10 August 2019 | New Scientist | 15Analysis Machine learning
ARTIFICIAL intelligence trained on
health records can now detect kidney
injury up to two days before it occurs.
In principle, an advance warning
could help doctors intervene earlier
to prevent irreversible damage
to the kidneys, but this hasn’t yet
been confirmed by rigorous tests.
Nenad Tomašev at research
firm DeepMind and his colleagues
trained an algorithm to predict
whether a person will develop
acute kidney injury (AKI). This results
in a drastic drop in the rate at which
the kidneys filter blood and can be
fatal if untreated.
Using blood tests from medical
records related to old hospital
admissions, the AI predicted whether
a kidney injury would occur within
the next 48 hours. Its accuracy
was determined by comparing the
prediction with the actual outcome
at the time - whether the records
showed the person was later
diagnosed with AKI.
The algorithm was fairly accurate
at predicting the most severe forms
of AKI. It correctly predicted 90 per
cent of the cases in which the
patient’s kidney function deteriorated
so severely that they eventually
required long-term dialysis.
It is hard for doctors to anticipate
kidney injury, so that level of accuracy
is significant, says Eric Topol at
Scripps Research in California.
“If we knew a patient was going to
have a kidney injury, there are many
things we can do,” says Topol. These
include adjusting the person’s fluid
levels, monitoring blood pressure
more closely and avoiding medicines
that can hamper kidney function.
However, the algorithm was far
less accurate for all forms of AKI,
correctly predicting only 56 per cent
of all episodes, with a ratio of two
false alerts for each correct prediction
(Nature, doi.org/c82m).
The AI was trained on more than
700,000 anonymised electronic
health records from US military
veterans, but only around 6 per cent
of these were from women, meaning
the AI may be more accurate at
predicting the condition in men.
The only way to be sure that the
technology will be helpful is to test
it in a clinical trial, where doctorsintervene depending on the advice
of the AI. So far, few algorithms have
gone through such rigorous tests, but
the new system may be well-suited
to this step. With many medical AIs,
there could be serious consequencesif they get things wrong. A false
negative could mean a missed
cancer diagnosis, for example,
while a false positive may lead to
unnecessary treatment.
But the potential consequences
are less serious for the AKI-predicting
AI, says Topol. A missed prediction
will eventually be picked up by a
blood test and false alerts aren’t likely
to have a negative impact, he says.
This is because managing AKI usually
involves checking fluid levels and
blood pressure. At worst, it may
mean temporarily withholding a
drug that is toxic to the kidneys,
which is unlikely to be catastrophic.
But some barriers remain. Withthese kinds of algorithms, it is hard
to be sure what their predictions are
based on. It is possible, for example,
that an AI can use its knowledge of
lab tests that were ordered to work
out what a doctor suspects is wrong.
Clinicians might be hesitant to take
the advice of such an algorithm if it
doesn’t provide clinical reasons for
a prediction, says John Prowle at
Queen Mary University of London.
A further hurdle to developing
such technology is that AIs need a lot
of data for the initial training, which
can be hard to come by. For instance,
in 2016, the Royal Free National
Health Service Trust in London
provided 1.6 million identifiable
medical records to DeepMind to test
a smartphone app called Streams,
which monitors people with kidney
disease. The UK data watchdog, the
Information Commissioner’s Office,
later ruled that the agreement failed
to comply with data protection law.
Streams is now used by four NHS
trusts in the UK. While this app
doesn’t currently make use of AI,
DeepMind has said it plans to add it
to Streams. ❚90%
The AI’s accuracy at spotting
cases of severe kidney injuryDeepMind’s medical AI The technology firm’s latest
algorithm can detect early signs of kidney injury, but it is
too soon to tell if it will save lives, says Donna Lu
Space
Will GaterSupercharged star
charts a course for
intergalactic void
WE HAVE seen a star zooming
through the Milky Way so fast that
it will eventually leave our galaxy.
The so-called hypervelocity
star, named S5-HVS1, is moving
at a blistering pace of more than
1700 kilometres per second.
Astronomers have spotted
other stars travelling faster than
this. But those were dying stars
thought to have been blasted
outwards by supernova explosions.
S5-HVS1 seems to have been set
on its way by an encounter with
the supermassive black hole at the
centre of our galaxy, Sagittarius A*.
The supercharged star was found
by the Southern Stellar Stream
Spectroscopic Survey using the
Anglo-Australian Telescope in
New South Wales, Australia.
The team behind the work thinks
that the object might once have
been one of a pair of stars orbiting
each another. When they neared
our galaxy’s central black hole, its
intense gravity would have ripped
them apart, flinging S5-HVS
onto its current course (arxiv.org/
abs/1907.11725).
It is usually hard to trace the
origin of hypervelocity stars, but not
in this case. “We can calculate quiteprecisely where this star is coming
from,” says Sergey Koposov at
Carnegie Mellon University in
Pennsylvania. “It looks like it is
coming from a tiny region that
includes the galactic centre.”
We could learn a lot from objects
like this. For example, by tracking
the star’s route, we might see how
it was affected by the halo of dark
matter near our galaxy’s centre.
S5-HVS1 is destined for a lonely
end. It is on course to spend its final
days sailing through intergalactic
space. The researchers estimate
that it will be several million light
years away from our galaxy when
it finally burns out and dies. ❚BABAK TAFRESHI/SCIENCE PHOTO LIBRARYThe Anglo-Australian
Telescope was used to
spot a speeding star