12 | New Scientist | 1 February 2020
AN ALGORITHM to predict which
people may experience a mental
health crisis has been trialled
in the UK and found effective
enough for routine use. A version
that would track people’s mobile
phone calls, messages and
location in a bid to improve
accuracy is now being considered.
Birmingham and Solihull
Mental Health NHS Foundation
Trust worked with Alpha, a
division of Spanish telecoms
firm Telefónica, which owns O2,
to see if there was any benefit
in automatically flagging to NHS
staff which people were thought
to be most at risk of experiencing
a mental health crisis. The results
of the Predictive Analytics project,
released under freedom of
information rules, suggest that
there is.
The project ran between
November 2018 and May 2019.
Alpha developed a machine-
learning algorithm fed with
historical patient data to predict
who could face an imminent crisis.
Once a fortnight, staff on four
community mental health teams
in Birmingham were presented
with what the system calculated
were the 25 people most at risk
of a crisis within the next 28 days.
In some cases, healthcare
professionals followed up with
individuals by phone or face to
face. Overall, the clinicians found
the tool useful in about 64 per cent
of the cases that were flagged.
An evaluation of the project
concluded that the tool “could
become a part of routine clinical
care”. That positive verdict could
help the NHS trust and Alpha
proceed with a mooted second
phase of the research, where
people’s mobile network datawould also be accessed in a bid
to improve the algorithm.
The team is examining whether
such data could reduce the
number of false positives, though
how this would do it is unclear.
It would involve Alpha having
access to “call/message recordsand location details”, according
to minutes of a board meeting
of Birmingham and Solihull
Mental Health NHS Foundation
Trust held last October.
The project used five years’
worth of historical patient
records and socio-economic
data from the NHS trust. The
information was pseudonymised,
in which names are replaced by
strings of letters and numbers.
Mining the data unearthed
22 indicators that the algorithm
could use to judge if someone
was at risk of an upcoming crisis,
though details of those indicators
are redacted in documentsreleased by the trust.
Staff were generally positive
about the project. “It highlights
people that would otherwise, in
my opinion, fall through the net
and get lost in the system,” one
said in the evaluation.
However, there were issues too.
Many of the people flagged by the
system were already known to
staff as being in a mental health
crisis, because such events can
last up to 28 days, which raised
concerns that clinicians could
become “dismissive” of the tool
and “less inclined” to follow up
people it flagged, said the report.
The consent of patients under
the care of the community mental
health teams wasn’t sought,
as the trust says it was advised
by the UK’s Health Research
Authority that it wasn’t needed.
“Intervening early helps prevent
people from experiencing a
mental health crisis and also
improves the chances of recovery,
so the results of this pilot are
interesting,” says Adrian James
at the UK’s Royal College of
Psychiatrists. “But much more
research and evaluation withinformed consent is needed as
using data in this way is very much
in its infancy.”
Allen Frances, a US psychiatrist,
is more sceptical. “Predictive tools
of all sorts have been available for
50 years, but have limited clinical
or public health utility because
they are so imprecise: identifying
a great many people who don’t
get into trouble, missing many
who do,” he says.
Sam Smith at MedConfidential,
a UK campaign group that
requested the documents from
the trust, urged the NHS trust to
think carefully before allowing
Alpha to combine people’s phone
data with health records.Phase two
A spokesperson for Birmingham
and Solihull Mental Health
NHS Foundation Trust says:
“We envisage the algorithm could
be used to enhance our existing
care and risk management
processes. It would enhance and
not replace clinical judgement
and decision-making.” The trust
is still consulting with partners
before making a decision on
phase two, they added.
In a statement, Alpha said:
“The phase one results have
demonstrated that the algorithm
has high predictive power, and
that most clinicians valued the
extra insights provided by the
algorithm to help inform their
decision-making. However, the
pilot study showed that there
is more work to do to improve
accuracy, with 7 per cent of
clinicians disagreeing with
predictions.
“Phase two aims to improve
the accuracy of the predictions
by adding complementary data
to the algorithm, however we
have not yet made a final decision
on how we would proceed.” ❚“Informed consent is
needed as using data
in this way is very
much in its infancy”News
AN
DREWAITCH
ISON/GET
TY
IMAG
ESCan phone data hold
clues to a person’s
mental health?Health dataQuestions over NHS data deal
An algorithm could track smartphone use to assess mental health
Adam Vaughan