Applying Machine Learning for the Prediction of Tuberculosis Among People Living with HIV from Health Facilities in Addis Ababa, Ethiopia

Technical Summary: Machine Learning for TB Prediction in PLWHIV in Addis Ababa

Problem Statement
Tuberculosis (TB) remains a leading cause of death globally, particularly among People Living with HIV (PLWHIV). In Ethiopia, TB is the primary cause of death for this demographic, accounting for approximately 40% of AIDS-related deaths. Early diagnosis is hindered by non-specific clinical symptoms, coexisting opportunistic infections, and limitations in conventional screening. While HIV status increases TB risk, it does not fully explain TB occurrence, necessitating the identification of additional clinical, immunological, and socio-demographic determinants. The study aims to develop and validate a machine learning (ML) model to predict TB cases among PLWHIV using routine health facility data to support early risk stratification and targeted intervention.

Methodology
The study employed a quantitative experimental research design utilizing secondary data from four public health facilities in Addis Ababa: Dagmawi Menilik II Hospital, Yekatit 12 Hospital, Zewuditu Memorial Hospital, and Entot 2 Health Center.

• Dataset: The analysis included 16,698 adult patients currently on Antiretroviral Therapy (ART) for at least six months. The dependent variable was the active TB diagnosis status (positive/negative).
• Features: Independent variables included sociodemographic factors (age, sex, weight, marital status, etc.), clinical factors (months on ART, WHO stage, functional status), hematological/immunological markers (CD4 count, viral load), and treatment factors (ART regimen, Tuberculosis Preventive Therapy [TPT], Cotrimoxazole Prophylaxis [CPT]).
• Data Preprocessing:
  • Cleaning: Eight features with >50% missing data (e.g., religion, occupation, viral load) were dropped. Missing values in numerical features (age, CD4) were imputed with the mean, and ART dose with the median. Categorical missing values were imputed using the mode.
  • Feature Selection: Correlation analysis identified high positive correlations between "days since HIV diagnosis" and "months on ART" ($r=0.81$), and between "age" and "months on ART" ($r=0.65$). Consequently, "days since HIV diagnosis" and "ART dose" were removed to improve model efficiency.
  • Scaling and Encoding: Categorical variables were label-encoded, and numerical features were normalized using Min-Max scaling (0 to 1).
  • Handling Imbalance: Given the high imbalance in the data (98% negative cases), the study evaluated various balancing techniques. The Hybrid method combining Synthetic Minority Over-Sampling Technique (SMOTE) with Edited Nearest Neighbors (ENN) was selected as the optimal approach.
  • Data Splitting: The dataset was split into 70% training, 15% validation, and 15% testing.
• Models: Eight algorithms were evaluated: Logistic Regression (LR), Decision Trees (DT), Random Forest (RF), XGBoost (XGClassifier), Support Vector Machines (SVM), k-Nearest Neighbors (k-NN), Multilayer Perceptron (MLP), and Convolutional Neural Network (CNN).
• Evaluation Metrics: Performance was assessed using Accuracy, Precision, Recall, F1 Score, F-Beta (with $\beta=2$ to weight recall higher), Matthews Correlation Coefficient (MCC), and Area Under the Receiver Operating Characteristic Curve (AUC).

Key Results

• Model Performance: No single model outperformed all others across every metric, but XGBoost demonstrated the strongest overall performance.
  • XGBoost Metrics: Accuracy (98%), Recall (80%), Precision (25%), F1 Score (38%), F-Beta (55%), MCC (52%), and AUC (96%).
  • Comparative Performance: Random Forest achieved high accuracy (97%) and AUC (95%). SVM and Logistic Regression showed higher recall (92% and 88%, respectively) but lower precision.
  • Precision vs. Recall: The study notes that while XGBoost had lower precision (25%), it balanced precision and recall effectively (F1: 38%) and maintained a high AUC, indicating strong class discrimination.
• Feature Importance (SHAP Analysis): Using SHAP (SHapley Additive exPlanations) values on the XGBoost model, the strongest predictors for TB occurrence or protection were identified as:
  1. ART Regimen (Mean absolute SHAP value: 1.4)
  2. Taking Tuberculosis Preventive Therapy (TPT) (1.2)
  3. Cotrimoxazole Prophylaxis (1.1)
  4. Months on ART
  5. Patient Weight
  6. Functional Status
  7. Sex, Age, and CD4 count followed in importance.

Significance and Claims
The paper claims that XGBoost is a viable tool for predicting TB cases among PLWHIV in the Addis Ababa context, outperforming other models in accuracy, AUC, and MCC. The study highlights that specific treatment factors—particularly the ART regimen, TPT initiation, and Cotrimoxazole prophylaxis—are critical determinants of TB risk, aligning with global WHO recommendations and previous regional studies (e.g., Debre Markos, China).

The authors position the study as a step toward improving early detection and targeted screening. By identifying complex, non-linear patterns in large-scale clinical data, the proposed ML approach aims to support risk stratification and timely intervention, potentially reducing TB-related morbidity and mortality in this vulnerable population. The study emphasizes that while precision remains a challenge due to data imbalance, the high recall and AUC of the XGBoost model make it suitable for identifying high-risk patients where missing a case (false negative) could be fatal.

Limitations and Modesty
The authors acknowledge that the low precision scores across models are likely a result of the highly imbalanced dataset. They note that while the model performs well on the available data, further testing with additional features and higher-quality data is required to refine precision. The study does not claim the model is a standalone diagnostic tool but rather a predictive support system for risk stratification.