Clinical protocols, e.g. for malaria quantitation and for differential blood counts, often specify small sample sizes for examination, giving large Poisson error for even a perfectly accurate human. This paper describes the relationship between Poisson error, algorithmic error, and total error in clinical tasks, using malaria as an example. It shows how increasing examined sample size reduces Poisson error, enabling an ML system to offset imperfect accuracy to meet clinical performance requirements.

A version is available online at arXiv

A local copy is here.