Bring Halo AI to your lab

The world’s fastest antibiotic testing for neonatal units.

Halo AI tests a positive blood culture against the 8 antibiotics used in neonatal units and gives a susceptibility result in about 90 minutes. Standard testing takes two days.

We have just raised £750k in seed funding, led by Kestrel Point Capital. The test works on a laboratory prototype, which we ran on 1,214 blood cultures in 6 NHS labs. We’re now building the benchtop Halo Reader for the first hospital pilots and hiring our first four people.

Bring Halo AI to your lab Read the study We’re hiring

Amoxicillin still dividing

Gentamicin stopped

E. coli from a positive blood culture, 38 minutes into a run. In amoxicillin the cells are still dividing. In gentamicin they have stopped. This was the fastest result in our study.
  • Still dividing: the antibiotic isn’t working.
  • Stalled and swelling: the antibiotic is working.

Run time from loading the cartridge, compared with standard testing. The bars are to scale.

It takes two days to find out whether a newborn’s antibiotics are working.

A baby born six weeks early develops a fever on day five, and overnight her blood culture flags positive. The team already has her on standard antibiotics. Whether those are the right ones won’t be known for up to two more days, because standard testing needs a day to grow a pure culture and then another night to run the test.

  • 3 million newborns develop sepsis each year.
  • Up to 570,000 of them die.
  • An estimated 214,000 of those deaths are attributed to resistant bacteria.

Each dot is 10,000 newborns a year.

Median time to a susceptibility result at our six study sites.

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.

Sources: Bulletin of the WHO, 2022; PLOS Medicine meta-analysis, 2021.

Built for the bacteria that infect newborns

Most rapid susceptibility systems are built for adult infections and test Gram-negative bacteria only. In newborns a lot of sepsis is Gram-positive: group B strep in the first days of life, staphylococci after that. We trained Halo AI on both.

Gram-positive, validated

  • Group B streptococcus
  • Staphylococcus aureus
  • Coagulase-negative staphylococci

Gram-negative, validated

  • E. coli
  • Klebsiella
  • Enterobacter

How the AI reads a channel

The software does two jobs. A convolutional neural network finds every cell in every image and follows it from one image to the next. A second model then looks at what the cells in each channel did, and classes the channel as still dividing or stopped.

The two channels from the top of the page, as the software reads them. Each box is a cell it has found, and each line is a cell it has followed from one image to the next.

  1. Find

    The first model marks every cell in each image. It was trained on cells of different shapes, including the chains and clusters that Gram-positive cocci form.

  2. Follow

    It links each cell to the same cell in the next image, taken two minutes later, and to its two daughters when it divides.

  3. Decide

    The second model reads the tracks from each channel and classes the channel as still dividing or stopped. Where it can’t tell, Halo AI gives no call and recommends standard testing.

How it was trained and tested

  • 768 combinations filmed

    The method began in the Okoro Laboratory at the Sir William Dunn School of Pathology in Oxford, with films of 96 E. coli isolates against 8 antibiotics.

  • 6 groups of bacteria

    The model was then trained on the bacteria that infect newborns: Group B streptococcus, the staphylococci, E. coli, Klebsiella and Enterobacter.

  • 1,214 cultures it had never seen

    Version 2.3.1 of the model was locked before our study in six NHS labs began, so every culture in that study was new to it. Later versions will be locked and checked against a frozen test set before release.

How it works

A render of the Halo Reader as designed: a light grey instrument with a graphite front plate, a cartridge slot and a small display.

  1. The instrument

    The Halo Reader

    A benchtop instrument that takes a single-use cartridge. The test itself already runs on a laboratory prototype, and we are now building the first Halo Readers.

  2. Inside

    Inside are phase-contrast optics and a camera, a heater that holds the cartridge at 37°C, a GPU that runs the AI on board, and the connection to your laboratory information system.

  3. On the bench

    The Halo Reader is designed to sit next to your blood culture incubators and run 24 hours a day.

  4. The cartridge

    The cartridge dilutes the sample, removes most of the blood cells and spreads the bacteria across about 500 microchannels. Eight sets of channels each hold one antibiotic. A ninth set has none and is the growth control.

  5. Running a test

    Load

    Pipette 50 µL of broth from the positive bottle into a cartridge and load it. Hands-on time is about 3 minutes.

  6. Watch

    The Halo Reader keeps the cartridge at 37°C and photographs every channel once every 2 minutes.

  7. Read

    The AI follows each cell from one image to the next. If the cells in a channel keep dividing, that antibiotic isn’t working. If they stall and swell, it is.

  8. Report

    Halo AI runs alongside the mass spectrometry or molecular panel your lab already uses to identify the organism, which takes 35 to 65 minutes. It then applies the EUCAST breakpoints for that organism and sends the result to your laboratory information system.

Antibiotics on the panel: Benzylpenicillin, amoxicillin, flucloxacillin, gentamicin, cefotaxime, vancomycin, meropenem, amikacin.

Tested on 1,214 blood cultures from newborns in 6 NHS labs.

See the full results Bring Halo AI to your lab

Roadmap

  1. Now

    Proven technology

    The cartridge, the imaging and the AI all work. On a laboratory prototype in 6 NHS labs they agreed with reference testing on 96.8% of results from 1,214 newborn blood cultures. We have designed the benchtop Halo Reader but haven’t built one yet.

  2. Next

    Build the Halo Reader

    Our prototype is a research microscope wired to a GPU workstation, and it can’t sit on a hospital bench running 24 hours a day. We raised £750k, led by Kestrel Point Capital, to build the Halo Reader, hire our first four people and start taking the test into hospital labs.

  3. Launch

    Newborn sepsis

    The first Halo Readers will go to NHS labs that serve neonatal units. A newborn with sepsis is started on standard antibiotics today and waits up to two days to find out whether they work. Halo AI will report on the 8 antibiotics used in newborns in about 90 minutes, for Gram-positive and Gram-negative bacteria.

  4. Then

    Adult sepsis

    Sepsis affects an estimated 49 million people of all ages each year, and 11 million of them die. An adult’s positive blood culture is the same kind of sample as a newborn’s and goes through the same labs, so it will run on the same cartridge and the same Halo Reader. The rapid systems those labs use on adult cultures take between three and eight hours, and several test Gram-negative bacteria only. Before we can offer it for adults we need to validate a longer list of species.

    Source: The Lancet, 2020

  5. Later

    Urinary tract infections

    There are an estimated 405 million urinary tract infections a year, and about 237,000 deaths. Standard testing waits for the bacteria in urine to grow into colonies, as it does with blood. Halo AI watches single bacteria, so it doesn’t need a culture first. The Halo Reader, the imaging and the AI would stay as they are, and we would design a new cartridge that takes urine straight from the sample pot.

    Source: Global Burden of Disease 2019

  6. Long term

    Other infections, starting with drug-resistant TB

    Watching single bacteria respond to a drug should work for many infections where the result now waits on a culture. Drug-resistant tuberculosis is first on our list. About 390,000 people develop it each year. TB bacteria divide slowly and can only be handled in a biosafety-rated lab, so it will be a separate product.

    Source: WHO Global Tuberculosis Report 2025

We plan to place the Halo Reader free and charge £45 a test.

The plan is a reagent-rental agreement. We put a Halo Reader on your bench at no charge and you pay for the cartridges you use. It is designed to sit next to your blood culture incubators, run 24 hours a day and send results to your laboratory information system.

You can’t order one yet. We are building the first Halo Readers now, and they will go to pilot labs. If your lab serves a neonatal unit and processes more than 20 positive blood cultures a month, register below and we’ll contact you when a pilot Halo Reader is ready.

The Halo Reader on a bench in a working hospital microbiology laboratory, a cartridge in its slot. Behind it, two scientists in lab coats work at a safety cabinet and a blood culture incubator.
A visualisation of the Halo Reader in a microbiology laboratory.

Register your lab

We need all four to register you.

So far Halo AI is one person, Ellie Rose. We’re hiring the next four.

See open roles

Ellie Rose, founder of Halo AI, outside the Radcliffe Camera in Oxford.

Ellie Rose

Founder and CEO

Ellie Rose developed the method behind Halo AI in the Okoro Laboratory at the Sir William Dunn School of Pathology, University of Oxford, where she filmed 96 bacterial isolates against 8 antibiotics, 768 combinations in all. She is first author of the study in six NHS laboratories and reads biology at Corpus Christi College, Oxford.

Ellie Rose on LinkedIn

Backed by

We raised a £750k seed round led by Kestrel Point Capital, a London investor in early-stage life sciences companies. The money goes on building the Halo Reader, hiring our first four people and getting the test into hospital labs.

Read the press release