Association between Premature Birth and Air Pollutants Using Fuzzy and Adaptive Neuro-Fuzzy Inference System (ANFIS) Techniques

Abstract

This article uses machine learning techniques as fuzzy and neuro-fuzzy ANFISs, to develop and compare prediction models capable of relating pregnant women’s exposure to air pollutants, such as Nitrogen Dioxide and Particulate Matter, the mother’s age, and the number of prenatal consultations to the incidence of premature birth. In the current literature, studies can be found that relate prematurity to the exposure of pregnant women to NO₂, O₃, and PM₁₀; to Toluene and benzene, mainly in the window 5 to 10 days before birth; and to PM₁₀ in the week before birth. Both models used logistic regression to quantify the effects of pollutants as a result of premature birth. Datasets from Brazil—Departamento de Informatica do Sistema Único de Saúde (DATASUS) and Companhia Ambiental do Estado de São Paulo (CETESB)—were used, covering the period from 2016 to 2018 and comprising women living in the city of São José dos Campos (SP), Brazil. In order to evaluate and compare the different techniques used, evaluation metrics were calculated, such as correlation (r), coefficient of determination (R²), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), Mean Square Error (MSE), and Mean Absolute Error (MAE). These metrics are widely used in the literature due to their ability to evaluate the robustness and efficiency of prediction models. For the RMSE, MAPE, MSE, and MAE metrics, lower values indicate that prediction errors are smaller, demonstrating better model accuracy and confidence. In the case of (r) and R², a positive and strong result indicates alignment and better performance between the real and predicted data. The neuro-fuzzy ANFIS model showed superior performance, with a correlation (r) of 0.59, R² = 0.35, RMSE = 2.83, MAPE = 5.35%, MSE = 8.00, and MAE = 1.70, while the fuzzy model returned results of r = 0.20, R² = 0.04, RMSE = 3.29, MSE = 10.81, MAPE = 6.67%, and MAE = 2.01. Therefore, the results from the ANFIS neuro-fuzzy system indicate greater prediction capacity and precision in relation to the fuzzy system. This superiority can be explained by integration with neural networks, allowing data learning and, consequently, more efficient modeling. In addition, the findings obtained in this study have potential for the formulation of public health policies aimed at reducing the number of premature births and promoting improvements in maternal and neonatal health.

1. Introduction

Preterm birth is defined as birth that occurs before 37 completed weeks of gestation. It is estimated that every year 15 million premature births result in one million neonatal deaths due to complications during childbirth [1]. In addition to being the cause of 75% to 95% of neonatal deaths worldwide, disregarding congenital malformations, prematurity is associated with 50% of neurological impairments. This condition can have a negative impact on an individual’s development, increasing the likelihood of serious illnesses, visual and hearing problems, learning difficulties, and disorders that persist into adulthood. The incidence of preterm birth is considered significant in underdeveloped and developing countries [2,3,4].

In Brazil, in 2000, there were 29 deaths for every 1000 births. Projections by the Instituto Brasileiro de Geografia e Estatística (IBGE) indicate a reduction in the infant mortality rate over the years, with the most recent data released in 2021 showing a rate of 11.2 deaths per thousand births in the country [5]. Factors such as socioeconomic status, the mother’s age and weight, the presence or absence of partners, hypertension, smoking, illicit drug use, schooling, and medical factors such as previous prematurity and multiple pregnancies are responsible for premature births [6,7].

Some studies have found an association between premature birth and maternal exposure to air pollutants such as Nitrogen Dioxide (NO₂), Ozone (O₃) and Particulate Matter with an aerodynamic diameter of less than 10 μm (PM₁₀), especially for female births. In a model developed using logistic regression with multi-pollutants, PM₁₀, O₃, NO (Nitrogen Oxide), and NO₂ (Nitrogen Dioxide), it was proven that the exposure of pregnant women to PM₁₀ seven days before delivery stands out as a promising factor for prematurity [8,9].

1.1. Artificial Intelligence

According to Kaplan and Haenlein (2019), artificial intelligence is defined as the ability of a system to correctly interpret external data in order to learn and to use this learning to achieve specific goals and tasks through flexible adaptations [10]. Artificial intelligence is increasingly being used in the medical field, helping professionals with diagnoses, analysis, and decision-making [11]. Its use has proven to be efficient in screening mammography exams, being able to identify cancer in its early stages, detect arrhythmias through outpatient electrocardiograms, and diagnose Alzheimer’s and lung infections caused by COVID-19 [12,13,14,15].

Computational intelligence seeks to develop systems that work in a similar way to biological systems or as closely as possible to natural systems. Its main paradigms include artificial neural networks, fuzzy systems, and evolutionary computing [16].

Unlike classical logic, in which each element does or does not belong to a particular set, in fuzzy logic, the element can belong to several sets, depending on the degree of pertinence, ranging from zero to one [17]. Based on the use of fuzzy logic and the need for new actions, a technique was created, called the neuro-fuzzy ANFIS system or hybrid system, which is the integration between the fuzzy model and the so-called artificial neural networks, efficient in solving complex problems through learning examples. The neuro-fuzzy ANFIS system can minimize the isolated disadvantages of each technique, highlighting and promoting the qualities of the unifications, as well as responding to more complex and non-linear behaviors [18,19].

Artificial intelligence, precisely fuzzy logic, was chosen as the method of this study because it has been used efficiently in medical diagnoses, as already mentioned, and because it allows analyzing large sets of data, identifying patterns with precision.

1.2. Bibliometric Research

A bibliometric search was carried out in the SCOPUS and Web of Science databases using the keywords “prematurity” or “premature birth” or “preterm birth” and “air pollution” or “air pollutants” between the years 2000 and 2024, for full articles only. The R Studio software version 2023.12.1-40 for MacOS was used, accompanied by the Bibliometrix/Biblioshiny tool, allowing more detailed analysis and visualization of the articles.

For the period from 2000 to 2020, 417 articles related to the keywords were found in the SCOPUS database and 608 articles in the Web of Science, totaling 776 non-duplicate articles from 2873 authors in 248 sources, with an annual growth rate of 16.21%.

When we changed the research in the most current period, 2021 to 2024, 236 articles were found in the SCOPUS database and 332 in the Web of Science, with a total of 439 non-duplicate articles from 2337 authors in 162 different sources. The majority of documents were produced in the United States (152 documents), followed by China (125) and then Australia and the United Kingdom, with 17 productions each. Regarding the most current documents (2021–2024), a word cloud is presented in Figure 1, with greater focus on the terms pregnancy, preterm birth, Particulate Matter, air pollution, and female.

From the bibliometric research carried out, studies were found in the literature that use fuzzy logic to estimate neonatal deaths, based on variables such as birth weight, gestational age, Apgar score, and stillbirth reports [20]. Other studies have used the same logic to predict hospitalizations for respiratory diseases caused by air pollution in Brazil, such as predicting the number of hospitalizations for respiratory diseases in children up to ten years old due to air pollution caused by PM₁₀, NO₂, wind speed, and temperature in the city of São José do Rio Preto (SP) [21]. In addition, a study developed a model based on fuzzy logic using data on exposure to PM₁₀, SO₂, minimum temperature, and wind speed in the city of São José dos Campos (SP), Brazil to predict the length of hospitalization for cardiovascular diseases [22].

When the bibliometric search was carried out in the SCOPUS and Web of Science databases with the keywords “prematurity” or “premature birth” or “preterm birth”, “air pollution” and “air pollutants” and “fuzzy”, no articles were found.

Although there is research exploring the exposure of pregnant women to air pollutants, no studies using fuzzy logic or neuro-fuzzy ANFIS as association techniques were found in the SCOPUS or Web of Science databases from 2000 to 2024. The use of these approaches in this field represents a completely new contribution to the literature. The aim of this study is to use computational intelligence techniques, such as fuzzy and neuro-fuzzy ANFIS, to estimate prematurity, taking as a starting point a pregnant woman’s exposure to atmospheric pollutants (NO₂ and PM₁₀) and other factors, such as the mother’s age and the number of prenatal consultations carried out. To achieve this, we used data on premature births in the city of São José dos Campos-SP, Brazil, between 2016 and 2018.

This article is organized into five section, the first of which contains the introduction with concepts of premature birth, air pollutants, fuzzy logic, neuro-fuzzy ANFIS, and a bibliometric review regarding the association of premature birth and air pollutants. Section 2 presents the study site, as well as the basis used to collect data on live births and air pollutants. This section also covers the use of the fuzzy model, in Section 2.1, and neuro-fuzzy ANFIS model, in Section 2.2, as well as the evaluation metrics in Section 2.3. In Section 3, the results of the study using both techniques are presented. In Section 4, a discussion is provided. Section 5 presents the limitations and opportunities for future work.

2. Materials and Methods

This section will discuss the materials and method used to develop this research, from the study site, data extraction, and collection to the development of fuzzy and neuro-fuzzy ANFIS models.

The city of São José dos Campos is in the interior of the state of São Paulo, in the metropolitan region of the Paraíba Valley, Brazil. According to the IBGE, the municipality has a land area of 1099.409 km², an estimated population of 737,310 in 2021, a Gross Domestic Product (GDP) per capita of 53,646.74, and a Municipal Human Development Index (IDHM) of 0.807. According to the distribution of the municipal GDP in 2020, the services sector is responsible for 52.9% of its economy, while industry accounts for 34.3% [23].

The Departamento de Informática do Sistema Único de Saúde (DATASUS) was created in 1992 and its mission is to “promote modernization through information technology to support the Sistema Único de Saúde (SUS)”. This department is responsible for providing information and support systems necessary for planning, operation, and control process, as well as collecting, processing, and disseminating health information in Brazil [24].

Through the DATASUS open access portal, it was possible to extract an annual database with information collected from all over the country. One of the data points available on the platform is the Declaração de Nascido Vivo (DNV) through the Sistema de Informações sobre Nascidos Vivos (SINASC). After being extracted from the portal, complete information about those born can be found, such as birth data, weight, city, sex, presence or absence of a partner, the mother’s education, the number of living and dead children of the pregnant woman, the number of consultations carried out, race, color, and other information.

From the extracted tables (2015, 2016, 2017, and 2018), information was selected on all live births of pregnant women living in the municipality of São José dos Campos between 1 January 2016 and 31 December 2018. Newborns with less than 22 completed weeks of gestation, births weighing less than 500 g because they are considered abortions, twin and triplet births, and fetuses with congenital malformations were excluded from the analysis. Births by cesarean section were also excluded to rule out possible iatrogenesis.

Data on pollutants, Nitrogen Dioxide (NO₂) and Particulate Matter with an aerodynamic diameter of less than 10 μm (PM₁₀), were collected through the portal of the Companhia Ambiental do Estado de São Paulo (CETESB). This state government agency is responsible for controlling, inspecting, monitoring, and licensing pollution-generating activities, with the fundamental concern of preserving and restoring the quality of water, air, and soil [25].

On the CETESB portal, through Qualidade do Ar (QUALAR), by creating a login, it was possible to view the air quality in real time, the location of the stations, and the daily pollutant reports. From 2015 to 2018, the city of São José dos Campos had two measuring stations for the pollutants under analysis. Daily emission values for each pollutant were extracted throughout the study period. For calculation purposes, the average values from the two stations over the gestational period were used.

2.1. The Fuzzy Model

The software used to apply the fuzzy system was MATLAB R2022b^®, plus the Fuzzy Logic Toolbox. This combination generates a powerful problem-solving tool using fuzzy logic [26]. The Mamdani inference method (maximum and minimum inference composition rule) was used to connect the model and the Centroid method (average of areas with degrees of pertinence) was used for defuzzification. In a fuzzy system, it is essential to specify some information, such as coherent inputs, as well as the number of functions for each input, precise parameters, and the number of rules [27].

The input and output variables were selected from the information extracted and compiled by the DATASUS and CETESB portals, according to the methodology described. To create the model, the input variables chosen were the mother’s age, the number of prenatal consultations, and the NO₂ and PM₁₀ concentrations, and for the output variable, prematurity.

Tests were carried out in order to find the fuzzy model with the best correlation. Initially, three input functions and one output function were considered, both of the triangular type. Other tests were carried out, changing the types of functions for the variables and making parameter adjustments. In the worst model developed, four membership functions were inserted for the mother’s age, four for the number of consultations, and two for each pollutant concentration and output with four membership functions, both of the trapezoidal type.

In the best fuzzy model found, the mother’s age was classified as low, medium, and high; the number of prenatal consultations as inadequate, regular, and adequate; the NO₂ concentration as acceptable and unacceptable; and the PM₁₀ concentration as acceptable and unacceptable.

Table 1. Classifications and parameters of the fuzzy model’s input variable.

Input Variable	Classification	Parameters
Mother’s age	Low	[−10 0 14 18]
	Medium	[15 19 33 36]
	High	[33 37 45 50]
Number of consultations	Inadequate	[−3 0 3 3 4]
	Regular	[3 4 6 7]
	Adequate	[6 7 10 15]
NO2 concentration	Acceptable	[−36 0 36 39]
	Unacceptable	[36 40 50 60]
PM10 concentration	Acceptable	[−20 0 20 24]
	Unacceptable	[20 24 30 40]

shows the classifications and parameters for each input variable in this model.

The output variable, prematurity, had its parameters classified according to the weeks of gestation: for extreme premature, the parameters were set to [−10 0 25 27]; very premature, [25 27 30 32]; moderate premature, [30 32 33 35]; and late premature, [33 34 36 40]. The fuzzy set for this variable is shown in Figure 2. The input and output variables responded best with trapezoidal pertinence functions. The “if-then” rule base in 3 × 3 × 2 × 2 format was validated by an expert (a pediatrician with experience in maternal and child health), covering the maximum number of possible combinations, totaling 36 rules for this system.

2.2. The Neuro-Fuzzy ANFIS Model

The neuro-fuzzy system, also called a hybrid system, is the combination of two modeling techniques, fuzzy and neural networks. From this integration, the disadvantages of isolated techniques are removed, and the construction of a fuzzy inference system (FIS) is improved by adjusting the parameters obtained by learning neural networks [18].

This type of technique was chosen because it is capable of modeling complex functions with greater precision. Furthermore, the neuro-fuzzy ANFIS system has the ability to recognize patterns and predict, crucial for this type of study [28,29].

The main disadvantage of the ANFIS neuro-fuzzy system compared to the fuzzy system is the need for more intense computational work and a significant amount of data to achieve good performance [30].

There are articles that use this technique to solve health problems, such as the prediction of diabetes mellitus [31], heart disease [32], chronic kidney disease [33], Alzheimer’s [34], and others.

To build this model, the Takagi–Sugeno Inference System was chosen; this type of system allows input functions to be used as rule responses and outputs [35].

Similar to the fuzzy model developed, MATLAB R2022b^® software was used, but with the addition of the ANFIS Toolbox. For a coherent comparison of the results, the same input and output variables were used. Forty-five tests were performed, with different numbers and types of membership function, a grid partition clustering technique, a hybrid optimization method, training in 20 epochs, zero tolerance for errors, and constant output with simple validation.

For all tests performed, the data were split and used in a 50:50 format—half for training and the rest for testing and the average testing error values observed.

However, this time, the Gaussian type with a constant output variable was chosen for the input functions, as they had better average test error returns in the software. In the most efficient neuro-fuzzy ANFIS model, the mother’s age was classified as low, medium, high, and very high (Figure 3a), while the number of consultations was considered inadequate, regular, adequate, and extrapolated (Figure 3b), the NO₂ concentration was considered low risk, acceptable risk, and high risk (Figure 3c), and the PM₁₀ concentration as low risk, acceptable risk, and high risk (Figure 3d).

The parameters of the output variables were created by loading the system itself, totaling 192 responses. This time, the rules were given in 4 × 4 × 4 × 3 format, with 192 rules generated by the neuro-fuzzy ANFIS system.

Figure 4 illustrates part of the architectural scheme of the ANFIS neuro-fuzzy model created. In this case, only the variable “mother’s age” is represented. “Input” illustrates the input variable and “inputmf” is the membership functions, followed by the rules, output membership functions, defuzzification, and the model’s output variable.

2.3. Evaluation Metrics

Metrics such as the Mean Absolute Percentage Error (MAPE), Mean Square Error (MSE), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE) can measure the performance of models in various studies [36]. According to Garson, “Pearson’s correlation (r) is a bivariate measure of association of the degree of relationship between two variables” [37]. Ranging from −1 to 1, this metric reveals the strength of the relationship between the variables, which can be perfect, negative, positive, or without any relationship at all [38]. An excellent correlation value for the biological area is around 0.60, as it rarely reaches a value close to 1.00 [39]. The coefficient of determination (R²) is also used as an evaluation measure, mainly to explain the regression line. This coefficient is calculated by squaring Pearson’s correlation and a value close to 1 indicates a model with a good fit [40]. The formulas for calculating the correlation and coefficient of determination are represented in Equations (1) and (2), respectively.

$$r={\displaystyle \frac{\sum _{i=1}^n\left(x-\overline{x}\right)(y_i-\overline{y})}{\sqrt{\left[{\sum _{i=1}^n(x_i-\overline{x})}^2\right]\left[{\sum _{i=1}^n(y_i-\overline{y})}^2\right]}}}$$

(1)

where the definitions of the variables are the following:

• i is the period;
• n is the sample quantity;
• $x_i$ and $y_i$ represent the variables;
• $\overline{x}$ and $\overline{y}$ are the sample means.

$${\mathit{R}}^2={\mathit{r}}^2$$

(2)

3. Results

In total, information was used from 1,096 premature babies, 327 born in 2016, 355 in 2017, and 414 in 2018 in the city of São José dos Campos (SP), Brazil. It is worth noting that the data are just numbers, with no personal identification of pregnant women or newborns. The youngest mother found in the database was 12 years old, and the oldest was 45 years old, with a mean of 26.4 years and a standard deviation of 6.9. The number of visits ranged from none to fifteen, with an average of four visits throughout pregnancy and a standard deviation of 2.7. The minimum NO₂ concentration during the period was 33.2 μg/m³, while the maximum was 51.2 μg/m³, with a mean of 41 μg/m³ and a standard deviation of 3.9. The concentration of PM₁₀ varied between 18.2 and 29.7 μg/m³, with an average of 23.5 μg/m³ and a standard deviation of 2.3. The youngest live birth was born at 22 weeks’ gestation, while the oldest occurred at 36 weeks’ gestation, averaging 33.9 weeks with a standard deviation of 3.1.

The structures of each model are shown in Figure 5 for the fuzzy model and Figure 6 for the neuro-fuzzy ANFIS model. From the structure, the system inputs can be seen, as well as the type and number of the pertinence function (3 × 3 × 2 × 2 trapezoidal for the fuzzy model and 4 × 4 × 4 × 3 for the neuro-fuzzy ANFIS), inference method (Mamdani type 1 for fuzzy and Sugeno type 1 for neuro-fuzzy ANFIS), and system output (4 trapezoidal type for fuzzy and 192 constant variables for neuro-fuzzy ANFIS).

Tests were carried out in the rule viewer of the MATLAB R2022b^® software to verify the veracity of the systems. For both models, taking as an example a 16-year-old mother with eight consultations and exposure to an average of 47.69 μg/m³ of NO₂ and 23.10 μg/m³ of PM₁₀ throughout her pregnancy, we obtained a predicted result at 35 weeks of gestation, a value identical to the real one.

For the fuzzy model, the correlation (r) between the real data of weeks of gestation and the data predicted by the system was 0.20 and the coefficient of determination (R²) was 0.04. As for the neuro-fuzzy ANFIS model, r = 0.59 and R² = 0.35.

The evaluation metrics, Mean Absolute Percentage Error, Mean Square Error, Mean Absolute Error, and Root Mean Square Error, were obtained through the Deep Learning Toolbox in MATLAB R2022b^®. These values were calculated for each system and are presented in

Table 2. Evaluation metrics.

Metric	Fuzzy	Neuro-Fuzzy
RMSE	3.29	2.83
MSE	10.81	8.00
MAPE	6.67%	5.35%
MAE	2.01	1.70

.

4. Discussion

The best results are for the lowest error values (RMSE, MSE, MAPE, and MAE) and the highest for Pearson’s correlation and the coefficient of determination. The ANFIS neuro-fuzzy model presented more satisfactory results in all calculated metrics, in relation to the developed fuzzy model, thus demonstrating greater accuracy and consistency in prediction, important factors that can directly impact subsequent policies to reduce polluting gasses.

In the first metric calculated, Pearson’s correlation, the fuzzy model presented a value of 0.20, indicating a positive correlation, although weak in relation to the real values, while the neuro-fuzzy ANFIS model, r = 0.59, represents greater strength between the predicted and real data, with a value considered close to an optimum value for biological areas. In addition to this metric, per the values of the coefficient of determination, square root of the squared error, mean squared error, Mean Absolute Percentage Error, and Mean Absolute Error, the ANFIS neuro-fuzzy model also presented more robust results compared to the fuzzy model (RMSE fuzzy = 3.29, RMSE neuro-fuzzy = 2.83; MSE = 10.81, MSE neuro-fuzzy = 8.00; MAPE fuzzy = 6.67%, MAPE neuro-fuzzy = 5.35%; MAE fuzzy = 2.01, MAE neuro-fuzzy = 1.70; R² fuzzy = 0.04, and R² neuro-fuzzy = 0.35).

The superiority of the ANFIS neuro-fuzzy model can be explained by its learning capacity with neural networks combined with fuzzy inferences, contrasting with the fuzzy model as it contains a defined rule base, limiting its adaptation.

This study, in addition to exploring prematurity and the exposure of pregnant women to air pollutants in an innovative way, highlights the importance of choosing artificial intelligence, especially the neuro-fuzzy ANFIS with Takagi–Sugeno inference and fuzzy with Mamdani methods, in data analysis.

Finally, the methods used represent an advance in the analysis of data related to premature birth and can be of great use in aiding medical and environmental decision-making.

5. Limitations and Future Opportunities

Among the limitations encountered was the possibility of incorrectly filled in or entered data in the DATASUS system, or even the absence of data on the portal. The concentration of pollutants was considered homogeneous, since in the city of São José dos Campos (SP), during the study period, there were only two measuring stations. As a result of this factor, the data collected may not adequately represent the spatial distribution of gasses; in addition, pregnant women who live close to avenues or industrial areas could be exposed to higher levels of air pollution compared to those who live in other areas.

In addition, for this study it was considered that the pregnant women were in the municipality for the entire gestational period. Environmental factors, such as pollution by Particulate Matter and Nitrogen Dioxide, have a direct influence on human health, and this study proves their effect on premature births. The development of public strategies aimed at reducing pollutant gasses that can affect maternal and neonatal health is recommended. The development of a system capable of verifying the incidence of premature births according to the mother’s age, number of prenatal consultations, and exposure to NO₂ and PM₁₀, besides being unprecedented in the literature, is of great value to society, as it can potentially promote improvements in the health of pregnant women and newborns.

Opportunities for future studies were found, such as the use of different validation techniques for the ANFIS neuro-fuzzy system, including cross-validation, hold-out, or even 70:30 validation. Furthermore, the fuzzy-2 or fuzzy type-3 techniques may be more effective tools for studying the association between premature birth and air pollutants.

In addition, there is an opportunity to explore new input variables, such as temperature, given the impacts that global warming and climate change have on human life. Furthermore, new geographic regions could be the target of studies.