Development of CART model for prediction of tuberculosis treatment loss to follow up in the state of São Paulo, Brazil: A case–control study

doi:10.1016/j.ijmedinf.2020.104198

International Journal of Medical Informatics

Volume 141, September 2020, 104198

https://doi.org/10.1016/j.ijmedinf.2020.104198 Get rights and content

Highlights

•
The CART model develop in this study can predict tuberculosis treatment loss to follow up.
•
As the model is a decision tree it can be used by healthcare professionals to prevent loss to follow up.
•
The Model had an accuracy of 0.76, F-measure of 0.77, sensitivity of 0.80 and specificity of 0.71.
•
The Model emphasizes the relation between several variables already identified in previous studies as related to the patient treatment loss to follow up or cure in the tuberculosis treatment.

Abstract

Background

Tuberculosis is the leading cause of infectious disease-related death, surpassing even the immunodeficiency virus. Treatment loss to follow up and irregular medication use contribute to persistent morbidity and mortality. This increases bacillus drug resistance and has a negative impact on disease control.

Objective

This study aims to develop a computational model that predicts the loss to follow up treatment in tuberculosis patients, thereby increasing treatment adherence and cure, reducing efforts regarding treatment relapses and decreasing disease spread.

Methods

This is a case-controlled study. Included in the data set were 103,846 tuberculosis cases from the state of São Paulo. They were collected using the TBWEB, an information system used as a tuberculosis treatment monitor, containing samples from 2006 to 2016. This set was later resampled into 6 segments with a 1-1 ratio. This ratio was used to avoid any bias during the model construction.

Results

The Classification and Regression Trees were used as the prediction model. Training and test sets accounted for 70% in the former and 30% in the latter of the tuberculosis cases. The model displayed an accuracy of 0.76, F-measure of 0.77, sensitivity of 0.80 and specificity of 0.71. The model emphasizes the relationship between several variables that had been identified in previous studies as related to patient cure or loss to follow up treatment in tuberculosis patients.

Conclusion

It was possible to construct a predictive model for loss to follow up treatment in tuberculosis patients using Classification and Regression Trees. Although the fact that the ideal predictive ability was not achieved, it seems reasonable to propose the use of Classification and Regression Trees models to predict likelihood of treatment follow up to support healthcare professionals in minimising the loss to follow up.

Introduction

Tuberculosis (TB) is the leading cause of infectious disease-related death, surpassing the immunodeficiency virus (HIV) [1]. TB control is based on early and appropriate diagnosis and treatment, interrupting the transmission chain [2]. The National Tuberculosis Control Plan (PNCT) is a strategy implemented by the Brazilian government. The Directly Observed Treatment Short-Course (DOTS), is a part of this strategy and has assisted in the country's reach to an 85% cure rate and has reduced the loss to follow up rate in TB treatments by 5% [2]. However, loss to follow up treatment and irregular medication use contribute to the persistence of morbidity and mortality, increasing drug resistance cases [3], [4].

TB is related to poverty and different social causes. Therefore, patients are difficult to monitor. Factors such as initial improvement, lack of patient follow-up and monitoring are all associated with poor treatment adherence [5]. After the first few months of using the correct medications during treatments, patients present improvements in their overall physical well-being and most of their symptoms are gone, which then leads to patients interrupting the treatment themselves, even though they are not cured. Social factors, treatment methodology and provided health services are factors associated with treatment loss to follow up [6], [7], [8].

In the last 10 years, computational techniques for predicting the outcomes of TB treatment have been explored. Neural networks have been applied to predict the TB treatment loss to follow up in some cities of Espírito Santo in Brazil [9]. A logistic regression model was used to select features [10] and predict loss to follow up TB treatment [5]. Another study compared the application of different machine learning methods to classify treatment outcomes, among which the C4.5 decision tree showed better performance [11].

With a computational model that is able to identify patients that are prone to not follow-up treatment, teams would then be able to focus their efforts in reducing the rate of relapse, dissemination and loss to follow-up treatment. Although there are studies about the subject [9], [10], [5], [11], no study has developed an effective human-readable model that can be applied by health professionals in TB treatment control. Decision trees can be represented as a set of rules that would assist healthcare professionals, the ones with and the ones without access to an information system, to control TB. Thus, this study aims to develop a CART predictive model to predict TB treatment loss to follow up cases.

Section snippets

Materials and methods

For the development of this case-control study, 208,620 TB treatment cases were used, all between the years of 2006 and 2016 and in the state of São Paulo. The sample had sensitive and drug-resistant tuberculosis cases. The state of São Paulo is one of 27 federal states of Brazil and is located in the southeast region of the country. It has 645 cities and an area of 248,219,491 km² [12]. It has the greatest population in the country, with around 45.9 million inhabitants [12].

The inclusion

Results

The final population of the study contained 91,823 controls and 12,023 cases (Fig. Figure 1 ). Table Table 2 shows the distribution for the population demographic data.

Most of the patients are male adults between the ages of 20 and 39 (24.67%). Less than 2% of the population represents the youth and elders.

The sampled population is mainly composed of white (39.84%) and brown (26.69%) ethnic groups. Only 0.52% of them were from an indigenous ethnic group. Most patients had basic education (less

Discussion

The decision tree created can be used as a guideline among health professionals to detect treatment loss to follow up. The accuracy of the model is similar to other models already presented in the literature [11]. For health professionals, the decision tree is a much more intuitive and understandable approach than a statistical model, since it can be visualized and explored by anyone, as well as be applied without a computational system. To exemplify, the decision tree for this study can be

Conclusion

It was possible to build the predictive model of treatment loss to follow up of tuberculosis through CART. During model development, new features for TB treatment loss to follow up predictors were identified. Although the fact that the ideal predictive ability was not achieved, it seems reasonable to propose the use of CART models. We aim in future studies to improve the model accuracy and implement the model in an informational system for tuberculosis treatment and management.

Summary Points

What was already known on the topic:

•
Treatment loss to follow up and irregular medication use contribute to increased bacillus drug resistance and has a negative impact on disease control.
•
Other studies apply different models to predict the TB treatment outcome.
•
There is evidence that, among different models, decision trees present better performance.

What this study added to our knowledge:

•
The CART model was able to perform equally or better than other models in the literature, additionally

Acknowledgements

We would like to thank the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior – Brasil (CAPES) – Finance Code 001 and São Paulo Research Foundation (FAPESP) (Grant Nos. 2018/23963-2 and 2018/00307-2).

References (15)

WHO
Global Tuberculosis Report 2019, Geneva
(2019)
Ministério da Saúde
Plano Nacional pelo Fim da Tuberculose como Problema de Saúde Pública
(2017)
L.M.M. Paixão et al.
Perfil de casos de tuberculose notificados e fatores associados ao abandono, Belo Horizonte, MG
Rev. Saúde Pública
(2007)
K.M.J. de Souza
Abandono do tratamento da tuberculose na Atenção Primária à Saúde: uma análise segundo o enfoque familiar do cuidado
(2008)
E.d.A. Silva et al.
Modelo preditivo ao abandono do tratamento da tuberculose
Saúde em Debate
(2014)
H.B. De Oliveira et al.
Abandono de tratamento e recidiva da tuberculose: Aspectos de episódios prévios, Campinas, SP, Brasil, 1993–1994
Rev. Saude Publica
(2000)
M.F. Rabahi et al.
Noncompliance with tuberculosis treatment by patients at a tuberculosis and AIDS reference hospital in midwestern Brazil
Braz. J. Infect. Dis.
(2002)

There are more references available in the full text version of this article.

Cited by (4)

Clinical pathways and hierarchical clustering for tuberculosis treatment outcome prediction
2023, Procedia Computer Science
Clinical pathways are chronological event series that happen throughout a patient's treatment. They can be extracted from the Electronic Health Record medical information and this can be used to correlate the pathway to possible healthcare outcomes. This can be applied to a wide variety of diseases to point pathways related to bad outcomes. These pathways can be audited and patients that start to follow such patterns can be put in special observation and care. Tuberculosis (TB) is one of the leading causes of death through infectious disease and its control is based on search for cases, accurate and premature identification, and treatment. The use of the aforementioned method can help in disease control and premature identifications of bad outcomes for ongoing treatments. Therefore, the current study goals are: 1) identify the existing clinical pathways; 2) group these pathways using hierarchical clustering; 3) create a classification model based on the generated clusters to predict bad outcomes. The dataset used consisted of 277,870 TB treatment cases from the state of São Paulo collected through TBWEB, a information system for monitoring and follow-up of TB cases. All cases with ongoing treatment were excluded from the study and the resulting dataset was splitted in training and test samples. To reduce bias due to imbalance the undersampling technique was applied to the training dataset resulting in a final sample size of 90,184. The test dataset had a size of 52,639 cases. Both datasets had 16 attributes describing the patient diagnosis and drug scheme evolution through the treatment. All attributes unique values were mapped and a representation character was assigned to each one. Later, these representation characters were concatenated in the chronological order of the events and diagnosis creating a representational string for the clinical pathway. The resulting pathways of the training dataset were used to build the clusters which were later used to build the classifier to predict the treatment outcome based on the test dataset clinical pathways. The final model overall accuracy is at 0.829. The model showed a significant improvement of accuracy from previous studies and had similar or better performance than others in the literature. We believe this model can be implemented to a informational system to further improve treatments management and tuberculosis control.
An information system for monitoring tuberculosis cases: Implementation research protocol using RE-AIM for a health region in Brazil
2023, Procedia Computer Science
Tuberculosis is an infectious bacterial disease and one of the biggest public health problems in the world despite being curable. Research is still needed that considers operational aspects of treatment and control of its spread. Considering this scenario, the need to develop software that provides real-time information about the patient's path along health information systems for clinical decision-making is fundamental. Because of this, a region made up of 26 municipalities will implement a computerized system to monitor tuberculosis cases, the SISTB. However, the implementation of new technology can result in the creation of new functions and cause changes in the existing ones. Implementation research can be used as an approach where evidence can be provided to guide the use of digital technology in tuberculosis care. The RE-AIM framework provides a model to inform this research. Thus, this work aims to report the protocol of a study to evaluate the implementation of a health information system to assist in the monitoring of tuberculosis treatment, using the RE-AIM. The implementation will be carried out in 5 phases: i) define the locations that will receive the intervention; ii) prepare people to receive the intervention iii) train key people; iv) adapt the SISTB and/or the training according to the discussions in the previous phase; v) follow up and monitor the support of the technology. The data collected for the evaluation will come from the database of the implemented system, questionnaires, and training meetings. The forecast for the study conclusion is the end of 2023. As a result, we hope that the SISTB implementation will increase the positive outcomes of tuberculosis patients' treatment.
Preliminary results to predict tuberculosis outcomes applying traditional and automated machine learning models
2023, Procedia Computer Science
Tuberculosis (TB) remains one of the most lethal infectious diseases in the world and, despite being preventable and curable, kills 4.500 people daily, according to the World Health Organization (WHO). Brazil, being a country heavily affected by TB, works to improve social intervention programs, since the decrease in the patients vulnerability seems to have a positive effect for the cure of TB. The Brazilian public health system records data on TB treatment that can guide actions and interventions. In this context, machine learning (ML) algorithms have been used successfully to analyze health and medicine (H&M) datasets. An emerging area of ML called Automated Machine Learning (Auto-ML) was tested in this analysis to predict the following TB results: good and bad outcomes. Our results indicate that it is possible to build reasonable ML models with the available data.
T-BACCO SCORE: A predictive scoring tool for tuberculosis (TB) loss to follow-up among TB smokers
2023, PLoS ONE

View full text

Development of CART model for prediction of tuberculosis treatment loss to follow up in the state of São Paulo, Brazil: A case–control study

Highlights

Abstract

Background

Objective

Methods

Results

Conclusion

Introduction

Section snippets

Materials and methods

Results

Discussion

Conclusion

Summary Points

Acknowledgements

Global Tuberculosis Report 2019, Geneva

Plano Nacional pelo Fim da Tuberculose como Problema de Saúde Pública

Perfil de casos de tuberculose notificados e fatores associados ao abandono, Belo Horizonte, MG

Rev. Saúde Pública

Abandono do tratamento da tuberculose na Atenção Primária à Saúde: uma análise segundo o enfoque familiar do cuidado

Modelo preditivo ao abandono do tratamento da tuberculose

Saúde em Debate

Abandono de tratamento e recidiva da tuberculose: Aspectos de episódios prévios, Campinas, SP, Brasil, 1993–1994

Rev. Saude Publica

Noncompliance with tuberculosis treatment by patients at a tuberculosis and AIDS reference hospital in midwestern Brazil

Braz. J. Infect. Dis.