How health tech is redefining the future of biopharma and care delivery

Howhealthtechis
redefiningthefuture
ofbiopharmaand
caredelivery
UNDERWRITTEN BY

TABLE OF CONTENTS
China is leapfrogging the U.S. in
using AI in medicine. These five
companies are leading the way
Google’s AI improves accuracy
of lung cancer diagnosis, study
shows
01
09
HOW HEALTH TECH IS REDEFINING THE FUTURE OF BIOPHARMA AND CARE DELIVERY INTRODUCTION
Digital health startups want to
help Medicare prevent diabetes,
if only it will let them
New voices at patients’
bedsides: Amazon, Google,
Microsoft, and Apple
How an IBM Watson Health
rescue mission collapsed — and
a top executive was ousted
The opportunity ‘is huge’:
Why tech developers are trying
to tackle mental health
As patients tell apps they’re
feeling suicidal, digital health
startups scramble to respond
29
40
45
54
34
SPONSOR CONTENT
The rise of privacy awareness
and data ethics can be a
marketer’s golden ticket
14
Hospitals look to computers
to predict patient emergencies
before they happen
There’s a massive new data set
that aims to help artificial intelli-
gence work better for biotech —
and it’s free
18
24

In medical parlance, “stat” means important and urgent, and that’s
what we’re all about — quickly and smartly delivering good stories.
We take you inside science labs and hospitals, biotech boardrooms,
and political backrooms. We dissect crucial discoveries. We examine
controversies and puncture hype. We hold individuals and institutions
accountable. We introduce you to the power brokers and personalities
who are driving a revolution in human health. These are the stories
that matter to us all.
Our team includes talented writers, editors, and producers capable
of the kind of explanatory journalism that complicated science issues
sometimes demand. And even if you don’t work in science, have never
stepped foot in a hospital, or hated high school biology, we’ve got some-
thing for you. The world of health, science, and medicine is booming
and yielding fascinating stories. We explore how they affect us all. And,
with our eBook series, we regularly do deep dives into timely topics to
get you the inside scoop you need.
BOSTON • WASHINGTON • NEW YORK
SAN FRANCISCO • LOS ANGELES • CLEVELAND
HOW HEALTH TECH IS REDEFINING THE FUTURE OF BIOPHARMA AND CARE DELIVERY INTRODUCTION

Tech is a powerful force of change in every sector of health care.
Some of the world’s largest companies, including Google, Apple, Amazon, and
Microsoft, along with emerging players from the U.S. to China, are poised to
transform everything from how care is delivered in hospitals and doctors’ offices
— and virtually — to how drugs are discovered and distributed. Advanced
artificial intelligence and novel software tools — from tech giants and scrappy
startups — are challenging the status quo in every corner of the industry.
STAT tracks these developments, examining their impact on the lives of doctors
and patients as well as the bottom lines of entrepreneurs and established com-
panies. Investors are trying to suss out which of the latest innovations are real
and which are hype. Washington regulators are watching, too.
Here are select STAT stories covering the great promise of tech innovation and
the many challenges of deploying it in an industry where human life is at stake,
not to mention a fifth of the U.S. economy and the fortunes of the nation’s
biggest companies.
Howhealthtechisredefining
thefutureofbiopharmaand
caredelivery
HOW HEALTH TECH IS REDEFINING THE FUTURE OF BIOPHARMA AND CARE DELIVERY LETTER FROM OUR EDITOR

The short answer is no. Data gives us
invaluable insights about how people feel,
what they believe, and how they behave.
But to create advertising and marketing that
moves them to act requires living, breathing,
and endlessly clever human creativity.
At W2O, that’s how we turn deep,
data-derived insights into human ideas
that inspire human action.
Data matters,
but can it
make you
fall in love?
Visit us at W2Ogroup.com

In the race to deploy artificial intelligence in medicine, China has surged ahead of the
U.S. and other countries, spawning technology companies that are using algorithms
and novel software tools to help treat millions of patients and relieve pressure on the
country’s overwhelmed hospital system.
One company, WeDoctor, uses a combination of AI and telemedicine to conduct
90,000 online consultations a day and screen patients for cancer and chronic diseases.
No company with anywhere near such reach exists in the U.S.
Another AI firm, Infervision, has stockpiled more than 1 million patient scans
from Chinese hospitals and is working with medical centers to detect lung cancer
and diagnose stroke, collapsed lung, and other conditions. U.S. tech firms typically
train medical imaging algorithms with thousands, or tens of thousands, of scans
from a single institution. Infervision says it is collaborating with more than 300
hospitals where its software is helping to make more than 33,000 diagnoses per day.
China’s government has been crucial to the early growth of these companies: It has
set a goal to become the world’s leader in AI and build a $150 billion industry by
2030, with a particular focus on health care as a way of addressing a national shortage
of physicians. China has only 1.8 doctors per 1,000 people, compared to 2.6 in the
United States and 4.2 in Germany.
By Casey Ross
ARTICLE 1 | 01
ChinaisleapfroggingtheU.S.in
usingAIinmedicine.Thesefive
companiesareleadingtheway
HEALTHTECH
@CASEYMROSS JUNE 5, 2019|
HOW HEALTH TECH IS REDEFINING THE FUTURE OF BIOPHARMA AND CARE DELIVERY

“One thing that differentiates China from the U.S. is how the government, smaller
startups, and big tech companies are working toward one agenda,” said Deepashri
Varadharajan, lead artificial intelligence analyst for CB Insights, a business research
firm. She noted that Tencent Holdings, a China-based technology company that op-
erates the giant social media platform WeChat, was hand-picked by the government
to lead the development of AI products in health care.
Some business leaders have fretted that the U.S. risks falling behind in the devel-
opment of artificial intelligence, which could cause business opportunities, and
investment dollars, to flow to Chinese companies. At a national health technology
conference earlier this year, the influential tech investor John Doerr said the U.S.
may be in a losing fight, given China’s lack of privacy restrictions and the extent of
the government’s investment in AI.
Hype surrounding Chinese AI companies is rampant, as it is in the U.S., and it
remains unclear how use of the technology is affecting patient outcomes. But there’s
no doubt that the development of medical artificial intelligence is progressing faster
in China. The government support helps, but AI companies have other advantages,
including fewer privacy regulations and other hurdles, and access to much larger
data sets that facilitate the training and refinement of algorithms.
With a population of 1.4 billion, the country is more than four times the size of
the United States, and its hospitals treat and collect data on many more patients.
“China is a great example of how AI can expand access” to care, said Terri Bresenham,
chief innovation officer of GE Healthcare, which has a business relationship with
Infervision. “The incredible unmet need in China is motivating the provider com-
munity to work on these solutions and address the uneven distribution of
health care resources.”
ARTICLE 1 | 02HOW HEALTH TECH IS REDEFINING THE FUTURE OF BIOPHARMA AND CARE DELIVERY

GE is among many U.S. companies with a large presence in China. Apple has major
business operations there, and Amazon recently opened an AI lab, although the ability
of U.S. companies to grow may be impaired by the Trump administration’s trade war
with China.
Likewise, it may be difficult for Chinese companies to gain access to U.S. patient data
needed to train their algorithms for use in America, an issue highlighted by the recent
flap involving the company PatientsLikeMe, which the U.S. government forced to
unwind a Chinese company’s investment.
Still, many of China’s digital health companies are already working with hundreds of
hospitals to build and test their products. Here is a look at five businesses at the fore-
front of the effort.
If you go by patient visits, WeDoctor is one of China’s largest medical providers
— and it operates primarily online. It sees about 90,000 patients a day, providing
diagnostic services, online chats with doctors, and prescriptions.
The company started in 2010 as an online appointment booking portal to help
navigate a badly broken system in which Chinese patients pay middlemen to get
access to doctors or wait in long lines outside hospitals. Since then, it has expanded
to provide direct consultations with doctors. Its AI software helps diagnose patients
and screens them for diseases like cervical cancer.
It also sells a $600 smart speaker, called WeDoctor Tong, that serves as a physician
hotline and can connect users to services at 2,700 hospitals and 20,000 pharmacies
across the country. The company sends patients automated reminders of checkups
and can help them arrange services, including surgeries and drug purchases.
WEDOCTOR

“This is an example where you see China leapfrogging the world,” said Nisa Leung,
managing partner of Qiming Venture Partners, a Hong Kong-based venture capital
firm that is an investor in WeDoctor.
She told STAT that, in addition to helping wealthier patients through its smart
speaker, the company is improving access for people of lesser means. “By using the
internet, patients from rural villages can get access to the top doctors,” Leung said.
“It’s really quite transformative.”
WeDoctor, which also sells insurance and data management capabilities to hospitals,
is backed by Tencent Holdings and has attracted investments from the insurance
company AIA Group and Shanghai Fosun Pharmaceutical Group. It was valued at
$5.5 billion in advance of an initial public offering expected by the end of this year.
Many of the world’s largest technology companies are using AI to analyze medical im-
ages. But few, if any, have the volume of data that Bejing-based Infervsion has amassed.
The company is ingesting tens of thousands of CT scans, MRIs, and other images per
day through its partnerships with hundreds of hospitals across China. Its InferRead
CT Chest product analyzes patient scans for lung nodules, bone tumors, emphysema,
and the early signs of arterial disease. It has also developed algorithms to detect bone
fractures and the onset of stroke.
Infervision is seeking FDA approval to begin selling its AI services in the U.S., where it
is testing its products with Stanford Children’s Hospital and a handful of other pro-
viders. The company’s CEO, Kuan Chen, is a frequent visitor to medical conferences,
where he emphasizes that the practical benefits of AI are just beginning to be realized.
INFERVISION

“When we get enough data to [replicate] the techniques of the best doctors at the
top hospitals, we might be able to provide this expertise for rural areas and smaller
hospitals,” he said during a panel discussion at the 2019 World Medical Innovation
Forum in Boston.
The company still has plenty of work to do to demonstrate the effectiveness of its
products, but its ability to collect imaging data through the Chinese health care
system is unsurpassed: So far, it has assisted doctors in analyzing images on 4.5
million patients.
Ping An Good Doctor has showcased one of the most novel creations of China’s
digital health sector: A tiny, unstaffed medical “clinic” in which patients can get
advice from an “AI doctor.”
The company, backed by Ping An Insurance Co., has declared its intention to open
hundreds of thousands of the 3-square-meter clinics across the country over the next
few years.
The clinics are designed to be a one-stop shop: Patients can give a short medical his-
tory, get a diagnosis, and then chat online with an in-house doctor from Ping An. The
booth even includes a medicine cabinet to supply patients with more than 100 drugs.
The company boasted 265 million registered users at the end of last year, giving it
a large supply of data it can use to refine its algorithms and build new products. Its
average daily number of patient consultations has surpassed 535,000, an annual
increase of 45 percent.
PING AN GOOD DOCTOR

Established in 2013, Yidu Cloud is using artificial intelligence to analyze hundreds of
millions of electronic medical records of patients across China. The company con-
verts the records into a standard format that can be used by doctors and researchers
to examine patient outcomes. Its software is also being used to help predict the path of
patients’ diseases and suggest treatment options, according to its founder, Rujing Gong.
“The machine can calculate on a real world dataset the possibility of certain treatment
outcomes or the choices of medicines for that particular patient,” Gong said during the
World Medical Innovation Forum.
Gong added that the company is working with the China Food and Drug Administra-
tion and other regulators to evaluate the effectiveness of certain drugs. “That research
can be done by our system much cheaper and much faster,” she said, before adding a
crucial caveat: “It may not be the most accurate, but it can give a quick perspective to
the researchers and the scientists.”
Gong said Yidu Cloud has spent $100 million to build its software and employs about
850 people, including 200 physicians, a number that would be cost prohibitive in the
United States. The company has also begun to reach out to clients beyond China’s
borders, including the U.K.’s National Health Service.
YIDU CLOUD
Not to be confused with Yidu Cloud, Yitu Technology has developed AI software
that analyzes medical images, pathology slides, written records, and other data to help
diagnose and treat cancer patients.
The Shanghai-based company got its start in security products using facial recognition
software and has built a health care business over the past several years.
YITU TECHNOLOGY

It showcased its products last November at the annual conference of the Radiological
Society of North America. It is focused on diagnosis and treatment of lung and breast
cancer, the two most common forms of cancer in China.
Yitu said it employs more than 400 physicians to help with the development of its
technology, which also includes a 4D chest scanning system that shows moving, re-
al-time images of lung nodules and other lesions.
The company is led by Leo Zhu, who studied at UCLA and the AI lab at the Massa-
chusetts Institute of Technology. It has announced research partnerships with large
hospitals across China, including a collaboration on lung cancer with West China
Hospital in Chengdu, giving it access to tens of thousands of images.
“Data collection is a lot faster in China because we have so many patients going into
each hospital,” said Leung, the managing partner at Qiming Venture Partners. “That’s
really the key.”

UP NEXT
Google’sAIimproves
accuracyoflungcancer
diagnosis,studyshows
By CaseyRoss
ARTICLE 2HOW HEALTH TECH IS REDEFINING THE FUTURE OF BIOPHARMA AND CARE DELIVERY

One of lung cancer’s most lethal attributes is its ability to trick radiologists.
Some nodules appear threatening but turn out to be false positives. Others escape
notice entirely, and then spiral without symptoms into metastatic disease.
On Monday, however, Google unveiled an artificial intelligence system that — in early
testing — demonstrated a remarkable talent for seeing through lung cancer’s disguises.
A study published in Nature Medicine reported that the algorithm, trained on
42,000 patient CT scans taken during a National Institutes of Health clinical trial,
outperformed six radiologists in determining whether patients had cancer. It detected
5% more cancers and cut false positives — when cancer is suspected though a nodule
is harmless — by 11% from reviewing a single scan. It performed on par with the
radiologists when prior images of patients were also included in the evaluation.
The results underscore AI’s potential to improve lung cancer screening and help
radiologists diagnose malignancies earlier and with greater accuracy — though the
research did not show whether the Google system would help patients live longer.
Lung cancer is by far the most common killer of Americans among cancers, resulting
in about 160,000 deaths in 2018.
By Casey Ross
Google’sAIimprovesaccuracy
oflungcancerdiagnosis,study
shows
HEALTHTECH
@CASEYMROSS MAY 20, 2019|

A large NIH study — the same one that supplied data for Google’s algorithm —
established that screening of high-risk patients such as long-term smokers can reduce
the risk of death by about 20%, but it also raised concerns about false positives leading
to unnecessary testing that can harm patients. It reported that several deaths resulted
from false positives that led patients to undergo invasive biopsies and other procedures.
Ongoing concerns about the overall accuracy and benefit of screening have led to low
rates of such testing. Meanwhile, about 50 percent of lung cancers are diagnosed after
they have already spread, when the disease is harder to effectively treat.
Experts who were not involved in the study said Google’s system could make
screening much more viable.
“These people have a technology that will improve the precision of screening tremen-
dously,” said Dr. Otis Brawley, a professor of oncology and epidemiology at Johns
Hopkins University and formerly executive vice president of the American Cancer
Society. He said he is generally a skeptic of lung cancer screening, but emphasized
that Google’s performance in reducing false positives was a significant step forward.
“It’s going to prevent more bad things from happening to people who are being
screened,” said Brawley, adding that the system’s strong performance in that regard
doesn’t necessarily mean it will save more lives from lung cancer.
Google’s system will require more rigorous testing — probably a randomized con-
trolled trial — before it could be put into medical practice. The study was limited to
patients who had already been treated, so it is impossible to tell whether the system,
when used on new patients, will result in more effective care and better outcomes.
Google executives acknowledged that point in the study and an accompanying
blog post, adding that they are working with clinical partners to further refine and
validate the system. “To do a full assessment of this, you have to work with research
organizations and run large-scale trials to understand how this technology will work
at scale and on broad populations,” said Daniel Tse, a Google product manager who
co-authored the study.

He added that the company has had pre-submission discussions with the Food and
Drug Administration to discuss approval criteria.
Google engineers who developed the AI system emphasized that it is not designed to
replace radiologists, but to improve their ability to detect nodules and determine if
they are dangerous. Existing computer-aided systems separate nodule detection and
diagnosis into different tasks. Google’s system performs both functions, honing in
on regions of interest within a scan and providing a risk score that indicates whether
a given patient’s nodules are cancerous.
The system uses convolutional neural networks, a type of AI architecture, to learn
the features of malignancy and point out the problematic areas by analyzing three-
dimensional CT scans. That task is difficult and time-consuming for radiologists
because they cannot examine three-dimensional scans the way a computer can; they
must review hundreds of individual slices of the scan to hone in on problems. But the
computer can review all the dimensions at once.
“We were able to train the [AI models] over the whole scan at a very high resolution,”
said Shravya Shetty, a Google engineer and technical lead on the study. “Although
radiologists can look at the slices, there are clear advantages to the model here.”
The authors reported that the system’s performance remained consistent when it was
exposed to patients outside of the NIH dataset on which it was trained. The system re-
viewed scans of 1,700 patients from Northwestern Memorial HealthCare in Chicago,
and produced similar results in classifying nodules and making diagnoses.
Experts said the Google software may be particularly useful for general radiologists,
who often review patient lung scans in community hospitals in the U.S. Most lack
the expertise of thoracic radiologists who specialize in lung disorders and are con-
centrated at large academic medical centers.

“To be able to make screening available to everyone, it can’t just be done by thoracic
radiologists. It has to be done by all radiologists,” said Dr. Jorge Gomez, a medical
oncologist at Mount Sinai Health System who serves as a national spokesman for
the American Lung Association.
At Mount Sinai, he said, thoracic radiologists regularly participate in meetings to
discuss patient care and point out the specific aspects of scans that indicate why
nodules may or may not be malignant.
“That’s an incredible resource to have, and computers might be able to do that well,”
he said. “This is a very important study that should prompt somebody with very
deep pockets to do a randomized trial.”
Tse, the Google product manager, said the company is actively pursuing that work.
“We are pushing things forward both internally within Google and with our part-
ners,” he said, adding that he could not yet identify the time frame for the follow-
up study.

SPONSOR CONTENT
By W2O

Thanks to Cambridge Analytica, the Marriott mega-breach, the Equifax breach,
British Airways, Facebook’s criminal investigation and many more newsworthy
data privacy events, consumers are starting to wake up and understand the scale of
digital data collection and how bad actors can use it to manipulate everything from
purchasing decisions to elections. Politicians have responded with legislation like
the California Consumer Privacy Act (CCPA), the EU’s General Data Protection
Regulation (GDPR), the Vermont data broker law, and 10 other (currently) pro-
posed state-level data privacy laws, designed to reign in an industry that arguably
abused an unwitting public.
The current paradigm shift is straight-forward: data is now owned by the people
it reflects, not by martech vendors and advertising platforms, and it should not be
sold, traded, or used in unexpected ways without informed consent. You would
think that this is our demise….
Actually, it’s our golden ticket. Privacy, and more importantly data ethics, are
now a key differentiators and offer an opportunity to deepen relationships with
our audiences.
Social science and psychology research show that when people know they are being
observed, they feel intruded and change their behavior.
SPONSORCONTENT
By W2O

For example, information that we willing to tell our spouse is usually quite different
than what we’d be willing to share with an advertiser. Which is why when the news
broke that one of the most popular ovulation tracker apps, Flo, was sharing data
with Facebook, it created an overwhelmingly negative reaction from users.
A few brands, most notably Apple, have begun to market their data protection and
privacy practices as a prominent selling point. Studies are now beginning to show
that strong data protection and ethics programs build trust with consumers and
help brands gain advantage over competitors. Strong data controls enable agility and
innovation, increase operational efficiency, mitigate losses from data breaches, and
increase appeal to investors, surveys also found.
Now is the time for marketers to embrace data ethics and privacy.
Marketers that proactively go beyond compliance requirements and build privacy
protection into the foundation will not only build trust with existing audiences but
also capture the attention of those that are becoming disillusioned with competitors.
Marketers and advertisers should question the data captured and ask what we should
do, instead of what we can do. Here are a few points to consider.
First, invest in a privacy program and map understand your data flow. Where is
your data coming from? How is it being stored? Are your vendors using your data in
ways you’re not aware of? Are you using platforms that don’t respect privacy ethics
and legislation? How are you using your data – and if you’re not using it, why are
you collecting it?
Having a clear line of sight into the data journey and map in addition to a robust
marketing operations controls will help reduce risk and ensure customer privacy.

Second, embrace transparency. Whether it’s required for legal compliance or not.
Ensuring your customers know why and how you’re collecting data, what you’ll use
it for and how they’ll benefit from it – will help deepen brand loyalty and trust.
Third, do better work. Give people a good reason to engage by developing interest-
ing and relevant content. Add value, entertain and enlighten them, and give them
a reason to give you their data. A fair-value exchange will develop an authentic
relationship between the consumer and a brand.
Finally, keep innovating! As the industry continues to change and more legislation
comes to fruition, we need to continue creating new ideas and spins on older ap-
proaches. Embrace the use of AI-driven contextual advertising, premium publishers,
better CRM programs, real 1-1 website personalization, better use first party data and
CDPs – the list goes on. You can do all of that and more with a privacy-first approach.
The above has already worked for marketers in the EU where post-GDPR studies
show that programmatic ad spend is growing by double digits as well as publisher’s
digital revenue. So remember, privacy protection isn’t just a legal compliance bur-
den, it’s an opportunity. Marketers who wait will be left behind whereas those who
embrace privacy and transparency will cultivate trust and engagement, win market
share against competitors, and build customer loyalty.

UP NEXT
Hospitalslooktocomputers
topredictpatientemergencies
beforetheyhappen
By CaseyRoss

CLEVELAND — Seven floors, and long odds, were stacked against John S.
He was undergoing a test on the first floor of a Cleveland Clinic hospital when his
nursing team — on the eighth floor — got a call, telling them the 57-year-old had
developed a dangerously rapid heartbeat that was spiraling toward cardiac arrest.
It is a predicament that often ends badly. Only about 25 percent of U.S. patients
survive when their hearts stop in hospitals. Crucial minutes elapse before help
arrives, sometimes because alarms are missed amid the din of beeping monitors.
But the call that day didn’t come from within the hospital. It came from a darkened
room in an office park several miles away, where a technician in the clinic’s Central
Monitoring Unit (CMU) was watching the patient’s vital signs on a computer
monitor and noticed the onset of ventricular tachycardia.
The subsequent emergency response pulled John (not his real name) back from
the brink, but the CMU’s success in this case was just the first step toward a more
ambitious goal to use artificial intelligence to predict such events ahead of time, so
that patients like John get effective treatment before a crisis arises in the first place.
By Casey Ross
Hospitalslooktocomputers
topredictpatientemergencies
beforetheyhappen
HEALTH
@CASEYMROSS MAY 13, 2019|

“This is where we think machine learning has the great opportunity to help us,”
said Dr. Daniel Cantillon, a cardiologist who serves as medical director of the CMU.
“The challenge is we have to be able to call our shots in advance, and that’s some-
thing we’re deeply invested in.”
Hospital command centers have proliferated across the country in recent years,
with medical centers from Oregon to Florida deploying them to tackle a range of
data-monitoring tasks, such as maximizing bed capacity, calibrating staffing levels,
and detecting the onset of sepsis, a life-threatening response to infection that is a
common killer in hospitals.
Recent advances in artificial intelligence promise to help hospitals identify new
warning signs of patient deterioration and intervene earlier in the process. Admin-
istrators of command centers at Johns Hopkins and Yale New Haven Hospital both
said they are exploring the use of machine learning to deliver more timely care.
The Cleveland Clinic’s ultimate goal is to give front-line clinicians notice of serious
cardiac events an hour or more before they happen. That would be a significant leap
forward from the system’s current capabilities. Right now the CMU can offer some
advance notice of cardiac emergencies, but it is heavily reliant on technicians to
pluck out the signals from massive data streams on hundreds of patients and quickly
route them to caregivers.
Nemours Children’s Hospital, which operates facilities in Delaware and Florida, is
a pioneer in the command-center model. The health system created a logistics center
more than a decade ago to track vital signs of non-ICU pediatric patients across its
system. Paramedics who work in the center monitor data feeds and details of patient
electronic health records to ensure timely follow up.
“If there is an abnormal laboratory result, the logistics center makes sure someone
has looked at it,” said Dr. Stephen Lawless, the health system’s chief clinical officer.
He added that Nemours is beginning to apply artificial intelligence to better predict
nursing demand on floors throughout its facilities.

Lawless has been tinkering with the logistics center for more than 15 years. The theory
behind it is derived from the military, where tactical command centers removed from
the chaos of battle are used to direct soldiers on the front lines and coordinate the use
of air power and other resources.
The need for such facilities in health care has increased in recent decades, as an array
of monitoring devices produces tens of thousands of alarms on a daily basis. The vast
majority of alarms do not require a clinical response, which makes it harder for nurses
and doctors to pick out the rare ones that do.
“We had reports of alarms going off for an hour,” Lawless said. “Most times it’s not a
harm. But when you realize that it takes that long to answer, you begin to think, ‘Oh
my God, how would they know if this is real or not?’”
It is an ideal task for artificial intelligence systems whose power of pattern recognition
can surface crucial information that otherwise would get lost in the noise. At the
Cleveland Clinic, a customized algorithm developed by clinicians crunches an array
of data, such as blood pressure, heart rate, and oxygen saturation levels, to flag the
patients that are at highest risk of deterioration.
On a recent afternoon, those patients were marked in red at the top of the computer
monitors inside the CMU, a dimly-lit conference room in a medical building in
Beachwood, a Cleveland suburb. Teams of three — usually made up of nurses and
emergency medical technicians — clustered in front of computers arranged in
parallel rows.
A radio played Billy Joel in the background while the workers clicked in and out of
patient reports and electronic records detailing their medical histories and lab results.
Each time they noticed a problematic heart rhythm, or blood pressure change, they
picked up the phone. In April alone, CMU workers logged 77,000 calls to nursing
units across the health system, according to Alicia Burkle, the unit’s program manager.

Most calls are routine, providing a quick nudge to a charge nurse to check on a
particular patient. But occasionally the technicians will notice a situation like John’s
— an urgent problem that requires activation of the emergency response team.
Kris Rhode, a registered nurse, was on the other end of the line when the CMU
called that day a few years ago to notify her of John’s decline. “We looked (at his
telemetry data) and said, ‘Oh my gosh, this patient is in a lethal arrhythmia,” Rhode
said, adding that John was off her floor having his heart imaged. “We figured out
where he was and got down there within two minutes.”
A 2016 paper published in the Journal of the American Medical Association
sought to quantify the performance of the CMU in notifying front-line caregivers
of impending emergencies. It found that the unit accurately alerted clinicians of
79 percent (772 of 979) of heart rate or rhythm changes that eventually led to the
activation of an emergency response team. It also provided notice of 27 cardiac
arrests, with caregivers restoring circulation in 25 of those patients.
The challenge now, Cantillon said, is to get outside that hour window, to give
doctors even greater visibility into future events. “If we can unlock the secrets to
the subtle pattern changes happening with the patient’s digital biomarkers, then
we can do a better job of getting out in front of those events,” he said.
That depends on being able to identify those changes and link them to actual
cardiac events. Cantillon said the clinic is analyzing an array of different measures,
such as variations in heart rate and cardiac repolarization (or the return of the heart
to its resting state during a heartbeat.)
“We have to weave them into an algorithm that is already very good at identifying
sick patients and we have to do that in real time and validate it,” Cantillon said.
“It’s a tremendous amount of work and development.”

He said the clinic has been doing all of the heavy lifting on its own but is now
working with external AI vendors to further develop the system’s capabilities.
“We need the right commercial partners to take us that last mile of this race,”
Cantillon said, adding that any future improvements to the algorithms will augment
— not replace — the workers in the CMU. “What we’re trying to do, “ he said, “is to
make the machines work better for them, so they’re able to more efficiently handle
all of the data coming their way.”
The leader of a command center at Johns Hopkins said the hospital is exploring the
use of machine learning to predict changes in patients’ conditions and monitor the
delivery of treatment. Currently, the center is focused on managing hospital capacity
to allow for timely transfers and prevent patients from languishing in the emergency
department.
But Jim Scheulen, chief administrative officer for emergency medicine and capacity
management, said the hospital is contemplating a variety of clinical uses. “If a patient’s
lab values are supposed to be heading in one direction, but they suddenly veer off in
another direction, should we get an alert for that,” he said. “But some of this is just
in discussion.”
At Yale New Haven Hospital, which also created a command center focused on
managing capacity, clinical leaders instituted a program to monitor use of Foley
catheters to prevent infections and other complications and are beginning to look
at possible predictors for patient deterioration.
Most of that work is now done by floating ICU nurses and other clinicians. “But the
goal is to bring that into the capacity coordination center,” said Dr. Robert Fogerty,
director of bed resources at the hospital. “It’s a natural fit.”

UP NEXT
There’samassivenewdata
setthataimstohelpartificial
intelligenceworkbetterfor
biotech—andit’sfree
ByKateSheridan

Artificial intelligence is the hot new thing in drug discovery and development.
AstraZeneca, Pfizer, and Merck have all relied on machine learning to advance their
research. Bristol-Myers Squibb and Boston-based Concerto HealthAI established
a new partnership in March. And Relay Therapeutics raised an astonishing $400
million series C round to support its efforts to use AI techniques to understand the
way proteins bend and twist and create new drugs.
But if artificial intelligence programs are actually going to make an impact, they’re
going to need a lot of data — and high quality data is hard to come by. Currently
available data sets aren’t ideal for machine learning, and relying on that data might
even set certain algorithms down the wrong path.
Salt Lake City-based Recursion Pharmaceuticals thinks they have a solution — and
they’re not charging for it. The company released more than 125,000 images of four
different cell types collected through a microscope on Monday, and the data set can
be downloaded by anyone, for free.
There’samassivenewdatasetthat
aimstohelpartificialintelligence
work better for biotech — and
it’s free
By Kate Sheridan @SHERIDAN_KATE MAY 6, 2019|
BIOTECH

“I don’t think it’s fair to say that the release of this one data set will end up impacting
hundreds of millions or billions of people — but hopefully we’re able to take a step
in the right direction,” said Recursion’s co-founder and CEO, Chris Gibson. “We
can get the ball rolling here to have more open sharing of data sets and information.”
To create the data set, called RxRx1, Recursion’s team exposed cells grown in a dish
to small pieces of RNA that were custom-designed to suppress one of 1,000 certain
genes. Then, a specialized microscope captured the RNA’s effect on various parts of a
cell, including those that are vital for the cell’s energy supply, movement and protein
production. The company repeated this process 51 times over the course of nearly a
year across various cell types, including those from an umbilical vein, the retina, liver
cancer, and bone cancer.
RxRx1 takes up about 300 gigabytes of space — about 400 CD-ROMs’ worth;
Recursion’s team claims RxRx1 is just 0.4 percent of the two petabytes of data the
company has generated.
Recursion isn’t just an artificial intelligence company; it’s also a biotech in their own
right. The company is running two ongoing clinical trials; its business model depends
in part on partnerships and, eventually, on commercializing drugs.
Biopharma companies are interested in artificial intelligence for one reason: finding
new drugs costs a ton of money. The exact cost to bring a new drug to market varies,
depending on who is doing the analysis, but figures range from $650 million to $2.6
billion — per drug. And the cost is so high in part because drug developers have to
churn through many potential drugs before they find one that works. In theory, arti-
ficial intelligence could help companies pick better drug candidates in the first place.
“Life is short. You never have the time and the budget to test everything you’d want to
test, so you have to make good decisions and be judicious and about what you bother
to work on,” said Abraham Heifets, the CEO and founder of San Francisco-based
Atomwise, an artificial intelligence-based company that has partnerships with Charles
River Laboratories, Merck and Y Combinator.

Companies and researchers working in drug discovery have tended to self-sort into
two groups: those working on designing drugs to fit targets — the chemists — and
those working on finding targets — the biologists. Recursion’s data set will help
the latter.
Applying artificial intelligence to biology problems has proven to be tricky, in part
because scientists in biology have complicated questions to answer and more com-
plicated data to analyze. People trying to model the diseases those drugs target
require lots of data collected in a consistent fashion over long periods of time, and
analyzing a collection of cells requires different tools than analyzing a photo to
detect a cat or dog.
“If we’re talking about cancer, if we’re talking about Alzheimer’s, if we’re talking
about diabetes, all of these diseases are actually like our own body slowly going to
hell over decades,” said Heifets. “You can eyeball some data when you have ten data
points. But when you have ten million data points, you can’t eyeball them anymore,”
he added.
Atomwise, which counts itself among the chemists, is unlikely to use Recursion’s
dataset — but Heifets was still enthusiastic about its release. “I think it’s super cool,”
he said. “This is the thing that’s toughest to get if you’re a machine learning research-
er, is real real-world datasets.”
Jason Yosinki, one of Recursion’s machine learning advisors, pointed to a 2009 data
set that catapulted forward the field of image analysis: the ImageNet set enabled
Facebook to auto-tag faces or Google to recognize the Eiffel tower in images, he
said. He’s hoping Recursion’s data set will do similar for biotech.
Recursion itself had relatively small beginnings. The company launched in 2014
with three founders and funds from friends, families and grants from the National
Institutes of Health. Since then, the company has raised more than $100 million,
most of which came in a series B in 2017 led by San Francisco-based venture firm
Data Collective DCVC. About 140 people now work at the company.

“We’ve worked really hard to convince the world that this thing we’ve been doing
is really useful,” Gibson said. “It feels like we’re far enough ahead now that we can
maintain the competitive advantages for our business while still being able to do this
service to the industry and the community around machine learning.”
And the company’s leaders and advisors anticipate this service will get a warm recep-
tion. “The best models we can train are still data-limited — they’re still hungry for
more data,” Yosinski said. “By training on more images, models will be able to learn
more subtle features.”

UP NEXT
Digital health startups
want to help Medicare
prevent diabetes, if only
it will let them
ByCaseyRoss

In rolling out a first-of-its-kind program to prevent diabetes, Medicare officials
dashed the hopes of many startups in 2017 by deciding to exclude online health
companies from participation.
Since the program’s launch, however, large cracks have formed in Medicare’s service
network. The program, which offers classes to improve diet and exercise habits, has
no providers in Alabama or Mississippi — states in the heart of the so-called diabetes
belt — and only one in Georgia, according to a federal database of more than
460 participants.
The spotty service is undermining the goal of creating a national dragnet to deliver
timely lifestyle interventions to some 84 million Americans at risk of developing
type 2 diabetes before it takes hold.
It is also spurring calls for Medicare to change course.
The American Medical Association is urging the government to include digital
providers in the program, arguing it is unlikely to reach large swaths of the population
if they are not allowed to join in.
By Casey Ross
Digitalhealthstartupswantto
helpMedicarepreventdiabetes,
ifonlyitwillletthem
HEALTH
@CASEYMROSS APRIL 4, 2019|

“Given the fact that in-person DPPs (diabetes prevention programs) aren’t available
and accessible to patients in every part of the country, it is essential that patients have
the option to participate in a virtual DPP,” the association’s president, Dr. Barbara
McAneny, said Wednesday.
The digital providers themselves, which help patients track weight loss and offer
nutrition and exercise coaching, have also stepped up the pressure, lobbying
members of Congress and other officials to allow them to participate. They stand
to gain financially from inclusion in the program, as it would allow them to collect
federal reimbursement and gain traction with private insurers that tend to follow
Medicare’s lead.
“Right now, access to this program is largely dictated by the ZIP code in which
you live,” said Adam Brickman, senior policy director for Omada Health, an online
provider of diabetes prevention services. “That’s certainly a difficult situation for
Medicare beneficiaries who could benefit from this program.”
It is unclear whether the federal Centers for Medicare and Medicaid Services is
meeting its targets for the program. The agency projected it would enroll between
65,000 and 110,000 patients per year, but it did not respond to a request for the
number of people it has reached since the official launch last April.
The program, which arose out of the Affordable Care Act, was tested between 2013
and 2016 by providing in-person diabetes education classes through YMCAs. Officials
then decided to expand it nationally after finding it saved about $2,650 in total cost
of care per enrollee over a 15-month period. That’s a potentially dramatic impact in
a disease that costs $327 billion annually in medical expenses and lost wages and is
especially prevalent in the Medicare population.
In an interview last week, Adam Boehler, the head of Medicare’s innovation institute,
said the agency is open to the possibility of including digital providers to expand access
to the program. But he suggested the government is not in a position to make such
changes on the fly.

“If you make too many changes, it actually disrupts our ability to see whether the
model is working,” Boehler told STAT. “I have to look at different [possibilities]
while maintaining our ability to measure.”
The debate over the diabetes program showcases the broader tensions Medicare
must balance to achieve its stated goal of leveraging technology to improve health
care services and cut costs, while ensuring that such services are fully validated and
don’t open the door to waste or fraud.
Boehler, who previously led a company called Landmark Health that uses telehealth
services to provide in-home care to people with chronic conditions, has said he wants
Medicare to make better use of technology to “blow up” the fee-for-service system of
paying for health care based on the amount of services provided to patients.
“I want to get out of saying, ‘I’m going to pay $10 for this, or $15 for this,’” Boehler
said. “Part of our whole goal is to say … ‘Let’s give accountability out, and let phy-
sicians make decisions.’”
The diabetes prevention program is among the first efforts to test that philosophy,
known broadly as value-based care. Medicare provides small payments for enrolling
members in the program and keeping them engaged in classes, but the bulk of re-
imbursement is available only if providers meet certain weight loss targets. (The
primary target is to get members to achieve 5% weight loss in a year.)
Ensuring access to the program, especially in rural areas, appears to have emerged as
a significant challenge. Several states in the South and Midwest have no providers or
only a small handful. Even in states that have better numbers — Ohio has 39 providers
— most are clustered around urban areas such as Columbus and Cleveland.
Digital providers assert that their platforms are tailored to reach patients in com-
munities that lack community health centers or YMCAs that typically provide
diabetes prevention classes.

They also argue that providing access to online services is necessary to establish equity
between rural and urban populations, and between people who want to get help in
their homes versus those who prefer public settings.
“People should have choice,” said Dr. Neal Kaufman, co-founder and chief medical
officer of Canary Health, which provides services to prevent and manage diabetes and
other chronic conditions. “We find some people love the idea of going to an in-person
group — they thrive on it. … Other people enjoy the anonymity of a digital program.”
There remains some debate over whether digital health providers can achieve weight
loss goals and other targets as well as in-person providers. But there is an increasing
body of evidence to back up the effectiveness of digital services. The Institute for
Clinical and Economic Review, a nonpartisan research organization, released a report
in 2016 recommending that insurers cover diabetes prevention programs without a
copay. Private Medicare Advantage plans have also contracted with digital diabetes
prevention providers, and reported positive results.
Meanwhile, digital health companies are producing their own research to help make
the case. Omada has published several peer-reviewed studies to show that participants
in their program achieved lasting reductions in body weight and blood sugar.
The question remains whether that will be enough to convince the leaders of Medi-
care to take the digital leap. Their hopes were buoyed by some recent news from the
United Kingdom, where the National Health Service said it will expand access to
diabetes prevention services offered by digital providers in July.
“We’re cautiously optimistic” that U.S. officials will follow suit, said Brickman,
the policy director for Omada. “I know there are folks within [the U.S.] Health
and Human Services Department who really want to empower technology to help
prevent and manage the growing population of folks with chronic conditions.”

UP NEXT
Aspatientstellappsthey’re
feelingsuicidal,digitalhealth
startupsscrambletorespond
ByRebeccaRobbins

SAN FRANCISCO — Digital health apps, which let patients chat with doctors
or health coaches or even receive likely medical diagnoses from a bot, are transform-
ing modern health care. They are also — in practice — being used as suicide crisis
hotlines.
Patients are confessing suicidal thoughts using apps designed to help them manage
their diabetes or figure out why they might have a headache, according to industry
executives. As a result, many digital health startups are scrambling to figure out how
best to respond and when to call the police — questions that even suicide prevention
experts don’t have good answers to.
“To be honest, when we started this, I didn’t think it was as big an issue as it obviously
is,” said Daniel Nathrath, CEO of Ada Health.
The European company built a chatbot to provide smartphone users with possible
explanations for their medical complaints. Since the app launched in late 2016,
people around the world have used it to complete more than 10 million health
assessments. In about 130,000 of those cases, users have told Ada that they’re
struggling with suicidal thoughts or behaviors, the company said.
Aspatientstellappsthey’re
feelingsuicidal,digitalhealth
startupsscrambletorespond
BUSINESS
By Rebecca Robbins @REBECCADROBBINS MARCH 25, 2019|

For digital health startups, suicidal patients present just one of a number of unforeseen
crisis situations. At the telemedicine company American Well, a physician once con-
ducted a video visit with a woman who said that she had been punched by her spouse;
he was shouting in the background of the call, while the concerned physician called
911. Another company, which spoke on condition of anonymity, said it had to get au-
thorities involved when one patient using its service threatened to hurt her own child.
Admissions of suicidal thinking and planning, though, come up more often. And it’s
not just digital health startups that are grappling with what to do. After Facebook
developed an algorithm to flag posts from potentially suicidal users for review, the
company called first responders about 3,500 times in 12 months to check in on users
deemed to be at high risk. Hospitals and doctors’ offices are dealing with their own
new challenges, as patients respond to their social media accounts with suicidal com-
ments or voice them on clinics’ private online portals.
The phenomenon is, in some respects, no surprise: There’s a large body of research
showing that people are more willing to confess potentially taboo thoughts to a
computer than to a fellow human a few feet away.
“People are going to express their suicidality. We’ve destigmatized it. What we’ve
not done is prepared everybody” to respond, said April Foreman, a psychologist
who works on digital crisis care at the veterans crisis line run by the Department of
Veterans Affairs.
If you open Ada’s app and tell it you’ve been having difficulty concentrating for
months, the chatbot will ask you: Have you been feeling low or depressed? Is it
impacting your daily life? Say yes, and you’ll be asked: Any thoughts or urges related
to ending your own life?
Say yes again, and the chatbot will ask you: Might you try to end your life today? If you
say you’ve already tried today, the chatbot will tell you to call an ambulance right away.
And if you say you have plans to do so today? You’ll be encouraged to call someone
and tell them about your location and emotional state.

That’s where the conversation about suicidality ends with Ada’s chatbot. But some
users don’t stop there. They send emails expressing suicidal thoughts that go to Ada’s
customer support team — which often then takes on the task of researching and
recommending local mental health resources.
All of the digital health companies that STAT surveyed for this story said they have
a response plan in place that they follow when patients express suicidal thoughts on
their platforms. Such incidents may be rare in the context of all the cases they see, but
the frequency — and the stakes — are high enough to demand protocols, executives
said. In situations deemed to be lower risk, that often means getting patients on the
phone, directing them to a crisis hotline, or encouraging them to contact a friend.
And in situations in which a patient is deemed to be at higher suicide risk? Some
startups say they don’t hesitate to get emergency medical services or police involved
to conduct a wellness check when they think it’s in a patient’s best interest — whether
or not the patient wants it.
San Francisco-based Omada Health approaches things differently. The startup relies
on a network of coaches to exchange messages with patients with chronic conditions
like hypertension and type 2 diabetes. While Omada makes a phone call to check in
on patients who express suicidal thoughts, calling the authorities in these cases “isn’t
within our scope of practice,” said Dr. Carolyn Jasik, the company’s vice president of
medical affairs.
In many of these cases, startups are effectively trying to assess the likelihood that
patients will try to harm themselves. The trouble is, not even the experts know.
“This has been a problem that people have been struggling with for a really long time
— and there’s just no science on this,” said Matthew Nock, a Harvard psychologist
who studies suicide and self-injury. “People are largely winging it and using their
clinical wisdom to try and figure out when and how to intervene.”

Nock was part of a team of researchers that published a 2016 analysis of the
past 50 years of studies trying to predict suicidal thinking, suicide attempts, and
suicide deaths. Nearly all of those studies looked at how to know if a person is at
risk for suicidality using relatively long prediction windows, such as one year from
now. Just 0.1 percent of those studies looked at a window of less than a month.
Even fewer looked at a window of days or hours — the type of data that might be
particularly useful for digital health startups trying to help their suicidal patients.
Nock said he wants to see digital health companies conduct research and evaluate
their own practices to gather data on what works and what doesn’t, at different levels
of suicide risk. Just as a decision not to respond carries risk, so too does a decision to
respond too aggressively. Calling an ambulance for patients who are not at immediate
risk could backfire, if it makes them hesitate to seek help in the future.
Still, as digital health companies and traditional clinicians alike try to assess and
mitigate a patient’s suicide risk, not everyone is convinced that chatbots and
messaging apps are ready to play a useful role.
Dr. Peter Antall is chief medical officer at Boston-based American Well, a 13-year-
old company where telemedicine visits between a patient and a physician happen
exclusively by video, or, when the internet connection is poor, by phone. Antall is
excited about the potential of medical chatbots — he even advises one such startup,
called Gyant, in San Francisco — but he’s worried about what gets lost when a clini-
cian can’t see patients’ faces or hear their voices when they might be suicidal.
“Given the acuity and the seriousness of somebody potentially trying to kill them-
selves, I don’t believe that any of those other technologies are there at this point,”
Antall said.

By way of contrast, Antall pointed to a visit that took place a few years ago on
American Well’s platform. A patient video-conferenced in from her home, com-
plaining of chronic pain and seeking pain medication. The physician on the other
side of the video chat started the conversation from that vantage point, but quickly
picked up on concerning signs that pointed in another direction: The patient had a
depressed affect. She spoke in a low, monotone voice. She described other signs of
major depression and eventually admitted having active thoughts of suicide — and
a plan to do so.
It took some persuading, but the physician was able to convince the patient to go to
the emergency department. They ended the video call so the patient could head out
on her way. But the physician called back by phone within the hour to check in. The
news was good: The patient was already being assessed at the hospital.

UP NEXT
Newvoicesatpatients’
bedsides:Amazon,Google,
Microsoft,andApple
ByCaseyRoss

At first it was a novelty: Hospitals began using voice assistants to allow patients
to order lunch, check medication regimens, and get on-demand medical advice
at home.
But these devices, manufactured by Amazon, Google, Apple, Microsoft and others,
are now making deeper inroads into patient care. Hospitals are exploring new uses in
intensive care units and surgical recovery rooms, and contemplating a future in which
Alexa, or another voice avatar, becomes a virtual member of the medical team — mon-
itoring doctor-patient interactions, suggesting treatment approaches, or even alerting
caregivers to voice changes that could be an early warning of a health emergency.
“Why not have a connected speaker in the room listening to conversations?” asked
John Brownstein, chief innovation officer at Boston Children’s Hospital, which is
piloting dozens of voice applications. Voice technology still remains at the edges of
patient care, he added, but the hospital is already using it to improve the efficiency
of ICU care and help prepare doctors for transplant surgeries.
In New York, Northwell Health is preparing to put Alexa in private rooms next month to
allow patients to tap into their medical records. And Mayo Clinic is using voice to deliver
wound care instructions to some surgical patients and is studying the technology’s ability
to diagnose cardiovascular disease and other conditions.
By Casey Ross
Newvoicesatpatients’
bedsides:Amazon,Google,
Microsoft,andApple
HEALTH
@CASEYMROSS FEBRUARY 6, 2019|

Underlying that work is an increasingly fierce competition for health care dollars
among giant technology companies and scores of startups that are investing heavily
in voice-enabled products and services. Clinicians are waiting to see which of the
largest companies will be the first to introduce a smart speaker that fully complies
with health care privacy laws, a step that would allow them to delve even deeper into
patient care.
With artificially intelligent voice assistants rapidly becoming pervasive in the con-
sumer marketplace, many people are familiar with the devices and may soon expect
the same convenience in health care settings.
More than 100 million Alexa-enabled devices have been sold, while Google Assis-
tant, which gets a big boost from its use in Android smartphones, was expected to
be in 1 billion devices by the end of January, according to a company blog post. A
recent study by Nielsen reported that nearly 25 percent of U.S. households own a
smart speaker, a rapid uptick during the past year.
“We believe that the technology that exists in patients’ homes will be a demand that
patients will have sooner than later,” said Dr. Vishwanath Anantraman, chief innova-
tion architect at Northwell Health, adding that the hospital is planning to introduce
several new uses for voice technology and bots, which run automated tasks over the
internet, during the next few months.
“Voice tech can help improve service requests and deliver real-time analytics to the
staff to ensure patient satisfaction and patient safety,” he said. For clinicians, the
hospital system is focusing on retrieving complex information instantly on voice and
mobile devices.
Nathan Treloar, president and co-founder of Orbita, a Boston-based company work-
ing to create voice-enabled services in health care, said he is starting to see a shift in
the way hospitals are using voice. Most began with simple consumer-facing services,
such as voice programs that educate patients on the symptoms of diseases, he said.
But now they are embedding them into the process of delivering care.

One of the emerging uses, Treloar said, is deploying a voice assistant as an alternative
to the nurse’s call button, so that a patient can ask for help with specific problems
and get more timely service.
“The virtual assistant can collect enough context about the patient’s need so it can be
properly prioritized,” Treloar said, adding that such uses are likely to expand rapidly
with the introduction of devices designed to comply with U.S. patient privacy laws.
Several startups have already created HIPAA-compliant voice software for use
with electronic medical records systems. Sopris Health, a Denver-based company,
developed a product designed to automatically convert a doctor-patient conver-
sation into text that is then loaded into a doctor’s note. Other competitors in the
field include Suki, Notable, Nuance, and Seattle-based SayKara, which is led by
former Amazon engineers.
Health care technology specialists said these types of products could make electronic
health records easier to use and may help ease documentation burdens. But they are
unlikely to result in a revolution in record-keeping any time soon.
“I do not think in the near, medium, or long future the EMR is going to be replaced
with a voice-enabled application,” said Darren Dworkin, chief of information at
Cedars-Sinai, a Los Angeles-based health system. He added: “Like many technologies
before this, the important part will be that we don’t get too far ahead with the hype.
Voice is a wonderfully empowering technology, but we have to figure out how it finds
its rightful place.”
In 2017, Cedars-Sinai invested in a voice company called Aiva Inc. and took a small
equity stake in the firm as part of an accelerator program the hospital created to ex-
plore new technologies. Dworkin said Cedars-Sinai is particularly interested in using
the technology to help deliver care instructions and notifications to patients at home.

Meanwhile, Aiva has also attracted capital from the Google Assistant Investment
program and Amazon’s Alexa Fund — vehicles the companies use to invest in
early-stage companies.
Mayo Clinic, one of the pioneers of voice in health care, built an Alexa-enabled
program to deliver first aid instructions to consumers. More recently, it has begun
piloting the use of the technology to deliver post-discharge instructions to patients
recovering from surgeries to remove skin lesions.
Dr. Sandhya Pruthi, medical director of global business solutions at Mayo, said its
study on the use of voice to diagnose cardiovascular disease — it found analyzing
speech signals such as tone and intensity could help detect coronary artery disease
— points to an exciting future for the technology. Several companies, such as Sonde
Health, are developing diagnostic tools based on changes in a person’s voice. The
hope is that analyzing subtle shifts in tone, clarity, and cadence will help predict the
onset of psychotic episodes, stroke, and other health emergencies.
“It opens possibilities to deliver care at a distance,” Pruthi said. “Think about people
living in small towns who aren’t always getting access to care and knowing when to
get health care. Could this be an opportunity if someone had symptoms to say, ‘It’s
time for this to get checked out?’”
She said she attended the recent Alexa voice conference in Chattanooga, Tenn.,
where other attendees were talking about the use of voice to detect symptoms of
Parkinson’s disease and autism, “It’s still very early,” Pruthi said. “You do need to
do the work. There has to be evidence that this makes sense.”

UP NEXT
HowanIBMWatsonHealth
rescue mission collapsed —
andatopexecutivewasousted
ByCaseyRoss&IkeSwetlitz

The elite team of engineers and medical specialists assembled by IBM’s Watson
Health division had the innocuous code name “Project Josephine,” but its mission
could not have been more urgent: to fix the artificial intelligence software at the core
of the company’s campaign to tackle the $7 trillion global health care market.
The predicament faced by IBM officials, STAT has found, was that it could not get
its software to reliably understand and analyze language in patient medical records.
That was critical for the company to deliver on multimillion-dollar contracts with
hospitals and drug companies.
To fix the problem, IBM recruited 10 of its smartest minds in the spring of 2017,
according to an internal slide deck obtained by STAT. Their charge was to quickly
improve Watson’s comprehension, and ensure it could perform consistently on a
range of tasks, from matching patients with clinical trials to suggesting optimal
cancer treatments. The team was to be given nine months to turn things around
and “improve accuracy,” “standardize terminology,” and “reduce overlap” among
the groups working on a dozen different health products.
But not long after the project’s conception, it began to fall apart. Managers of Watson
Health didn’t make the project a priority, failing to invest necessary time and money
to pursue its proposed solutions.
HowanIBMWatsonHealth
rescuemissioncollapsed—and
atopexecutivewasousted
EXCLUSIVE
By Casey Ross and Ike Swetlitz@CASEYMROSS NOVEMBER 1, 2018|

Within a year, most members of the team had left the company, according to former
employees and social media postings.
Among the team members who departed were Dr. Andrew Norden, deputy chief
health officer of Watson Health; Dr. Patrick McNeillie, clinical lead of Watson for Ge-
nomics; and Meena Nagarajan, an expert developer of artificial intelligence software.
The failure to achieve the team’s mission was but one factor that undermined business
performance and led to the removal last month of Deborah DiSanzo after three years
as general manager of Watson Health. DiSanzo headed the division that oversaw Proj-
ect Josephine. Former employees familiar with her role said she decided against invest-
ing the resources the team needed to fix the problems with the software.
The issues with Project Josephine were hardly the only difficulties plaguing IBM’s
Watson Health division, but the ability to read medical records and other clinical doc-
uments is critical to the success of its products. The company’s struggle to solve prob-
lems with this technology contributed to missed contract deadlines, souring client
relationships, and difficulties attracting new customers.
As the problems persisted, one of Watson Health’s earliest and largest customers, Teva
Pharmaceuticals, scaled back its contract with the company after delays in providing
data services designed to help it develop treatments for chronic conditions such as
asthma, pain, and neurodegenerative diseases, according to two people familiar with
the matter.
One former IBM employee, speaking on the condition of anonymity, said: “I expect-
ed Teva to walk away from the contract last year or the year before actually because of
how long it was taking. They stopped paying at a certain point — period.”
In an emailed statement, Teva spokesperson Doris Saltkill wrote that the company’s
work with IBM was scaled back as part of company-wide restructuring that began in
December 2017.

“Many programs were resized to meet our financial objectives,” the statement said.
“This is the case with our business with IBM Watson who remains a valued partner for
our development programs in digital health.”
In interviews with STAT, former and current IBM employees pointed to major lapses
that went far beyond Watson’s struggle with reading clinical records, including the
decision to spend billions of dollars acquiring health data companies without a clear
plan for integrating the information into a comprehensible format. Several people also
cited hamstrung efforts to create cloud infrastructure to store these vast data streams
and deliver promised insights.
“One by one, the customers were leaving,” said one former Watson Health employee.
“And the reason most of them left is that we never invested in the technology needed
to do the job we were being asked to do.”
IBM spokesman Edward Barbini did not directly answer questions about how its tech-
nical challenges have affected client relationships. In an email, he said Watson Health
has conducted research to standardize its technology for reading and analyzing clinical
records, a capability known as natural language processing. “As a result of that work,
IBM is on the path to creating a company-wide state-of-the-art framework for natural
language processing,” Barbini wrote.
He also wrote that the company is still working on cloud infrastructure to serve
its health clients. The statement said the goal is to create a “hybrid IBM cloud” to
allow clients to keep some data on their premises and protect access in line with
U.S. privacy rules.
DiSanzo could not be reached for comment. She was ousted shortly after the company
disclosed a disappointing financial performance in the third quarter for the cognitive
services division, which includes Watson Health.

The lackluster performance of the health business led to a series of tense encounters
in which IBM executives, including CEO Ginni Rometty, grilled DiSanzo about its
struggles, according to the employees, who asked for anonymity out of fear of reprisals
from the company. But the company’s challenges in health care run far deeper than
DiSanzo, who was unable to fix technical and business problems that took root before
she was hired to lead the Cambridge, Mass.-based health division in September 2015.
IBM is trying to regain its momentum in health care following a series of strategy
missteps and technical problems. The health sector is seen as a key market for IBM,
which is transitioning from an old-line hardware company to one that operates in
the cloud and along the cutting edge of artificial intelligence.
In recent years, IBM has begun selling an array of products using Watson to help
doctors and drug companies analyze genomic data and match patients to the best
cancer treatments and clinical trials. It also developed products to help hospitals
manage their pay-for-performance contracts and improve patient safety, among
other tasks.
To some degree, its difficulty selling the products reflects broader challenges associ-
ated with using artificial intelligence to deliver insights on health care data that are
collected for divergent purposes and housed in different formats. But it is also clear
that IBM failed to appreciate the extent of those challenges and made promises to
clients that it has struggled to keep.
Many of its problems date back to the acquisition of four companies with access to
different types of health care data in 2015 and 2016.
The company bought Explorys, a Cleveland Clinic spinoff, and Phytel, a Dallas
maker of health software, to get access to their large clinical datasets and expertise.
ACQUISITIONS WERE FOLLOWED BY LAYOFFS

It also acquired Merge Healthcare, a Chicago company with a trove of medical
imaging data, and Truven, which holds hundreds of millions of records on patient
health care claims.
Former employees who worked with these business units said the overarching
strategy behind the acquisitions — to merge their data to drive insights for Watson
products — faced significant technical hurdles that IBM failed to overcome.
The employees said managers at the company became aware of these problems, but
failed to invest the time and resources to resolve them. Instead, these people said,
they became trapped by the company’s focus on quarterly earnings, making it harder
to justify expenses that would sacrifice short-term revenue gains to achieve long-
term success.
In June, IBM laid off employees across the acquired companies and decided to scale
back its services to help hospitals manage pay-for-performance contracts. The extent
of the layoffs are unclear, as IBM declined to provide exact numbers, but they affected
employees in offices across the country.
The layoffs occurred about 14 months after the conception of Project Josephine,
which was designed to solve a more systemic problem with Watson’s natural language
processing capability. Customers were telling IBM its products simply weren’t working
as intended, according to the internal slide deck obtained by STAT.
“Clients report attributes found with NLP are often incorrect or missed, preventing
sales,” the internal document states, using an abbreviation for natural language process-
ing. The document, dated April 26, 2017, also describes a troublesome “lack of stan-
dardization” among teams working on different projects, noting as an example, that
four of Watson’s cancer products each defined a particular type of leukemia in different
ways. “NLP developed by one team cannot be used by another,” the document states.
TROUBLES WITH NATURAL LANGUAGE PROCESSING

It also cites duplication within IBM, stating that “25 percent of developers across 12
projects are independently developing the same NLP.”
Project Josephine’s leaders used the document to explain technical problems and
propose a cost-effective solution for improving accuracy. It describes a plan to deploy
employees to annotate medical records to develop a standardized dictionary that
could be used to train a machine-learning-based system for use across the organiza-
tion, instead of developing separate algorithms for each product.
The document states that Project Josephine was planned to unfold over nine months,
with a goal to annotate between 1,200 and 2,000 clinical records to establish its profi-
ciency. A separate group of employees, working under the moniker “Tiger Team,” was
conceived to implement the software solutions across Watson Health’s products.
But the work never moved forward, according to two former employees familiar
with the matter, even as clinical leaders within Watson Health flagged the challenges
with natural language processing. In June 2017, months after the Project Josephine
team was formed, its highest-ranking member, Norden, gave a separate presentation.
It detailed the shortcomings of one of the company’s flagship products, Watson for
Oncology, which advises doctors how to treat cancer patients.
“Limited NLP currently in use,” the presentation reads. “Output errors not un-
common.” The presentation was shared widely among Watson Health management
at the time.
Indeed, hospitals around the world told STAT last year that they were not actually
using Watson’s natural language processing capabilities, opting instead to enter the
data by hand. Meanwhile, Watson’s marketers emphasized the software’s ability to
automatically read medical records.
DiSanzo opted not to proceed with the Project Josephine effort, according to the for-
mer employees, which would have required significant investments of time, money,
and people.

The troubles with natural language processing have continued. One former employee
said the company has taken only “baby steps” toward addressing the problem in the
past year. Nearly every member initially appointed to Project Josephine has since left
the company, including Norden. Most have taken jobs with other technology com-
panies, including Facebook and Syapse, a San Francisco-based company that also uses
artificial intelligence to provide personalized cancer treatment recommendations,
according to their LinkedIn profiles.
IBM’s troubles with natural language processing have made it harder to execute on
its contracts and keep customers happy. Among Watson Health’s earliest clients was
Froedtert Health, a Wisconsin-based hospital system that adopted Watson for Clinical
Trials Matching, a product designed to comb patients’ medical records to help link
them with opportunities to participate in trials.
An initial press release about the deal said Froedtert was planning to initiate the pro-
gram in the fall of 2016. But Froedtert’s website now says it plans to begin using IBM’s
software on cancer patients sometime in 2018.
A hospital spokeswoman declined STAT’s requests to speak with an oncologist work-
ing with the product or answer questions about the status of the work and its contract
with IBM. The hospital is one of a handful of U.S. clients to adopt the clinical trials
matching software.
A former IBM employee said implementation of the product is time-consuming for
hospitals because it requires customizing Watson to understand the language and
context of an organization’s electronic health records. Such record systems vary wide-
ly based on the vendor of the system and the hospital’s approach to treating patients,
among other factors.
Among the other users of the system is Mayo Clinic, which released a study earlier this
year showing that early use of the clinical trials matching software was associated with
a significant increase in patients enrolled in Mayo’s trials for breast cancer.

However, a former IBM employee familiar with the work at Mayo said getting the
product to perform effectively has been a long and frustrating process.
“That’s been going on for five years at Mayo,” this person said. “People were pulling
their hair out. It was very difficult to use.”
A spokeswoman for Mayo Clinic said the doctor in charge of the work with IBM was
unavailable to speak with STAT, but added that Mayo continues to use IBM’s software
to match breast cancer patients to trials and is still working with the company to devel-
op the product to help patients with other types of cancer.

UP NEXT
Theopportunity‘ishuge’:
Whytechdevelopersaretry-
ingtotacklementalhealth
ByRebeccaRobbins

SAN FRANCISCO — Here in the technology epicenter of the world, developers
are increasingly writing code and launching products to try to disrupt yet another
field: mental health.
Even as big tech players have conquered the markets in industries like transportation
and lodging, they’ve largely steered clear of mental health treatment. Now, however,
with an influx of funding, companies are revamping pills with digital sensors, de-
signing virtual reality worlds to treat addiction and other conditions, and building
chatbots for interactive therapy.
The excitement in the field was apparent this week at the inaugural Anxiety Tech
summit, a day of TED-Talk-style speeches that felt much like any of the other hack-
athons and demo days that clutter the calendar here. The venue was scattered with
signs that trumpeted the event’s unusual mission: “Inspiring developers to build for
mental health.”
The summit drew 150 attendees both from small startups and Silicon Valley giants
including Apple and Google, according to co-organizer Kari Ferguson, a writer who
has blogged about her struggle with obsessive-compulsive disorder.
When it comes to bringing technology to mental health, the opportunity “is huge,”
said Nicolas Rosencovich, CEO and co-founder of a startup called MindCotine.
Theopportunity‘ishuge’:
Whytechdevelopersaretrying
totacklementalhealth
BUSINESS
By Rebecca Robbins @REBECCADROBBINS JULY 20, 2019|

The company, which set up a booth at the summit, is developing a smartphone-pow-
ered virtual reality program for nicotine addiction that it plans to soon start market-
ing to corporate wellness programs in the U.S.
In the first half of this year, startups working on behavioral health collectively raised
$273 million — more than in any prior six-month period on record, according to
a recent report from the digital health venture firm Rock Health. More than half
of the 15 companies that brought in that funding are working on products that are
virtual or on-demand.
Among the companies funded this year: the telepsychiatry startup Regroup Therapy
and Woebot Labs, which is working on a mental health services chatbot branded as
“Your charming robot friend who is ready to listen, 24/7.”
Entrepreneurs pointed to a number of factors driving the surge in investment and
activity. The cost of virtual reality and other hardware has plummeted, while the
technical quality of the technology has improved. And the stigma around mental
health is lessening as conversations about mental illness move into the mainstream.
Still, it’s not clear whether insurers and employers will be willing to pay for these
products, especially with a relatively thin track record of evidence. There are big reg-
ulatory question marks as many companies seeking to treat or manage mental health
conditions market their products as an unregulated wellness play.
Another problem: “Some of these things are overkill,” said Amber Case, a researcher
who studies the impact of technology on culture. While she’s pleased to see more
commercial activity to develop tech for mental health, some applications don’t seem
likely to help, she said. Case points to the example of an smartphone alert telling
users to drink water: It might interrupt them in the wrong situation, or when they
don’t have water handy. And it could make them feel guilty that they haven’t hydrat-
ed enough, spurring more anxiety.
Despite all the obstacles, entrepreneurs and investors see a wide-open field.

“Nobody’s won in mental health yet,” said Jonathan Sockell, chief operating officer
of Limbix, a startup gearing up to start selling its virtual reality goggles to mental
health clinics, hospitals, and rehab clinics. Among the company’s VR programs is an
exposure therapy for patients with phobias or trauma associated with driving. While
patients strap on the headset, clinicians can work with them to introduce different
conditions (a clear or rainy day) or different road situations (a bridge or a tunnel or
blind left turns).
Both large pharma stalwarts and new startups are also trying to bring new technology
to the medicines used to treat mental illness.
The Japanese drug maker Otsuka Pharmaceutical and its Silicon Valley partner
Proteus Digital Health last November got marketing approval from the Food and
Drug Administration for the first so-called digital pill, designed to alert your doctor
when you swallow it. The pill is an upgraded version of Abilify, the antipsychotic long
approved treat schizophrenia, bipolar disorder, and depression.
A San Francisco startup called Kick plans to soon start selling a cardiac drug meant
to be popped like a mint to people anxious about public speaking and first dates.
The company intends to launch its telemedicine platform this quarter, according
to Justin Ip, Kick’s founder and CEO.
In many cases, the new technology being brought to mental health carries with it
the cultural quirks of Silicon Valley.
Consider one of the exhibitors at this week’s summit: a virtual reality startup with
the creatively misspelled name Happinss. The company, which licenses its virtual
reality programs to employer wellness programs to try to monitor and reduce stress,
set up a table at the event to display its VR headsets. Plopped on the floor below
were two neon green and blue beanbags, for people to lounge in while trying out the
VR experience.
“Part of the solution is to have beanbags on site,” co-founder Wang-Tsu Liu explained.

How health tech is redefining the future of biopharma and care delivery

Recommandé

Recommandé

Contenu connexe

Tendances

Tendances (20)

Similaire à How health tech is redefining the future of biopharma and care delivery

Similaire à How health tech is redefining the future of biopharma and care delivery (20)

Plus de Δρ. Γιώργος K. Κασάπης

Plus de Δρ. Γιώργος K. Κασάπης (20)

Dernier

Dernier (20)

How health tech is redefining the future of biopharma and care delivery