There’s A 90% Chance Your Son Is Pregnant

Понравилась презентация – покажи это...

Слайд 0

Predicting The Future Of Predictive Analytics In Healthcare There’s A 90% Chance Your Son Is Pregnant

Слайд 1

Presenter and Contact Information 2 Dale Sanders Senior Vice President, Strategy, Health Catalyst 801-708-6800 dale.sanders@healthcatalyst.com @drsanders www.linkedin.com/in/dalersanders/

Слайд 2

Acknowledgements David Crockett, PhD, Health Catalyst Eric Siegel, PhD, Columbia University Ron Gault, Aerospace Corporation, Northrup-Grumman, TRW Wikipedia 3

Слайд 3

The Goal Today I hope you leave this webinar with… Informed Expectations and Opinions: Be generally aware of the realistic possibilities for predictive analytics in healthcare, over the next few years The Right Questions: To be conversant in the concepts of predictive analytics and be able to ask reasonably well-informed questions of your analytics teams, especially vendors, during the strategic process of developing your organization’s predictive analytics strategy 4

Слайд 4

Agenda 5 Basic Concepts, Fundamental Assertions Predictive Analytics Outside Healthcare Predictive Analytics Inside Healthcare Key Questions To Ask Vendors And Your Analytics Teams

Слайд 5

Sampling of My Background In Predictive Analytics 6

Слайд 6

Gartner 2014 Hype Cycle for Emerging Technology 7 Predictive Analytics in Healthcare, according to Dale Sanders

Слайд 7

“Beyond math, there are no facts; only interpretations.” - Friedrich Nietzsche 8

Слайд 8

Challenge of Predicting Anything Human 9

Слайд 9

What Should We Expect In Healthcare? Machines are predictable; humans aren’t 10 “People are influenced by their environment in innumerable ways. Trying to understand what people will do next, assumes that all the influential variables can be known and measured accurately. People's environments change even more quickly than they themselves do. Everything from the weather to their relationship with their mother can change the way people think and act. All of those variables are unpredictable. How they will impact a person is even less predictable. If put in the exact same situation tomorrow, they may make a completely different decision. This means that a statistical prediction is only valid in sterile laboratory conditions, which suddenly isn't as useful as it seemed before.” Gary King, Harvard University and the Director of the Institute for Quantitative Social Science

Слайд 10

Healthcare Analytics Adoption Model Level 8 Level 7 Level 6 Level 5 Level 4 Level 3 Level 2 Level 1 Level 0 Personalized Medicine & Prescriptive Analytics Clinical Risk Intervention & Predictive Analytics Population Health Management & Suggestive Analytics Waste & Care Variability Reduction Automated External Reporting Automated Internal Reporting Standardized Vocabulary & Patient Registries Enterprise Data Warehouse Fragmented Point Solutions Tailoring patient care based on population outcomes and genomic data. Fee-for-quality rewards health maintenance. Organizational processes for intervention are supported with predictive risk models. Fee-for-quality includes fixed per capita payment. Tailoring patient care based on population metrics. Fee-for-quality includes bundled per case payment. Reducing variability in care processes. Focusing on internal optimization and waste reduction. Efficient, consistent production of reports & adaptability to changing requirements. Efficient, consistent production of reports & widespread availability in the organization. Relating and organizing the core data content. Collecting and integrating the core data content. Inefficient, inconsistent versions of the truth. Cumbersome internal and external reporting. © Sanders, Protti, Burton, 2013 11

Слайд 11

Concepts & Principles of Predictive Analytics 12

Слайд 12

Semantics, Ssschmantics Predictive Analytics and Predictive Models: These terms have their origins in statisticians; e.g., understanding real-world phenomena such as healthcare, retail sales, customer relationship management, voting preferences, etc. Machine Learning Algorithms: This term has its origins in computer scientists; e.g., natural language processing, speech recognition, image recognition, adaptive control systems in manufacturing, robots, satellites, automobiles and aircraft, etc. 13 As it turns out, the latter can be applied to the former, so the two schools of thought are now generally interchangeable. Don’t let vendors fool you into thinking that “machine learning” is more sophisticated or better than predictive modeling.

Слайд 13

For Now, Just Know The Terms 14 And Know Where To Go For Details

Слайд 14

For Now, Just Know The Terms And Know Where To Go For Details MachineLearningMastery.com

Слайд 15

The Basic Process of Predictive Analytics 16

Слайд 16

A Big & Common Mistake: Over Fitting 17 You train the model to be very specific on a given data set, but the model cannot adapt to a new, unknown data set

Слайд 17

Specificity vs. Sensitivity: Trading One For Another Specificity: The true negative rate. For example, the percentage of diabetic patients identified who will not have a myocardial infarction Sensitivity: The true positive rate. For example, the percentage of diabetic patients that will have a myocardial infarction

Слайд 18

Receiver Operating Characteristic (ROC) Plot 19 Tuning radar receivers in WWII Maximum radar receiver sensitivity led to many false positives… too many alarms Lower radar receiver sensitivity led to many false negatives… missed threats Same challenge in airport security screening systems and spam filters Concept has been applied heavily in diagnostic medicine True Positive Rate vs. False Positive Rate

Слайд 19

Data Volume vs. Predictive Model “But invariably, simple models and a lot of data trump more elaborate models based on less data.” “The Unreasonable Effectiveness of Data”, March 2009, IEEE Computer Society; Alon Halevy, Peter Norvig, and Fernando Pereira, Google

Слайд 20

The Human Data Ecosystem 21

Слайд 21

We Are Not “Big Data” in Healthcare Yet 22

Слайд 22

Predictive Precision vs. Data Content 23

Слайд 23

Thank you for the graphs, PreSonus Healthcare and patients are continuous flow, analog process and beings But, if we sample that analog process enough, we can approximately recreate it with digital data 24 Remember Your Calculus Digital Sampling Theory?

Слайд 24

We are asking physicians and nurses to act as our “digital samplers”… and that’s not going to work 25

Слайд 25

Predictive Analytics Outside Healthcare Predictive Analytics Outside Healthcare 26

Слайд 26

“Mr. Sanders, while your 9-year tenure as an inmate has been stellar, our analytics models predict that you are 87% likely to become a repeat offender if you are granted parole. Therefore, your parole is denied.” - 2014, 80% of parole boards now use predictive analytics for case management* * The Economist, “Big data can help states decide whom to release from prison” April 19, 2014 27

Слайд 27

Thank you Sonja Star, New York Times “Evidence Based” Sentencing 20 states use predictive analytics risk assessments to inform criminal sentencing. 28 “Evidence Based” Sentencing

Слайд 28

Recidivism Risk Assessment: Level of Service/Case Management Inventory (LS/CMI)* 29 15 different scales feed the PA algorithm Criminal history Education/employment Family/marital Leisure/recreation Companions Alcohol/drug problems Antisocial patterns Pro-criminal attitude orientation Barriers to release Case management plan Progress record Discharge summary Specific risk/needs factors Prison experience - institutional factors Special responsivity consideration 42.2% of high-risk offenders recidivate within 3 years *Nov. 2012, Hennepin County, Minn. Department of Community Corrections and Rehabilitation

Слайд 29

30 “Since the publishing of Lewis' book, there has been an explosion in the use of data analytics to identify patterns of human behavior and experience and bring new insights to fields of nearly every kind.”

Слайд 30

eHarmony Predictions “Heart” ? of the system: Compatibility Match Processor (CMP) 320 profiling questions/attributes per user 29 dimensions of compatibility ~75TB 20M users 3B potential matches daily 60M+ queries per day, 250 attributes 31 Thank you, Thod Nugyen, eHarmony CTO

Слайд 31

Thank you, Ryan Barker, Principal Software Engineering – Matching, eHarmony 29 Dimensions of Compatibility 32

Слайд 32

Predictive Analytics Inside Healthcare 33

Слайд 33

What Are We Trying to Predict? Common applications being marketed today Identifying preventable re-admissions: COPD, MI/CHF, Pneumonia, et al Sepsis Risk of decubitus ulcers LOS predictions in hospital and ICU Cost-per-patient per inpatient stay Cost-per-patient per year by disease and comorbidity Risk of ICU mortality Risk of ICU admission Appropriateness of C-section Emerging: Genomic phenotyping 34

Слайд 34

True Population Predictive Risk Management Thank you, for the diagram, Robert Wood Johnson Foundation, 2014 Very Little ACO Influence Very Little ACO Influence >/=30% Waste* 100% ACO Influence *Congressional Budget Office, IOM, “Best Care at Lower Cost”, 2013 True Population Health Management 35

Слайд 35

Not all patients can functionally participate in a protocol At Northwestern (2007-2009), we found that 30% of patients fell into one or more of these categories: Cognitive inability Economic inability Physical inability Geographic inability Religious beliefs Contraindications to the protocol Voluntarily non-compliant Socioeconomic Data Matters 36

Слайд 36

The key to predictive analytics in the future of health care will be the ability to answer this two-part question: What’s the probability of influencing this patient’s behavior towards our desired outcome and how much effort (cost) will be required for that influence? Return on Engagement (ROE)

Слайд 37

Return on Engagement (ROE)

Слайд 38

Socioeconomic Data Matters 39

Слайд 39

Development Partner 40

Слайд 40

Flight Path “Outcomes” 41 Examples Diabetes Cohort Good Flight Path Poor Flight Path $ COST Per Member Per Year (Charges) For > 1 year of encounters (~5 yrs and 26k patients) These aren’t really outcomes… they are proxies for outcomes

Слайд 41

True Outcomes 42 Good Flight Path Poor Flight Path Absence of: Cardiovascular disease (angina, MI, stroke) Nephropathy/End stage renal Diabetic retinopathy Glaucoma Cataracts Lower extremity tissue narcosis, foot ulcers Peripheral neuropathy Diabetic ketoacidosis Diabetic preeclampsia GI complications (nausea, constipation) Erectile dysfunction Presence of: Cardiovascular disease (angina, MI, stroke) Nephropathy/End stage renal Diabetic retinopathy Glaucoma Cataracts Lower extremity tissue narcosis, foot ulcers Peripheral neuropathy Diabetic ketoacidosis Diabetic preeclampsia GI complications (nausea, constipation) Erectile dysfunction Diabetes Cohort (~5 yrs and 26k patients)

Слайд 42

43 Two Layers of Predictive Function Risk scores Simulation

Слайд 43

Microsoft Azure: Cloud-Based Algorithms 44

Слайд 44

Allina Health Readmissions Model* Variables Considered *- Thank you, Jonathan Haupt

Слайд 45

Allina Compared To Other Models Multiple logistic regression 5.2% of discharged patients in high risk category

Слайд 46

Allina’s Intervention To Reduce Risk Transition of Care “Conferences” Patients, families, care givers 15% reduction in readmissions 100+ APR-DRGs affected More patients utilizing post-acute care Skilled Nursing Facility Home Health TCU

Слайд 47

Predictive and prescriptive (suggestive) analytics in the same user interface The efficacy and costs of antibiotic protocols for inpatients Thank you, Dave Claussen, Scott Evans, et al, Intermountain Healthcare 48 The Antibiotic Assistant

Слайд 48

The Antibiotic Assistant Impact Complications declined 50% Avg. number of doses declined from 19 to 5.3 The replicable and bigger story Antibiotic cost per treated patient: $123 to $52 By simply displaying the cost to physicians 49

Слайд 49

Wrapping Up 50

Слайд 50

Key Questions To Ask Of Vendors and Your Analytics Teams 51 What is your formal training, education, and practical experience in this field? What are the input variables to the model? What model and/or algorithms are you using and why? How are you going to train the model? Are you using our data or other organizations’ data for training? Why? If you are using other organizations’ data, how are you going to customize the model to our specific data environment?

Слайд 51

52 Action matters: What is the return in investment for intervention? Are we prepared to invest more... or say “no”… to patients who score low on predicted engagement? Human unpredictability: The mathematical models of human behavior are relatively immature. Socio-economics: Can today’s healthcare ecosystem expand to make a difference? Missing data: Without patient outcomes, the PA models are open loop. Social controversy: How much do we want to know about the future of our health, especially when the predictive models are uncertain? Wisdom of crowds: Suggestive analytics from “wise crowds” might be easier and more reliable than predictive analytics, until our data content improves Closing Thoughts and Questions

Слайд 52

Q & A 53 Submitted prior to the webinar Submitted through the webinar chat box

Слайд 53

Thank You For questions and follow-up, please contact me dale.sanders@healthcatalyst.com @drsanders Upcoming Educational Opportunities An Overview of the Healthcare Analytics Market Date: January 21, 2015, 1-2pm, EST Host: Jim Adams, Executive Director, The Advisory Board A Pioneer ACO Case Study: Quality Improvement in Healthcare Date: January 28, 2015, 1-2pm, EST Hosts: Robert Sawicki, MD, Senior Vice President of Supportive Care, OSF HealthCare Roopa Foulger, Executive Director Data Delivery, OSF HealthCare Linda Fehr, RN, Division Director of Supportive Care, OSF HealthCare