elmer bernstam - personalized medicine
TRANSCRIPT
-
8/7/2019 Elmer Bernstam - Personalized Medicine
1/12
3/9/20
Personalized Medicine Requires
Personalized Clinical Decision
Support
Elmer V. Bernstam, MD, MSE
Our view of a car
Ferrari Enzo 400 produced
6.0 L V12 putting out651 hp
12 mpg city18 mpg hwy
2
-
8/7/2019 Elmer Bernstam - Personalized Medicine
2/12
3/9/20
Roads view of a car
0-60 mph in 3.14
sec
0-100 mph in 6.6
sec
Top speed 222 mph
3
Our view vs. Roads view
The roads view is what determinesperformance
Similarly
The effects of personalized medicine (genomics)on health depend on what the clinician can do
Not on what the scientist can do
(Selected) Information challenges topersonalized medicine from the clinicalperspective
4
-
8/7/2019 Elmer Bernstam - Personalized Medicine
3/12
-
8/7/2019 Elmer Bernstam - Personalized Medicine
4/12
3/9/20
From clinical perspective
Personalized medicine not new
We always individualized therapy using
Patient preferences
Non-genomic risk profile
Smoking, exposure to asbestos
Available facilities and resources
Co-morbidities: many patients have more thanone problem
Burke W, Psaty BM. Personalized Medicine in the era of genomics. JAMA. 2007;298(14):
1682-4.7
Mulrow C, Cook D, et al. (1997) Systematic reviews: critical links in the great chain of evidence. Annals of Internal
Medicine, 126: 389-91.8
-
8/7/2019 Elmer Bernstam - Personalized Medicine
5/12
3/9/20
Whats new? quantitative difference
The amount of evidence exploding How much you have to know that you dont learn in
the encounter and is not common knowledge
Personalized medicine = dividing the populationinto smaller groups Non-genomic medicine: divide population into few
categories (2-4) with many members based on clinicalcharacteristics (e.g., age groups, gender, exposures,histology) then individualize based on what youlearn in the encounter (e.g., patient preferences)
Genomic medicine: divide population into manycategories with very few (1) members
9
Many categories, few members
Nature (2000)
10
-
8/7/2019 Elmer Bernstam - Personalized Medicine
6/12
3/9/20
Subdividing diseases
Diffuse Large B Cell Lymphomamultiple
diseases
Survival correlates with gene expression
Genomic medicine Biomedical
Scientist
This is cool!
We study diseases now we understand them
better
Getting at the fundamentals, rather than epi-
phenomena
Lets do more of this if we only had morefunding
12
-
8/7/2019 Elmer Bernstam - Personalized Medicine
7/12
3/9/20
Genomic medicine -- Clinician
Personalized medicine is what weve been doing all along Genomics is just one more variable
Difference is quantitative not qualitative
We couldnt manage the information before How are we supposed to manage more of it?
I dont have the time to look it up in a book Thats out of date by the time its published anyway
As an aside: Policy Privacy Discrimination
How will insurance work without (or with greatly reduced)uncertainty?
13
Genomic medicine -- Informatician
Its an opportunity
Genomic data tends to be discrete and well-defined Easier to represent in a computer compared to
traditional clinical data
How do we Still need to represent traditional clinical data thats
hard
Provide data within encounters, but not dominate theencounter?
14
-
8/7/2019 Elmer Bernstam - Personalized Medicine
8/12
3/9/20
Incidentalome
False positives
16
-
8/7/2019 Elmer Bernstam - Personalized Medicine
9/12
3/9/20
The problem
If you do enough tests you get false
positives
Gene chips can do 500,000 tests
No test has 100% sensitivity/specificity
Suppose that you had a VERY good test
100% sensitivity
0.01% false positive rate (i.e., specificity of 99.9%)
17
18
-
8/7/2019 Elmer Bernstam - Personalized Medicine
10/12
3/9/20
Solution
Increase the prevalence
Only test a population that is at high risk of
disease
However, need to determine a pre-test
probability
What is the pre-test probability of a 45 year old
female with no affected 1st degree relatives, no
smoking history and two affected 2nd degree
relatives having pancreatic cancer?
19
Humans bad at bedside math
At the bedside, not practical to
Retrieve relevant information
Crunch the Bayes theorem calculation
Prevalence
Probability of finding
Probability of disease given finding
Need decision support
20
-
8/7/2019 Elmer Bernstam - Personalized Medicine
11/12
3/9/20
Suppose that your test was entirely
accurate
Positive predictor test increasesprobability by 10-40%1
May have multiple predictors
Suppose you have 20 predictors for pancreaticcancer
Still need to do bedside math based on toomuch data
1. Feero WR, Guttmacher AE, Collins FS. The genome gets personal almost. JAMA (2008) 299:11, p. 1351-2.
21
Incidentalome -- Biomedical scientist
Some might say Is this my problem?
If it is
Need more precise data
Account for more variables
More precise predictions
22
-
8/7/2019 Elmer Bernstam - Personalized Medicine
12/12
3/9/20
Incidentalome -- Clinician
Bad test I wont use it.
Too many false positives, too little useful
information
23
Incidentalome -- Informatician
Computers ARE good at math
Trouble is, garbage in garbage out Results are only as good as the inputs
Cant rely on clinician to enter the data during 10 minuteencounter
Must determine prevalence in specific populationsmake them available during the encounter Drawing on real time data
National Regional Institutional my patients Thats hard need to represent clinical data, not just
genetic data Outcomes, demographics, treatments,
24