A deep
learning algorithm performed as well as human clinicians when detecting the
presence of cancer in radiology reports.
By Jessica
Kent
July
30, 2019 - A team from Dana-Farber Cancer Institute has developed
a deep learning toolthat performed as well
as human reviewers in extracting clinical information regarding changes in
tumors from unstructured radiology reports for patients with lung cancer.
EHRs
collect vast amounts of information on cancer patients, but unless these
patients are enrolled in clinical trials, information about their outcomes and
whether their tumors shrink or grow in response to treatment is recorded only
in the text of the medical record.
This
unstructured information is historically difficult to analyze, and therefore
cannot be used for research into the effectiveness of treatment.
In
past research, Dana-Farber scientists have collected an enormous amount of
molecular information about patients’ cancers, but this data isn’t immediately
helpful to providers unless they can access a patient’s entire medical
history.
“It
can be difficult to apply this information to understand what molecular
patterns predict benefit from treatments without intensive review of patients’
medical records to measure their outcomes. This is a critical barrier to
realizing the full potential of precision medicine,” said Kenneth Kehl, MD,
MPH, a medical oncologist and faculty member of the Population Sciences
Department at Dana-Faber.
To
determine whether artificial intelligence tools could extract the most
high-value cancer outcomes from radiology reports, researchers obtained and
manually reviewed over 14,000 imaging reports for 1,112 patients.
Human
reviewers analyzed the imaging text reports and noted whether cancer was
present, whether it was worsening or improving, and whether it was spreading to
different parts of the body. The team then used these reports to train a deep
learning algorithm to recognize these outcomes from text reports.
“Our
hypothesis was that deep learning algorithms could use routinely generated
radiology text reports to identify the presence of cancer and changes in its
extent over time,” researchers said.
The
group compared the human and computer measurements of outcomes, including
disease-free survival, progression-free survival, and time to improvement or
response. Researchers found that the deep learning tool could replicate human
assessment of these outcomes.
The
team then had the algorithms analyze another 15,000 reports for 1,294 patients
whose records had not been manually reviewed. The group found that computer
outcome measurements among these patients predicted survival with similar
accuracy to human assessments among manually reviewed patients.
Human
reviewers were able to annotate imaging reports for about three patients per
hour, a rate at which one reviewer would need about six months to annotate all
the nearly 30,000 imaging reports for the patients in the cohort.
In
contrast, the deep learning model was able to annotate the imaging reports for
the patient cohort in about ten minutes.
Artificial
intelligence technologies have proven to be adept at extracting relevant
information from clinical notes and unstructured data. Amazon’s Comprehend
Medical, an advanced machine learning tool, allows developers to comb through
unstructured EHR data and pull out key clinical terms related to patient
diagnoses, medications, treatments, symptoms, and other interactions with the
healthcare system.
Natural
language processing tools have also demonstrated their ability to generate
actionable insights from unstructured information. Studies have shown that the
technology can sift through clinical notes and EHRs to flag care guideline
adherence, adverse drug events (ADEs), and key terms related to the social
determinants of health.
In
future research, the Dana-Farber team will test the deep learning method on EHR
data from other cancer centers and use the data to find out which treatments
work best for which patients.
“To
create a true learning health system for oncology and to facilitate delivery of
precision medicine at scale, methods are needed to accelerate curation of
cancer-related outcomes from electronic health records,” researchers
said.
“This
technique could substantially accelerate efforts to use real-world data from
all patients with cancer to generate evidence regarding effectiveness of
treatment approaches.”
No comments:
Post a Comment