Predicting Tuberculosis Incidence in Adult HIV Patients on ART in Debre Markos, Ethiopia: A Machine Learning Approach

Tuberculosis (TB) is the commonest comorbidity among individuals with HIV/AIDS, especially in low- and middle-income nations such as Ethiopia. Early diagnosis of TB infection in HIV-infected patients is crucial for effective management of opportunistic infections that can result in mortality. Early identification of TB in HIV/AIDS patients plays a significant role in reducing morbidity and mortality. Machine learning algorithms have a significant role in detecting TB occurrences among HIV/AIDS patients.

In this study, we used 5,392 HIV-infected individuals’ medical records. Techniques such as SMOTE and ADASYN were employed to adjust data imbalance between positive and negative TB status. Random forest, decision tree, logistic regression, gradient boosting, K-nearest neighbors, and XGBoost were evaluated to predict TB incidence.

Of the total records, 3,440 (63.8%) were female patients, and the remaining 1,952 (36.2%) were male patients. 3,715 (68.9%) were labeled as green records addresses, while 1,677 (31.1%) had yellow records. The XGBoost algorithm is the best-performing model to predict TB incidence. Among the features included in this study is the most important classifier for feature selection. Among all the features, CD4 count and patient age were found to be the most important predictors of TB incidence among adult HIV patients.

This study demonstrates that the XGBoost model was the most effective model for predicting tuberculosis incidence among HIV patients, utilizing features such as low CD4 counts, age, duration on ART, weight, sex, and WHO clinical stage.

Author Summary

Tuberculosis is the main opportunistic infection among HIV-infected individuals. The comorbidities of HIV and TB increase the risk of mortality among HIV patients. Early detection of TB infection from people living with HIV is a crucial step for increasing HIV patients’ life expectancy. A machine learning approach plays a vital role in detecting TB infection among HIV patients. This study aims to predict tuberculosis (TB) occurrence in adult HIV patients on antiretroviral medication (ART) in Debre Markos City, Ethiopia. Using a retrospective dataset of 5,392 patients, researchers compared seven methods. The XGBoost model fared best, earning 82% accuracy and a 90% AUC after resolving class imbalance using MOTE+ENN. Key predictors revealed were low CD4 count, age, time on ART, sex, WHO clinical stage, address status, DSD category, and TB preventative treatment (TPT status). Machine learning models will help health providers to predict the risk of TB infection and make early intervention in high TB-HIV co-infection loads.