Lesson 2 Artificial Intelligence Lesson 2 Artificial Intelligence.
AI: Artificial or Augmented Intelligence - SAS: Analytics, Artificial … · 2017-11-28 · should...
Transcript of AI: Artificial or Augmented Intelligence - SAS: Analytics, Artificial … · 2017-11-28 · should...
© Kaiser Permanente 2016, Internal Use Only 1© Kaiser Permanente 2016, Internal Use Only 1
AI: Artificial or Augmented Intelligence
Jason Jones, PhDKaiser Permanente
© Kaiser Permanente 2016, Internal Use Only 2
I have none
No images or other references to products are an endorsement
Disclosures2
© Kaiser Permanente 2016, Internal Use Only 3
Last updated March 2017 page 3
Northwest
552,651
Northern
California
3,992,501
Southern
California
4,264,119
Hawaii
249,687Colorado
663,240
Georgia
284,213
Mid-Atlantic
States
665,402
Membership by region(As of January 2017)
Washington
651,000
Kaiser Permanente Integrated Care and Coverage
>200K Employees
>52K Nurses
>20K Physicians
38 Hospitals
>650 Medical office buildings
© Kaiser Permanente 2016, Internal Use Only 4
AI: “Artificial” or “Augmented” Intelligence
Since at least the 1950’s there has been some debate about whether AI should replace or enhance human judgment. The focus on artificial intelligence has resulted in computing power and algorithms that have commoditized predictive modeling. We will focus on some practical examples in healthcare—aspirations and solvable gaps that remain in achieving the quadruple aim:
• Vision: Computers as a foundation for quality medical care as a right.
• Experience: Algorithms can improve clinician work life.
• Hard parts: Data preparation and workflow.
• Solvable gaps: Tools to respect preference and provide agency.
Overview4
© Kaiser Permanente 2016, Internal Use Only 5
History Why It Matters
“Instruments are at hand which…will give [us] access to and command over the inherited knowledge of the ages.”Editor in Bush (1945)
“The doubling time of medical knowledge in 1950 was 50 years; in 1980, 7 years; and in 2010, 3.5 years. In 2020 it is projected to be 0.2 years.”Densen (2011)
“’Problem solving’ is largely…a matter of appropriate selection…and as we know that power of selection can be amplified.”Ashby (1956)
“Two 1998 reports validate…40% discordance…diagnoses…compares with 35% in 1938, 39% in 1959, 43% in 1974, and 47% in 1983”Lundberg (1998)
“Quality medical care as a right cannot be achieved unless we can establish need, separate the well from the sick and do that without wasting [clinicians’] time.”Garfield (1970)
Some History on “AI” and Computers in Care5
• Bush (1945): https://www.theatlantic.com/magazine/archive/1945/07/as-we-may-think/303881/• Densen (2011): https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3116346/• Ashby (1956): http://pespmc1.vub.ac.be/books/IntroCyb.pdf • Lundberg (1988): http://jamanetwork.com/journals/jama/fullarticle/188042• Garfield (1970): https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3076970/
© Kaiser Permanente 2016, Internal Use Only 6
AI: “Artificial” or “Augmented” Intelligence
Since at least the 1950’s there has been some debate about whether AI should replace or enhance human judgment. The focus on artificial intelligence has resulted in computing power and algorithms that have commoditized predictive modeling. We will focus on some practical examples in healthcare—aspirations and solvable gaps that remain in achieving the quadruple aim:
• Vision: Computers as a foundation for quality medical care as a right.
• Experience: Algorithms can improve clinician work life.
• Hard parts: Data preparation and workflow.
• Solvable gaps: Tools to respect preference and provide agency.
Overview6
© Kaiser Permanente 2016, Internal Use Only 7
• Watson: By Source, Fair use, https://en.wikipedia.org/w/index.php?curid=31142331• Google Car: By Grendelkhan - Own work, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=47467048
Recent Successes of AI7
© Kaiser Permanente 2016, Internal Use Only 8
• Level 0 – No Automation: …
• Level 1 – Driver Assistance: …
• Level 2 – Partial Automation: …
• Level 3 – Conditional Automation: The driving mode-specific performance by an
Automated Driving System of all aspects of the dynamic driving task with the expectation that
the human driver will respond appropriately to a request to intervene
• Level 4 – High Automation: The driving mode-specific performance by an Automated
Driving System of all aspects of the dynamic driving task, even if a human driver does
not respond appropriately to a request to intervene
• Level 5 – Full Automation: The full-time performance by an Automated Driving System
of all aspects of the dynamic driving task under all roadway and environmental conditions that
can be managed by a human driver
Levels of AutomationSAE (Society of Automotive Engineers)
8
• https://www.sae.org/news/3544/
© Kaiser Permanente 2016, Internal Use Only 9
Patient Experience
Population Health
Reduced Cost
Provider Work Life
Quadruple Aim in Health/CareBodenheimer (2014)
• http://www.annfammed.org/content/12/6/573.full
© Kaiser Permanente 2016, Internal Use Only 10
As you know I'm a big fan of clinical decision support. It's not just an abstract notion in my opinion but instead translates into real improvement in the daily lives of our physicians and more importantly in the patients we serve.
A simple example recently occurred with CURB 65. I was working a night shift with a Hospitalist consultant with which I don't always have the best working relationship (not necessarily his fault, we just don't seem to be on the same page sometimes). I had a patient with pneumonia that at first glance didn't appear very ill, and given the clinical appearance and the relationship with the consultant my mind started mentally preparing for discharge. Then I applied the tool, and it reminded me that this patient had a CURB score of 2.
The score changed my direction and I decided to call a consult. The impressive thing was that the interaction with the Hospitalist, with whom I was expecting push back, went
very smoothly once I mentioned the CURB score. It occurred to me that the objectivity of this tool not only made me think twice about my disposition but it also gave a common language for the consultant and I to speak to each other, one that overcame the historical communication gap we shared.
I have found that this type of situation repeats itself. Clinical decision support, when used correctly, can make us better physicians and colleagues as well. In this case it not only saved me from making a possible error in judgment, but also helped facilitate what may have been a difficult conversation. Just thought you and your staff would want to know. Thanks for all that you do.
Todd Newton, MD Director Medical Services, SCPMG Regional Chief - Emergency Medicine
© Kaiser Permanente 2016, Internal Use Only 11
AI: “Artificial” or “Augmented” Intelligence
Since at least the 1950’s there has been some debate about whether AI should replace or enhance human judgment. The focus on artificial intelligence has resulted in computing power and algorithms that have commoditized predictive modeling. We will focus on some practical examples in healthcare—aspirations and solvable gaps that remain in achieving the quadruple aim:
• Vision: Computers as a foundation for quality medical care as a right.
• Experience: Algorithms can improve clinician work life.
• Hard parts: Data preparation and workflow.
• Solvable gaps: Tools to respect preference and provide agency.
Overview11
© Kaiser Permanente 2016, Internal Use Only 12
• https://hbr.org/2012/10/data-scientist-the-sexiest-job-of-the-21st-century
Developed Tools12
© Kaiser Permanente 2016, Internal Use Only 13
Developed Tools(Though we still spend a lot of time hunting and gathering)
13
© Kaiser Permanente 2016, Internal Use Only 14
Authoring
Point of Care Decisions (Inside Transactions)
Config FileLimited
Deploy
Full
Deploy
Engine
End User Interaction
Deployment Decisions (Outside Transactions)
Parameter
ExplorationFunction/
Formula
Case Reviews Cases/
Context
Operational
CharacteristicsPopulation/
System
Data Stores
Backend Analytics
Prediction
Modeling
Performance
Improvement
Formal
Evaluation
Tools Oriented to Augmented Intelligence
© Kaiser Permanente 2016, Internal Use Only 15
Authoring
Point of Care Decisions (Inside Transactions)
Config FileLimited
Deploy
Full
Deploy
Engine
End User Interaction
Deployment Decisions (Outside Transactions)
Parameter
ExplorationFunction/
Formula
Case Reviews Cases/
Context
Operational
CharacteristicsPopulation/
System
Data Stores
Backend Analytics
Prediction
Modeling
Performance
Improvement
Formal
Evaluation
Tools Oriented to Augmented Intelligence
© Kaiser Permanente 2016, Internal Use Only 16
• The best medicine is useless in the cabinet
• Self-serviceo Leverages critical knowledge and preference
oProvides agency
oMust have the same underlying quality
Examples:
• Reports with embedded rigor
• Preference sensitive predictive parameters
• Decision support deployment
• Forecasts change the discussion
Underdeveloped Tools16
© Kaiser Permanente 2016, Internal Use Only 17
Facilitate Learning—Embed Appropriate Methods17
© Kaiser Permanente 2016, Internal Use Only 18
Building Predictive Models
Viewable from mobile device
Preferences:• Vital signs highest• Early values higher• Labs a little lower• Prior conditions lower still
© Kaiser Permanente 2016, Internal Use Only 19
Clinical Decision Support by and for Clinicians19
© Kaiser Permanente 2016, Internal Use Only 20
Decision Support Tools as Apps
© Kaiser Permanente 2016, Internal Use Only 21
Changing the Discussion21
© Kaiser Permanente 2016, Internal Use Only 22
• http://www3.weforum.org/docs/WEF_GAC15_Deep_Shift_Software_Transform_Society.pdf
Who Is (Willing To Be) Augmented?Maybe we should ask: which task and under what circumstances?
22
© Kaiser Permanente 2016, Internal Use Only 23
Appendix
© Kaiser Permanente 2016, Internal Use Only 24
Building Predictive Models
0.5
0.6
0.7
0.8
0.9
1.0
CURB-65 + SpO2 eCURB CURB-65 + SpO2 eCURB KP Model
Pe
rfo
rman
ce
"EBM" Applied to KP
Improve Data Quality
"EBM" as Published
Improve ProcessRedefine "EBM"Redefine "Risk"
Interdependence
© Kaiser Permanente 2016, Internal Use Only 25
SIRS ERT(Earliest Recognition Time)The first time 2 or more SIRS criteria are met within 2 hours of each other
© Kaiser Permanente 2016, Internal Use Only 26
SIRS ERT-A Consolidated ViewHow do I know when SIRS ERT was met?
If you hover over the SIRS ERT icon with your mouse in a Patient List Column in your Patient List or on a TrackBoard, you will see the time SIRS ERT was met as a pop-up. This is shown below:
How do I know when SIRS findings are "positive" for my patient?
If you double click on the SIRS ERT icon shown in the Patient List or on a TrackBoard, it will take you to the Clinical Risk Scores Report. SIRS related vital signs and lab results are displayed here. Values that meet SIRS criteria for the specified parameters are highlighted in red.
© Kaiser Permanente 2016, Internal Use Only 27
Severe Sepsis Risk Probability Model
• Leaves parameters continuous
• Considers them together
• Allows finer grained control over alerting
Understanding How A Model Works
© Kaiser Permanente 2016, Internal Use Only 28
Severe Sepsis Risk Probability Model
• Leaves parameters continuous
• Considers them together
• Allows finer grained control over alerting
Understanding How A Model Works
© Kaiser Permanente 2016, Internal Use Only 29
Severe Sepsis Risk High at TriageSIRS ERT delayed (CBC slow)
© Kaiser Permanente 2016, Internal Use Only 30
- Span clinical quality and cost
- Span encounter level interaction with population management
30
Operating Characteristics
© Kaiser Permanente 2016, Internal Use Only 31
Supporting Purposeful Deviation
31
31
© Kaiser Permanente 2016, Internal Use Only 32
Visualization for Clinicians & Patients