dependent on the presence or absence of asthma locations (B) in the study area is as follows in this model (Equations (5) and (6)).

$$W^{+}=\ln \frac{P\left\{B|A\right\}}{P\left\{B|\overline{A}\right\}}$$

(5)

$$W^{-}=\ln \frac{P\left\{\overline{B}|A\right\}}{P\left\{\overline{B}|\overline{A}\right\}}$$

(6)

A positive weight ($W^{+}$) indicates a positive relationship between the presence of an influential factor, and a negative weight ($W^{-}$) indicates that the level of the relationship is negative. $B$ and $\overline{B}$ show the presence and absence of asthma factors, respectively. $A$ and $\overline{A}$ indicate the presence and absence of asthma, respectively. The difference between positive and negative weight is parameter $C$ (Equation (7)) [52]. The standard deviation $\left(S\left(C\right)\right)$of $W$ is determined by Equation (8):

$$C=W^{+}-W^{–}$$

(7)

$$S\left(C\right)=\sqrt{S^2\left(W^{+}\right)+S^2\left(W^{–}\right) }$$

(8)

where $S^2\left(W^{+}\right)$ is the variance of $W^{+}$ and $S^2\left(W^{–}\right)$ is the variance of$W^{-}$. Weight variances are calculated as follows (Equations (9) and (10)):

$$S^2\left(W^{+}\right)=\frac{1}{N\left\{B\cap A\right\}}+\frac{1}{B\cap \overline{A}}$$

(9)

$$S^2\left(W^{-}\right)=\frac{1}{N\left\{\overline{B}\cap A\right\}}+\frac{1}{\overline{B}\cap \overline{A}}$$

(10)

The final weight of each category is calculated using Equation (11):

$$W_{final}=\frac{C}{S\left(C\right)}$$

(11)

2.7. Bagging Algorithm

Breiman developed the bagging algorithm in 1996 to increase the classification and generalization of data [53]. This algorithm consists of a group of tree-based classifiers. This algorithm is a meta-algorithm based on the concepts of bootstrapping and combination to improve machine learning. Ensemble machine learning algorithms combine several weak learners to achieve a strong learner [54]. Bagging also helps to reduce variance and avoid over fitting. Bagging can be applied to any model, but decision trees are the most common. Bagging is a particular model of the average trend. In bagging, different training subsets are randomly selected by replacing all the training data. These individual predictors are combined using a method of averaging their decisions. For a test sample, the prediction value will be equal to the value obtained by averaging all predictors [55].

2.8. AdaBoost Algorithm

Freund and Schapire [56] proposed AdaBoost, an iterative algorithm for building a “powerful” classifier as a linear combination classification. Boosting is an ensemble meta-algorithm in machine learning used to reduce imbalances and variances. This method is based on combining the results of different categories to transform weak learning methods into strong ones [57]. A series of decision trees are created using the boosting method, with each tree attempting to reduce the error rate of incorrect classification. Then, each tree makes a prediction, and from these predictions, a vote is derived. Finally, a prediction with the highest number of options is selected as the final prediction [58].

2.9. Stacking Algorithm

The Stacking algorithm was developed in 1992 by Wolpert [59]. Stacking uses heterogeneous-based learning algorithms to implement ensemble learning. The Stacking algorithm structure consists of two levels: base-learners (level-0) and meta-learners (level-1). Meta-learners generalize the predictions of several base-learners by using the low-level output as the high-level input for relearning [60]. The three stages of the stacking algorithm are as follows: (1) by K-fold cross-validation, train various base classifiers using the training set; (2) to create a new reorganized training data set, gather the output predictions of these base classifiers; (3) train the meta-classifier with the new training data set. The stacking algorithm uses meta-learning steps to reduce estimation residuals [61].

2.10. Validation Metrics

• Receiver operating characteristic (ROC) curve

The ROC curve is one of the most important criteria for evaluating the performance of classified or multilayer models. This curve can measure models at different thresholds, and this curve is based on probability [62]. The TPR (True Positive Rate) is on the Y-axis, and the FPR (False Positive Rate) is on the X-axis in the ROC curve [63]. Based on a set threshold value such as T, a sample is considered positive if X > T and negative if X ≤ T. The random variable x here has a probability density function of $f_1\left(x\right)$ for the time it is in the positive group; otherwise, its probability density function is specified by $f_0\left(x\right)$ [64]. Therefore, TPR and FPR are calculated using Equations (12) and (13), respectively.

$$\mathrm{TPR} \left(\mathrm{T}\right)={{\displaystyle \int }}_{\mathrm{T}}^{\infty }{f}_1\left(x\right)\mathrm{dx}$$

(12)

$$\mathrm{FPR} \left(\mathrm{T}\right)={{\displaystyle \int }}_{\mathrm{T}}^{\infty }{f}_0\left(x\right)\mathrm{dx}$$

(13)

The best model is one in which the area under the curve (AUC) is close to one. This means that the closer to one, the more accurate and appropriate the measurement [65].

• Prediction error metrics

The performance of prediction models was assessed using the RMSE and mean absolute error (MAE) indices [66]. The MAE index was calculated using Equation (14):

$$\mathrm{MAE}=\frac{{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}\left|{\mathrm{p}}_{\mathrm{i}}-{\mathrm{o}}_{\mathrm{i}}\right|}{\mathrm{n}}$$

(14)

where ${\mathrm{p}}_{\mathrm{i}}$ is the calculated value of the model, ${\mathrm{o}}_{\mathrm{i}}$ is the value of the observational variables, and n is the number of observations.

3. Results

3.1. Result of Multicollinearity Analysis

Table 2. Result of multicollinearity analysis.

Independent Variables	VIF
CO	1.438
Altitude	2.689
Humidity	1.427
NDVI	1.148
NO2	1.550
O3	1.977
PM10	1.212
PM2.5	1.388
Rainfall	1.301
Distance to road	1.143
SO2	1.304
Temperature	1.440
Traffic volume	1.103
Wind speed	1.939

shows the results of the multicollinearity analysis. The VIF value for all independent variables was less than five, according to the findings. It indicates that there was no multicollinearity in the independent factors used. As a result, modeling should include all independent variables.

3.2. Result of Gini Index

Figure 4 shows the importance of effective factors in modeling with the Gini index. According to the results, the distance to the street (0.45), NDVI (0.4), traffic volume (0.38), SO₂ (0.33), NO₂ (0.3), CO (0.29), O₃ (0.27), PM_2.5 (0.26), PM₁₀ (0.21), temperature (0.18), wind speed (0.17), altitude (0.14), humidity (0.12), and rainfall (0.1) are the most important in modeling, respectively.

3.3. Result of WOE Model

Table 3. Result of WOE model.

Factors	Total Area (pixels)	Asthma Patients	${\mathit{W}}^{+}$	${\mathit{W}}^{-}$	${\mathit{W}}_{\mathit{f}\mathit{i}\mathit{n}\mathit{a}\mathit{l}}$
Altitude (m)
1032–1185.72	247,478	335	0.56	−0.42	12.13
1185.72–1311.21	236,755	159	−0.13	0.05	−2.05
1311.21–1449.25	141,573	43	−0.93	0.12	−6.66
1449.25–1609.25	115,491	45	−0.68	0.08	−4.91
1609.25–1828.86	52,535	29	−0.33	0.01	−1.85
Rainfall (mm)
229.1–265.45	272,915	129	−0.48	0.18	−6.77
265.45–303.98	198,666	137	−0.1	0.03	−1.48
303.98–338.15	133,321	180	0.56	−0.16	8.19
338.15–374.51	130,669	121	0.18	−0.04	2.22
374.51–414.49	58,263	44	−0.01	0.001	−0.13
Temperature (°C)
14.45–15.16	30,509	17	−0.32	0.01	−1.35
15.16–15.6	44,501	21	−0.48	0.02	−2.3
15.6–16.07	165,866	196	0.42	−0.15	6.7
16.07–16.59	258,261	145	−0.31	0.12	−4.6
16.59–17.19	294,697	232	0.02	−0.01	0.43
Humidity (%)
36.58–38.11	187,149	157	0.08	−0.028	1.23
38.11–39.27	193,668	113	−0.27	0.075	−3.38
39.27–40.48	181,871	125	−0.11	0.031	−1.44
40.48–41.59	124,203	160	0.51	−0.13	7.05
41.59–43	106,943	56	−0.38	0.04	−3.09
Wind speed (m/s)
12.69–14	95,796	106	0.36	−0.06	3.97
14–15.04	267,619	352	0.53	−0.44	12.01
15.04–16.11	318,928	143	−0.54	0.24	−8.24
16.11–17.5	44,686	1	−3.53	0.05	−3.59
17.5–18.88	66,805	9	−1.74	0.07	−5.4
CO (mol/m2)
0.031–0.034	159,554	85	−0.51	0.11	−5.34
0.034–0.036	150,954	175	0.26	−0.08	3.98
0.036–0.038	120,295	101	−0.05	0.01	−0.61
0.038–0.04	140,418	140	0.11	−0.03	1.53
0.04–0.042	116,890	110	0.05	−0.01	0.66
NO2 (mol/m2)
0.0004–0.0005	120,420	52	−0.72	0.1	−5.68
0.0005–0.0006	146,419	144	0.1	−0.02	1.38
0.0006–0.0007	120,760	134	0.22	−0.05	2.83
0.0007–0.00079	153,663	109	−0.22	0.05	−2.65
0.00079–0.00089	146,849	172	0.27	−0.09	4.08
O3 (mol/m2)
0.1331–0.1332	80,632	90	0.22	−0.034	2.3
0.1332–0.1333	145,443	188	0.37	−0.13	5.77
0.1333–0.1338	180,368	153	−0.04	0.015	−0.65
0.1338–0.1344	169,154	117	−0.24	0.069	−3.1
0.1344–0.1355	112,514	63	−0.46	0.069	−3.99
SO2 (mol/m2)
0.0001–0.00016	57,938	17	−1.1	0.059	−4.74
0.00016–0.0002	124,441	64	−0.54	0.08	−4.8
0.0002–0.00023	196,937	193	0.098	−0.042	1.62
0.00023–0.00026	156,211	212	0.42	−0.16	6.97
0.00026–0.00031	152,584	125	−0.08	0.021	−1.02
PM2.5 (µg/m3)
22.14–28.85	114,191	24	−1.29	0.11	−6.78
28.85–31.76	242,056	56	−1.2	0.26	−10.48
31.76–34.1	227,113	305	0.55	−0.35	11.26
34.1–36.7	141,277	185	0.53	−0.16	7.9
36.7–44.24	69,197	41	−0.26	0.02	−1.75
PM10 (µg/m3)
59.14–69.96	139,877	29	−1.31	0.14	−7.65
69.96–76.7	222,377	203	0.17	−0.07	2.86
76.7–83.85	262,926	246	0.19	−0.11	3.73
83.85–93.24	96,331	51	−0.37	0.04	−2.84
93.24–111.21	72,323	82	0.38	−0.048	3.67
Distance to street (m)
0–100	306,513	265	0.11	−0.08	2.41
100–200	186,271	193	0.29	−0.11	4.7
200–300	98,020	55	−0.31	0.03	−2.5
300–400	67,266	41	−0.23	0.01	−1.56
>400	135,764	57	−0.6	0.08	−5.003
Traffic volume
0–1112	163,808	7	−0.82	0.01	−2.22
1112–2636	1,129,079	78	−0.34	0.06	−3.39
2636–4634	1,553,159	134	−0.12	0.03	−1.68
4634–7348	3,351,979	369	0.11	−0.15	3.3
7348–59258	43,025	23	1.69	−0.03	8.13
NDVI
0.043–0.18	279,652	304	0.2	−0.16	4.65
0.18–0.29	164,231	133	−0.08	0.02	−1.15
0.29–0.42	139,223	111	−0.1	0.02	−1.21
0.42–0.57	72,949	40	−0.47	0.04	−3.18
0.57–0.92	35,133	23	−0.3	0.01	−1.47

shows the results of the relationship between independent and dependent variables using the WOE model. In the altitude criterion, the highest weight belongs to class 1032–1185.72. According to the rainfall results, asthma occurrence rises initially with rising rainfall, then declines with high rainfall. The results of the temperature criterion show that the class 15.6–16.07 has the highest weight (WOE value = 6.7). In the humidity, the highest incidence of asthma occurs at high humidity (40.48%–41.59%). Asthma is more likely to arise at lower wind speeds. Class 14–15.04 m/s has the highest weight in the wind speed factor (WOE value = 12.01). The results of the CO factor show that as the levels of this pollutant increase, the probability of asthma increases. However, most of the weight of this criterion is related to the middle classes (WOE value = 3.98). The highest weight of the WOE model (WOE value = 4.08) for the NO₂ factor occurs at high values of this parameter. Factor O₃ is inversely related to the occurrence of asthma in the study area. As a result, asthma is more prone to be created in small concentrations of this pollutant. The results of the SO₂ factor show that the highest weight is related to high levels of this pollutant (WOE value = 6.97). According to PM_2.5, the highest weight belongs to the class 31.76–34.1 (WOE value = 11.26). The highest weight of the PM₁₀ factor is in the 76.7–83.85 class (WOE value = 3.73). At distances close to the street, the highest incidence of asthma occurs, so the highest weight (WOE value = 4.7) is related to the class of 100–200 m. The spatial relationship between the occurrence of asthma and the volume of traffic shows the probability of asthma occurring in high amounts of this factor (WOE value = 8.13). The results of the NDVI factor reveal that asthma is more likely to occur with lower values of this parameter (0.043–0.18).

3.4. Result of Modeling and Mapping

For modeling, a spatial database containing asthma data and influencing factors were built. In addition to occurrence data (value 1), we require nonoccurrence data (value 0) for improved network training in machine learning models. Nonoccurrence data were collected at random in the study area, much like the number of occurrence data. Therefore, spatial databases including dependent data (872 asthma occurrence locations and 872 non-asthma occurrence locations) and independent data (WOE model weight for factors) were considered modeling input. Seventy percent of the database was used as training data and 30% as validation data. The Waikato Environment for Knowledge Analysis (WEKA) software was used to implement the three ensemble algorithms. The parameters used in each algorithm are shown in

Table 4. Parameters used by ensemble algorithms.

Algorithms	Parameters
AdaBoost	Number of iterations = 10; seed = 1; batch size = 100; weight threshold = 100; use a base classifier (Random Forest)
Bagging	Number of iterations = 10; seed = 1; number of execution slots = 1; batch size = 100; percentage of bag size = 100; use a base classifier (Random Forest)
Stacking	Seed = 1; number of execution slots = 1; batch size = 100; use a base classifier (Random Forest)

.

After training the three algorithms, RMSE and MAE indices were used to evaluate the accuracy of the algorithms. The results of the RMSE and MAE indices are presented in

Table 5. Result of metrics indices.

Algorithm	Train		Validation
	RMSE	MAE	RMSE	MAE
AdaBoost	0.1678	0.0572	0.252	0.2049
Bagging	0.2169	0.1531	0.3241	0.2773
Stacking	0.2353	0.1555	0.3488	0.3073

. Based on the results, the RMSE index values for training and validation data are for AdaBoost (0.1678, 0.252), Bagging (0.2169, 0.3241), and Stacking (0.2353, 0.3488) algorithms, respectively. According to the results, the AdaBoost (0.0572, 0.2049), Bagging (0.1531, 0.2773), and Stacking (0.1555, 0.3073) algorithms have the highest accuracy based on MAE index values for training and validation data, respectively. The graph of the error rate between the predicted values and the actual data for each algorithm for training and test data is shown in Figure 5. The results show that algorithms AdaBoost, Bagging, and Stacking have the highest accuracy in modeling asthma-prone areas, respectively.

After training the three algorithms, the fitted model for each algorithm was generalized to the whole study area. The output of the three algorithms were converted from WEKA software to ArcGIS 10.3 and were used for asthma-prone area mapping. The spatial mapping of the asthma was divided into five classes based on the natural breaks classification method: very low, low, moderate, high, and very high. The asthma-prone areas mapping using AdaBoost algorithm is shown in Figure 6.

3.5. Result of Validation

Thirty percent of the data on asthma occurrence and nonoccurrence were used to test the map of asthma-prone areas. The ROC curve and AUC in MedCalc software were used for validation. The results of the validation with the ROC curve are shown in Figure 7. The AUC values for AdaBoost, Bagging, and Stacking algorithms are 0.849, 0.82, and 0.785, respectively. The results show that the AdaBoost algorithm is more accurate than the other two algorithms in modeling asthma-prone areas. The results showed good accuracy of AdaBoost and Bagging algorithms and relatively good accuracy of the Stacking algorithm in modeling asthma-prone areas.

According to the AdaBoost algorithm, 16.83% of the area is situated in the very high class, 19.1% in high, 18.5% in moderate, 13.82% in low, and 31.75% in the very low class. The Bagging algorithm assigns 19.22, 21.12, 17.4, 23.99, and 18.27% to the very high, high, moderate, low, and very low categories, respectively. For the Stacking algorithm, similar classes are 17.81, 20.35, 18.43, 27.54, and 15.87%.

4. Discussion

As there is no cure for asthma, it might be helpful to analyze the environmental factors that influence the incidence of this disease to prevent and manage it. Therefore, the aim of this study was spatial modeling of asthma-prone areas with ensemble machine learning algorithms. The WOE model was used to investigate the spatial relationship between independent and dependent data and the input of machine learning algorithms. The WOE model is a useful tool for dealing with nonlinearities between predictor and target [67]. According to the results of the WOE model, lower altitude values had a greater effect on the occurrence of asthma. Low-level pollution, such as that generated by transportation, generally decreases with altitude. This implies that areas will mostly hang suspended mid-air or build up into dense clouds at lower altitudes [68]. Rainfall results showed that at low and high levels of this factor, the probability of asthma is low. Rainfall can have a variety of effects on people with asthma. Pollen may be washed away by light rain, which may help with asthma symptoms. However, heavy rain can scatter pollen quickly into the air. On the other hand, heavy rainfall reduces air pollutants [69]. Owing to the different effects of rainfall, a moderate amount of rainfall can help reduce asthma [70]. Temperature factors, such as rainfall factor, has a different effect on the occurrence of asthma. Cold air can dry out the tissues of the airways and cause them to become more sensitive and closed [71]. When the air temperature is cooler, exhaust pollutants may become trapped at the surface under a layer of dense, cold air [72]. Warm air rises throughout the summer months, dispersing contaminants from the Earth’s surface into the upper troposphere, while more sunlight causes O₃ to form [73]. In this study, moderate values of temperature (15 °C) have a greater effect on the occurrence of asthma. The results of humidity analysis showed that higher values of this factor increase the risk of asthma. High humidity increases pollutants in the air. The O₃ pollutant rises when humidity rises [74]. Asthma is more likely to arise with lower amounts of wind speed, according to the results. Because winds carry pollutants around, wind patterns influence air quality. High values of wind speed play an effective role in reducing air pollutants [75]. According to the results of the WOE model, increasing the CO factor reduces the incidence of asthma. In this study, CO did not play its role well in modeling asthma-prone areas. The results of SO₂ showed that higher values of this factor are more likely to cause asthma. Sulfur oxides, in combination with suspended particles and moisture, increases air pollution. The most common source of SO₂ is fossil fuel combustion [76]. In this research, increasing the amount of NO₂ pollutants increases the risk of asthma. The consumption of fuels at higher temperatures in refineries, petrochemicals, power plants, and household and commercial heating systems are all sources of NO₂ [77]. The results of the O₃ factor showed that this factor has no positive effect on the occurrence of asthma in the study area. O₃ factor occurs more in summer and has less effect on increasing air pollution in cold seasons. As a result, asthma appears to be more common in Tehran during the cold seasons, while O₃ pollution appears to have minimal influence on the prevalence of asthma throughout these seasons [36]. The findings of PM_2.5 and PM₁₀ revealed that as these factor values rise, the probability of asthma in the study area rises. Combustion processes produce a major portion of the PM in urban air. The size of airborne particles affects the respiratory system, and as the particle size decreases, the symptoms increase more severely [78]. Based on the results of the Gini index in air pollutants, SO₂, NO₂, CO, O₃, PM_2.5, and PM₁₀ factors are the most important in modeling, respectively. The results of this study are not consistent with Razavi-Termeh et al.’s [7] research. In a study conducted by Razavi-Termeh et al. [7], all air pollutants were prepared based on ground station data, and PM_2.5 and PM₁₀ factors were the most important in modeling. However, in this research, because of the use of remote sensing images to produce air pollutants (SO₂, NO₂, CO, and O₃), the SO₂ factor is the most important. This method has the advantage of allowing for more accurate air quality measurements in urban areas with few monitoring stations [79]. The distance to the street factor indicates that shorter distances are more likely to induce asthma. Additionally, at higher levels of traffic volume factor, asthma is more likely to occur. According to the Gini index, these two factors were significantly relevant in the occurrence of asthma. In the distances close to the street, due to the traffic of cars, the air pollution is higher and also the high traffic causes more stopping of the cars and increasing the emission of air pollutants [80]. The results of the NDVI factor showed that in smaller amounts of this factor, asthma is more likely to occur. Additionally, according to the Gini index, this factor is of great importance in modeling asthma-prone areas. The concentrations of air pollutants were comparatively low in areas with high NDVI. Because vegetation has a dust-blocking effect, places with less vegetation are more likely to create particulate matter [81]. In urban environments, high population density causes more traffic, less green space and thus increases air pollution [43]. Therefore, living in densely populated areas causes infectious rhinitis, respiratory infections, and asthma [82].

AdaBoost, Bagging, and Stacking algorithms, respectively, had the highest accuracy in predicting asthma-prone areas, according to the findings of assessment indicators. The AdaBoost algorithm was more capable of modeling asthma-prone areas in the study area than the other two algorithms. Advantages of the AdaBoost algorithm over two algorithms is [83]: (1) the ability to merge different types of predictors; (2) decreases bias; (3) models are weighed in Boosting according to their performance; (4) when dealing with bias or under fitting in a data set, the boosting approach comes in useful. The Bagging algorithm has higher accuracy than the stacking algorithm owing to reducing the variance and solving the over fitting problem [84]. Because the Stacking algorithm does not use sampling in the training dataset and does not use a sequence of models to correct the predictions of prior models, it is less accurate than the other two algorithms [85].

The innovations of the present study including, the use of remote sensing to prepare criteria affecting the occurrence of asthma, the use of the WOE statistical method to determine the spatial relationship between independent and dependent criteria, and the use of ensemble machine learning algorithms to map asthma-prone areas. The limitations of the present study included the lack of access to up-to-date population density data, and the lack of direct access to PM_2.5 and PM₁₀ pollutant data using remote sensing images. For future research, it is suggested that PM_2.5 and PM₁₀ pollutants and climatic parameters be prepared using remote sensing images.

5. Conclusions

A combination of GIS, RS, and ensemble machine learning algorithms were employed in this work to propose a strategy for the prevention and management of asthma in urban areas. The results showed that the ensemble machine learning algorithms have good accuracy in modeling asthma-prone areas, where the AdaBoost algorithm showed higher accuracy than the other two algorithms. Low altitude, high rainfall and humidity, moderate temperature, low wind speed, high levels of air pollutants (except O₃), shorter street distance, high traffic volume, and less vegetation all played a part in the prevalence of asthma in Tehran, according to the results of the WOE model. Factors including distance to the street, traffic volume, and NDVI were all important in modeling. Less distance from the street, high traffic volume, and less vegetation increase the emission of pollutants. It seems that air pollution is the primary cause of asthma attacks in Tehran. Remote sensing factors such as NDVI and four air pollutants (SO₂, NO₂, CO, and O₃) played an essential role in modeling asthma-prone areas. Remote sensing factors such as NDVI and four air pollutants played an essential role in modeling asthma-prone areas. Remote sensing images have an excellent ability to integrate with GIS in spatial modeling of diseases owing to the monitoring of environmental factors anywhere in the world at any time. The center and southern parts of Tehran are more in danger, according to asthma risk maps. These regions have a significant impact on increasing levels of air pollution due to their high population density and transportation. Community planners and administrators will be aided using maps of asthma-prone regions for the management and presentation of asthma.