Predicting communities with high tuberculosis case-finding efficiency to optimise resource allocation in Pakistan:...

Area of Science:

• Epidemiology
• Public Health
• Biostatistics

Background:

• An estimated 183,000 tuberculosis (TB) cases may remain undiagnosed in Pakistan annually.
• Targeting active case finding (ACF) to high-risk populations is crucial for improving TB diagnosis rates.
• Developing predictive models can optimize the allocation of resources for TB control.

Purpose of the Study:

• To compare the predictive accuracy of a Negative Binomial Regression (NBR) model with a Bayesian Machine Learning (BML) model.
• To identify predictors of TB positivity in active case finding (ACF) settings in Pakistan.
• To evaluate the utility of statistical models for targeting TB hotspots.

Main Methods:

• Retrospective analysis of cross-sectional data from 414 ACF events (September 2020 - January 2022).
• Development of a Negative Binomial Regression (NBR) model incorporating spatial autocorrelation.
• Comparison of NBR and BML models using Root Mean Square Error (RMSE) and Akaike Information Criterion (AIC).

Main Results:

• 407 (1.9%) bacteriologically confirmed TB cases were detected among 21,227 visitors.
• Spatial lag variables significantly explained variation in TB positivity rates within the NBR model.
• Both NBR and BML models showed similar predictive performance at the sub-district level, with NBR having a slightly better fit (AIC).

Conclusions:

• Statistical models are effective for predicting TB hotspots to guide ACF planning.
• A simpler NBR model offers comparable performance to a more complex BML model for TB prediction.
• Model predictions are robust across different frameworks, supporting their use for targeted ACF in underserved areas.