The study

Between 15 May and 30 November 2025, six NHS microbiology laboratories ran Halo AI on positive blood cultures from their neonatal units. It was a prospective, multicentre diagnostic accuracy study. Halo AI ran as a laboratory prototype and was compared with ISO 20776-1 broth microdilution, read against EUCAST breakpoints. Clinicians did not see its results.

Rose EG, Mensah DK, Raman P, Fenwick A, Okoro A. Single-cell phenotypic antimicrobial susceptibility testing of neonatal positive blood cultures in under two hours: a prospective, multicentre diagnostic accuracy study. Preprint, January 2026.

Download the paper (PDF) (opens in a new tab)

Headline figures

1,214 cultures 4,842 organism–antibiotic pairs

96.8%

Categorical agreement 95% CI 96.3–97.3

0% 50% 100%
0.7%

False-susceptible results 8 of 1,11695% CI 0.4–1.4 Halo AI called a resistant organism susceptible. The paper uses the testing standard’s term for this, very major errors. The whole confidence interval is below the accepted limit of 1.5%.

Accepted limit, 1.5%

0% 0.5% 1% 1.5% 2%
1.6%

False-resistant results 60 of 3,726 Halo AI called a susceptible organism resistant.

0% 0.5% 1% 1.5% 2%

Agreement by organism

Few *Klebsiella* and *Enterobacter* isolates were resistant, so the false-susceptible rates for those two organisms are imprecise. No group B streptococcus isolate was resistant to any of the antibiotics tested.
Organism	Pairs	Agreement	False susceptible	False resistant
Coagulase-negative staphylococci	1,648	96.2%	4 of 721	22 of 927
S. aureus	672	97.3%	2 of 152	7 of 520
Group B streptococcus	588	98.0%	none resistant	4 of 588
E. coli	1,270	97.3%	2 of 200	14 of 1,070
Klebsiella	448	95.8%	0 of 35	8 of 413
Enterobacter	216	95.4%	0 of 8	5 of 208
All organisms	4,842	96.8%	8 of 1,116	60 of 3,726

Time to result

Each bar is the interquartile range and the line across it is the median, in minutes from loading the cartridge. The red edge marks 2 hours.

Organism	Median (IQR)	Reported within 2 h
Coagulase-negative staphylococci	102 min (89–118)	83.3%
S. aureus	74 min (65–86)	98.2%
Group B streptococcus	86 min (76–97)	95.9%
E. coli	52 min (46–61)	100%
Klebsiella	58 min (51–67)	100%
Enterobacter	63 min (55–72)	100%
All organisms	88 min (63–104)	93.4%

Compared with routine testing

From the moment a culture flagged positive, routine testing at the study sites took a median of 46 hours to give a result. Halo AI took a median of 88 minutes from loading the cartridge.

Routine testing 46hours

Halo AI 88minutes

Routine testing is counted from the moment the culture flagged positive, and Halo AI from loading the cartridge. The bars are to scale.

Limitations

We did not report results to clinicians, so the study can’t show any effect on treatment or outcomes. That will take a randomised evaluation, which is the next study we plan.
Coagulase-negative staphylococci are common contaminants, so some of the cultures we included will not have been real infections.
Few Klebsiella and Enterobacter isolates were resistant, so the error estimates for those two organisms are imprecise.
Every site was in England, where resistance is lower than in the countries where neonatal sepsis kills the most babies.
The study ran on a laboratory prototype. The benchtop Halo Reader has not been built yet, and when it is we will have to confirm that it performs as the prototype did.

Bring Halo AI to your lab Download the paper (PDF) (opens in a new tab)