ABSTRACT

OBJECTIVE

This study aims to integrate the concepts of planetary health and big data into the Donabedian model to evaluate the Brazilian dengue control program in the state of São Paulo.

METHODS

Data science methods were used to integrate and analyze dengue-related data, adding context to the structure and outcome components of the Donabedian model. This data, considering the period from 2010 to 2019, was collected from sources such as Department of Informatics of the Unified Health System (DATASUS), the Brazilian Institute of Geography and Statistics (IBGE), WorldClim, and MapBiomas. These data were integrated into a Data Warehouse. K-means algorithm was used to identify groups with similar contexts. Then, statistical analyses and spatial visualizations of the groups were performed, considering socioeconomic and demographic variables, soil, health structure, and dengue cases.

OUTCOMES

Using climate variables, the K-means algorithm identified four groups of municipalities with similar characteristics. The comparison of their indicators revealed certain patterns in the municipalities with the worst performance in terms of dengue case outcomes. Although presenting better economic conditions, these municipalities held a lower average number of community healthcare agents and basic health units per inhabitant. Thus, economic conditions did not reflect better health structure among the three studied indicators. Another characteristic of these municipalities is urbanization. The worst performing municipalities presented a higher rate of urban population and human activity related to urbanization.

CONCLUSIONS

This methodology identified important deficiencies in the implementation of the dengue control program in the state of São Paulo. The integration of several databases and the use of Data Science methods allowed the evaluation of the program on a large scale, considering the context in which activities are conducted. These data can be used by the public administration to plan actions and invest according to the deficiencies of each location.

Epidemiology; Dengue; Outcome and Process Assessment, Health Care; Data Science

INTRODUCTION

The number of dengue cases and deaths has been increasing every year, surpassing historical records. According to the Brazillian Ministry of Health Bulletin for the 52nd epidemiological week of 2022¹1. Ministério da Saúde (BR). Secretaria de Vigilância em Saúde. Monitoramento dos casos de arboviroses até a semana epidemiológica 51 de 2022. Brasília, DF: Ministério da Saúde; 2022., 992 confirmed dengue deaths were reported, a higher number than in 2015 (986), when the highest number had been recorded so far²2. Preite Sobrinho W. Mortes por dengue disparam em 2022 e batem recorde histórico no Brasil. Portal UOL. 2022 dec 28.. According to the same bulletin, the state of São Paulo presented the highest number of deaths that year (278), about 28% of the total registered in the country.

In Brazil, vector control of Aedes began in 1976 with sporadic actions, and since 2002, with the creation of the Programa Nacional de Controle da Dengue (PNCD – Brazilian Dengue Control Program), routine activities have been developed to combat the vector. Some of these activities, such as larval sampling research in households, visits to strategic points, home visits aimed at guiding the residents and eliminating immature forms of the vector by implementing transmission blocking agents when necessary, among others, are routinely developed in the state of São Paulo, following national guidelines.

The dengue-transmitting vector (the Aedes aegypti mosquito) finds favorable conditions for its development in places with higher temperatures and highly urbanized environments. According to projections, these places are likely to increase in the coming years³3. Messina JP, Brady OJ, Golding N, Kraemer MU, Wint GR, Ray SE, et al. The current and future global distribution and population at risk of dengue. Nat Microbiol. 2019 Sep;4(9):1508-15. https://doi.org/10.1038/s41564-019-0476-8
https://doi.org/10.1038/s41564-019-0476-... . Given the strong influence of changes in the environment for diseases such as dengue⁴4. Marinho RD, Duro RL, Mota MT, Hunter J, Diaz RS, Kawakubo FS, et al. Environmental changes and the impact on the human infections by dengue, chikungunya and zika viruses in Northern Brazil, 2010-2019. Int J Environ Res Public Health. 2022 Oct;19(19):12665. https://doi.org/10.3390/ijerph191912665
https://doi.org/10.3390/ijerph191912665... , climate change may contribute to the increased proliferation of vectors. Urbanization without adequate planning represents another threat to several infectious diseases⁵5. Neiderud CJ. How urbanization affects the epidemiology of emerging infectious diseases. Infect Ecol Epidemiol. 2015 Jun;5(1):27060. https://doi.org/10.3402/iee.v5.27060
https://doi.org/10.3402/iee.v5.27060... , and, with the worsening of living conditions, the migratory flow may increase even more. Over the past decade, an average of 24 million people per year have been displaced from their homes due to climate-related events⁶6. Adger WN, Crépin AS, Folke C, Ospina D, Chapin FS 3rd, Segerson K, et al. Urbanization, migration, and adaptation to climate change. One Earth. 2020 Oct;3(4):396-9. https://doi.org/10.1016/j.oneear.2020.09.016
https://doi.org/10.1016/j.oneear.2020.09... .

Although urbanization can be seen as a reflection of economic development, it does not always occur adequately⁷7. Burmann T, Censi D. Urbanização e cidadania: conflitos e contradições da cidade contemporânea. Revista Direito UFMS. 2017;3(2):85-102. https://doi.org/10.21671/rdufms.v3i2.5006
https://doi.org/10.21671/rdufms.v3i2.500... . The destruction of ecosystems and the disorderly occupation of spaces cause several impacts on the living conditions of the population⁷7. Burmann T, Censi D. Urbanização e cidadania: conflitos e contradições da cidade contemporânea. Revista Direito UFMS. 2017;3(2):85-102. https://doi.org/10.21671/rdufms.v3i2.5006
https://doi.org/10.21671/rdufms.v3i2.500... . While some indicators have improved, such as life expectancy, others have worsened, such as the frequency of disease outbreaks⁸8. Soares JAS, Alencar LD, Cavalcante LPS, Alencar LD. Impactos da urbanização desordenada na saúde pública: leptospirose e infraestrutura urbana. Polêm!ca. 2014;13(1):1006-20..

This complex relationship between human health and environmental conditions is the subject of a study in the area called Planetary Health⁹9. Whitmee S, Haines A, Beyrer C, Boltz F, Capon AG, Dias BFS, et al. Safeguarding human health in the Anthropocene epoch: report of The Rockefeller Foundation-Lancet Commission on planetary health. Lancet. 2015 Nov;386(10007):1973-2028. https://doi.org/10.1016/S0140-6736 (15)60901-1
https://doi.org/10.1016/S0140-6736 (15)6... , in which research seeks to assess the impact of the environment on health. Health conditions, which are often a reflection of human activity, are related to this context. Dengue cases are, among other factors, related to climatic conditions¹⁰10. Morin CW, Comrie AC, Ernst K. Climate and dengue transmission: evidence and implications. Environ Health Perspect. 2013;121(11-12):1264-72. https://doi.org/10.1289/ehp.1306556
https://doi.org/10.1289/ehp.1306556... that, in turn, can be related to human activity⁹9. Whitmee S, Haines A, Beyrer C, Boltz F, Capon AG, Dias BFS, et al. Safeguarding human health in the Anthropocene epoch: report of The Rockefeller Foundation-Lancet Commission on planetary health. Lancet. 2015 Nov;386(10007):1973-2028. https://doi.org/10.1016/S0140-6736 (15)60901-1
https://doi.org/10.1016/S0140-6736 (15)6... .

Thus, the assessment of health conditions may be related to the context in which the activities are performed. The Donabedian model¹¹11. Donabedian A. The definition of quality and approaches to its assessments. Ann Arbor: Health Administration Press, 1980. (Explorations in quality assessment and monitoring, v. 1). considers that health outcomes depend on processes and structure. In this model, evaluations can include analyses of three components that are related to each other: structure, process, and outcome.

Evaluating the indicators of the structure component allows us to analyze the situation of the health service structure in the area: the number of professionals, care units, investments, among others. Process indicators, in turn, evaluate whether care processes are being performed appropriately, for example. These two components can be evaluated in conjunction with outcome indicators, allowing for an integrated evaluation of the conditions under which health outcomes are produced.

However, the conditions and actions required for health outcomes may also depend on the context in which they are inserted¹²12. Contandriopoulos AP, Champagne F, Denis JL, Pineault R. A avaliação na área da saúde: conceitos e métodos. In: Hartz ZMA, org. Avaliação em saúde: dos modelos conceituais à prática na análise da implantação de programas. Rio de Janeiro: Fiocruz; 2000. p. 29-47.. Thus, health evaluations should also consider context-related variables.

Considering the complex relationship between health determinants discussed in the Planetary Health research area, health evaluation methods, such as the Donabedian model, could use data from different dimensions, such as the economy, living conditions, and urban development. However, the increasing complexity of the health evaluation process can hinder this process due to cost and time constraints, especially in the context of Big Data, in which multiple sources of data are generated at ever-increasing speeds.

In this sense, by adopting Data Science techniques, such as the application of machine learning algorithms, and statistical analysis, this study aims to aggregate context in the evaluation of the Brazilian Dengue Control Program based on the Donabedian model. Given the diversity of variables related to dengue, the use of Data Science can bring great benefits to the subject of health evaluation, such as allowing evaluation processes to be performed on a large scale. In this study, data from several variables of the 645 municipalities in the state of São Paulo, from 2010 to 2019, were analyzed.

METHODS

Research based on Data Science considers, in addition to computational and statistical techniques, the inclusion of domain experts¹³13. Viaene S. Data scientists aren't domain experts. IT Prof. 2013 Nov;15(6):12-7. https://doi.org/10.1109/MITP.2013.93
https://doi.org/10.1109/MITP.2013.93... . This interaction adds the vision/human insight of a field to the research, including identifying the necessary data sources, defining the problem, and evaluating the outcomes. Thus, this study is divided into six stages, described below, from the definition of the evaluation model to the analysis of the outcomes.

Definition of the evaluation model

The main indicators for the evaluation of the Brazilian Dengue Combat Program were identified. This process relied on existing literature to add formal evidence for the use of the indicators¹⁴14. Bicalho D, Santos TS, Slater B, Lima TM. Evaluation of quality indicators for management of the National School Feeding Program in Brazil: a systematic review. Cien Saude Colet. 2021 ago;26(8):3099-110. https://doi.org/10.1590/1413-81232021268.03802020.
https://doi.org/10.1590/1413-81232021268... , as well as validation by domain experts in the matrix shown in Chart 1. For this research, indicators related to the structure and outcome components of the Donabedian model were used, in addition to those related to the context.

The evaluation model for this research contains 19 indicators, 14 of which relate to context, three to structure, and two to outcomes (Table 1). The indicators cover six areas: climate (three indicators), socioeconomic (three), sanitation (two), demographics (one), land use (four), and health (six). The health indicators correspond mainly to the two components of the Donabedian model evaluated in this study (structure and outcomes). Infant mortality was adopted as a context indicator, as this index can be used to assess the living conditions of the population¹⁵15. Reidpath DD. Infant mortality rate as an indicator of population health. J Epidemiol Community Health (1978). 2003 May;57(5):344-6. https://doi.org//10.1136/jech.57.5.344
https://doi.org//10.1136/jech.57.5.344... .

Regarding the two outcome indicators, the guidelines from the Government of the State of São Paulo were used as references, which define an annual incidence coefficient according to the population size of the municipality¹⁶16. Secretaria de Estado da Saúde (SP). Diretrizes para a prevenção e controle das arboviroses urbanas no estado de São Paulo. São Paulo: Secretaria de Estado da Saúde; 2017.. This coefficient is used to classify municipalities according to the history of dengue transmission, a method that defines which means is used to assess the situation (histogram or control diagram).

Data collection and storage

After defining the evaluation model, data were collected from public data sources (Chart). All data for the period from 2010 to 2019 were available except for climatic variables (2010 to 2018), HDI (2010 only), and Gini Index (2010 only).

The datasets were stored in their original forms on a cloud platform. The only modifications to data in this stage were related to climate data, which was only available in raster format. For this data, we extracted the mean temperature and precipitation values for each municipality, which were stored in comma-separated values (csv) format.

Data Warehouse Creation

To perform the analyses, the data were integrated into a Data Warehouse (DW), a form of storage prepared for analytical operations. The operations are usually conducted using the extract-transform-load (ETL) model, in which the data is extracted from its original sources, transformed for necessary adjustments, such as standardization, and loaded into the DW for further analysis.

For this stage, Python scripts were developed and run in a cloud environment for each data set. The standard procedure for all scripts consisted of the following stages: collecting the dataset file from the cloud, performing the necessary modifications, and loading the data into the DW.

To integrate the data, the Brazilian Institute of Geography and Statistics (IBGE) code was used to identify each municipality since differences in the spelling of municipal names were identified between data sources. Even for this code, differences were found in the data sources, with some reporting the code in six digits and others in seven. These differences in data sources can often hinder data integration processes, making it necessary to develop automated procedures to transform them.

After the appropriate modifications, the data was loaded into a database hosted on the Google BigQuery, a cloud analytics platform that provides access to queries and machine learning algorithms.

Data selection and preparation

To perform the analyses, automated queries were performed to extract the data stored in the DW created in the previous stage. With the collection, the following actions were performed:

• Standardization: as the variables are available in absolute values, operations were performed to standardize them in terms of the population;

• Calculation of incidence: the annual incidence of dengue cases was calculated for each municipality;

• Evaluation of the municipalities: for each year of the study period, the municipalities were evaluated according to two criteria: 1) whether the incidence rate was equal to or greater than 20% of the expected value for their population size¹⁶16. Secretaria de Estado da Saúde (SP). Diretrizes para a prevenção e controle das arboviroses urbanas no estado de São Paulo. São Paulo: Secretaria de Estado da Saúde; 2017.; and 2) whether the incidence rate was equal to or greater than 100% of the expected value for their population size¹⁶16. Secretaria de Estado da Saúde (SP). Diretrizes para a prevenção e controle das arboviroses urbanas no estado de São Paulo. São Paulo: Secretaria de Estado da Saúde; 2017.;

• Classification of municipalities: an indicator was created related to the total number of years for each criterion previously described. Each municipality was evaluated on both indicators according to two classes: 1) the class that indicates that the municipality recorded less than five years of incidence below the criterion; and 2) five or more years of incidence above the adopted criterion. The indicators generated were named coefficient20 and coefficient100 for criterion A and B, respectively.

The adoption of this classification criterion is based on the guidelines of the state of São Paulo to control dengue. Municipalities with no historical series are those whose incidence was 20% lower than expected according to population size, in at least five of the last ten years. Thus, these municipalities could perform better than those that have had an incidence higher than this threshold for more than five years. The second criterion (above or below 100%) was defined as an additional ranking metric, that is, to verify those municipalities that exceeded this expected incidence by at least five years.

Data Analyses

Initially, statistical analyses were performed according to the two outcome indicators. The mean values of each context and structure indicator were calculated for each class to verify possible patterns.

Then, the K-Means algorithm was used to identify the groups according to the climatic conditions. The characteristic of this algorithm is the previous definition of the number of groups to be created¹⁷17. Sinaga KP, Yang MS. Unsupervised K-means clustering algorithm. IEEE Access. 2020;8:80716-27. https://doi.org/10.1109/ACCESS.2020.2988796
https://doi.org/10.1109/ACCESS.2020.2988... . According to the previous evaluation and with the aid of the Elbow method¹⁸18. Syakur MA, Khotimah BK, Rochman EM, Satoto BD. Integration K-means clustering method and elbow method for identification of the best customer profile cluster. IOP Conf Ser Mater Sci Eng. 2018 abr;336:012017. https://doi.org/10.1088/1757-899X/336/1/012017
https://doi.org/10.1088/1757-899X/336/1/... , it was defined that there would be four groups. As climatic conditions greatly influence the proliferation of mosquitoes, we sought to create groups to make comparisons in municipalities with similar climatic conditions. An advantage of this method is that analyses can be performed on large volumes of data.

Once the groups had been created, statistical analyses were conducted for each one, estimating mean values for each indicator within the class of outcome indicators. The objective of this stage is to verify possible differences in context and structure in relation to the performance of municipalities in dengue control.

Outcome Analyses

Finally, the outcomes were evaluated for the extraction of relevant information. In addition to the statistical analyses, spatial visualizations of the information were performed to identify possible patterns.

OUTCOMES

General Analysis

Table 1 shows the mean values for each indicator from 2010 to 2019 for the two classes of municipalities. With these data, some important differences between the two classes of municipalities can be observed.

For the coefficient100 indicator, it is noted that the municipalities of class 2 (worst performance) hold, on average, better socioeconomic indicators (HDI, GDP, and GDP per capita). In addition, these municipalities present lower infant mortality rates and better sanitation conditions. Despite this, it is observed that they have more favorable conditions for the proliferation of mosquitoes, as they show a larger urban population and higher temperatures. Only these factors could explain the worse performance of class 2 municipalities in relation to this outcome indicator.

However, despite showing better socio-economic indicators and more favorable conditions, these municipalities presented lower numbers in the three indicators related to health structure per capita: number of community health agents, number of basic health units, and investments in health. The same pattern is observed in the analysis of the second outcome indicator (coefficient20), except for indicators related to the annual precipitation volume and investments in health per capita. In this case, the worst-performing municipalities (class 2) presented, on average, lower annual precipitation volume and higher investments in health per capita.

Intergroup analysis

Then, the municipalities were grouped according to their climatic conditions. According to Figure 1, it is possible to see a regional pattern between the groups, with Group 0 corresponding mainly to municipalities in the central region of the state; Group 1 to the North and West regions; Group 2 to the region that includes the Metropolitan Region of São Paulo and the Paraíba Valley; and Group 3 corresponding to the state’s coastline. The spatial visualization of the groups is a way to verify the validity of the clustering method employed. This pattern is similar to that found in another study conducted in the state of São Paulo, in which the relationship between the increase in temperature and the territorial expansion of the dengue vector over the years was verified¹⁹19. Fonseca Júnior DP, Serpa LLN, Barbosa GL, Pereira M, Holcman MM, Voltolini JC, et al. Vetores de arboviroses no estado de São Paulo: 30 anos de Aedes aegypti e Aedes albopictus. Rev Saúde Pública. 2019;53:84. https://doi.org/10.11606/s1518-8787.2019053001264
https://doi.org/10.11606/s1518-8787.2019... .

For each group, the mean values of each indicator were calculated, as shown in Table 2. In addition, the percentage of the total number of worst-performing municipalities in each group was calculated (with indicators coefficient100 = 2 and coefficient20 = 2). From this data, the two groups with the worst performance in both indicators (Clusters 1 and 3) perform differently in terms of the indicators. Regarding infant mortality, Group 1 obtained the best outcomes among the four groups, while the municipalities in Cluster 3 scored the highest. Regarding the Gini index, the opposite is true: the municipalities in Group 1 scored the lowest average value for this indicator, while the municipalities in Group 3 scored the highest values.

While the HDI remained practically unchanged for the four groups, with only Group 2 showing a lower value, there was a great disparity in the GDP per capita values. The municipalities in Group 3 showed, on average, higher values, followed by groups 0 and 2 and, finally, Cluster 1. In terms of urban population, almost 90% of the population of Group 3 municipalities lived in urban areas compared to about 85% in Groups 0 and 1.

For water and sewerage, there are similar values for the indicator for water supply, whereas for sewage there are significant differences, especially for municipalities in groups 2 and 3, which mainly include municipalities in the Metropolitan Region of São Paulo, Vale do Paraíba, and Litoral Paulista.

Regarding climate variables, more favorable conditions (higher temperature and lower precipitation) are observed for the group of municipalities with the worst performance in the coefficient20 indicator (Group 1). Despite these municipalities accounting for the smallest percentage of urban areas, the indicator for area modified by human activities recorded the highest values.

Regarding structural indicators, there were higher investments and higher numbers of community health workers per capita in the municipalities with the worst outcomes (groups 1 and 3).

Intragroup analysis

Finally, the outcomes were evaluated internally in the groups, according to the classes of the outcome indicators. For this analysis, we estimated the mean values for the coefficient20 (Table 3), a measure adopted in the guidelines of the São Paulo State Health Department to classify municipalities with or without a historical transmission series.

According to Table 3, in general, the worst-performing municipalities (Class 2) in the four groups follow the same pattern: better socioeconomic conditions and sanitation and larger urban population, but with lower values for the health structure component indicators.

For Group 2, it can be observed that the municipalities with the worst performance showed better socioeconomic conditions but a larger urban population and a significantly higher value for urbanization. Although investment in health is higher, these municipalities show a worse structure for the two primary care indicators. Regarding the municipalities in groups 0 and 3, the same differences identified for Group 2 are observed: better general conditions and greater investment in health but worse health structure.

A similar pattern is observed in Group 1, except for the indicator referring to investments in health. The worst-performing municipalities also showed lower average investment in health per inhabitant.

Finally, the outcomes for the two outcome indicators according to the performance classes can be visualized spatially in Figure 1. The first map shows the municipalities according to the four groups identified in the algorithm. In the following two maps, we have a classification (Best/Worst) for the two coefficients of outcomes defined in Chart 1. It can be observed, in general, that the worst performances are in the municipalities located further east of the state. By analyzing the other indicators spatially, we identified which range comprises the municipalities with the highest urban population rates.

DISCUSSION

In this study, it was possible to aggregate contextual data for an evaluation of the Brazilian Dengue Control Program based on the Donabedian model for the 645 municipalities in the state of São Paulo. Although the program is the same for everyone, there are different municipal configurations, such as the number of professionals working against the vector, which does not always meet the standards of the national program. Although it is difficult to evaluate the quality of municipal work, these differences can influence the final quality of dengue control program.

Using the concepts of Planetary Health and Data Science techniques, data from various areas were integrated into a Data Warehouse to conduct the evaluations.

The context of Big Data, with greater availability of data sources, presents several opportunities to advance knowledge in healthcare. However, the extraction of useful information in Big Data requires extra methods to be efficient. The methodology adopted in this study allowed the evaluation of the dengue control program on a large scale, in addition to the integration of context data into a widespread model for health evaluations.

In the program’s evaluation, the deficiencies in health structure stood out in all three analyzed indicators for the municipalities with the best socioeconomic conditions, which did not necessarily imply better conditions in primary care and performance in terms of dengue cases. Although it cannot be said that these are the only reasons for the higher incidence of dengue, the differences observed in the indicators of these municipalities may suggest that economic development has not been accompanied by improvements in health conditions, especially considering that these places have a larger urban population.

Additionally, Table 1 suggests that, even with more favorable conditions for mosquito proliferation, such as higher temperatures and urban population, the municipalities with the worst performance in terms of outcome indicators showed a worse health structure than those with less favorable conditions, when the opposite would be expected. Considering only the outcome indicator coefficient20, the data in Table 1 show that about 38 million people live in areas more susceptible to mosquito proliferation and with worse health conditions in the two indicators related to primary care. Public management could consider this information to better target investments on the major needs of each location in the fight against dengue and other diseases related to the same factors.

As the Planetary Health studies suggest, unsustainable development comes at a cost to human health. Although some indicators may have improved, others have regressed, as can be seen in the indicators related to dengue.

A limitation of this study regards the observed data period. We used data up to 2019, which may be out of date, also considering the changes in health systems caused by the COVID-19 pandemic. In addition, due to delays in the Brazilian Census, some data from 2010, such as the HDI and the Gini Index, were used. Given the absence of the Census, which was to be conducted in 2020, population data were used based on estimates. Therefore, once updated data are available, new analyses can be conducted using the method presented in this study.

REFERENCES

Publication Dates

History