Data mining technique for identifying mortality patterns in a pulmonary tuberculosis cohortAbstract
Tuberculosis (TB) deaths occur early, mainly during the two months comprising the intensive treatment phase compared to other causes of death. Understanding the distinct TB mortality patterns provides more effective health care for specific groups. This study aimed to apply a data mining technique to examine whether there are distinct mortality patterns associated with diverse patient profiles within a cohort notified with TB in the city of Rio de Janeiro, Brazil. This was a retrospective cohort study using a probabilistic linkage between the Brazilian Information System for Notifiable Diseases (SINAN) and the Brazilian Mortality Information System (SIM). An association rule data mining technique was applied using the Apriori algorithm to evaluate patterns associated with causes of death. TB deaths were associated with adult men with short survival time. Deaths from AIDS with TB occurred during the treatment period and were more associated with women. In contrast, deaths from AIDS without a mention of TB were more frequent among younger men and Black/Mixed-race adults. Deaths from external causes occurred later, one year after diagnosis, and were related to young men. Deaths from other causes that were linked to longer survivals were associated with older white individuals with higher education. The Apriori algorithm provided the extraction of relevant knowledge from secondary data, such as identifying distinct subgroups of people who can benefit from specific actions, with a consequent positive effect for TB control.
Keywords:
Tuberculosis; Mortality; Information Systems; Medical Record Linkage
Resumo
As mortes por tuberculose (TB) ocorrem precocemente, principalmente durante os dois meses que compõem a fase de tratamento intensivo, em comparação com outras causas de morte. Compreender os padrões distintos das pessoas que morrem devido à TB permite melhor eficácia do atendimento à saúde para grupos específicos. O objetivo foi aplicar a técnica de mineração de dados para explorar se existem padrões distintos de mortalidade associados a perfis diversos de pacientes dentro de uma coorte de indivíduos notificados com TB na cidade do Rio de Janeiro, Brasil. Realizamos um estudo de coorte retrospectivo usando uma ligação probabilística entre o banco de dados de TB do Sistema de Informação de Agravos de Notificação (SINAN) e o Sistema de Informações sobre Mortalidade (SIM). A técnica de mineração de dados foi aplicada usando o algoritmo Apriori para avaliar padrões associados às causas de morte. Mortes por TB estavam associadas a homens adultos com tempo de sobrevivência curto. Mortes por aids com TB ocorreram durante o período de tratamento e foram mais associadas às mulheres. Em contraste, mortes por aids sem menção à TB foram mais frequentes entre homens mais jovens e adultos negros/pardos. Mortes por causas externas ocorreram mais tarde, um ano após o diagnóstico, e estavam relacionadas a homens jovens. Mortes por outras causas ligadas à sobrevivência mais longa estavam associadas aos brancos idosos com Ensino Superior. O algoritmo Apriori permitiu a extração de conhecimento relevante a partir de dados secundários, como identificar subgrupos distintos de pessoas que podem se beneficiar de ações específicas, com um consequente efeito positivo para o controle da TB.
Palavras-chave:
Tuberculose; Mortalidade; Sistemas de Informação; Registro Médico Coordenado
Resumen
Las muertes por tuberculosis (TB) se producen de forma temprana, principalmente durante los dos meses que comprende la fase de tratamiento intensivo, en comparación con otras causas de muerte. Comprender los distintos patrones de las personas que fallecen por TB permite mejorar la eficacia de la atención sanitaria para grupos específicos. El objetivo del estudio era aplicar una técnica de minería de datos para explorar si existen patrones de mortalidad distintos asociados a diversos perfiles de pacientes dentro de una cohorte de personas notificadas con TB en la ciudad de Río de Janeiro, Brasil. Realizamos un estudio de cohorte retrospectivo utilizando un enlace probabilístico entre la base de datos del Sistema de Información de Agravios de Notificación (SINAN) y el Sistema de Informaciones sobre Mortalidad (SIM). Se aplicó una técnica de minería de datos en reglas de asociación utilizando el algoritmo Apriori para evaluar los patrones asociados con las causas de muerte. Las muertes por TB se asociaron con hombres adultos con un tiempo de supervivencia corto. Las muertes por SIDA con TB se produjeron durante el periodo de tratamiento y se asociaron más con las mujeres. Por el contrario, las muertes por SIDA sin mención de TB fueron más frecuentes entre los hombres más jóvenes y los adultos negros/morenos. Las muertes por causas externas se produjeron más tarde, un año después del diagnóstico, y se relacionaron con los hombres jóvenes. Las muertes por otras causas relacionadas con una mayor supervivencia se asociaron con personas mayores, blancas y con un nivel educativo superior. El algoritmo Apriori permitió extraer información relevante de datos secundarios, como la identificación de subgrupos distintos de personas que pueden beneficiarse de medidas específicas, con un consecuente efecto positivo para el control de la TB.
Palabras-clave:
Tuberculosis; Mortalidad; Sistemas de Información; Registro Médico Coordinado
Introduction
Tuberculosis (TB) continues to pose a substantial public health challenge. Brazil is one of the 30 countries with the highest rates of TB 1. Despite advancements in diagnosis, prevention, as well as effective, free, and widely available treatment provided by the Brazilian Unified National Health System (SUS, acronym in Portuguese), over 80,000 new cases are diagnosed and approximately 6,000 deaths are recorded annually 2. This burden places the country, along with Peru, among those with the highest rates of TB in the Americas 1.
In recent years, several key health care interventions have been implemented in Brazil to improve TB control. These include the directly observed treatment implemented in the health care system, the expansion of the Rapid Molecular Testing Network for TB (RMT-TB), enhancements to the quality of active TB treatment, the expansion of preventive TB treatment (TPT), and the strengthening of primary health care (PHC) for the early identification of individuals with TB and the improvement of treatment adherence and monitoring 3. However, TB control remains unsatisfactory. Cure rates consistently fall short of World Health Organization (WHO) recommendations. Adverse outcomes such as treatment nonadherence and mortality are frequently observed. The death of a TB patient is considered a preventable event, with the WHO’s goal for Brazil by 2035 being a 95% reduction in TB-related fatalities.
Globally and in Brazil, several TB-associated factors have been linked to mortality, including advanced age, low educational level, use of alcohol and other drugs, TB-HIV coinfection, and loss to follow-up 4,5,6,7, as well as characteristics of health care services. Health care service-related factors, such as having been treated in an emergency hospital, are also linked to an increased risk of death 5. However, it is currently not clear whether or not the factors vary with the causes of death, during and after TB treatment, and can be associated with the burden of TB and HIV in the investigated region 4. Some studies in South India (Asia), Spain (Europe), Tanzania (Africa), and Brazil (South America) indicate TB deaths occurring early, mainly during the two months comprising the intensive treatment phase 6,7,8,9,10. However, such studies are scarce in Brazil.
Understanding the mortality patterns in TB patients and their association with specific risk profiles is essential for guiding public health surveillance, management, and care. This information would facilitate the prioritization of targeted interventions, aiming to enhance health care effectiveness and reduce disease-related fatalities. A study in the same municipality found that the primary underlying causes of death were TB and AIDS 11. Among other causes, the most common were neoplasms, circulatory diseases, and respiratory system diseases.
Data mining algorithms are valuable tools for the identification of patterns associated with a variety of health outcomes. When compared to traditional statistical methods, such tools can provide new insights, especially when studying many covariates that yield the dispersion of observations in multiple profiles 12. Association rule algorithms are used to extract key patterns from data that can identify patient profiles with different risks of death. Initially used in market basket analysis for identification of items that tend to be purchased together, these algorithms have been increasingly used in health studies 13,14,15,16,17,18. However, there are few TB studies, and those are predominantly for diagnosis 16,18.
The city of Rio de Janeiro is the second-largest economy in Brazil and has one of the highest TB mortality rates, more than double the country’s average 2. In light of this, this seeks to apply a data mining technique to examine whether there are distinct mortality patterns associated with diverse patient profiles within a cohort of individuals notified with TB in the city of Rio de Janeiro. The ability to distinguish specific groups of individuals could facilitate the establishment and prioritization of targeted actions, thereby improving the effectiveness of health care and reducing TB-related mortality.
Method
Data sources and study population
This study used a linked database from a previous investigation 19. This database was composed of a cohort of individuals with TB reported in the Brazilian Information System for Notifiable Diseases (SINAN, acronym in Portuguese) between 2009 and 2012. The cohort was passively followed for at least one year after notification through record linkage with the Brazilian Mortality Information System (SIM, acronym in Portuguese) database, covering the period from 2009 to 2013. The study population consisted of individuals living in Rio de Janeiro, diagnosed with pulmonary TB, who died between 2009 and 2013. When a person had more than one treatment reported at SINAN, only the most recent record was analyzed because the information was closer to the death. Subjects aged under 15 years of age and those with HIV infection were excluded. Also, records of individuals with HIV/AIDS were also excluded due to their elevated risk of TB mortality, unlike other patient groups. We restricted the analysis to cases of pulmonary TB in adults given their higher prevalence and the unexpected nature of death as an outcome in this specific group.
Probabilistic record linkage
We carried out a probabilistic record linkage using the patient’s name, mother’s name, sex, date of birth, and address 20.
We used 22 blocking steps for indexing the database to improve sensitivity, using a combination of variables, as follows: soundex of the person’s given name; soundex of the person’s last name; sex; year of birth; city of residence; soundex of the mother’s given name, and soundex of the mother’s last name 21. We used the person’s full name, mother’s full name, and birth date to compute each pair of records’ match score 20.
We empirically defined the match score threshold to classify a pair of records as belonging to the same person. This procedure involved a top-down inspection of record pairs, ordered by their scores, within the OpenReclink environment 22. We continued this inspection until we encountered pairs for which we could not confidently determine if they belonged to the same person. These pairs, along with all others with a score from this threshold down to zero, formed a “gray area” of uncertain classification. All pairs within this gray area subsequently underwent clerical review to enhance accuracy, based on the person’s full name, mother’s full name, date of birth, and address. When in doubt, we classified the pair of records as a non-match, adopting a more conservative classification procedure 23.
Apriori algorithm
Association rule data mining is a technique used to discover meaningful patterns and extract valuable new knowledge from large datasets. The Apriori algorithm is a widely used method for this end, as it efficiently identifies all frequent itemsets in a dataset to then form relevant association rules. Each category of a variable is equivalent to an item of the Apriori algorithm. An association rule relates two sets of data (called items) as a conditional sentence IF (X) THEN (Y) (short form X → Y), in which X is a set of binary items joined by the condition “and”, called antecedent, and Y is a set of binary items, disjunctive from X, also joined by the condition “and”, called consequent 12. Y is defined as a fixed set of items (an example would be the items represented by the treatment outcomes), whereas X are sets formed by combinations of all items xi in set X = {x1, x2, ..., xn}. As an example, these items would be the factors that can determine the outcomes.
The purpose of association rules is to find all possible sets of items X that come together with those of set Y 12. Different metrics are used to assess the rules’ frequency and strength 24. The support of a rule is the proportion of items that appear in the antecedent and consequent concomitantly, concerning the total records of the database. It can also be defined as the frequency of patterns in the database, for instance, the proportion of a certain patient profile that died from a specific cause in the total number of records in the database.
The confidence of a rule measures the proportion of people who died from a certain cause (consequent), among a certain profile of individuals (antecedent) 18. Lift is a measure that assesses the dependency between items in a rule. Lift evaluates the degree to which the occurrence of one item leads to the occurrence of the other 14, that is, it indicates dependence, the more frequent Y becomes when X occurs, given an association rule X - > Y. For instance, the lift of a rule indicates how much greater is the occurrence of a given profile with a cause of death, compared to the expected, if the given profile was independent of that cause of death.
Minimum support is required to assure the statistical significance of the generated standards, restricting the number of rules originated by the algorithm. High confidence values are vital to ensure a high relationship between the items analyzed. The more relevant rules are those that have a high lift value 25.
Variables and items used
The antecedent items were defined by several variables, including sociodemographic aspects, the status of diagnostic laboratory tests for TB and HIV, the type of health care service, and time until death (Box 1). The laboratory test variable had categories for whether the test was performed, not performed, or if the result was not registered in the database. The consequents were grouped according to the underlying cause of death and categorized into deaths from TB, external causes, AIDS without a mention of TB, and AIDS with TB, considering the codes of the International Classification of Diseases - 10th revision (ICD-10) registered in the primary cause: A15-A19, V01-Y98, B20.1-B24, and B20.0, respectively. Deaths from other causes were all deaths different from those mentioned above.
We included the HIV test variable to account for individuals whose underlying cause of death was AIDS but who had no recorded HIV test result. This procedure enabled us to identify possible undetected HIV patients at the time of notification of TB. No filtering was applied for prior causes of death to ensure a complete list of all potential recorded causes.
Survival time was defined as the number of days elapsed from the date of diagnosis in SINAN to the date of death registered in SIM.
Data analysis
We defined a minimum support of 0.5% to enable the generation of much less common rules. We eliminated redundancies considering confidence 24 and also rules with a lift below 1.1. Three experts in TB surveillance convened to select the optimal rules, based on both predefined metrics and their professional expertise. The selection criteria prioritized confidence and lift, followed by support or the more informative rule, as indicated by the specialists’ assessment.
The analyses used the R statistical package version 3.3.1 (http://www.r-project.org) with the libraries “arules” 26,27 and “arulesViz” 28. The probabilistic record linkage used the OpenReclink software 22.
Results
The analysis included 1,819 individuals diagnosed with pulmonary TB who died during the follow-up of the cohort. Table 1 shows the absolute and relative frequencies of the 36 items from 10 variables. Approximately three-quarters were new TB cases, male, and about half were aged 35 to 59 years, had less than nine years of education, Black/Mixed-race skin color, did not undergo an HIV test, and underwent sputum smear testing for TB diagnosis. Almost half of the deaths occurred within the treatment period, with the majority (30.6%) in the first two months after diagnosis. Most deaths were due to other causes (52%), followed by TB (36%) and AIDS (3.5%) (Table 1).
We generated 3,590 association rules without redundancy: 1,222 for deaths from TB, 73 for deaths from AIDS and TB, 23 for deaths from AIDS without TB, 612 deaths from external causes, and 1,660 deaths from other causes.
Table 2 shows the rules selected for deaths from TB. The rules selected had maximum support of 1.4%, a maximum lift of 2.8%, and maximum confidence of 100%. The items male and short survival periods are noted. The rules with higher confidence (100%) and lift (2.8) referred to death during the first two months of treatment (intensive phase) of men, aged 35 to 59 years, new TB cases or relapse, diagnosed in hospital (rules 1 to 4). Rule 5 has the greater support (1.4%) and refers to men, aged 35 to 59 years, Black/Mixed-race skin color with HIV-negative test who died in the intensive treatment phase. Among the other rules that included death during the first two months, two also had men, new case, and were treated at a PHC unit (rules 6 and 7). Rule 6 also included low education and HIV-negative test, whereas rule 7 also included young age. Rules 8 and 9 referred to women, who were diagnosed at a hospital. The latter rules differed because rule 8 included relapse, and rule 9 included young age. Rules that indicated the diagnosis of TB near the time of death (rules 10 to 15) also included patients diagnosed in hospital and with a laboratory test not performed (smear test or HIV). Male, aged 35 years or more and new case occurred more frequently in such rules. The two rules that included deaths during the maintenance treatment phase or at most within one year after diagnosis (rules 16 and 17) also had men that returned after default and an HIV-negative test. Finally, rule 18 referred to a more vulnerable to death profile consisting of Black or Mixed-race men and lower educational level, who returned after default.
Support, confidence and lift of top association rules selected by experts for deaths caused by tuberculosis (TB). Rio de Janeiro, Brazil, 2009 to 2012.
Rules related to deaths from TB-HIV/AIDS coinfection (Table 3) had maximum support of 0.8%, maximum lift of 6%, and maximum confidence of 12.5%. The rules with a higher lift (≥ 5.4) referred to women not tested for HIV (rules 19 to 21), new TB cases diagnosed in hospital (rule 19), aged 35 to 59 years (rule 20), with educational level ignored (rule 21). Three other rules (rules 22, 23, 24) included being treated in PHC and not presenting a sputum smear test. Rules 25 and 26 referred to deaths that occurred during treatment. The last rule (27) included only age (16 to 34 years). For deaths from AIDS without TB, we selected six rules (Table 3, rules 28 to 33) which presented maximum support of 0.6%, maximum lift of 1.9%, and maximum confidence of 2.7%. The combinations of items were more varied. Only one rule selected included male sex, which differs from deaths from TB-HIV/AIDS coinfection, in which we only selected rules that included women.
Support, confidence and lift of top association rules selected by experts for AIDS deaths with and without tuberculosis (TB). Rio de Janeiro, Brazil, 2009 to 2012.
Table 4 shows the four rules selected for deaths from external causes. Three of them included young men (rules 34 to 36) who died after treatment. The rules selected had maximum support of 0.7%, a maximum lift of 9.9%, and maximum confidence of 84.6%. Ten of the twelve chosen rules for deaths from other causes presented 100% confidence and a 1.9 lift (rules 38 to 49). We selected 12 rules related to deaths from other causes (Table 4). Such rules had maximum support of 2% and a maximum lift of 1.9. Ten of them had 100% confidence. Most rules included deaths one year after the diagnosis, older people, and those diagnosed in PHC. Differently from the observations in the other groups, the category white predominated when the rule selected included skin color.
Support, confidence and lift of top association rules selected by experts for deaths from external causes and other causes. Rio de Janeiro, Brazil, 2009 to 2012.
Discussion
Using the Apriori algorithm enabled us to identify various mortality patterns related to different patient profiles. The TB profiles differed according to the cause of death and considered the survival time after TB diagnosis. Overall, TB deaths were associated with adult men with shorter survival times. Deaths from AIDS with TB occurred within the treatment period and were more associated with women. In contrast, deaths from AIDS without mention of TB were associated with young men. Deaths from external causes occurred later, one year after diagnosis, and young men were the most common characteristic found in this profile. Deaths from other causes, related to longer survivals, were associated with older adults and white skin color.
Even more important than the identification of those relevant characteristics associated with the underlying cause of death was the identification of patient profiles. Studies using traditional regression models show that all TB mortality causes were associated with old age 29,30, male sex 31, low socioeconomic status 32,33, presence of comorbidities 30,34,35, abuse of alcohol or other illicit drugs 36,37 and previous treatment for TB 30,36,38. However, one limitation of such traditional approaches is that they make it hard to identify specific subgroups of more vulnerable people for unfavorable outcomes. Traditional studies on risk factors for treatment outcomes generally produce measures of association for each factor individually, after adjusting for others. Association rules, on the other hand, identify which factors - whether individually or in combination - are most relevant in defining profiles with high risks of distinct outcomes, including both frequent and rare patterns. In our study, using association rules, we identified profiles of patients that could benefit from greater consideration in health care services, aiming at better outcomes in their TB treatment.
Death from TB is a preventable event 39. However, more than one-third of people who died with pulmonary TB had this condition as the underlying cause of death, a result that was also observed in a previous study 11. Among TB deaths, the shorter survival time was highlighted as well as those patients diagnosed at death. Men, especially adults and older adults, were the most affected. Enhancing the management of respiratory symptomatic individuals in PHC would promote a more timely diagnosis of TB and, consequently, reduced mortality.
Shorter survival time may indicate a barrier to access to timely treatment. This hypothesis is reinforced by the lack of HIV and sputum smear tests, leading to the death of new TB cases occurring even before the diagnosis of the disease 8,40.
Diagnostic delay may increase disease severity and lead to a higher risk of death 41. Diagnostic delay is a function of the interval time that comprises the onset of symptoms, decision to seek care, obtaining medical care, and receiving diagnosis of the disease. Failure in diagnosis accounts for an important portion of the delay to treatment 42. Factors associated with the subject’s vulnerability that influence the outcome add to explanations for early death. Malnutrition 43 and poor performance status 44 were associated with early death in hospitalized people. The presence of some comorbidities in HIV-negative subjects with pulmonary TB in a Brazilian study highlighted those related to infectious diseases, those related to blood diseases, immunological disorders, smoking, and lung diseases 45.
The presence of young men and men with low levels of education treated at a Municipal Health Center was notable in deaths occurring in the 2-month intensive phase of TB treatment. The decentralization of TB treatment and diagnosis to PHC is one of the measures against the disease in Brazil 3. Moreover, the male sex is related to lower use of health care services, considering that men and young people are not regular visitors to health care facilities. We found a relevant proportion of young men dying from TB with markers of low access to health care. On the other hand, some young women with recurrent TB fell in the group that died from TB less than two months after diagnosis. They composed a group of subjects diagnosed in general hospitals, which may have contributed to the occurrence of an unfavorable outcome.
Differently from new TB cases, deaths from TB in HIV-negative men who defaulted treatment occurred during the maintenance treatment phase (60-180 days) or after the end of treatment, but before one year after diagnosis. It is known that treatment default is related to an increased risk of dying from TB and is often related to socially vulnerable people 46. However, the association rules elicited here did not include deprived subjects, such as Black/Mixed-race skin color subjects with low education levels. This fact may indicate that not only these groups had the possibility of a worse prognosis.
For deaths from TB-HIV coinfection, there was a predominance of females, without sputum smear and HIV tests performed, Black or Mixed-race skin color, aged 16 to 59 years, treated at a Municipal Health Care Center and dying within the treatment period. Women of reproductive age visit health care facilities more frequently than men, which should have helped the earlier detection of TB. However, women are not a predominant subject in TB and AIDS 47, which may have contributed to misdiagnosis at PHC facilities, differently from expectation. On the other hand, men were predominant in deaths from AIDS without mention of TB. They were young adults, Black/Mixed-race skin color, treated at a Municipal Health Care Center, who did not undergo diagnostic tests for HIV and TB, and there was no association with the survival time after TB treatment initiation. They probably had undiagnosed HIV infection, and both TB and HIV disease should be detected and treated in due time. In Brazil, every person diagnosed with TB must be tested for HIV 3. The loss of opportunity to detect HIV may have contributed to the occurrence of preventable deaths.
In deaths from external causes, we note young men with high school education or Blacks with low education, whose deaths occurred later after the end of TB treatment. TB is associated with poverty, which, in turn, is also associated with violence. Young men compose a vulnerable group in this regard in Rio de Janeiro 48,49. Also, Mixed-race men were more likely to die from external causes (28%) than Black, Yellow, Indigenous, and White men (8%). Deaths from external causes are associated with young men living in areas of social vulnerability in large urban cities 50.
It is widely known that poor living conditions increase TB morbidity and hinder access to health care services. Black and Mixed-race people had a lower probability of cure, suggesting untimely diagnosis and inadequate treatment 51. Therefore, it is essential to implement social protection policies that address systemic racism, stigma, and access barriers in order to reduce the current structural inequality and promote equity.
Studies on multiple causes of death involving individuals with TB report that the main underlying causes - when not TB itself - are AIDS, neoplasms, circulatory diseases and respiratory system diseases 11,52. In our study, deaths from other causes occurred one year after TB diagnosis and were associated with older adults without an HIV test and to white people with high education, the latter a group typically considered less vulnerable to TB. These findings are consistent with the understanding that preexisting diseases and a declined immune response can explain the onset of TB in older individuals. This also might highlight a key concern, as TB may worsen existing conditions and increase the risk of death, even after successful treatment. Patients treated at TB clinics, a health care service with a secondary level of complexity for TB care, may indicate difficulty in treatment management, increasing the risk of death even after the cure of the TB episode 9. On the other hand, Black/Mixed-race older adults died earlier, between six months and one year after diagnosis, and TB could be considered as a contributing cause of death.
The use of metrics such as lift and confidence was essential to choose rules that indicated specific profiles, instead of support, more indicated to highlight the most common and general patterns 12. The lift measures the dependency between items in a data set and the highlighted rules had values above two in most. Although some rules have high values of confidence and lift, they can be considered rare and, therefore, some consider them to be of little use 53. However, since TB is a preventable and curable disease, understanding the factors associated with death, combined or isolated, can improve health care initiatives.
The findings of this study may also benefit other populations and geographical contexts, considering there are no significant differences in the profile of TB-related deaths nationwide. This is further supported by the fact that recommendations for TB prevention, diagnosis, and treatment are standardized throughout the country 3. Additionally, the study’s methodology can be used to assist with the diagnosis and outcome analysis of specific patient groups, such as individuals with extrapulmonary TB, drug-resistant TB, HIV, or diabetes. Furthermore, health authorities can use this tool to guide targeted strategic measures aimed at increasing the effectiveness of health care.
The results of this study highlight that prevention strategies within PHC, in close collaboration with surveillance, can be expanded or strengthened at the local level. These strategies can be applied to different patient groups to varying degrees and include contact investigation, expansion of TB preventive treatment, active case search for respiratory symptomatic individuals, strengthening of intersectoral programs to reduce treatment interruption, ongoing education for health care professionals, development of health education initiatives for the community using social media and accessible communication.
This study is subject to several key limitations. The time period analyzed may not fully account for recent changes, such as data enhancements in the SINAN-TB database and improvements in health care in Rio de Janeiro, both of which could influence the mortality profile. An example is the lack of specific variables for incarcerated and homeless populations, which prevents the identification of other patient profiles. The findings do not account for the potential impact of the RMT-TB, which has influenced early diagnosis and detection of rifampicin resistance since its implementation in 2014. The extended effects of the COVID-19 pandemic beyond 2021 were not considered, a factor that has impacted health care services and altered global mortality trends for individuals with TB. Furthermore, a limitation of the study is that some individuals had a passive follow-up period of only one year, which may have constrained the observation of later-onset mortality.
Even though the data are not recent, it is worth noting that despite the availability of RMT-TB and improvements in TB care and surveillance in the city of Rio de Janeiro over the years, the cure rate has remained stable. Meanwhile, the mortality rate showed a slight increase between 2012 and 2014, with a gradual reduction until 2016, followed by a plateau, except during the pandemic period 2. Furthermore, between 2012 and 2023, the case profile did not show any significant changes, and the rate of laboratory confirmation decreased. This suggests that, even when referring to the 2009-2012 period, the improvements mentioned likely did not lead to substantial changes in the mortality rate.
Another limitation of this study is the non-inclusion of other variables that may influence mortality, such as quality of care, multidrug-resistant TB, and comorbidities. Furthermore, the study’s reliance on routine data from health information systems is a limitation, as these data are subject to the quality of completion by professionals. As a result, skin color and education data had high percentages of missing values. To minimize possible bias in excluding records with missing values, a category of non-information was added to these variables. Possible errors in the linkage process could also be noted, such as pairs of records that were not found. However, the conservative approach of considering doubtful links as non-pairs minimizes potential bias in the association measures (e.g., odds ratio and relative risk), and, by extension, in the findings of this study’s association rules 23.
Another limitation was the low support of the rules, highlighting patterns with few records. Association rule algorithms can find associations that other methods, such as regression models, do not easily detect. However, there is still no technique for statistical validation of the rules found, whose selection is based empirically using metrics and expert judgment. However, this aspect can only be evaluated in a future study that consists of databases even larger than the one used in the present study, and that repeated the health care system conditions observed in the analyzed period.
Conclusion
The Apriori algorithm was able to extract relevant knowledge from secondary databases, identifying mortality patterns, which indicated distinctive subgroups of subjects, who can benefit from specific measures, with a positive impact on TB control. The Apriori algorithm can be used for knowledge acquisition in the public health care sector, to monitor actions, and to support the surveillance and decision-making process.
Acknowledgments
This work was supported by the Health Surveillance Department, Brazilian Ministry of Health/National Health Foundation (TC 234/2012). R. S. Pinheiro [processes n. 310173/2015-9 and n. 316755/2021-4] and C. M. Coeli [process n. 306668/2023-8] received grants from the Brazilian National Research Council (CNPq). C. M. Coeli was also partially supported by research fellowship grant from the Rio de Janeiro State Research Foundation (FAPERJ; process n. 200.003/19) The Brazilian Coordination for the Improvement of Higher Education Personnel (CAPES) supported the study with a scholarship.
References
- 1 World Health Organization. Global tuberculosis report 2024. Geneva: World Health Organization; 2024.
-
2 Secretaria de Vigilância em Saúde e Ambiente, Ministério da Saúde. Boletim Epidemiológico, Tuberculose 2025; número especial. https://www.gov.br/saude/pt-br/centrais-de-conteudo/publicacoes/boletins/epidemiologicos/especiais/2025/be-de-tuberculose-numero-especial-mar-2025.pdf
» https://www.gov.br/saude/pt-br/centrais-de-conteudo/publicacoes/boletins/epidemiologicos/especiais/2025/be-de-tuberculose-numero-especial-mar-2025.pdf - 3 Departamento de Doenças de Condições Crônicas e Infecções Sexualmente Transmissíveis, Secretaria de Vigilância em Saúde, Ministério da Saúde. Manual de recomendações para o controle da tuberculose no Brasil. 2nd Ed. Brasília: Ministério da Saúde; 2019.
- 4 Waitt CJ, Squire SB. A systematic review of risk factors for death in adults during and after tuberculosis treatment. Int J Tuberc Lung Dis 2011; 15:871-85.
- 5 Rocha MS, Oliveira GP, Saraceni V, Aguiar FP, Coeli CM, Pinheiro RS. Effect of inpatient and outpatient care on treatment outcome in tuberculosis: a cohort study. Rev Panam Salud Pública 2018; 42:e112.
- 6 Jonnalagada S, Harries AD, Zachariah R, Satyanarayana S, Tetali S, Chander GK, et al. The timing of death in patients with tuberculosis who die during anti-tuberculosis treatment in Andhra Pradesh, South India. BMC Public Health 2011; 11:921.
- 7 Lopes VDS, Couto RM, Luz APD, Silva PRS, Lima JC. Factors associated with deaths by tuberculosis in the state of Mato Grosso, 2011-2020: retrospective cohort study. Epidemiol Serv Saúde 2024; 33:e20231402.
- 8 Rodrigo T, Casals M, Caminero JA, García-García JM, Jiménez-Fuentes MA, Medina JF, et al. Factors associated with fatality during the intensive phase of anti-tuberculosis treatment. PLOS One 2016; 11:e0159925.
- 9 Bukundi EM, Mhimbira F, Kishimba R, Kondo Z, Moshiro C. Mortality and associated factors among adult patients on tuberculosis treatment in Tanzania: a retrospective cohort study. J Clin Tuberc Other Mycobact Dis 2021; 24:100263.
- 10 Dos Santos DT, Arroyo LH, Alves YM, Alves LS, Berra TZ, Crispim JA, et al. Survival time among patients who were diagnosed with tuberculosis, the precocious deaths and associated factors in southern Brazil. Trop Med Health 2021; 49:31.
- 11 Rocha MS, Oliveira GP, Aguiar FP, Saraceni V, Pinheiro RS. Do que morrem os pacientes com tuberculose: causas múltiplas de morte de uma coorte de casos notificados e uma proposta de investigação de causas presumíveis. Cad Saúde Pública 2015; 31:709-21.
- 12 Agrawal R, Srikant R. Fast algorithms for mining association rules in large databases. In: Proceedings of the 20th International Conference on Very Large Data Bases. San Francisco: Morgan Kaufmann Publishers; 1994. p. 487-99.
- 13 Raorane AA, Kulkarni RV, Jitkar BD. Association rule - extracting knowledge using market basket analysis. Res J Recent Sci 2012; 1:19-27.
- 14 Kang W, Liu S, Du J, Tang P, Chen H, Liu J, et al. Epidemiology of concurrent extrapulmonary tuberculosis in inpatients with extrapulmonary tuberculosis lesions in China: a large-scale observational multi-centre investigation. Int J Infect Dis 2022; 115:79-85.
- 15 Souza KKG, Mendes MS, Alencar VC, editors. Análise epidemiológica da tuberculose no estado da Paraíba utilizando big data analytics. Curitiba: CRV; 2018.
- 16 Kang W, Yu J, Liang C, Wang Q, Li L, Du J, et al. Epidemiology and association rules analysis for pulmonary tuberculosis cases with extrapulmonary tuberculosis from age and gender perspective: a large-scale retrospective multicenter observational study in China. Int J Clin Pract 2023; 2023:5562495.
- 17 Jahantigh F, Ameri H. Evaluation of TB patients characteristics based on predictive data mining approaches. J Tuberc Res 2017; 5:13-22.
- 18 Lakshmi KR, Krishna MV, Kumar SP. Utilization of data mining techniques for prediction and diagnosis of tuberculosis disease survivability. International Journal of Modern Education and Computer Science 2013; 5:8-17.
- 19 Rocha MS, Oliveira GP, Guillen LCT, Coeli CM, Saraceni V, Pinheiro RS. Uso de linkage entre diferentes bases de dados para qualificação de variáveis do Sinan-TB e a partir de regras de scripting. Cad Saúde Pública 2019; 35:e00074318.
- 20 Camargo Jr. KR, Coeli CM. Reclink: aplicativo para o relacionamento de bases de dados, implementando o método probabilistic record linkage. Cad Saúde Pública 2000; 16:439-47.
- 21 Santos ML. Fatores associados à subnotificação de TB e AIDS, durante os anos de 2001 a 2010, a partir do SINAN [Dissertação de Mestrado]. Recife: Centro de Pesquisas Aggeu Magalhães, Fundação Oswaldo Cruz; 2014.
- 22 Camargo Jr. KR, Coeli CM. Going open-source: some lessons learned from the development of OpenRecLink. Cad Saúde Pública 2015; 31:237-63.
- 23 Coutinho ESF, Coeli CM. Acurácia da metodologia de relacionamento probabilístico de registros para identificação de óbitos em estudos de sobrevida. Cad Saúde Pública 2006; 22:2249-52.
- 24 Bürkle PY. Um método de pós-processamento de regras de associação com base nas relações de dependência entre os atributos [Masters Thesis]. Niterói: Programa de Pós-graduação em Computação, Universidade Federal Fluminense; 2006.
- 25 Zaki MJ. Generating non-redundant association rules. In: Proceedings of the sixth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. New York: Association for Computing Machinery; 2000. p. 34-43.
-
26 Hahsler M, Buchta C, Gruen B, Hornik K. arules: mining association rules and frequent itemsets. https://CRAN.R-project.org/packAge=arules (accessed on 15/May/2017).
» https://CRAN.R-project.org/packAge=arules - 27 Hahsler M, Gruen B, Hornik K. arules - a computational environment for mining association rules and frequent item sets. J Stat Softw 2005; 14:1-25.
-
28 Hahsler M. arulesViz: visualizing association rules and frequent itemsets. https://CRAN.R-project.org/package=arulesViz (accessed on 10/Jun/2017).
» https://CRAN.R-project.org/package=arulesViz - 29 Shen X, Yang C, Wu J, Lin S, Gao X, Wu Z, et al. Recurrent tuberculosis in an urban area in China: relapse or exogenous reinfection? Tuberculosis (Edinb.) 2017; 103:97-104.
- 30 Domingos MP, Caiaffa WT, Colosimo EA. Mortality, TB/HIV co-infection, and treatment dropout: predictors of tuberculosis prognosis in Recife, Pernambuco State, Brazil. Cad Saúde Pública 2008; 24:887-96.
- 31 Lefebvre N, Falzon D. Risk factors for death among tuberculosis cases: analysis of European surveillance data. Eur Respir J 2008; 31:1256-60.
- 32 Basta PC, Marques M, Oliveira RL, Cunha EAT, Resendes APC, Souza-Santos R. Desigualdades sociais e tuberculose: análise segundo raça/cor, Mato Grosso do Sul. Rev Saúde Pública 2013; 47:854-64.
- 33 Duarte EC, Bierrenbach AL, Barbosa da Silva Jr. J, Tauil PL, Duarte EF. Factors associated with deaths among pulmonary tuberculosis patients: a case-control study with secondary data. J Epidemiol Community Health 2009; 63:233-8.
- 34 Dooley KE, Tang T, Golub JE, Dorman SE, Cronin W. Impact of diabetes melitus on treatment outcomes of patients with active tuberculosis. Am J Trop Med Hyg 2009; 80:634-9.
- 35 Seiscento M. Tuberculose em situações especiais: HIV, diabetes mellitus e insuficiência renal. Pulmão RJ 2012; 21:23-6.
- 36 Kliiman K, Altraja A. Predictors and mortality associated with treatment default in pulmonary tuberculosis. Int J Tuberc Lung Dis 2010; 14:454-63.
- 37 Christensen A-SH, Roed C, Andersen PH, Andersen AB, Obel N. Long-term mortality in patients with pulmonary and extrapulmonary tuberculosis: a Danish nationwide cohort study. Clin Epidemiol 2014; 6:405-21.
- 38 Nahid P, Jarlsberg LG, Rudoy I, de Jong BC, Unger A, Kawamura LM, et al. Factors associated with mortality in patients with drug-susceptible pulmonary tuberculosis. BMC Infect Dis 2011; 11:1.
- 39 Malta DC, França E, Abreu DX, Oliveira H, Monteiro RA, Sardinha LMV, et al. Atualização da lista de causas de mortes evitáveis (5 a 74 anos de idade) por intervenções do Sistema Único de Saúde do Brasil. Epidemiol Serv Saúde 2011; 20:409-12.
- 40 Oliveira GP, Pinheiro RS, Coeli CM, Barreira D, Codenotti SB. Mortality information system for identifying underreported cases of tuberculosis in Brazil. Rev Bras Epidemiol 2012; 15:468-77.
- 41 Abós-Hernández R, Ollé-Goig JE. Patients hospitalized in Bolivia with pulmonary tuberculosis: risk factors for dying. Int J Tuberc Lung Dis 2002; 6:470-4.
- 42 Enarson DA, Grzybowski S, Dorken E. Failure of diagnosis as a factor in tuberculosis mortality. Can Med Assoc J 1978; 118:1520-2.
- 43 Zachariah R, Spielmann MP, Harries AD, Salaniponi FML. Moderate to severe malnutrition in patients with tuberculosis is a risk factor associated with early death. Trans R Soc Trop Med Hyg 2002; 96:291-4.
- 44 De Vallière S, Barker RD. Poor performance status is associated with early death in patients with pulmonary tuberculosis. Trans R Soc Trop Med Hyg 2006; 100:681-6.
- 45 Oliveira SP, Silveira JTPS, Beraldi-Magalhães F, Oliveira RR, Andrade L, Cardoso RF. Early death by tuberculosis as the underlying cause in a state of Southern Brazil: profile, comorbidities, and associated vulnerabilities. Int J Infect Dis 2019; 80S:S50-7.
- 46 Naidoo P, Dick J, Cooper D. Exploring tuberculosis patients' adherence to treatment regimens and prevention programs at a public health site. Qual Health Res 2009; 19:55-70.
- 47 Cavalin RF, Pellini ACG, Lemos RRG, Sato APS. Coinfecção TB-HIV: distribuição espacial e temporal na maior metrópole brasileira. Rev Saúde Pública 2020; 54:112.
- 48 Campos MR, von Doellinger VR, Mendes LVP, Costa MFS, Pimentel TG, Schramm JMA. Morbidity and mortality associated with injuries: results of the Global Burden of Disease study in Brazil, 2008. Cad Saúde Publica 2015; 31:121-36.
- 49 Araujo EM, Costa MCN, Hogan VK, Mota ELA, Araújo TM, Oliveira NF. Diferenciais de raça/cor da pele em anos potenciais de vida perdidos por causas externas. Rev Saúde Pública 2009; 43:405-12.
- 50 Zaluar A. Integração perversa: pobreza e tráfico de drogas. Rio de Janeiro: Editora FGV; 2004.
- 51 Viana PVS, Paiva NS, Villela DAM, Bastos LS, Bierrenbach ALS, et al. Factors associated with death in patients with tuberculosis in Brazil: competing risks analysis. PLOS One 2020; 15:e0240090.
- 52 Santo AH. Causas múltiplas de morte relacionadas à tuberculose no Estado do Rio de Janeiro entre 1999 e 2001. J Bras Pneumol 2006; 32:544-52.
- 53 Berry MIA, Linoff G. Data mining techniques for marketing, sales, and customer support. New York: John Wiley & Sons; 1997.
The research data are not available.
