Predicting multi-drug resistant tuberculosis using machine learning on genomic and clinical data.

Area of Science:

• Genomics
• Machine Learning
• Medical Diagnostics

Background:

• Tuberculosis (TB) remains a leading global cause of death, disproportionately affecting low- and middle-income countries.
• Multi-Drug Resistant Tuberculosis (MDR-TB) presents significant treatment and control challenges.
• Conventional MDR-TB diagnostics are time-consuming, resource-intensive, and often unavailable in resource-limited settings.

Purpose of the Study:

• To develop and evaluate machine learning models for rapid and accurate detection of MDR-TB.
• To assess the utility of integrating whole-genome sequencing data with clinical factors for MDR-TB diagnosis.

Main Methods:

• Utilized a dataset of approximately 5000 TB patient samples with diverse drug resistance profiles.
• Applied feature selection and normalization to whole-genome sequencing and clinical data.
• Trained and validated various machine learning models, including Gradient Boosting and Deep Neural Networks, using stratified cross-validation.

Main Results:

• Gradient Boosting and Deep Neural Network models achieved high prediction accuracy (92.3% and 93.1%) and AUC-ROC scores (94.7% and 95.4%).
• Combined genomic and clinical data improved model performance compared to genomic data alone.
• Identified key genetic mutations and clinical factors influencing drug resistance through feature importance analysis.

Conclusions:

• Machine learning models integrating genomic and clinical data show significant promise for fast and accurate MDR-TB detection.
• This approach is particularly valuable for improving diagnostics in resource-limited areas.
• Future work should focus on expanding datasets and simplifying models to enhance diagnostic accuracy and clinical applicability.