Background: Public health agencies suggest targeting "hotspots" to identify individuals with undetected HIV infection. However, definitions of hotspots vary. Little is known about how best to target mobile HIV testing resources.
Methods: We conducted a computer-based tournament to compare the yield of 4 algorithms for mobile HIV testing. Over 180 rounds of play, the algorithms selected 1 of 3 hypothetical zones, each with unknown prevalence of undiagnosed HIV, in which to conduct a fixed number of HIV tests. The algorithms were: 1) Thompson Sampling, an adaptive Bayesian search strategy; 2) Explore-then-Exploit, a strategy that initially draws comparable samples from all zones and then devotes all remaining rounds of play to HIV testing in whichever zone produced the highest observed yield; 3) Retrospection, a strategy using only base prevalence information; and; 4) Clairvoyance, a benchmarking strategy that employs perfect information about HIV prevalence in each zone.
Results: Over 250 tournament runs, Thompson Sampling outperformed Explore-then-Exploit 66% of the time, identifying 15% more cases. Thompson Sampling's superiority persisted in a variety of circumstances examined in the sensitivity analysis. Case detection rates using Thompson Sampling were, on average, within 90% of the benchmark established by Clairvoyance. Retrospection was consistently the poorest performer.
Limitations: We did not consider either selection bias (i.e., the correlation between infection status and the decision to obtain an HIV test) or the costs of relocation to another zone from one round of play to the next.
Conclusions: Adaptive methods like Thompson Sampling for mobile HIV testing are practical and effective, and may have advantages over other commonly used strategies.
Keywords: Bayesian search theory; HIV testing; bandit algorithms.