1. Applying AI to Healthcare’s Biggest Opportunity: Clinical Variation
Dr. Todd Stewart, Mercy
Lonny Northrup, Intermountain
November 18, 2018

2. The challenges associated with healthcare data are tailor made for AI.

3. The Characteristics of Healthcare Data
Big + getting bigger
Noisy + sparse
Complex

4. Intermountain’s Definition of Big Data
Using additional data sources and new analytic tools to produce superior, actionable analytic insights (not previously possible or cost effective) leading to:
Improved Healthcare Outcomes
Reduced Cost
Healthier People
Value = Results / Costs
NOTE: Volume, Variety and Velocity (and sometimes Veracity) are frequently used to describe big data. For Intermountain, our primary measure is Value.
Confidential

5. What is Machine Learning in Healthcare
Machine learning, cognitive computing, artificial intelligence and deep learning are related terms. Collectively, they refer to the ability of computers to learn from data how to replicate and improve human predictions and decisions. In healthcare this means consuming a variety of data (clinical, cost, claims, patient characteristics, etc.) to produce actionable insights leading to lower cost and more effective healthcare outcomes.

6. The Challenge of Clinical Variation
“ An inefficient, broken system based on perverse incentives that cripples innovation and wastes $800 billion annually on health care that doesn’t make us healthier.”
– Bernard J. Tyson
PAYMENT METHODS
Type of Waste
% of all waste
Cost-plus
Fee for service
Per case
Population- based payment
Production level Inefficient production of individual care units, such as drugs, tests, nursing support 5%
Payer
Provider
Case level
Use of unnecessary or suboptimal services in treating a case
50%
Population level
Unnecessary or avoidable patient cases
45%
Clinical variation is a $400B problem for healthcare with no principled way to determine best practices across a hospital
Ayasdi’s award winning application surfaces best practices from across the entire healthcare data chain
Learns complex patterns within patient data to create nuanced patient cohorts, leading to better, more accurate prediction
Collaboratively create standard pathways tailored to organization’s patient population, physicians, and specific care parameters to reduce costs and improve outcomes.
Source: The Case for Capitation, Harvard Business Review

7. The Value of CVM is Broad
Unbiased approach surfaces all variation, good + bad across the care continuum.
Care paths are only the beginning – measurement/ adherence matters.
CVM needs to reflect the mission of the organization and how it practices medicine
Always remember – it is about enhancing patient outcomes

8. What Makes CVM So Difficult?
Multi-Event time series for every procedure
Sparsity of individual events
Physician buy-in hampered by “my patients are different” problem
Tracking adherence to care process models immensely complex
Confidential

9. A Framework for Intelligence
Discover
Predict
Justify
Act
Learn
Confidential

10. Discover
Analyze all EMR + financial data, representing thousands of patient procedures + millions of individual events to automatically surfaces groups of similar patient procedures.

11. Predict
Powerful prediction draws on discovery and in the case of clinical variation management allows healthcare organizations to accurately predict the quality and cost for the desired treatment outcomes. 

12. Justify
For AI to deliver on its promise in healthcare it must be able to justify and document its recommendations. 

13. Act
Integrate with EMR systems to facilitate the rapid deployment of intelligence across the organization.

14. Learn
Intelligent applications are always looking at new data as it comes in, finding emerging patterns that reflect the current practice within the organization.  

15. Mercy’s Application of Intelligence: CVM
A Truven top 5 hospital, Mercy operates 43 acute care + specialty hospitals, employing over 700 physicians.
Mercy has developed over 40 clinical carepaths since our first work together on total knee replacement.
Mercy saves $1M per year on TKR and projects savings of $45M over 3 years ending in FY2018 on other carepaths.

16. Driving Variance Out of Care Process Models w/ AI
Nearly 60 CPM’s Today
Behavioral health
Cardiovascular
Collaborative Pharmacy
Imaging Services
Intensive Medicine
Musculoskeletal
Oncology
Pain Services
Pediatric
Primary Care
Surgical Services
Women and newborns
Etc.
Reducing variance from knee and hip replacement surgeries alone resulted in $60M savings over 3 years AND delivered improved outcomes
* Link to Intermountain Care Process Models:
Confidential

17. Colon Surgery – Developing Care Models with AI
Alvimopan
Ketorolac
Ambulation
Foley Catheter
Oral fluids well tolerated
Intermountain Colon Surgery (~4500)
Fac: 128
Lowest LOS
Found
Ayasdi 2015 Colon Surgery (~530)
3 groups
>4 day LOS (~200)
Found in Best group
More frequent ambulation
Removed earlier in best group
Found – trend to earlier fluids for best group

18. AI Requires Analytical Structure and Organizational Buy In
Confidential

19. Intermountain’s Data + Analytics Structure
Data Warehousing and Integration
Data Governance
Data Semantics and
Clinical Modeling
Analytics and Data Science
Analytic Technology Services
Data & Analytics Advancement
Data Stewardship / Ownership – train and enable business to engage in data management processes
Data Quality Management
Tools for defining business
terms + metrics (metadata)
Data Standardization,
Shared Data Services
Reference Data (ICD codes,
Zip codes, etc.) Management
Master Data (provider, patient,
locations) Management
Clinical Data Modeling
Organizing Data for Analysis
Integration of Disparate
Data (EDW)
Data Movement
Enterprise Data Analyst
Coordination
Data Analyst Best Practices
Data Science = predictive
analytics, integration into
decision processes
Technologies & Tools to
support data analysts
Enterprise-level dashboard
and analytic solution dev.
Coordinate innovation efforts for data and analytics

20. Intermountain’s Infrastructure

21. Augmented Medical Intelligence
Machine Learning (cognitive computing, artificial intelligence, deep learning, etc.) assists humans to make better decisions and take better actions, but cannot completely replace people in the processes of achieving the very best outcomes.

22. Machine Learning in Healthcare – Additional Results
Machine Learning driven Patient Engagement
Type 2 Diabetes Prevention: 58% to 85% risk reduction for over 70% of participants (Omada)
Diabetes: Average 3.2 drop in HbA1c in 3 months (typical drop is 0.5 to 1.0 in 1 year) (Twine Health)
Congestive Heart Failure: 4% readmission rate compared to national average of 26-28% (Sensely)
Remote Patient Monitoring: 89% reduction in inpatient visits, 70% reduction in emergency department visits (Vivify)
Chronic Obstructive Pulmonary Disease (COPD): 87% adherence to care plan, 92% medication adherence and more than 70% reduction in hospitalizations (Senscio)
Real Time Machine Learning driven Emergency Department Optimization
20% reduction in door to doc time (Qventus)
30% reduction in Leave Without Being Seen (LWBS) rate (Qventus)
13% reduction in length of stay - (Qventus)

23. Machine Learning in Healthcare – Additional Results
Machine Learning driven Personalized Treatment Recommendations
$895,000 in savings in under 3 months and significantly reduced readmissions (Health First, Florida - Jvion)
$4M in savings from readmission reductions (University of Tennessee Medical Center - Jvion)
Reduced Catheter Associated Urinary Tract Infections (CAUTI) (Jvion)
Over 60 areas of clinical outcomes improvement (Jvion)

24. $$$ 24/7 N of 1 Where Do We Go Next?
Individual Cost Prediction
Individual “Next Best Action”
Care Plan
Continuous Connection to Care Team
$$$
N of 1
24/7
The promise of personalized medicine
Confidential

25. Questions