Machine learning to improve HIV screening using routine data in Kenya.

Area of Science:

• Public Health
• Machine Learning
• Epidemiology

Background:

• Optimal HIV testing resource allocation is crucial for global HIV control.
• Kenya's national HIV testing positivity rate is 2.8% via the electronic medical record (EMR) system.
• Machine learning (ML) shows promise for identifying individuals with undiagnosed HIV for targeted testing.

Purpose of the Study:

• To apply ML to routine EMR data in Kenya to predict undiagnosed HIV positivity.
• To develop a real-time clinical decision support system for prioritizing HIV testing.

Main Methods:

• Utilized de-identified EMR data from 167,509 individuals tested between June-November 2022.
• Included demographics, clinical history, behavioral data, and population-level data; addressed missing data with multiple imputations.
• Trained and evaluated four ML algorithms (logistic regression, Random Forest, AdaBoost, XGBoost) using Area Under the Precision-Recall Curve (AUCPR).

Main Results:

• All ML models surpassed the current HIV testing positivity rate.
• XGBoost achieved the highest AUCPR, demonstrating a 10.5-fold improvement over the baseline positivity rate.

Conclusions:

• ML applied to routine HIV testing data can serve as an effective clinical decision support tool.
• The developed ML model can be integrated into EMR systems for real-time testing decision support.
• Data quality and missing data challenges can be mitigated with robust data preparation techniques.