Machine learning model development for predicting pulmonary tuberculosis in symptomatic patients from Ethiopia’s Mid-Rift Valley

Tuberculosis continues to pose a serious health threat to high-burden countries, including Ethiopia. Complex interactions among multiple risk factors challenge disease control. Conventional epidemiological methods often fail to capture the complex interactions among risk factors; as a result, targeted interventions to control and prevent the disease are often less effective. The present study attempted to develop and test eight supervised machine learning models to address the methodological gap using comprehensive data collected from tuberculosis-suspected individuals attending public hospitals in the Mid-Rift Valley area of Ethiopia. A cross-sectional study using consecutive sampling was carried out between December 30, 2023, and January 2025 at five hospitals in the Mid-Rift Valley, Ethiopia. The required sociodemographic, clinical, and behavioral data, along with sputum samples, were collected from study subjects with signs and symptoms suggestive of TB using a structured questionnaire. Sputum samples collected from 2025 individuals suspected of having TB at OPDs were tested using the GeneXpert MTB/RIF assay for TB diagnosis. Data were split into training (70%), validation (15%), and test (15%) sets before applying the models. Eight supervised ML models were trained and tuned. Model performance was assessed using metrics: Area Under the Curve (AUC), accuracy, sensitivity, and specificity. 95% confidence intervals were computed using bootstrap resampling (1000 replicates; percentile interval method) for AUC and F1-score, and the Wilson score interval for binomial proportions. We interpreted the best model using SHAP (SHapley Additive exPlanations) and permutation importance. The data were analyzed using Jupyter Notebook, Python Version 3.12.4. Tuberculosis was diagnosed in 280 (13.8%) of the 2025 participants in the study. The Random Forest model performed best at predicting tuberculosis positivity. The final test set showed an AUC of 0.94 (95% CI 0.90–0.97), with an accuracy of 0.88 (95% CI 0.84–0.91), sensitivity of 0.83 (95% CI 0.69–0.92), and specificity of 0.89 (95% CI 0.84–0.92). HIV status, age, Body Mass Index (BMI), Bacillus Calmette-Guérin (BCG) vaccination history, night sweats, and weight loss were identified as strong predictors of tuberculosis disease in the Random Forest model. The Random Forest algorithm showed the best performance in predicting tuberculosis status, based on actual patient data in a high-burden setting. It also showed good promise as a clinical decision-support tool. However, further external validation in independent populations is required before routine clinical implementation.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-13
DOI
https://doi.org/10.1038/s41598-026-78644-6
Primary Topic
Tuberculosis Research and Epidemiology
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Machine learning model development for predicting pulmonary tuberculosis in symptomatic patients from Ethiopia’s Mid-Rift Valley

Ketema Tafess, Dereje Bekele, Getu Diriba, Dawit Wolday et al.
1 links
Scientific Reports
Tuberculosis Research and Epidemiology
article

Machine learning model development for predicting pulmonary tuberculosis in symptomatic patients from Ethiopia’s Mid-Rift Valley

Ketema Tafess, Dereje Bekele, Getu Diriba, Dawit Wolday, Melat Mezemir, Gadissa Gutema, Desye Melese Sisay, Gilman Siu, Mujib Abdulkadir, Tobias F. Rinke de Wit, Helen Nigussie, Tariku Dejene
article en

Abstract

Tuberculosis continues to pose a serious health threat to high-burden countries, including Ethiopia. Complex interactions among multiple risk factors challenge disease control. Conventional epidemiological methods often fail to capture the complex interactions among risk factors; as a result, targeted interventions to control and prevent the disease are often less effective. The present study attempted to develop and test eight supervised machine learning models to address the methodological gap using comprehensive data collected from tuberculosis-suspected individuals attending public hospitals in the Mid-Rift Valley area of Ethiopia. A cross-sectional study using consecutive sampling was carried out between December 30, 2023, and January 2025 at five hospitals in the Mid-Rift Valley, Ethiopia. The required sociodemographic, clinical, and behavioral data, along with sputum samples, were collected from study subjects with signs and symptoms suggestive of TB using a structured questionnaire. Sputum samples collected from 2025 individuals suspected of having TB at OPDs were tested using the GeneXpert MTB/RIF assay for TB diagnosis. Data were split into training (70%), validation (15%), and test (15%) sets before applying the models. Eight supervised ML models were trained and tuned. Model performance was assessed using metrics: Area Under the Curve (AUC), accuracy, sensitivity, and specificity. 95% confidence intervals were computed using bootstrap resampling (1000 replicates; percentile interval method) for AUC and F1-score, and the Wilson score interval for binomial proportions. We interpreted the best model using SHAP (SHapley Additive exPlanations) and permutation importance. The data were analyzed using Jupyter Notebook, Python Version 3.12.4. Tuberculosis was diagnosed in 280 (13.8%) of the 2025 participants in the study. The Random Forest model performed best at predicting tuberculosis positivity. The final test set showed an AUC of 0.94 (95% CI 0.90–0.97), with an accuracy of 0.88 (95% CI 0.84–0.91), sensitivity of 0.83 (95% CI 0.69–0.92), and specificity of 0.89 (95% CI 0.84–0.92). HIV status, age, Body Mass Index (BMI), Bacillus Calmette-Guérin (BCG) vaccination history, night sweats, and weight loss were identified as strong predictors of tuberculosis disease in the Random Forest model. The Random Forest algorithm showed the best performance in predicting tuberculosis status, based on actual patient data in a high-burden setting. It also showed good promise as a clinical decision-support tool. However, further external validation in independent populations is required before routine clinical implementation.

Scientific Reports
Hong Kong Polytechnic University (HK), Ethiopian Public Health Institute (ET), Wollo University (ET), Amsterdam Institute for Global Health and Development (NL), Addis Ababa City Administration Health Bureau (ET), Oromiyaa Regional Health Bureau (ET), Addis Ababa University (ET), Adama Science and Technology University (ET), McMaster University (CA)
Addis Ababa University
Good health and well-being
Tuberculosis Research and Epidemiology
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon