nueva página del texto (beta)
Artículo en XML
Referencias del artículo
Traducción automática
Enviar artículo por email
Citado por SciELO 
Similares en
SciELO 
Permalink
1 Introduction
Non-communicable diseases (NCDs) are not spread through infection or other people but are typically caused by unhealthy behaviors. In the last two decades, there has been an increase in the incidence of NCDs in most countries of the world. According to the World Health Organization (WHO), 70% of the 56.4 million deaths in 2016 were caused by NCDs. Furthermore, 75% of NCD-associated deaths occurred in low- and middle-income countries, reflecting the seriousness of the problem in countries such as Mexico [21]. Among NCDs, chronic vascular diseases are the leading cause of death in the world, including Systemic Arterial Hypertension (SAH), Ischemic Heart Disease, Heart Failure, Degenerative Calcific Aortic Valve Stenosis, and Congenital Heart Disease [21].
Among these diseases, SAH has shown an accelerated epidemiological change [22]. In Mexico, approximately 450 thousand new cases of SAH are diagnosed annually.
In the past two decades, it has remained among the first nine causes of death in Mexico, and in the past six years, the mortality rate from this cause has increased by 29.9% [22].
In addition, Mexico has the highest prevalence of arterial hypertension in the world, according to the National Survey of Health and Nutrition 2012.
This paper proposes to collect data from public records provided by official sources to concentrate them in a database. The data obtained allowed the generation of a predictive model to analyze the behavior of the disease (SAH) in a period of one year using artificial intelligence tools.
Through the use of data mining techniques, the system can also establish a relationship between variables such as socioeconomic status (SES) with SAH. The impact of SES on hypertension has been reported in several studies with conflicting results [15, 17].
2 Data Analysis and Hypertension
BNP-PL registry is the first multicenter and prospective registry of adult and pediatric patients with PAH and CTEPH created in any of the Central-Eastern European countries [14]. In Greece, a large-scale database of ageing participants was integrated to evaluate the long-term hypertension risk of men and women aged older than 50 years with Machine Learning [8].
In Latin America, Cuba used their data to design a predictive model for the diagnosis of arterial hypertension in adulthood from records taken from adolescence; Perez et al. [18] obtained an effectiveness of 80%, and Gonzalez et al. [2] analyzed the importance of data engineering and obtained a methodology that delivers more effective results 88.41% success in grouping instances. They also conclude that the risk of becoming hypertensive in the trajectory from adolescence to adulthood is mainly due to birth and family factors.
Machine learning classification techniques have been proposed to help healthcare professionals in diagnosing heart disease. Some works focus on the identification of risk patterns associated with hypertension.
In this sense, [19] uses Statistical Tree algorithm, Classification, and Regression, and [20] uses algorithms like K-star, J48, SMO, Naïve Bayes, MLP, Random Forest, Bayes Net, and REPTREE.
It is also important to study the correlation of the disease with other factors. For example with demographic elements [4], Body Mass Index (BMI) [1], race [17], socioeconomic status (SES) [15, 17, 10], or COVID [24]. For this study, we focus on SES.
2.1 Analysis of Hypertension in Mexico
The Center for Research in Nutrition and Health and the National Institute of Public Health describes the prevalence, distribution, and trends of arterial hypertension (HTN) in Mexican adults aged 20 years or older, considering 10,898 adults.
The project performed a time trend analysis using statistical analysis based on the data reported in the National Health Survey (2000), the National Health and Nutrition Survey (2006), and the National Health and Nutrition Survey (2012). They found stabilization of the problem between 2000, 2006, and 2012, but better control of the disease is required [5].
A comparative cross-sectional study was carried out in 627 rural communities of Durango, Mexico, and determined blood pressure figures and sociodemographic variables through the IMSS-Solidarity program. For the statistical analysis, the confidence interval was calculated in the binomial distribution for the dichotomous variables (prevalence estimate).
For nominal variables, the
The solution designed in this document works with a larger number of samples collected over a longer time interval and considers using data mining tools.
3 Methodology: Data Mining Approach with Hypertension Data
This section describes the different stages of the proposed solution: data collection and analysis, pre-processing, data mining application, and data interpretation. The proposed project includes data on hypertension and socioeconomic indicators. The information on each of these variables was subjected to the following stages: data collection, data selection, data pre-processing, application of data mining, and data interpretation, as we can see in Figure 1.
3.1 Data Collection and Selection for HBP
We use data about hypertension and socioeconomic status. For the first one, the Mexican government keeps track of patients with hypertension; and for the second one, data was collected from Instituto Nacional de Estadística y Geografía (INEGI) [11] and Open Data of Mexican government [6]. The original data was presented in multiple formats, with different characteristics, and in some cases present inconsistencies. The data collection period comprises from August 2019 to May 2020. Databases were consulted and collected from the official sites of ENSANUT, IMSS, ISSSTE, INEGI, and the Health Sector.
These data was available to the Mexican population in files with different formats, distributed in portals of each institution. The selection considered mainly the completeness of data. The final data set considers data from the institutions of ENSANUT, IMSS, INEGI, and the Health Sector. A deeper description of the data set in presented in Section 4.
3.2 Preprocessing
First, data exploration of the datasets was performed. Then an iterative data cleaning process was executed using data reduction and data transformation in order to generate an integral data set.
Data wrangling was performed, it allowed to transform data into a unified format. This process reduced the dimensionality of each database. Then, we looked for null and empty data that is presented in 4. If a column has more than 30% of null or empty data records, we drop it. Some other inconsistencies as dates and different indicators for the same thing were solved. A unification and transformation of data types was also required.
The preprocessing stage was applied separately to the records corresponding to Hypertension, Nutrition, and Socioeconomic level. Then, integration stage was executed, joining the datasets of the same institution with data transformation in some cases looking for the consistency of the data.
3.3 Forecast and Correlation with Socioeconomic Factors
A pattern means that the data are correlated, have a relationship, or are predictable. Data mining builds models to identify patterns among the variables in a data set. Some of these patterns are predictive (projecting future values), whereas others are explanatory (explaining the interrelationships among the variables) [7]. Using the created data set, we identify predictive and explanatory patterns. We predict the number of cases for subsequent years, and we study which socioeconomic indicators are more related to the registered hypertension cases in Mexico.
3.3.1 Forecast of the Number of Cases
Using the historical data since 2000, we predict the years 2019 and 2020 for each estate of Mexico and for the total number of detected cases in Mexico. We compare ARIMA [16] and CNN-QR [3] to forecast the number of hypertension cases.
Time series analysis can be viewed as the task of detecting patterns in data and predictability of values. Pattern detection can include: trend, cyclical, periodic, or outliers [9]. ARIMA obtains the representation of the series in terms of the temporal interrelation of its elements, characterizing the series as sums or differences, weighted or not, of random variables or the resulting series.
The fundamental element when analyzing the properties of a time series is the autocorrelation coefficient, which measures the degree of linear association that exists between observations separated by k periods. The ARIMA model is commonly used in infectious disease time series prediction[23], and recently used for the prediction of COVID-19 cases [12].
Convolutional Neural Networks (CNN) are the evolution of Artificial Neural Networks, they use supervised learning that processes their layers to identify different characteristics in the inputs to identify objects. The CNN contains several specialized hidden layers with a hierarchy that become more specialized until reaching deeper layers.
Using causal CNN, the Convolutional Neural Network - Quantal Regression (CNN-QR) is an algorithm for predicting scalar time series. This supervised learning algorithm trains a collection of time series and uses a quantal decoder to make probabilistic predictions.
3.3.2 Correlation of BPH with Socioeconomic Factors
We used clustering for the purpose of analyzing the socioeconomic characteristics associated with the HBP cases detected. Clustering is used to group existing data whose common characteristics are unknown or are to be discovered.
The clusters generated are based on the similarity between states of the number of the proportion of people (since they are normalized data) that are part of these indicators. Two clustering algorithms were tested: k-means and hierarchical clustering.
k-means groups objects into k groups depending on their characteristics. k-means clustering is an unsupervised learning technique to classify unlabeled data by grouping them by features, rather than pre-defined categories. The variable k represents the number of groups or categories created.
The goal is to split the data into k different clusters and report the location of the center of mass for each cluster. Then, a new data point can be assigned a cluster (class) based on the closed center of mass. The main advantage is that the machine creates its own clusters based upon empirical proofs, rather than assumptions.
Hierarchical Clustering (HCA) constructs a tree that represents the similar relationships between the different elements. The exploration of all possible trees is computationally intractable. We used agglomerative hierarchical clustering, it starts with as many clusters as there are individuals and consists of forming groups according to their similarity.
According to the selected IMSS data, there are 46 different socioeconomic indicators. Let
Let the cluster
4 Generated Data Set
We cover from the year 2000 to 2019, containing data related to Hypertension, nutrition, and socioeconomic indicators. Those repositories that contemplated less than three years of registration were eliminated since, based on the surveys, information was collected at intervals of two years. Finally, the analysis set was reduced to 46 databases and similar databases were integrated in 27 final CSV files.
The data was explored in search of cases such as repeated records, incomplete records, repeated fields, and fields with invalid data (null) or unintelligible content. Table 2 shows the numer of NaN and Nulls detected in the HAS datasets. To unify the information, the types of data used in different databases for fields with the information of the same type were identified, for example, for fields such as date were used, data types such as:
Table 1 Evaluation of WAPE and RMSE from 1 to 5 years prediction
| Number of Predicted Years | WAPE | RMSE |
| 2019 (1) | 0.0825 | 150505.70 |
| 2020 (2) | 0.1029 | 213206.77 |
| 2021 (3) | 0.1517 | 331492.10 |
| 2023 (5) | 0.1736 | 356471.12 |
Table 2 Numer of null and NaN
| DataSet | No. of cells | No. of NaN | No. of NULL |
| Ensanut (Economy) |
22,487,304 | 224,198 | 3,699,427 |
| Ensanut (Food) |
1,556,334 | 25,476 | 30,915 |
| Ensanut (Activity) |
1,447,800 | 453,171 | 96,923 |
| Ensanut (Teenagers) |
1,261,248 | 0 | 661,073 |
| Ensanut (Adults) |
1,596,425 | 108,310 | 756,868 |
| INEGI (Deaths) |
8,745,373 | 0 | 0 |
| Health Sector (HAS-Obesity) |
37,510,485 | 0 | 14,580,308 |
| IMSSS (HAS) |
756 | 0 | 0 |
To get data integrity, some values were modified; for example, in the databases of ENSANUT, there are interviews of patients, they used the value ”0” for negative responses in most of the cases but in 2016 and 2018 they used ”2”. Also, the periodicity of the records was evaluated to enter into the final system only those databases whose records coincided in the periods. For data integration, the datasets of the same institution were integrated as follows:
– From ENSANUT, the datasets of Surveys of Tennagers and Adults were integrated by years.
– From ENSANUT, the datases about Anthropometry Adults, Frequency of Adult Consumption, Food Safety, Information about the Household, and Information about the members of the household were joined considering a common number (folio) of register.
– From INEGI, we only consider the registers with direct cause of death of cardiac problems, the datasets of each year were integrated in one and a column of ”year of registration” was added.
– From the socioeconomic data from datos.gob, a manual exploration of 27 CSV files was performed, we integrated and merged the databases considereing to regional (municipality, state) aspect.
Then the data set includes the tables:
– IMSS with: year, state, and number of cases.
– INEGI-deaths with: registration entity, living entity, municipality, number of cases.
– INEGI with: registration entity, age, municipality of registration, detailed cause of death, mexican list of cause of death, sex, day of registration, month of registration, year of registration, occupation, escolarity, CIE detailed cause by chapter, CIE detailed cause by group, list of tabulation for CIE mortality, mexican list of disease.
– Socioeconomic-datos.gob with categories (with
– ENSANUT-Nutrition with: entity, sex, age, weight, high, waist circumference, IMC, category of IMC, worry about food at home, home without food, homer without healthy food, low variety of food, one time food missing, eat less than you should, hungry, home with persons with less that 18 years, detected HAS, house construction material, house floor construction material, kitchen room, sleeping rooms, numer of rooms, electric light, water, number of days with eater, type of toilet, share bathroom, toilet flush type, cooking fuel, trash, owner, construct, were to cook, stove type, heating, number of bulb, more houses, car, truck, moto, other, TV, payed TV, radio, consols, iron, blender, refrigetaron, washing machine, computer, internet, microwave, phone, water tank, cistern, light, celphone, air conditioner, native, native language, school grade, civil status, number of persons at home, and share expenses of food.
Subsequently, the final representation of the data was generated in JSON, which is a lightweight data exchange format completely independent of language. The new representation allows hierarchical visualization and makes the data easier to read. This representation was chosen in order to speed up computational processes.
4.1 Forecast and Correlation with Socioeconomic Factors
A pattern means that the data are correlated, have a relationship, or are predictable. Data mining builds models to identify patterns among the variables in a data set. Some of these patterns are predictive (projecting future values), whereas others are explanatory (explaining the interrelationships among the variables) [7].
Using the created data set, we identify predictive and explanatory patterns. We predict the number of cases for subsequent years, and we study which socioeconomic indicators are more related to the registered hypertension cases in Mexico.
5 Evaluation of Patterns in Hypertension Data
There are two main patterns presented in this paper. The first one is focused on the forecast of cases of hypertension with two models that are evaluated with RMSE and WAPE. The second one evaluates the correlation of BPH with socioeconomic indicators using clustering (k-means and HCA) and identifying the indicators with the highest weight for the cluster with more HAS; we also used Pearson Correlation Coefficient to identify the principal indicators in order to compare them with the results given by the clusters using Recall.
5.1 Metrics
Two analysis are presented. To evaluate the prediction of cases we used RMSE and WAPE. To evaluate the correlation of BPH with socioeconomic factors we use recall and PCC.
5.1.1 Root Mean Square Error (RMSE)
It is a standard way to measure the error of a model in predicting quantitative data. It compares the observed value
This metric uses the squared value of residuals, amplifying the impact of outliers, then:
Such that
5.1.2 WAPE o Weighted Absolute Percentage Error
It is one of the most common methods to measure the overall deviation of forecast values from observed values. It measures the percentage error of the sum of the observed values versus the sum of predicted values. A lower value indicates a more accurate model. Let
5.1.3 Recall
The completeness metric will inform us of the number of samples that the machine learning model can identify. For this paper, recall refers to the percentage of important indicators we identify:
where,
5.1.4 Pearson Correlation Coefficient (PCC)
It is a measure of linear correlation between two sets of data. By definition PCC
where,
where
5.2 Evaluation of the Prediction of the Number of Cases of BPH
To develop the prediction of the number of HBP cases, several tests were performed by generating the models, as it is known efficiency decreased as the prediction range increased (backtest window). The predictions made at one year showed the lowest WAPE and RMSE, while for predictions of 2, 3, and 5 years they increased.
Table 1 shows this behavior and its error parameters. WAPE is more robust to outliers than Root Mean Square Error (RMSE) because it uses the absolute error instead of the squared error.
For the evaluation, a one-year backtest window was performed, i.e., the algorithm predicted 2019, which is the year for which data is available.
It should be noted that the same tests were performed for the prediction of 2020; however, since it was a pandemic year, the cases of hypertension recorded were below those recorded 20 years ago; since it was an atypical year, the data obtained from official sources were not consistent.
We compare two algorithms: Mean ARIMA and CNN-QR (Convolutional Neural Network - Quantile Regression). The results given by ARIMA were appropriate compared with some others; nevertheless, CNN-QR showed better results, as we can see in Table 3.
5.3 Evaluation of the Correlation of BPH with Socioeconomic Indicators
In order to compare the number of HBP cases with socioeconomic indicators, the required data sets were normalized. We used clustering to identify the cluster where hypertension is more important (highest weight). Then, we identify the socioeconomic indicators with the highest presence in the same cluster, such that those indicators are associated with hypertension as it was explained in Section 3.3.2.
We evaluated two clustering methods: Hierarchical Clustering (HCA) and k-means. Let
The input vector for each observation has 47 variables, such that
We compare the indicators
For the HCA the distances between each observation in the hierarchical grouping, defined by the linkage matrix, were calculated and the correlation of distances were obtained to determine the accuracy. While for K-means, the sum of squared distances from the center to the nearest cluster were calculated, which can be interpreted as the error range of the same model. The accuracy of hierarchical clustering is measured through the generated distance matrix and the original data, so the variation in the number of clusters did not greatly affect the result. Within these tests, it was decided to use five clusters for both groupings.
The expected value for hierarchical clustering accuracy is 1, while for K-means error it is zero (it is important to remember that the latter is a sum of squared distances). The accuracy obtained was above 0.75 (75% success rate), i.e., for the 32 data vectors entered corresponding to each state, at least 24 were correctly clustered according to their socioeconomic characteristics. The classification results are presented in Table 4.
Table 4 Clustering algorithms precisión evaluated by the the error range for k-means and by the accuracy given by correlation of distances for HCA
| Precision | ||
| Year Data Set | k-means Error | HCA Accuracy |
| 2008 | 11769.69 | 0.7854 |
| 2010 | 10436.07 | 0.7902 |
| 2012 | 9526.41 | 0.7658 |
| 2014 | 9991.22 | 0.7684 |
| 2016 | 8348.98 | 0.7925 |
| 2018 | 9780.10 | 0.7932 |
Once the clusters are obtained, we evaluated the output vectors
Let
Table 5 Recall of socioeconomic indicators in the same cluster (two cluster algorithms) compared with PCC
| Recall | ||
| Year Data Set | k-means | HCA |
| 2008 | 0.9166 | 0.8888 |
| 2010 | 0.9000 | 0.8750 |
| 2012 | 0.6000 | 0.8888 |
| 2014 | 0.9090 | 1.0000 |
| 2016 | 0.4444 | 1.0000 |
| 2018 | 0.8888 | 0.7500 |
Some of the most frequent socioeconomic indicators correlated with BPH are population vulnerable by income, non-poor and non-vulnerable population, contribution of educational backwardness to poverty, contribution of access to health care to poverty, contribution of access to social security to poverty, contribution of access to food to poverty, population affiliated to IMSS, population with private medical insurance, population with access to indirect social security health services and population food insecurity.
6 Discussion and Further Analysis
Data were collected, integrated, normalized, and preprocessed by the team. This resulted in an unpublished dataset, so the results of socioeconomic factors correlated with hypertension were validated using a ranking derived from the Pearson coefficient through recall. There are some other socioeconomic indicators that are not predominant in the cluster with the highest registered hypertension, like the population in poverty, with educational lagging, and without access to social security. It was deduced that hypertension itself could be associated with these factors; however, the databases considered are those of registered cases of hypertension.
Therefore, in people with the aforementioned indicators, the disease has not yet been identified, but that does not mean that they do not have it. An early diagnosis of this type of disease is vital; people with deficiencies often do not have access to medical services and therefore are not detected in time.
The final step of the data mining process is the data representation. Some graphic patterns were generated. Figure 2 shows the correlation of the number of BPH cases and two indicators. Each point represents a state and a year located on the plane that correlates the number of hypertension cases and the indicator.
Once we are evaluating the number of cases of BPH in Mexico for each state, Figure 3 shows the country; darker states have the highest register of BPH cases, like Mexico City in the middle of the map and Nuevo Leon in the northeast.
The generated data set could also be used to analyze some correlations inside the socioeconomic factors; for example, Figure 4 shows two evident clusters if we correlate vulnerable population by income and poor population for access to basic housing services.
7 Conclusions
The initial phase of this project facilitated the creation of a comprehensive database consolidating data from diverse public sources on High Blood Pressure (HBP or hypertension), including socioeconomic indicators of the patients. The collected data underwent a meticulous process of organization and standardization, resulting in the establishment of a robust data repository.
The standardized data enabled the prediction of registered cases of high blood pressure and an assessment of the potential influence of factors not traditionally considered by medical professionals in the development of this disease. The algorithm takes into account variables such as food insecurity, susceptibility due to income inadequacy, and affiliation to health services.
Drawing insights from information spanning the past 10 years, a projection was generated concerning the trajectory of HTN diagnoses, exhibiting promising outcomes until 2019. However, the results for 2020 were compromised due to the lack of records during the COVID-19 pandemic.
