This document discusses using machine learning and analytics to improve clinical decision making. It outlines some of the shortcomings of current evidence-based practices, including biases and limited data. The document proposes that machines could help overcome these issues by analyzing large amounts of data from real-world practice to uncover patterns that may improve care delivery and outcomes. It presents a model for developing analytics as reusable products that learn over time from new data, to help create a sustainable "learning health system." Examples of potential applications discussed include analyzing clinical variation and developing condition-specific insights.
PMED: APPM Workshop: Evidence-Based Practice v Practice-Based Evidence - Jason Burke, March 15, 2019
1. UNCHEALTHCARESYSTEMUNCHEALTHCARESYSTEM DRAFT – FOR DISCUSSION PURPOSES ONLY – INFORMATION SUBJECT TO CHANGE
Evidence-based Practice vs.
Practice-based Evidence
March 2019
Innovations in Precision Medicine
within Health Care Delivery
Jason Burke
Chief Analytics Officer & System VP
@jaburke
jasonburke
3. 3
UNCHEALTHCARESYSTEM DRAFT – FOR DISCUSSION PURPOSES ONLY – INFORMATION SUBJECT TO CHANGE
How do you make clinical decisions?
Source Short-comings (not exhaustive)
Published Research
Other patients in practice
Gut instinct
Sales reps
• Sample bias
• Publication bias
• Information overload / innovation pace
• Increasing specialization / complexity exceeds cognition
• Data sparsity / completeness
• Psychological / cognitive biases
• Commercial bias
• Limited scopes of inquiry
• Bench to bedside timeframes
What if machines could help overcome these shortcomings?
• Crunch the sources and factors
• Model more representative populations
• Uncover patterns too complex for the human brain to readily identify
• Serve as an aide
4. 4
UNCHEALTHCARESYSTEM DRAFT – FOR DISCUSSION PURPOSES ONLY – INFORMATION SUBJECT TO CHANGE
Image via Flickr user horiavarlan
PRACTICE-BASED
EVIDENCE
Growing: Expanding the Opportunity Funnel
global
institutional
individual
multi-institutional
SCOPE
controlled semi-controlled real world
cohorts
CONSTRAINTS
QUALITY
IMPROVEMENT
CLINICAL
RESEARCH
5. 5
UNCHEALTHCARESYSTEM DRAFT – FOR DISCUSSION PURPOSES ONLY – INFORMATION SUBJECT TO CHANGE
Growing: Increasing Analytical Complexity for Population Health
Utilization Explorer Care Variation Analytics
6. 6
UNCHEALTHCARESYSTEM DRAFT – FOR DISCUSSION PURPOSES ONLY – INFORMATION SUBJECT TO CHANGE
Example: Bundled Payments
The “obvious answers” are sometimes
neither obvious nor answers.
So what is needed to create
sustainable change?
Increasing complexity
%target
7. 7
UNCHEALTHCARESYSTEM DRAFT – FOR DISCUSSION PURPOSES ONLY – INFORMATION SUBJECT TO CHANGE
Rethinking Analytics in Health Care
• Stakeholders are project area experts
• Effort is focused on predefined questions
• Work is relevant to project team
• Timeline is project driven
• Data definitions are project specific
• Data structured for single use
• Little-to-no analytical code reuse
• Release available to project stakeholders
• Stakeholders are functional experts
• Questions are not predefined
• Work must be relevant to multiple customers
• Timeline is engineering driven
• Data definitions are enterprise-wide
• Data is structured for broad re-use
• Analytical models are built for multiple projects
• Release available to entire enterprise
Most health industry analyses are
conducted as a PROJECT
Our work is managed
more like a PRODUCT
8. 8
UNCHEALTHCARESYSTEM DRAFT – FOR DISCUSSION PURPOSES ONLY – INFORMATION SUBJECT TO CHANGE
COGNITIVEPHASEASSOCIATIVE
PHASE
AUTONOMOUS
PHASE
Learning Health System “Learning Model”
COLLECT
Gather
Aggregate
INFLUENCE
Iteration
Implementation
Automation
Validation
OPTIMIZE
Refactoring
Tuning
Maintenance
PLAY
Explorations
Novel data
What-ifs
Discovery
New Research
SYNTHESIZE
Factors
Models
Relationships
Patterns
DerivedfromFitts&Posner’s3StagesofCognitiveLearing(1967)
OBSERVE
Describe
Profile
Report
EXTEND
New data
Modeling
Integration
Simplification
Jason Burke
9. 9
UNCHEALTHCARESYSTEM DRAFT – FOR DISCUSSION PURPOSES ONLY – INFORMATION SUBJECT TO CHANGE
Clinical Variation: Supporting Value-Based Care
Opportunity
Identification
Cohort
Insights
Outcomes
Comparison
Cost
Comparison
Condition Specific
Insights
10. 10
UNCHEALTHCARESYSTEM DRAFT – FOR DISCUSSION PURPOSES ONLY – INFORMATION SUBJECT TO CHANGE
Discussion
Jason Burke
@jaburke
jasonburke
Editor's Notes
Health care’s ongoing evolution (e.g., accountable care, risk delegation, population health management, consumer technologies, Internet of Things) is creating new imperatives for the way both physicians and patients approach heath care delivery. Clinicians are faced with competing demands: standardize care across patients, but ensure optimal individual health outcomes. Care pathways targeting broad-based patient populations derived from clinical research (i.e., evidence-based practice) must be reconciled with real-world observations (i.e., practice-based evidence). The intersection of these two forces create innovative opportunities for transforming patient care through richer understandings of patients, clinicians, health outcomes, quality measures, processes, and costs. Capturing these opportunities requires new competencies in harnessing data sciences and analytics to make more tailored care decisions, rationalize sparse delivery resources, manage risks, and drive efficiencies that reduce the overall cost of care.