Machine Learning: Risk Scoring on Steroids?
Machine learning sounds like another one of those ambiguous high-tech
terms that most of us tune out until we are forced to pay attention to it, but
Amazon, Google, and Facebook are all making big bets on the field. And
that means it’s likely a future driver in the tech industry, which according to
many a modern thinker means it will probably be a future driver in medicine
too. Still mostly in the realm of academia and with numerous hurdles to overcome, machine learning risk prediction holds interesting promise.
In short, machine learning is human recognition done by computers. Used
for risk prediction, machine learning has “similar goals to regression-based
approaches but different motivating philosophies,” wrote Benjamin A. Goldstein, PhD, and colleagues in a July review on the topic.1 Dr. Goldstein is an
assistant professor in the Department of Biostatistics and Bioinformatics at
Duke University, Durham, North Carolina.
Rather than relying on a pre-specified theory or set of assumptions, the
machine models can “search for the optimal fit within certain constraints
(specific to the individual algorithm).” The benefit of this is a better final
prediction model, but at the “sacrifice of interpretability of how risk factors
relate to the outcome of interest.”
According to Dr. Goldstein, one major advantage of the technology is
that it may allow for standardized approaches and techniques that can
then be applied to specific patient populations. So, for instance, while a big
complaint about the Framingham Risk Score was that it was derived from a
very specific and homogenous patient population limiting its applicability, a
machine learning technique can be developed globally but applied locally.
“Traditionally, this wasn’t very feasible because we didn’t have large
enough cohorts of specific populations,” said Dr. Goldstein, but EHRs have
changed this.
“So, I can come up with the risk score that is highly relevant and tailored
for the Duke Medical Center population, which has a sizeable African-Amer-
ican population, a large low-income population, and a decent-size Latino
population,” he said in an interview with CSWN. “These are cohorts which
may have their own idiosyncrasies in terms of their health care that can and
should be taken into account and that may not be relevant to other locales,
which will have their own health concerns that need to be managed.”
Also, as new information on novel risk markers comes in, there is more opportunity to easily integrate it, said Dr. Goldstein, or even as policies change,
like the recent push to reduce 30-day readmissions.
“What machine learning says is I don’t care about the causal relationship to cardiovascular disease, this information is useful for prediction, for
identifying risk subgroups, and I will incorporate all those different pieces of
information regardless of their underlying mechanisms,” said Dr. Goldstein.
“So, something like medication use, from a causal perspective isn’t going
to tell you who has a risk of developing cardiovascular disease, but from an
identification perspective, it may be of greater use.”
During the “training your model” phase of algorithm development, the
different variables are fed into the model and assessed for bias and variance
and fit into the algorithm. Once the algorithm is chosen, the model can be
applied to different databases.
Dr. Goldstein’s group at Duke is already using a machine learning paradigm to identify individuals at risk of 30-day readmission and flag them for
further attention.
“We’ve collected a fairly large database, around 130 pieces of information for each patient at time of discharge, and now we’re working to come
up with a model that doesn’t use all 130 variables but uses some of them to
identify those who are at risk.” With an electronic health record, this model
can readily be applied and tailored to specific patient populations, he said.
REFERENCE
1. Goldstein BA, et al. Eur Heart J. 2016 July 19. [Epub before print]
RISKY BUSINESS
Follow the money and we see that in the first half
of 2016 alone, according to recent research from
Rock Health, (a venture fund dedicated to digital
health), that digital health and data management
vendors amassed $2 billion in funding.10 Almost
25% of this money went to
companies dedicated to data
The SYNTAX Score is
aggregation and/or anal ysis
broadly used today
and to population health
for making decisions
regarding the optimal
management, or “comprehenrevascularization
sive delivery system tools to
strategy for a given
manage the health of populapatient. Now there
is a SYNTAX 2 (and a
tions under the shift to ACO
SYNTAX 3 is around
models,” said Rock Health.
the corner). CSWN
Indeed, according to Jentalked with Patrick
Serruys, MD, PhD, a
nifer Bresnick, lead editor for
world renowned interHealthITAnalytics, it’s been preventional cardiologist
and currently profesdicted that the population health
sor of Cardiology in
management market will grow
the Cardiovascular
from $14 billion to $31.9 billion
Science Division of
the National Heart
by 2020, “driven largely by risk
and Lung Institute
stratification technologies and
(NHLI) of the Imperial
predictive analytics services.”11
College in London
(UK). Use the QR code
In other words, using risk preto access this video,
diction to help prevent adverse
taped at ESC.16.
events is big money.
SHARED DECISIONMAKING AIDS
For many clinicians, risk
scores are primarily a means
30 CardioSource WorldNews
of involving patients in their own care and motivating adherence and lifestyle change.
“If you just say, ‘your risk is high,’ that doesn’t
mean much to a patient, but if you say your risk for
having a heart attack or stroke in the next 10 years
is 15% and the national guidelines say anything
above 7.5% is high enough that we should consider
putting you on a statin because is lowers risk by at
least one third, then they say, ‘yeah, OK, I get it. I
want that benefit.”
Perhaps more helpful is showing a patient how
they can reduce their risk with behavioral change-initiating exercise, losing weight, tobacco cessation.
“The rate-limiting step in cardiovascular disease
prevention is the implementation and maintenance
of healthy lifestyle behaviors,” wrote Nicole D.
White, PharmD, and colleagues, from the Creighton University School of Pharmacy and Health
Professions (Omaha, NE).12 White and colleagues
demonstrated that well-designed risk assessment
education—showing a patient, for example, that she
can lower her risk score by three points by quitting
smoking—can motivate behavioral change.
“I think the more patients understand what we’re
thinking and the risk we’re trying to estimate, the
more on board they are with out strategies to lower
risk,” said Dr. Eagle.
Finding The Sweet Spot
New scores, old scores with new components, EHRbased automatic scoring, convenience, inconvenience. In the end, even the best risk score will not
replace experience and intuition. Risk scores are
only estimates of risk based on population statistics. They are not fate, nor can they stand alone or
replace physician experience and intuition.
As with any clinical tool or new technology, there
is big hype and then pull back, and then the trick is
to find “the sweet spot of its use,” said Dr. Goldstein.
And certainly there is little risk in saying that
technology—mobile devices, apps, EHRs—is helping physicians find a comfort zone where they can
utilize risk prediction tools to their best avail. ■
REFERENCES
Pell JP. Heart. 2012;98:1272-7.
Goff DC Jr, et al. J Am Coll Cardiol. 2014;63:2935-59.
DeFilippis AP, et al. Ann Intern Med. 2015;162:266-75.
Ridker PM, Cook NR. Lancet. 2013;382:1762-5.
Stone NJ, Lloyd-Jones DM. Mayo Clin Proc. 2016;91:692-4.
Levine GN, et al. J Am Coll Cardiol. 2016; doi: 10.1016/j.
jacc.2016.03.513.
7. Goldstein BA, et al. Eur Heart J. 2016 July 19. [Epub
before print]
8. 2016 Outpatient Practice Management (PM) and
Electronic Health Record (EHR) Essentials Brief. Available at www.himssanalytics.org/news/2016-studyadoption-of-outpatient-solutions?utm_source=pr&utm_
medium=news&utm_campaign=2016outpatient.
Accessed August 8, 2016.
9. Goldstein BA, et al. J Am Med Inform Assoc. 2016 May
17. [Epub ahead of print]
10. Digital Health Funding 2016 Midyear Review. Available
at https://rockhealth.com/reports/digital-health-funding2016-midyear-review/. Accessed August 9, 2016.
11. Bresnick J. “How the search for smart data drives healthcare IT investment.” Available at healthitanalytics.com/
news/how-the-search-for-smart-data-drives-healthcare-itinvestment. Accessed August 9, 2016.
12. White ND, et al. Psychol Res Behav Manag. 2013;6:55-63.
1.
2.
3.
4.
5.
6.
September 2016