Evaluating AI-Driven TB Case Finding in Pakistan, a prospective study

Tuberculosis (TB) remains a major public health challenge in Pakistan, one of the highest TB burden countries globally. Despite large-scale efforts, a substantial proportion of TB cases remain undetected each year, highlighting the need for more efficient and targeted approaches to case finding. To address this, EPCON developed the MATCH-AI platform (also referred to as the Epi-control platform), based on the MATCH framework developed by KIT Royal Tropical Institute and implemented in close collaboration with Mercy Corps. The prospective study in Pakistan was conducted by Mercy Corps, with independent evaluation support from the Center for Global Public Health (CGPH).

Study Objective and Approach

The MATCH-AI evaluation (SPOT-TB trial) aimed to assess whether AI-guided site selection could improve TB case detection compared to conventional approaches based on local knowledge, historical data, and field experience.

The study was implemented across 68 districts in Pakistan and embedded within routine programmatic activities. A stepped-wedge trial design was used, in which mobile screening teams gradually transitioned from conventional planning to MATCH-AI–guided targeting.

MATCH-AI applies a Bayesian modeling approach to estimate TB risk and predict areas with higher expected yield. The model integrates multiple data sources, including TB notification data, previous screening results, demographic characteristics, and contextual risk factors, to identify priority locations for screening.

Key Findings

The study demonstrated that AI-guided targeting can improve the efficiency of TB case finding in programmatic settings.

When screening activities were conducted in alignment with MATCH-AI recommendations, results showed an approximate 31% increase in bacteriologically confirmed TB yield per camp compared to conventional site selection approaches. This confirms that more precise geographic targeting can translate into meaningful gains in case detection.

At the same time, results from the intention-to-treat analysis did not show statistically significant differences between the intervention and control groups. This was largely explained by implementation variability, as a proportion of screening camps were conducted outside the recommended locations.

These findings highlight that the impact of AI-driven approaches depends not only on model performance, but also on the degree to which recommendations are operationally followed in the field.

Operational Insights

The evaluation provided important insights into the practical use of AI in public health programs:

• Improved targeting: AI-supported planning enabled identification of high-risk areas that may not be captured through conventional approaches.

• Reduction of bias: The use of algorithm-based recommendations reduced reliance on subjective decision-making.

• System strengthening: The project contributed to improvements in data systems, including the use of electronic case-based surveillance and structured data workflows.

• Capacity building: Training and engagement of implementing partners strengthened local capacity in data use and analysis.

At the same time, operational challenges were identified, including variability in adherence to AI recommendations, data quality constraints, and the need for continued user training and system integration.

Towards a Hybrid Approach

Findings from the study indicate that AI-based targeting is most effective when combined with local expertise. While the model provides a systematic and data-driven approach to identifying high-risk areas, local knowledge remains essential to account for contextual factors such as accessibility, community engagement, and operational feasibility.

A hybrid approach—integrating AI-generated insights with field experience—emerged as the most practical and effective strategy for improving TB case finding.

Implications for Scale-Up

The MATCH-AI evaluation provides evidence that AI-driven targeting can enhance the efficiency of TB programs when implemented within routine systems.

Key considerations for scale-up include:

• Strengthening adherence to AI recommendations

• Continuous improvement of data quality and integration

• Ongoing capacity building for program staff

• Embedding AI tools within national planning and digital health systems

With these elements in place, AI-supported approaches such as MATCH-AI can contribute to closing the TB detection gap and improving the overall effectiveness of national TB responses.

Conclusion

The prospective evaluation of MATCH-AI in Pakistan demonstrates that artificial intelligence can support more targeted and efficient TB case-finding strategies. While implementation challenges remain, the study provides a strong foundation for further integration of AI into public health programs.

EPCON, together with KIT and Mercy Corps, will continue to build on these insights to support data-driven decision-making and strengthen TB programs in high-burden settings.