These are slides from a talk that I gave at Strata Rx in San Francisco on 10/17/2012.
Abstract:
It is clear that data are core to solving big problems in health care, and data science is the skill set needed to extract insights and make them actionable. Using lessons from experience from consumer internet (LinkedIn & online advertising) and a large dataset of clinical and claims data from across the US, we will discuss results from efforts to increase the quality of care, decreasing cost, and increasing hospital efficiency. Real-world use cases will be presented detailing the use, implementation and impact of deploying predictive analytics.
Examples of use cases to be discussed: - predictive modeling around identifying patients at high risk for overutilization (e.g., many return visits to the ED), allowing for proactive and less expensive care to be provided - using recommendation systems to identify procedures and charges missed during billing, resulting in recovered revenue for the hospital - identifying payer claims likely to be denied and why, to enable more efficient coding of charges - providing rich contextual data for physicians to allow them to maintain or increase the quality of care while decreasing cost
9. EHR integration barriers
Legal/compliance/privacy
Innovations very hard to Barriers to quick
scale
deployment, itera
Need technology + onsite tion,
ops
and impact
Open source does not play
well with others
10. Think like a startup: bias
towards customer feedback,
solving for a need, & iteration
Different hats: product How to
manager, biz dev, sales, data,
engg
overcome?
Think like a data
Work closely with customers scientist
(docs, patients, hosp. execs...)
Leverage expertise to build
better models (and be
compliant)
16. My definition of a
data scientist:
Someone who uses data
to solve problems end-to-
end, from asking the right
questions to making
insights actionable.
17. End-to-end data science: five stages
Ask the Leverage Extract &
Build a
right other clean Deploy
model
questions solutions your data
18. One of the hardest
Phase 1 things to find in a
data scientist
Ask the
right Health Care: Even for
the good ones, have
questions to work closely with
clinician partners
40. Just as physicists moved to
Wall Street to be quants and
then on to online advertising
and consumer web, there will
be a significant talent
migration into health care in
the next few years.
I’m in my 5th month in the health care industry and am really excited about this talk. I’ve given a lot of other talks at data conferences, but this is my first healthcare talk. Before we get started, a little about me: have a PhD in economics from Stanford, studied behavioral economics and did some applied econometric work around understanding biases and fatigue in decision makingWent to online advertising startup, crash course in how to work with dataThen LinkedIn, lead a team of data scientists, worked on user engagement, content & community teams, also worked on economic insights like predicting the unemployment rateNow health care, where I am Chief Data Scientist at Accretive Health. Our main focus is to help hospitals be smarter, from helping them manage their revenue and billing operations on one end to implementing population health management solutions to increase quality and decrease cost on the other.Why do I mention these things? Because it tells you about what is important to me…
Might need some clinicians on hand, but that’s one of the lessons…need to work with clinicians, can’t throw tech at the problemMy professional direction is to help….People, cognitive overload, limited information, risk aversion, To show you how legit I am, I’m going to show you the tattoo I got from this brand…j/k, was thinking about bringing it here with a butane torch to heat it up so that we could brand the true data believers, but thought that TSA wouldn’t be so happy with that. So, instead I tried it on steak, didn’t work so well.Data insights are really about overcoming complexity by distilling down information into some simple, actionable form.Two slightly incongruous part of this talk
Lesson 1: common sense, subject matter expertise and intuition are crucial for the successful deployment of predictive modelingExample: linkedin scoring models for subscriber churn. Spend a lot of time cleaning, preparing, extracting data. Running models, feature selection, robustness checks, determine best modelProblems: basic insights could have been tested more quickly: credit card expirationProblems: implementation of insights from modelParallel: population health managementCan go out and build a great predictive model for understanding who is most at-risk for ED over-utilization in the next 60 daysWho is going to pay attention to it? Where are the insights going to be surfaced? Will physicians get this infoObservation #1: most of the time the EHR is the devil.Small example how tweaking the decision context can create huge results: petrified forest, social pressure,
Been in the industry 5 months. First got here and thought:Bad Uis, poor data visuazliation. Not a lotDearth of predictive models to help doctors in real timeSome good graphs could make a huge impact“Oh wow, the EHR looks like THAT?” we can do that, too.Kid in a candy store
Lots of good work is being done but not on a scale outside of startups that have a hard time to scale or the elite research institutions. What about critical care hospitals?I’m going to get rich AND help my grandma get better treatment during that next knee replacement…can someone have more than 2 knees?
Wait a minute…no easy BI for docs to give them feedback on their decisions? High priority patients?OK I got this, good intuition for data, some hacky engineering skills, can build predictive models…I got this!
Good luck starting a company in health care analytics.Plug for accretive here
Good luck starting a company in health care analytics.Plug for accretive here
Let’s provide a bit more detail about what “end-to-end” means and compare/contrast my experience at LinkedIn to healthcare
Lesson 1: common sense, subject matter expertise and intuition are crucial for the successful deployment of predictive modelingExample: linkedin scoring models for subscriber churn. Spend a lot of time cleaning, preparing, extracting data. Running models, feature selection, robustness checks, determine best modelProblems: basic insights could have been tested more quickly: credit card expirationProblems: implementation of insights from modelParallel: population health managementCan go out and build a great predictive model for understanding who is most at-risk for ED over-utilization in the next 60 daysWho is going to pay attention to it? Where are the insights going to be surfaced? Will physicians get this infoObservation #1: most of the time the EHR is the devil.Small example how tweaking the decision context can create huge results: petrified forest, social pressure,
Lesson 1: common sense, subject matter expertise and intuition are crucial for the successful deployment of predictive modelingExample: linkedin scoring models for subscriber churn. Spend a lot of time cleaning, preparing, extracting data. Running models, feature selection, robustness checks, determine best modelProblems: basic insights could have been tested more quickly: credit card expirationProblems: implementation of insights from modelParallel: population health managementCan go out and build a great predictive model for understanding who is most at-risk for ED over-utilization in the next 60 daysWho is going to pay attention to it? Where are the insights going to be surfaced? Will physicians get this infoObservation #1: most of the time the EHR is the devil.Small example how tweaking the decision context can create huge results: petrified forest, social pressure,
Parallel for health care: do we need a bunch of fancy machine learning models? Or will a nice data visualization do the job to get people to change their behavior? Context, social relativity,
Parallel for health care: do we need a bunch of fancy machine learning models? Or will a nice data visualization do the job to get people to change their behavior? Context, social relativity, ALSO: can you easily deploy some A/B tests to learn some things?
Lesson 1: common sense, subject matter expertise and intuition are crucial for the successful deployment of predictive modelingExample: linkedin scoring models for subscriber churn. Spend a lot of time cleaning, preparing, extracting data. Running models, feature selection, robustness checks, determine best modelProblems: basic insights could have been tested more quickly: credit card expirationProblems: implementation of insights from modelParallel: population health managementCan go out and build a great predictive model for understanding who is most at-risk for ED over-utilization in the next 60 daysWho is going to pay attention to it? Where are the insights going to be surfaced? Will physicians get this infoObservation #1: most of the time the EHR is the devil.Small example how tweaking the decision context can create huge results: petrified forest, social pressure,
When people ask me what I do, they first are fascinated by the title. Then I tell them I’d love to build a robot that helps doctors make better decisions. And they’re thinking, wow, this guy’s a genius!I’m thinking are you kidding me? This is really what I’m doing most of the timePeople don’t want to know about the sausage makingLESSON: need to get other business partners and stakeholders on board. They need to understand that this stuff takes TIME
80% of the workBut you’re having fun doing it…except for this guy. Everyone is sometimes that guy.
Parallel for health care: do we need a bunch of fancy machine learning models? Or will a nice data visualization do the job to get people to change their behavior? Context, social relativity,
Data in different places, storage types change over time, not communicated, etcWill clean up this imageOpen source…health care has not caught on yetClaudia perlich…doesn’t work with data that she hasn’t pulled herself
Lesson 1: common sense, subject matter expertise and intuition are crucial for the successful deployment of predictive modelingExample: linkedin scoring models for subscriber churn. Spend a lot of time cleaning, preparing, extracting data. Running models, feature selection, robustness checks, determine best modelProblems: basic insights could have been tested more quickly: credit card expirationProblems: implementation of insights from modelParallel: population health managementCan go out and build a great predictive model for understanding who is most at-risk for ED over-utilization in the next 60 daysWho is going to pay attention to it? Where are the insights going to be surfaced? Will physicians get this infoObservation #1: most of the time the EHR is the devil.Small example how tweaking the decision context can create huge results: petrified forest, social pressure,
Online advertising: logistic regression in production at Yahoo for a long timeAgile data. Focus on quick solutions to identify bogeys and get feedbackAndrew Ng: get first model shipped in 24 hours regardless of what it looks like
Lesson 1: common sense, subject matter expertise and intuition are crucial for the successful deployment of predictive modelingExample: linkedin scoring models for subscriber churn. Spend a lot of time cleaning, preparing, extracting data. Running models, feature selection, robustness checks, determine best modelProblems: basic insights could have been tested more quickly: credit card expirationProblems: implementation of insights from modelParallel: population health managementCan go out and build a great predictive model for understanding who is most at-risk for ED over-utilization in the next 60 daysWho is going to pay attention to it? Where are the insights going to be surfaced? Will physicians get this infoObservation #1: most of the time the EHR is the devil.Small example how tweaking the decision context can create huge results: petrified forest, social pressure,
With the missing charges problem, you need qualified nurse auditors to sign off on suggestions. Kinda like crowdsourcingWhat about in consumer internet? We are crowdsourcing the evaluation of algorithms all the time when people either accept, reject or ignore a recommendation. That exhaust is then used to feed back into the algorithm to improve recommendationsOnly the experts are really anyone and they provide the semi-supervised part of the algorithm
With the missing charges problem, you need qualified nurse auditors to sign off on suggestions. Kinda like crowdsourcingWhat about in consumer internet? We are crowdsourcing the evaluation of algorithms all the time when people either accept, reject or ignore a recommendation. That exhaust is then used to feed back into the algorithm to improve recommendationsOnly the experts are really anyone and they provide the semi-supervised part of the algorithm
TRANSITIONSummarize, where we areWell that’s great but who is going to do all of that work?
One of the fundamental problems of our time18% of GDP! 0.01% is giant revenue potentialData availability and richness only increasingThe right people are realizing data and data science are core to the solution.The best data scientists see the world through the eyes of how data can help solve problems. They are less about a specific algorithm or industry or tool. Thus, their background are all over the placeHIGHLIGHT SOME KEY PROBLEMS IN HEALTH CARECan’t deploy solutions to scale- You have very talented researchers like Pete Szolovits and the CSAIL lab doing great work on decision support. Robots that work with the physician. Even if you can do this pilot in one hospital (MIT), it’s hard to implement this at some meaningful scale. Years vs weeks in consumer internetImmense privacy issuesMost hospitals are non-profit and are focusing on providing care to the community, not on impleeData products are built off of the “data exhaust”, which is easily accessible. In health care, it’s difficult to get the dataData aggregation across hospitals difficult because of competitive concernsIn consumer internet, users get something in return for their data (free product), or they can pay to restrict the usage of their data. What do patients get out of data mining in health care? Not obvious.
Plug for accretive
Bioinformatics and data science roles. Work on improving the quality of care for patients and making hospitals and physicians smarter. Recently engaged in a really exciting partnership with a nationwide group of private-practice oncologists to help them provide higher quality care while containing cost for patients and payers.