Bayesian Hierarchical Models for the Design and Analysis of Studies - - PowerPoint PPT Presentation

Bayesian Hierarchical Models for the Design and Analysis of Studies to Individualize Healthcare

Scott L. Zeger

John C. Malone Professor of Biostatistics and Medicine @ScottZeger September 19, 2019

2

Scott Zeger

Disclosures

Rela latio ionship ip Company ny(ies es) Speakers Bureau Advisory Committee Embold Health Board Membership Consultancy Review Panel PCORI Funding PCORI ME-1408-20318 Honorarium Ownership Interests

3

Outline

• A learning healthcare system aspires to:

Improve each clinical decision for this patient by learning from the experiences

• f prior similar patients: population individual
• Bayes rule is a logic for learning
• Prostate cancer application
• Lessons learned from implementation of learning systems within a

major academic health center

4

5

Healthcare Decisions: Be A CLINICIAN for this moment

• Presentation: 40-year-old man, no family history, tests positive for a life-threatening

disease in a routine screen

• Clinical Questions: What is his disease state? What action do you recommend?
• Decision Support: Data from prior population of similar people

True disease status Exam result Yes No Total Positive 15 985 1,000 Negative 5 8,995 9000 Total 20 9,980 10,000

6

Bayes Rule

7

Why Bayes?

• Focus on each patient
• Use probability as a natural measure of uncertainty
• Integrate population-based evidence with expert judgement
• Reflects how clinicians reason
• Earlier rule-based expert systems largely failed

8

Learning from prior patients’ experience

• Using Bayes rule, create the computational analogue of the 2x2

table for any complex measurements Population  Individual

• Build capacity to make tables for ever-narrower sets of

“otherwise-similar” individuals Subset, Subset, Subset

9

10

Prostate Cancer Application

(Bal Carter, Yates Coley, Ken Pienta, Mufaddal Mamawala, Scott Zeger, TIC, APL, IT@JH, JHTV)

Clinical questions about active surveillance:

• 1. Given the data collected to date on this

individual, should we do another biopsy today?

• 2. If we remove his prostate today, what is the

probability the tumor is aggressive vs indolent?

11

14

Learning Health System Steps Prostate Cancer Active Surveillance Example Challenges Frame unmet health/clinical need Half of active surveillance prostatectomies yield indolent cancers Specify biomedical model Predictors of indolence: PSA, biopsies, family history, genomic score, MRI Poor understanding of mechanisms Wrangle relevant data into a clinical cohort database (CCDB) from which to learn through careful analysis Brady Institute Active Surveillance clinical cohort database with 1300 men; Precision Medicine Analytics Platform (PMAP) Learning-grade data not collected; Data collected but “locked-up in EHR; HIPAA “minimum necessary standard” Design and test decision tool Coley, et al (a, b): Bayesian hierarchical model Inadequate predictive power; External validity checks not made Design and test users’ interface for population health manager, clinician, and/or patient PCORI ME-1408-20318 / TIC EHR has limited capacity for visualization, calculation, but ”owns the workflow”; $300K for two pages in EPIC Design and test on-going curation JHM Committee No standards; must create policies and procedures Devise business model to sustain/improve tool ?? New methods improve outcomes at lower costs; providers lose money Scale up and out for broad use CoE in a Box; Partners Takes capital investments and time

15

Learning Healthcare System

Learning Healthcare System of Systems – JHM Precision Medicine Centers of Excellence (PMCOEs)

17

PMAP – Precision Medicine Analytics Platform

18

Learn More

• Scott L. Zeger, PhD