Transcript View Learning - BYU Computer Science

Machine Learning
for Healthcare
David Page
Dept. of Biostatistics & Medical Informatics
and Dept. of Computer Sciences
University of Wisconsin-Madison
Electronic Medical Record
PatientID Gender Birthdate
P1
M
PatientID Date
P1
P1
3/22/63
Lab Test
1/1/01 blood glucose
1/9/01 blood glucose
PatientID Date Physician Symptoms
P1
P1
Result
42
45
PatientID Date Prescribed Date Filled
P1
5/17/98
5/18/98
1/1/01
2/1/03
Smith
Jones
Diagnosis
palpitations hypoglycemic
fever, aches influenza
PatientID SNP1 SNP2 … SNP500K
P1
P2
AA
AB
AB
BB
BB
AA
Physician Medication Dose
Jones
prilosec
10mg
Duration
3 months
Predictive Personalized
Medicine
Genetic,
Clinical,
&
Environmental
Data
Repeat for thousands of patients
State-of-the-Art
Machine
Learning
Individual
Patient
G+C+E
Predictive
Model for
Disease
Susceptibility
& Treatment
Response
Personalized
Treatment
Repeat for hundreds of diseases and treatments
3
Estimation of the Warfarin Dose
with
Clinical and Pharmacogenetic
Data


International Warfarin Pharmacogenetics
Consortium
 (IWPC)
NEJM, February 19, 2009, vol. 360, no. 8
Motivation





“In Milestone, FDA Pushes Genetic Tests
Tied to Drug”
Where: Front-page article, Wall Street
Journal, August 16, 2007
Why: FDA released new warfarin product
labeling with pharmacogenomics dosing
recommendations
What: New pharmacogenetics section and
changes in initial dosage section with
pharmacogentics in the warnings section
http://www.fda.gov/cder/foi/label/2007/009218s105lblv2.pdf
“In Milestone, FDA Pushes Genetic Tests
Tied to Drug”
Initial dosing (warfarin package insert)
“The dosing of COUMADIN must be individualized
according to patient’s sensitivity to the drug as
indicated by the PT/INR….. It is recommended
that COUMADIN therapy be initiated with a dose of
2 to 5 mg per day with dosage adjustments based
on the results of PT/INR determinations. The lower
initiation doses should be considered for patients
with certain genetic variations in CYP2C9 and
VKORC1 enzymes as well as for elderly and/or
debilitated patients….”
http://www.fda.gov/cder/foi/label/2007/009218s105lblv2.pdf
Clinicians’ responses to FDA labeling
change for warfarins


How, exactly, would I use this information?
Nice science, but prove to me that it’s better
than what we already do


i.e., I have to see a randomized trial comparing
genotype-guided versus usual dosing
Summer 2009: the NHLBI Clarification of Optimal
Anticoagulation through Genetics (COAG) trial (PI:
Stephen Kimmel, MD)
Current warfarin pharmacogenetics
information limitations


Clinical utility (or a randomized trial) will
require dosing equation that incorporates
genetic and non-genetic, demographic
information.
Numerous such equations have been
proposed, but:
•
•

most are highly geographically confined
none were developed from robust data in
Asians, Caucasians, and Africans
Thus, an equation derived from a large,
geographically and ethnically diverse
population was needed to help insure global
clinical utility.
IWPC - 21 research groups
4 continents and 9 countries
 Asia


Europe


Sweden, United Kingdom
North America


Israel, Japan, Korea, Taiwan, Singapore
USA (11 states: Alabama, California, Florida,
Illinois, Missouri, North Carolina, Pennsylvania,
Tennessee, Utah, Washington, Wisconsin)
South America

Brazil
Dataset






5,700 patients treated with warfarin
Demographic characteristics
Primary indication for warfarin treatment
Stable therapeutic dose of warfarin
Treatment INR
Target INR



5,052 patients with a target INR of 2-3
Concomitant medications

Grouped by increased or decreased effect on INR

CYP2C9 (*1, *2 and *3)
Presence of genotype variants

VKORC1 (one of seven SNPs in linkage disequilibrium)

blinded re-genotyping for quality control
Age, height and weight
Average warfarin doses for stable
INR (median – 2.5)
Race, inducers and amiodarone
CYP2C9 and VKORC1 genotypes
Weekly dose by CYP2C9 genotype
CYP2C9 genotype by race
Weekly dose by VKORC1 -1639 genotype
VKORC1 -1639 genotype by race
Modeling of VKORC1 SNPs

Missing values of VKORC1 -1639 G>A
(rs9923231)


Imputed based on race and VKORC1 SNP data at
2255C>T (rs2359612), 1173 C>T (rs9934438), or
1542G>Crs8050894
If the VKORC1 genotype could not be imputed, it
was treated as “missing” (a distinct variable) in
the model.
Data Analysis Methodology

Derivation Cohort


4,043 patients with a stable dose of warfarin and
target INR of 2-3 mg/week
Used for developing dose prediction models
Validation Cohort
•
•
1,009 patients (20% of dataset)
Used for testing final selected model
Analysis group did not have access to
validation set until after the final model was
selected
Real-valued prediction methods used

Included, among others
 Support vector regression
 Regression trees
 Model trees
 Multivariate adaptive regression splines
 Least-angle regression
 Lasso
 Logarithmic and square-root transformations
 Direct prediction of dose
Support vector regression and Ordinary least-squares
linear regression gave the lowest mean absolute error


Predicted the square root of the dose
Incorporated both genetic and clinical data
IWPC pharmacogenetic dosing algorithm


**The output of this
algorithm must be
squared to compute
weekly dose in mg
^All references to
VKORC1 refer to
genotype for
rs9923231
+
+
-
5.6044
0.2614
0.0087
0.0128
0.8677
1.6974
0.4854
-
0.5211 x
0.9357 x
1.0616 x
-
1.9206 x
2.3312 x
0.2188 x
+
=
x
x
x
x
x
x
Age in decades
Height in cm
Weight in kg
VKORC1 ^ A/G
VKORC1 A/A
VKORC1 genotype
unknown
CYP2C9 *1/*2
CYP2C9 *1/*3
CYP2C9 *2/*2
CYP2C9 *2/*3
CYP2C9 *3/*3
CYP2C9 genotype
unknown
0.1092 x
Asian race
0.2760 x
Black or African
American
0.1032 x
Missing or Mixed
race
1.1816 x
Enzyme inducer
status
0.5503 x
Ami odarone status
Squa re root of weekly warfarin dose**
IWPC clinical dosing algorithm

**The output of this
algorithm must be
squared to compute
weekly dose in mg
+
+
+
+
+
=
4.0376
0.2546 x
Age in decades
0.0118 x
Height in cm
0.0134 x
Weight in kg
0.6752 x
Asian race
0.4060 x
Black or African American
0.0443 x
Missing or Mixed race
1.2799 x
Enzyme inducer status
0.5695 x
Ami odarone status
Squa re root of weekly warfarin dose**
Results
Inclusion of genotypes for CYP2C9 and
VKORC1, in addition to clinical variables, are
significantly closer to estimating the
appropriate initial dose of warfarin than just
a clinical or fixed-dose approach
46.2% of the population with ≤21 mg/wk or
≥49 mg/wk benefit the most

These are the patients for whom an underdose or
overdose could have adverse clinical
consequences.
Patients requiring an intermediate dose are
likely to obtain little benefit including
genotypes
Model comparisons
Warfarin doses predicted for the clinical and
PGx algorithms with and without amiodarone
50 yr old
White
Male
175 cm
80 kg
Genotypes can change the recommended dose from
>45 mg/wk to <10 mg/wk when all other factors equal!
Warfarin doses predicted for the clinical and
PGx algorithms based on race and genotype
50 yr old
Male
175 cm
80 kg
Racial differences in the estimated dose are insignificant when
genotypes included. Clinical algorithm may substantially overestimate
or underestimate the dose.
% Patients with
dose estimates
within 20% of
actual dose
• Comparison of PGx, clinical
and fixed dose approaches
• 3 dose groups shown (mg/wk)
• low (≤21)
• intermediate (>21 to <49)
• high (≥49)
• Fixed dose (35 mg/wk)
• None of the estimates for
low and high dose groups were
within 20% of actual dose
Limitations of this study
Did not address the issue of whether a
precise initial dose of warfarin translates into
1.

improved clinical end points reduction in time
needed to achieve a stable therapeutic INR, fewer
INRs out of range, reduced incidence of bleeding or
thromboembolic events
Did not have sufficient data across the 21
groups to include potentially important
factors such as
2.

smoking status, vitamin K intake, alcohol
consumption, other genetic factors (e.g., CYP4F2,
ApoE, GGCX), environmental factors
New England Journal of Medicine, Feb 2009
Data available at PharmGKB
•
www.pharmgkb.org
•
Accession number: PA162355460
IWPC Authors
Writing committee: Teri E. Klein, Russ B. Altman, Niklas Eriksson, Brian F. Gage, Stephen E.
Kimmel, Ming-Ta M. Lee, Nita A. Limdi, David Page, Dan M. Roden, Michael J.
Wagner, Michael D. Caldwell, Julie A. Johnson
Data Contributors:
Academic Sinica, Taiwan, ROC: Ming-Ta M. Lee, Yuan-Tsong Chen
Chang Gung Memorial Hospital, Chang Gung University, Taiwan, ROC: Ming-Shien Wen
China Medical University, Graduate Institute of Chinese Medical Science, Taichung, Taiwan,
ROC: Ming-Ta M. Lee
Hadassah Medical Organization, Israel: Yoseph Caraco, Idit Achache, Simha Blotnick,
Mordechai Muszkat
Inje University, Korea: Jae-Gook Shin, Ho-Sook Kim
Instituto Nacional de Câncer, Brazil: Guilherme Suarez-Kurtz, Jamila Alessandra Perini
Instituto Nacional de Cardiologia Laranjeiras, Brazil: Edimilson Silva-Assunção
Intermountain Healthcare, USA: Jeffrey L. Anderson, Benjamin D. Horne, John F. Carlquist
Marshfield Clinic, USA: Michael D. Caldwell, Richard L. Berg, James K. Burmester
National University Hospital, Singapore: Boon Cher Goh, Soo-Chin Lee
Newcastle University, United Kingdom: Farhad Kamali, Elizabeth Sconce, Ann K. Daly
University of Alabama, USA: Nita A. Limdi
University of California, San Francisco, USA: Alan H.B. Wu
University of Florida, USA: Julie A. Johnson, Taimour Y. Langaee, Hua Feng
University of Illinois, Chicago, USA: Larisa Cavallari, Kathryn Momary
University of Liverpool, United Kingdom: Munir Pirmohamed, Andrea Jorgensen, Cheng Hok
Toh, Paula Williamson
University of North Carolina, USA: Howard McLeod, James P. Evans, Karen E. Weck
University of Pennsylvania, USA: Stephen E. Kimmel, Colleen Brensinger
University of Tokyo and RIKEN Center for Genomic Medicine, Japan: Yusuke Nakamura, Taisei
Mushiroda
University of Washington, USA: David Veenstra, Lisa Meckley, Mark J. Rieder, Allan E. Rettie
Uppsala University, Sweden: Mia Wadelius, Niclas Eriksson, Håkan Melhus
Vanderbilt University, USA: C. Michael Stein, Dan M. Roden, Ute Schwartz, Daniel Kurnik
Washington University in St. Louis, USA: Brian F. Gage, Elena Deych, Petra Lenzini, Charles
Eby
Wellcome Trust Sanger Institute, United Kingdom: Leslie Y. Chen, Panos Deloukas
Statistical Analysis:
University of Alabama, USA: Nita A. Limdi
Marshfield Clinic, USA: Michael D. Caldwell
North Carolina State University, USA: Alison Motsinger-Reif
Stanford University, USA: Russ B. Altman, Hersh Sagrieya, Teri E. Klein, Balaji S.
Srinivasan
Uppsala University, Uppsala Clinical Research Center, Sweden: Niclas Eriksson
University of California, San Francisco, USA: Alan H.B. Wu
University of North Carolina, USA: Michael J. Wagner
University of Florida, USA: Julie A. Johnson
University of Pennsylvania, USA: Stephen E. Kimmel
University of Wisconsin-Madison, USA: David Page, Eric Lantz, Tim Chang
Vanderbilt University, USA: Marylyn Ritchie
Washington University in St. Louis, USA: Brian F. Gage, Elena Deych
Genotyping QC of IWPC Samples:
Academic Sinica, Taiwan, ROC: Ming-Ta M. Lee, Liang-Suei Lu
Genotype and Phenotype QC:
Inje University, Korea: Jae-Gook Shin
Marshfield Clinic, USA: Michael D. Caldwell
Stanford University, USA: Teri E. Klein, Russ B. Altman, Balaji S. Srinivasan
University of Alabama, USA: Nita A. Limdi
University of Florida, USA: Julie A. Johnson
University of Pennsylvania, USA: Stephen E. Kimmel
University of North Carolina, USA: Michael J. Wagner
University of Wisconsin-Madison, USA: David Page
Washington University in St. Louis, USA: Brian F. Gage
Vanderbilt University, USA: Marylyn Ritchie
Data Curation:
Stanford University, USA: Teri E. Klein, Russ B. Altman, Balaji S. Srinivasan
University of North Carolina, USA: Michael J. Wagner
Washington University in St. Louis, USA: Elena Deych
Application: Mammography


Provide decision support for radiologists
Variability due to differences in training and
experience… to get 90% of cancers, have high
false positive rate

Experts have higher cancer detection and fewer
benign biopsies

Shortage of experts
Bayes Net for Mammography

Kahn, Roberts, Wang, Jenks, Haddawy (1995)

Kahn, Roberts, Shaffer, Haddawy (1997)

Burnside, Rubin, Shachter (2000)

Note: not CAD (computer-assisted diagnosis),
which circles abnormalities in an image… this is
based on data entered into National
Mammography Database schema by radiologists
Mass Stability
Mass Margins
Ca++ Lucent Milk of
Centered Calcium ++
Ca Dermal
Ca++ Round
Mass Density
Ca++ Dystrophic
Mass Shape
Mass Size
Breast
Density
Mass P/A/O
Benign v.
Malignant
Skin Lesion
Tubular
Density
Architectural LN Asymmetric
Distortion
Density
Ca++ Popcorn
Ca++ Fine/
Linear
Ca++ Eggshell
Ca++ Pleomorphic
Age
FHx
HRT
Ca++ Punctate
Ca++ Amorphous
Ca++ Rod-like
Mammography Database
Patient
Abnormality Date
Calcification …
Fine/Linear
Mass
Size
Loc
Benign/
Malignant
P1
1
5/02
No
0.03
RU4
B
P1
2
5/04
Yes
0.05
RU4
M
P1
3
5/04
No
0.04
LL3
B
P2
…
4
…
6/00
…
No
…
0.02
…
RL2
…
B
…
Level 1: Parameters
P(Benign) =
??
.99
Benign v.
Malignant
Calc Fine
Linear
Mass
Size
P(Yes| Benign) =
.01
??
P( size > 5| Benign)
=
P(Yes| Malignant) =
.55
??
P(size > 5| Malignant) =
.33
??
.42
??
Level 2: Structure + Parameters
Benign v.
Malignant
Calc Fine
Linear
P(Yes|
Benign)
P(Yes)
= .02 = .01
P(Yes| Malignant) = .55
P(Benign) = .99
Mass
Size
P( size > 5 )= .1
P(size > 5| Benign ^ Yes) = .4
P( size > 5| Benign) = .33
P(size > 5| Malignant ^ Yes) = .6
P(size > 5| Malignant) = .42
P(size > 5| Benign ^ No)
= .05
P(size > 5| Malignant ^ No) = .2
Data


Structured data from actual practice
National Mammography Database


Our dataset contains



Standard for reporting all abnormalities
435 malignancies
65,365 benign abnormalities
Link to biopsy results

Obtain disease diagnosis – our ground truth
Hypotheses


Learn relationships that are useful to
radiologist
Improve by moving up learning hierarchy
Results (Radiology, 2009)



Trained (Level 2, TAN) Bayesian network model
achieved an AUC of 0.966 which was
significantly better than the radiologists’ AUC of
0.940 (P = 0.005)
Trained BN demonstrated significantly better
sensitivity than the radiologist (89.5% vs.
82.3%—P = 0.009) at a specificity of 90%
Trained BN demonstrated significantly better
specificity than the radiologist (93.4% versus
86.5%—P = 0.007) at a sensitivity of 85%
ROC: Level 2 (TAN) vs. Level 1
Precision-Recall Curves
Mammography Database
Patient
Abnormality Date
Calcification …
Fine/Linear
Mass
Size
Loc
Benign/
Malignant
P1
1
5/02
No
0.03
RU4
B
P1
2
5/04
Yes
0.05
RU4
M
P1
3
5/04
No
0.04
LL3
B
P2
…
4
…
6/00
…
No
…
0.02
…
RL2
…
B
…
Statistical Relational Learning


Learn probabilistic model, but don’t
assume iid data: there may be relevant
data in other rows or even other tables
Database schema: defines set of features
SRL Aggregates Information from
Related Rows or Tables


Extend probabilistic models to relational
databases
Probabilistic Relational Models
(Friedman et al. 1999, Getoor et al. 2001)



Tricky issue: one to many relationships
Approach: use aggregation
PRMs cannot capture all relevant concepts
Aggregation Function:
AggregateMin,
Illustration
Max, Average,
etc.
Patient
Abnormality Date
Calcification …
Fine/Linear
Mass
Size
Loc
Benign/
Malignant
P1
1
5/02
No
0.03
RU4
B
P1
2
5/04
Yes
0.05
RU4
M
P1
3
5/04
No
0.04
LL3
B
P2
…
4
…
6/00
…
No
…
0.02
…
RL2
…
B
…
New Schema
Patient
Abnormality Date
Size
Calcification … Mass Avg
Avg
Size Loc
this Date
Fine/Linear
Size this date
Benign/
Malignant
P1
1
5/02
No
0.03 0.03 RU4
B
P1
2
5/04
Yes
0.05 0.045
0.045 RU4
M
P1
3
5/04
No
0.04 0.045 LL3
B
P2
…
4
…
6/00
…
No
…
0.02
…
B
…
0.02
0.02
…
…
RL2
…
Level 3: Aggregates
Benign v.
Malignant
Calc Fine
Linear
Avg
Size
this
date
Mass
Size
Note: Learn parameters for each
node
Database Notion of View



New tables or fields defined in terms of
existing tables and fields known as views
A view corresponds to alteration in
database schema
Goal: automate the learning of views
Possible View
Patient
Abnormality Date
Calcification …
Fine/Linear
Mass
Size
Loc
Benign/
Malignant
P1
1
5/02
No
0.03
RU4
B
P1
2
5/04
Yes
0.05
RU4
M
P1
3
5/04
No
0.04
LL3
B
P2
…
4
…
6/00
…
No
…
0.02
…
RL2
…
B
…
New Schema
Patient
Abnormality Date
Calcification … Mass Increase
Increase Loc
Benign/
Fine/Linear
Size
in
Malignant
In size
Size
P1
1
5/02
No
0.03 No
No
RU4
B
P1
2
5/04
Yes
0.05 Yes
Yes RU4
P1
3
5/04
No
0.04
No
No
LL3
B
P2
…
4
…
6/00
…
No
…
0.02
…
No
No
…
…
RL2
…
B
…
M
Level 4: View Learning
Increase
in Size
Benign v.
Malignant
Calc Fine
Linear
Avg
Size
this
date
Mass
Size
Note: Include aggregate features
Learn parameters for each node
Level 4: View Learning

Learn rules predictive of “malignant”


Treat each rule as a new field



We used Aleph (Srinivasan)
1 if abnormality matches rule
0 otherwise
New view consists of original table
extended with new fields
Experimental Methodology




10-fold cross validation
Split at the patient level
Roughly 40 malignant cases and 6000
benign cases in each fold
Tree Augmented Naïve Bayes (TAN) as
structure learner (Friedman,Geiger & Goldszmidt ’97)
Sample View
[Burnside et al. AMIA05]
malignant(A) :birads_category(A,b5),
massPAO(A,present),
massesDensity(A,high),
ho_breastCA(A,hxDCorLC),
in_same_mammogram(A,B),
calc_pleomorphic(B,notPresent),
calc_punctate(B,notPresent).
All Levels of Learning
1
Level 4 (View)
Precision
0.9
0.8
Level 3 (Aggregate)
0.7
Level 2 (Structure)
0.6
Level 1 (Parameter)
0.5
0.4
0.3
0.2
0.1
0
0
0.1
0.2
0.3
0.4
0.5
Recall
0.6
0.7
0.8
0.9
1
View Learning: First Approach
[Davis et al. IA05, Davis et al. IJCAI05]
Step 1
Step 2
Target
Predicate
Rule
Learner
Rule 1
Learn
Step 3
Rule 2
…
Rule N
Select Build Model
Drawback to First Approach

Mismatch between



Rule building
Model’s use of rules
Should Score As You Use (SAYU)
SAYU
[Davis et al. ECML05]

Build network as we learn rules
[Landwehr et al. AAAI 2005]


Score rule on whether it improves network
Results in tight coupling between
rule generation, selection and usage
SAYU-NB
Class
Value
0.0
0.1
Score = 0.3
2
0
5
…
Rule 14
seed 2
1
Rule 1
2
Rule 3
Rule N
SAYU-View
[Davis et al. Intro to SRL 06]
Class
Value
…
…
Feat
1
Feat
N
Agg
1
…
Agg
M
Rule
1
Rule
L
Parameter Settings





Score using AUC-PR (recall >= .5)
Keep a rule: 2% increase in AUC
Switch seeds after adding a rule
Train set to learn network
structure and parameters
Tune set to score structures
Relational Learning Algorithms
1
SAYU-View
0.9
Precision
0.8
Initial Level 4 (View)
0.7
Level 3 (Aggregates)
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
Recall
0.7
0.8
0.9
1
Electronic Medical Record
PatientID Gender Birthdate
P1
M
PatientID Date
P1
P1
3/22/63
Lab Test
1/1/01 blood glucose
1/9/01 blood glucose
PatientID Date Physician Symptoms
P1
P1
Result
42
45
PatientID Date Prescribed Date Filled
P1
5/17/98
5/18/98
1/1/01
2/1/03
Smith
Jones
Diagnosis
palpitations hypoglycemic
fever, aches influenza
PatientID SNP1 SNP2 … SNP500K
P1
P2
AA
AB
AB
BB
BB
AA
Physician Medication Dose
Jones
prilosec
10mg
Duration
3 months
Cox Inhibition


Non-steroidal anti-inflammatory drug
Cox-2 goal: reduce stomach trouble
Cox-1
Aspirin, Aleve,
Ibuprofen, etc
block both pathways
Cox-2
Vioxx,
Bextra,
Celebrex
block this
pathway
Cox-2 Timeline
Dec. 1998-May 1999,
Celebrex, Vioxx approved
2002,
Beginning of
APPROVe Study
Dec. 2004,
FDA issues warning
2001,
Cox-2 sales top
$6 billion/year in US
Sept 2004,
Vioxx voluntarily
pulled from market
April 2005,
FDA removes
Bextra from market
Predicting Adverse Reaction
to Cox-2 Inhibitors
Given: A patient’s clinical history
Do: Predict whether the patient will have a
myocardial infarction (MI)
Note: This is work in progress
Data


492 patients who took Cox-2, MI
77077 patients who took Cox-2, no MI


Sub-sampled 651 patients
Relational tables for




Lab tests
Drugs taken
Diagnoses
Observations
Q: What Data to Use?


All data for a patient? Many perfect
predictors
Cut off data right before MI



Model not relevant pre-Cox2ib
Uniformly more data for non-MI cases
Our choice: cut off data for each patient
at first Cox2ib prescription
Approaches Tried



Propositional: Linear SVM, naïve Bayes,
TAN, trees, boosted trees, boosted rules
Relational: Inductive Logic Programming
(ILP) system Aleph
SRL: View learning with SAYU
Experimental Methodology

10-fold cross validation

Feature selection pick top 50/fold

ROC curves to evaluate

Paired t-test for significance
Algorithms Compared

Naïve Bayes
Best feature vector
approaches

Boosted rules (C5)

SAYU-TAN (w/initial feature set)
Note: Preliminary results with Aleph were poor/slow
Algorithm Comparison
ROC Area
0.9
0.8
SAYU-TAN
Naive Bayes
Average AUC-ROC
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
Boosted Rules
Sample Rule

myocardial_infarction(A) :hasdrug(A, GLUCOSE),
diagnosis(A, ischemic heart disease).
Sample Rule

myocardial_infarction(A) :diagnosis(A,B, INFECTIOUS AND
PARASITIC DISEASES),
before(B,10/26/1982),
age(A,B,C),
younger(C, 51).
Lingering Questions

Are we predicting predisposition to MI?

Can we do better with data we have?

How much will genotype data help?
Conclusions

EMRs and genotyping give machine
learning a new opportunity for great
impact on healthcare in next few years




Personalized medicine
Pharmacovigilance (FDA’s Sentinel, OMOP)
Decision support
Statistical relational learning helps for
some tasks (but not all)
Conclusions (Continued)

Fancy new algorithms not always the
best… healthcare applications raise other
issues



Missing data (not missing at random)
Need simple, comprehensible models…
clinicians may prefer slightly less accurate
model if it makes more sense to them
Different evaluation metrics
Thanks









Jesse Davis
Beth Burnside
Vitor Santos Costa
Michael Caldwell
Peggy Peissig
Eric Lantz
Jude Shavlik
IWPC
WGI (Wisconsin Genomics Initiative)