Current methods for the diagnosis of sepsis have insufficient precision, causing regular misdiagnoses. Microbiological tests can help to diagnose sepsis, but are usually too slow to have an impact on timely clinical decision-making. Neutrophils have a high sensitivity to infections, yet measurements of neutrophil surface markers, genomic changes and phenotype alterations have had only a marginal effect on sepsis diagnosis. Here, we report a microfluidic assay that measures, from one droplet of diluted blood, the spontaneous motility of neutrophils in the presence of plasma. We measured the performance of the assay in two independent cohorts of critically ill patients suspected of sepsis. Using data from a first cohort, we developed a machine-learning-based scoring system (sepsis score) that segregated patients with sepsis from those without sepsis. We then validated the sepsis score in a double-blind, prospective case–control study. For the 42 patients across the two cohorts, the assay identified sepsis patients with 97% sensitivity and 98% specificity. The neutrophil assay could potentially be used to accurately diagnose and monitor sepsis in larger populations of at-risk patients.