AI's Health Exam
Analysis of artificial intelligence and machine learning in healthcare
PitchBook is a Morningstar company. Comprehensive, accurate and hard-to-find data for professionals doing business in
the private markets.
Credits & Contact
Analyst
CAMERON STANFILL Analyst
cameron.stanfill@pitchbook.com
Contact PitchBook
pitchbook.com
RESEARCH
reports@pitchbook.com
EDITORIAL
editorial@pitchbook.com
Contents
Key Takeaways
1
Introduction
2
Mapping the healthcare
AI environment
3
Large-cap outlook
4
Which segments are
prime for near-term
success?
5
Headwinds
8
Looking forward
9
Key Takeaways
Artificial intelligence and machine learning (AI/ML) is beginning to resonate
with VC investors in the healthcare sector, recording 48 deals in 2017 raising
$718 million, representing 65% and 126% growth YoY, respectively.
Within healthcare, we expect AI to be used mainly to augment workers'
skills, as opposed to displacing them. Personal AI assistants for physicians/
nurses can recall relevant medical research, compile population patient data,
provide initial possible diagnoses, and optimize workflow. Time saved by these
innovations represent countless hours that can be used more efficiently by
industry professionals.
Deep learning algorithms that leverage image recognition are already more
accurate than panels of specialists in detecting and diagnosing pneumonia, TB,
certain early stage cancers, and other diseases.
COPYRIGHT 2018 by PitchBook Data, Inc. All rights reserved. No part of this publication may be reproduced in any form or by any meansgraphic,
electronic, or mechanical, including photocopying, recording, taping, and information storage and retrieval systemswithout the express written permission
of PitchBook Data, Inc. Contents are based on information from sources believed to be reliable, but accuracy and completeness cannot be guaranteed.
Nothing herein should be construed as any past, current or future recommendation to buy or sell any security or an offer to sell, or a solicitation of an offer
to buy any security. This material does not purport to contain all of the information that a prospective investor may wish to consider and is not to be relied
upon as such or used in substitution for the exercise of independent judgment.
2
Analysis of AI/ML in Healthcare
Introduction
In our initial coverage of the artificial intelligence and machine learning (AI/ML)
vertical, we concluded that the technology has the potential to impact nearly
every industry by automating a range of tasks. Due to this widespread potential,
we want to provide deeper analyses into specific industries to highlight the
rapidly growing and transformative use cases for AI/ML. The potential for AI/ML
is beginning to resonate with venture investors particularly in healthcare, as 2017
deal count and value grew approximately 65% and 126% YoY, respectively.
Healthcare AI deal activity by year
The healthcare market is giganticrepresenting 17.9% of US GDP in 20161
and produces a massive amount of data, positioning the industry to receive
significant benefits from further integration of AI/ML. Highlighting some of the
overarching inefficiency, healthcare's percentage of GDP has doubled since 1980,
but health outcomes haven't improved at the same rate. While this technology
can increase business efficiency and profitability, AI/ML advancements in
healthcare also aim to improve patient outcomes and enhance the overall patient
experience. The widespread data digitization in healthcare is the backdrop
for all potential use cases in the space, but data structure and fragmentation
(e.g. isolated within hospitals, insurance groups) serve as significant hurdles to
widespread adoption.
Source: PitchBook
0
10
20
30
40
50
60
$0
$100
$200
$300
$400
$500
$600
$700
$800
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
Deal Value ($M)
Deal Count
The widespread data digitization in
healthcare is the backdrop for all
potential use cases in the space, but
data structure and fragmentation (e.g.
isolated within hospitals, insurance
groups) serve as significant hurdles to
widespread adoption.
1 Centers for Medicare and Medicaid Services
3
Analysis of AI/ML in Healthcare
Mapping the healthcare AI environment
Upon examining the deal flow data, we have segmented the current VC
investment in AI healthcare market into five distinct subcategories:
1. Automation & Assistance This segment represents 40% of deals and
includes those companies focusing on assisting or augmenting day-to-day
processes for medical professionals and patients.
2. Computer Vision This segment represents 19% of deals and includes
companies pursuing applications of computer vision of medical images to aid
in diagnosis, preventative medicine, etc.
3. Population Health & Predictive Analytics This segment represents 15%
of deals and includes companies creating platforms focusing on improving
population health using aggregated genome, patient, and treatment data, as
well as businesses seeking to improve the healthcare system at a more macro
level.
4. Personalized Medicine This segment represents 18% of deals and includes
companies working to create tailored healthcare analytics or devices to
improve personal health outcomes. AIenabled personal wearables and
sensors are included in this subcategory.
5. Drug Discovery This segment represents 8% of deals and includes
companies using AI/ML to research and develop novel drugs.
Automa on &
Assistance
Computer Vision
Drug
Discovery
Personalized
Healthcare
Popula on Health &
Predic ve Analy cs
VC investment in healthcare-focused AI (#) by sector2
Source: PitchBook
2 All data in this section is for the period from 20062017
4
Analysis of AI/ML in Healthcare
Large-cap outlook

While we can't know exactly what is mentioned behind closed doors, AI/
ML hasn't quite broken into the public commentary from executives of public
healthcare firms. Companies have instead opted for more broad classifiers
like "novel technology", big data/analytics, personalization of treatment and
commitments to "business development strategies". One explicit mention
came from the insurer United Healthcare, which stated they were continuing to
support "existing initiatives in AI" that could benefit both the insurance practice
as well as their healthcare solutions subsidiary Optum. Given the nascence of the
drug discovery space within healthcare AI/ML, the relative lack of penetration
into the large-cap pharmaceutical market was fairly expected.
Some technology giants and AI pioneers also have ambitions within the
healthcare sector. Most notably, one of the earliest movers is IBM's Watson
Health, which according to the company is beginning to scalereaching 115,000
patients and being used in 15 hospitals. Watson's breadth of applications has
also increased over the past year; the program has now been trained to identify
13 types of cancer, up from four a year ago. Google's DeepMind also has a health
initiative that has partnered with the NHS in the UK on an app called Streams
that uses technology in mobile phones to alert healthcare staff when a patient's
health deteriorates. As companies in the healthcare AI/ML field continue to
grow and prove their commercial capabilities, we expect the attention from
large corporates to increase over the next 10 years, including some notable
acquisitions.
IBM's Watson Health, which
according to the company is
beginning to scalereaching
115,000 patients and being used in
15 hospitals.
5
Analysis of AI/ML in Healthcare
Which segments are prime for near-term success?
Automation & Assistance
0
2
4
6
8
10
12
14
16
18
$0
$50
$100
$150
$200
$250
$300
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
Deal Value ($M)
Deal Count
Source: PitchBook
The first broad application of AI/ML in healthcare in which we expect to
see widespread adoption is the automation of menial or recurring tasks via
personal assistance for both patients and practitioners. These will be focused
on increasing workers' knowledge and improving day-to-day efficiency to allow
more focus on value-additive tasks, with the ultimate goals of achieving better
health outcomes and cost savings. With over $200 million of VC investment in
the last two years, this is a large and growing niche of the healthcare AI space.
We believe the automation function is key because of the nature of labor in
healthcare. AI/ML-enabled personal assistants for physicians/nurses could
recall relevant medical research, compile patient data, provide initial possible
diagnoses, and perform scribing duties, among others. These innovations
should allow medical practices to reallocate countless worker hours toward
more efficient tasks. One specific example is the automation of data entry into
electronic health records (EHRs) via speech recognition technology.
Similar benefits can be realized for patients, as AI/ML applications can improve
ease of access through initial contact points like chatbots. Digital preliminary
diagnoses can also limit unnecessary clinic visits and re-admissions, decreasing
costs for patients and freeing up capacity in the healthcare system. Once
patients enter a medical facility, AI/ML can be incorporated into initial touch-
points for patients to assess the severity of their condition to optimize physician
scheduling.
Another costly and time-consuming task within the healthcare industry is the
administration of clinical trials. AI/ML are aiming to help solve frictions in the
drug approval process, from finding eligible patients to participate in trials, to
aggregating and interpreting results. This is the niche that Clinithink, a UK-
based startup with over $22 million in VC funding, aims to serve by working to
transform unstructured clinical trial data into structured datasets to reduce risk
and accelerate go-to-market time for drugs.
VC activity in automation & assistance-focused AI companies
6
Analysis of AI/ML in Healthcare
Computer vision
0
2
4
6
8
10
12
$0
$20
$40
$60
$80
$100
$120
2010
2011
2012
2013
2014
2015
2016
2017
Deal Value ($M)
Deal Count
Computer vision is the second area of healthcare AI in which we've seen
extensive growth. A major use of AI/ML technology is applying it to medical
imaging analysis, which has the potential to deliver staggering improvements
in speed and accuracy. With the success of many general image recognition
algorithms created by Google, Clarafai and others, we see opportunity in this
space as nearing to commercial realization. The healthcare computer vision
space saw a significant influx in VC investment over the last two years, tallying
over $100 million in 2017 alone.
AI/ML algorithms built to analyze medical imaging have already exceeded
panels of specialists in diagnosing pneumonia, TB, certain early stage cancers.
AI implementation in diagnosis could drastically increase the amount of patient
data that could be processed, expediting test results while increasing accuracy.
Computer vision also applies to other visual-based diagnoses/assessments
like dermatology. For example, Stanford's skin cancer classification algorithm
matched the performance of dermatologists based on pictures of skin lesions,
while a separate group of scientists developed another neural network created
by that identifies onychomycosis (nail fungus).
These photo-based diagnosis applications could extend the health benefits
beyond the clinic and serve as a preliminary point of contact via patients' mobile
phones. While early successes of AI/ML programs for medical imaging have
fueled significant growth in financing, this market is still relatively nascent. As
such, we expect this space will continue to receive more funding over the near
term, and consequentially attract more new entrants.
VC activity in computer vision-focused AI companies
Source: PitchBook
These photo-based diagnosis
applications could extend the health
benefits beyond the clinic and serve
as a preliminary point of contact via
patients' mobile phones.
7
Analysis of AI/ML in Healthcare
Personalized healthcare

AI/ML in the healthcare industry will also allow the opportunity for greater
personalization of care. Armed with more detailed data on individuals, healthcare
can be tailored based on personal history, genetics, demographics and real-
time inputs. A key use for this is changing the way we prescribe medications.
Currently, many treatments are given out in a standardized fashion, especially
regarding the size of the dose. This can lead not only to less effective treatments,
but also the potential for adverse reactionsthe cause of thousands of deaths
every year3. Chemotherapy treatments, for example, are largely thought of as
a one-size-fits-all option; however, personalization in this instance has huge
potential by determining the biomarkers found in the patient's tumor and
tailoring molecular-based therapies or the most effective chemotherapeutic
agents depending on that patient's specifics. Siris Medical is already providing
personalized cancer treatments that should accelerate the treatment planning
process to streamline workflow and strengthen health outcomes.
Reinforcement learning algorithms, such as contextual bandits, provide a
promising solution to personalization interfaces. These programs use contextual
and historical data to decide among the available options, then receives reward-
based feedback based on decision it has made. The algorithm then learns to
optimize its decision to the most favorable outcomes by repeating the process
a multitude of times. This feedback process is not perfect and the bandit may
discard a viable action during the course of its learning, so the algorithm must
strike a balance between exploiting what it believes is current optimal decision
while still exploring all available actions.
The reward function, especially when dealing with biology and health, is
fairly abstract because there aren't clean, quantifiable data points or easily
determinable right or wrong answers, especially not relayed in a timely fashion.
This is where the integration of real-time data from wearables and other sensors
begins to play a key role in proving the feasibility of many of these applications
and, ultimately, making the analytics more accurate and valuable. For example, a
wearable device could track blood sugar data to provide feedback to the bandit
on the selected treatment, then determine a positive or negative reward in near
real-time that serves to improve the program.
Note: Contextual bandits are, in
essence, computerizations of a
slot machine with multiple arms
corresponding to the number of
available options where every decision
returns a reward to the program.
Before the program (i.e., bandit)
makes its next decision, it considers
all the rewards received previously
in conjunction with information
about the current context (external
data, multi-world testing, etc.). For
example: programmatically generating
personalized headlines on a news site
to optimize clicks, the bandit reviews
past decisions/rewards, headlines
from other sources, what the user has
clicked on before, etc. to come to a
decision on which option to select.
3 Shepherd, Greene & Mohorn, Phillip & Yacoub, Kristina & Williams May, Dianne. (2012).
Adverse Drug Reaction Deaths Reported in United States Vital Statistics, 1999-2006.
The Annals of pharmacotherapy. 46.
8
Analysis of AI/ML in Healthcare
Headwinds

Although deployment of AI/ML in the healthcare industry might deliver better
health outcomes in a more efficient manner, many factors will impose limitations
on the adoption and scalability of these technologies. The implementation of
large-scale AI programs is likely to be mired by prevalent data fragmentation
issues, as the healthcare sector produces large amounts of unstructured
data that are handled differently by various institutions. As more data are
collected (via wearables, IoT, etc.), more emphasis will be put on the collection
and cleaning of data before it can be used to train algorithms. Additionally,
companies that choose to operate in this space must consider the regulatory and
security issues around healthcare data, a barrier that has kept many companies
from entering the industry. Clean datasets and the ability to securely store and
share this data are critical to expanding capabilities of the healthcare AI field in a
sustainable manner.
Humans innately like to know the reasoning behind decision making, especially
when it concerns our health. As AI/ML programs start assisting with diagnoses,
more questions will arise about the rationale behind recommendations.
The "black-box problem" as this phenomenon is known is an important
consideration in the AI/ML field. However, as there is already a relative lack of
interpretability of human intuition relating to complicated medical situations,
we believe this fear in healthcare may be slightly overhyped. When an AI agent
makes diagnoses, it could have the benefit of drawing on the complete patient
history, aggregation of other relevant patients, relevant studies and other sources
to come to an evidence-based conclusion. That said, it will still be important
to continue to have humans make the final decisions as a sanity check and for
patient confidence, because the importance of the physician's experience and
direct contact with the patient cannot be denied. This is particularly pertinent
when it comes to assigning culpability for misdiagnoses.
On a related note, we find another hurdle to widespread adoption of AI/ML
in healthcare to be the preference for human relationships. Automating initial
contact points with both physicians or administrative workers in clinics may not
be well received by the broader patient population. However, this should be
mitigated partially because in general we see AI/ML technology in this industry
as working much more to assist workers rather than replace them.
9
Analysis of AI/ML in Healthcare
Looking forward

On the frontier of the AI/ML healthcare field is the possibility of new drug
discovery by machines. The spontaneity and creativity required in novel
discovery adds an extra layer of complexity to this use case of AI/ML. Bringing a
new drug to market is a notoriously time-consuming journey that can take more
than a decade and between $650 million and $2+ billion of investment4. Many
times, large-cap pharma & biotechs try to avoid this commitment by acquiring
smaller companies with promising drugs already in the clinical trial process. AI/
ML programs may shine in this context due to their ability to model a multitude of
outcomes which help predict side effects or lack of efficacy, thus reducing failure
rates and rapidly shortening R&D timelines, while potentially allowing more
efficient in-house drug development processes.
Though drug discovery with AI/ML is still in a nascent stage, algorithms have
already been used to determine whether a drug molecule will bind with the
target protein with up to 99% accuracy and are working to predict the outcomes
of CRISPR-Cas9 treatments. These predictive modeling techniques can also
be applied to genome data to find novel patterns and approaches to treating
genetic diseases.
Conclusion
Due to the healthcare industry's size and depth of data, it's clearly a huge market
opportunity for AI/ML. As many applications remain in the proof-of-concept
phase or singleclinic partnerships, the speed of adoption and total market
penetration remains uncertain. Over the next several years, we expect the more
developed areas like assistants, medical imaging, and diagnosis to increase
reliability and drive further use of these platform services, with the winners
enjoying positive network effects. As the industry receives more funding to
continue to expand research efforts, we expect the pace of breakthroughs to
accelerate. The prospect of improved health at the population level provided
more efficiently and lower cost is an enticing outcome that AI/ML can provide
and cannot be ignored.
3 Tufts Center for the Study of Drug Development