Harvard-MIT Division of Health Sciences and Technology
HST.951J: Medical Decision Support, Fall 2005
Instructors: Professor Lucila Ohno-Machado and Professor Staal Vinterbo
6.873/HST.951 Medical Decision Support
Fall 2005
Biomedical Decision Support
Lucila Ohno-Machado
Staal Vinterbo
Pete Szolovits
Medical decisions
Maximize value:
• Prolong life
• Increase quality of life
• Minimize pain
• Minimize cost
• Match available
resources
SMART = Scalable Medical Alert Response Technology
HIS
SMART CENTRAL
Figures by MIT OCW.
Patient PDA
1 2 3
Cricket listener
Sensors
SpO2
ECG
Location
Equipment
Defibrillator
Location (via RFID)
SMAR
T
Central
Cricket
beacons
Recognition
Location
SDM
- SpO2
- ECG
- Location
DSM
- Data analysis
Caregiver
PDA
Cricket
listener
Pattern
LSM
- priorities
- directions
Decision Theory
• Game theory
– Statistics
– Operations research
• Maximize utility
– In many domains, this means maximize $$$
Example of a Decision Problem
• College athlete considering knee surgery
• Uncertainties:
– success in recovering perfect mobility
– infection in surgery (if so, needs another surgery and may loose more mobility)
– survive surgery
Knee Surgery
Death
0.05
Death Surgery II
Surgery
Infection
0.05
Survival
0.95
No infec.
0.95
0.05
Survival
0.95
Full mobility
0.6
Poor mobility
Death
0
Death
0
Wheelchair
3
Full mobility
10
Poor mobility
6
Poor mobility
6
0.4
No Surgery
Expected Value of Surgery
Death
0.05
Death Surgery II
Surgery
Infection
7.7
0.05
Survival
0.95
No infec.
0.95
0.05
Survival
0.95
Full mobility
0.6
Poor mobility
Death
0
Death
0
Wheelchair
3
Full mobility
10
Poor mobility
6
Poor mobility
6
0.4
No Surgery
Sensitivity Analysis
• Effect of probabilities in the decision
10
Surgery
Expected
Values
6
No surgery
0
P(Death)
Sensitivity Analysis
• Effect of probabilities in the decision
10
Surgery
Expected
Values
6
No surgery
0
P(Full Mobility)
Knee Surgery
Death
0.05
Death Surgery II
Surgery
Infection
0.05
Survival
0.95
No infec.
0.95
0.05
Survival
0.95
Full mobility
0.6
Poor mobility
Death
0
Death
0
Wheelchair
3
Full mobility
10
Poor mobility
6
Poor mobility
5
0.4
No Surgery
Predictive Models Model
Building
System
Evaluation
Data
Experiment
Pattern
Discovery
Hypothesis
Objectives
• Build models from existing data
– Pattern recognition
• Apply model to new data to predict an
unknown feature such as:
– Diagnosis
– Prognosis (outcome)
Figures removed due to copyright reasons.
Chromosomes
Genome
Cell
Genes
T
T
C
G
C
G
DNA
Proteins
C
A
T
A
A
A
A
T
A
G
T
T
A
Genes contain instructions
for making proteins
G
C
G
C
Proteins
Proteins act alone or in complexes to perform many cellular functions
Figure by MIT OCW.
What kind of data?
DNA
Genome
RNA
Transcriptome
Transcription
DNA bases
Nucleus
Clinical Data
Cell
membrane
mRNA
Phenome
Chain of
amino acids
Gene
DNA
Ribosome
Protein
Translation
Physiology
Physiome
Protein
Figure by MIT OCW.
Proteome
Coronary Disease
45
ECG Interpretation
QRS amplitude
R-R interval
SV tachycardia
QRS duration
Ventricular tachycardia
AVF lead
S-T elevation
LV hypertrophy
RV hypertrophy
Myocardial infarction
P-R interval
Risk Score of Death from Angioplasty
Unadjusted Overall Mortality Rate = 2.1%
3000
60%
53.6%
62% Number
of Cases
Number of Cases
2500
50%
Mortality
Risk
2000
40%
1500
30%
21.5%
26%
1000
20%
12.4%
500
0
10%
7.6%
0.4%
0 to 2
1.4%
3 to 4
2.2% 2.9%
5 to 6
7 to 8
Risk Score Category
1.6%
9 to 10
1.3%
>10
0%
Predicting Individual Outcome
in Coronary Intervention
Logistic
Regression Model
Age > 74yrs
B2/C Lesion
Acute MI
Class 3/4 CHF
Left main PCI
IIb/IIIa Use
Stent Use
Cardiogenic Shock
Unstable Angina
Tachycardic
Chronic Renal Insuf.
Odds
Ratio
p-value
2.51
2.12
2.06
8.41
5.93
0.57
0.53
7.53
1.70
2.78
2.58
0.02
0.05
0.13
0.00
0.03
0.20
0.12
0.00
0.17
0.04
0.06
Prognostic Risk
Score Model
beta
Risk
coefficient Value
0.921
0.752
0.724
2.129
1.779
-0.554
-0.626
2.019
0.531
1.022
0.948
2
1
1
4
3
-1
-1
4
1
2
2
Informed consent
"Informed consent and good clinical practice require a discussion of
these risks and benefits, but there is very little data on the degree to
which patients comprehend the specifics of this information,"
The researchers found that, of the patients who received angioplasty
42 percent could not identify any risks, and 41 percent could not
identify any benefits. For the surgery patients, 45 percent could not
identify any risks and 22 percent could not identify any benefits.
Furthermore, when asked to quantify the risks of the procedure, 78
percent of the angioplasty and 57 percent of the surgery patients
could not.
Alexander et al, 52th ACC meeting
Overview of this Course
Individualized prediction
for decision support in
medical/biological
problems
•Theory -- how it works
•Practicality -- when to
apply
•Implementation -- how to
apply
Pre-Requisites
6034 -- Intro to AI (Machine Learning)
basic statistics, including linear regression
If needed, we will consider optional refresher
recitations:
• basic linear algebra (mostly notation)
• basic statistical tests
• set theory
Course Structure
•
•
•
•
Homeworks, individual (30%)
Midterm
(30%)
Final Project
Presentation and write-up – 5 pages plus references, figures, tables on the web
(40%)
• No final exam
Slides available online.
Office hours by arrangement. Password protection for posting articles: Username and
password.
Intro to Decision Theory and Decision Analysis
– Optimal classification
performance of a model
– Cost functions
– Individualized decisions
• Confidence in predictions
• Decision trees
Source: DOE
Simple Models
– Artificial Intelligence
• Nearest neighbors
• Association rules
• Learning from experts
– Statistics
• Linear regression
• Linear discriminant analysis
Analysis of Failure Times
• Survival analysis • Cox model
• Assumptions required for
models
• Alternatives
Supervised Methods I
• Logistic Regression
– interpretation of coefficients
– limitations
asymmetry
asymmetry
• Classification Trees
– splitting functions
– pruning
– forests
<2
color
color
border
border
A
R
border
border
detail
detail
<2
Y
<2
“malig”
“malig”
detail
detail
detail
detail
“benigh”
“benigh”
Y
> 10
“malig”
“malig”
detail
detail
“benign”
“benign”
Supervised Methods II
• Neural networks
– Regularization
– Mixture of experts
.6
34
.2
Σ
.
4
.5
.1
• Support Vector
Machines
– VC dimension
– Soft margins
2
.2
.3
Σ
.7
4
.2
Weights
.8
Weights
Σ
Supervised Methods III
• Rule-based approaches – Rough sets
– Fuzzy sets
Figures removed due to copyright reasons.
Unsupervised Learning
Clustering
– Agglomerative/divisive
– Hierarchical/nonhierarchical
– K-means, k-medoids
– Multidimensional scaling
– Visualization
Dimensionality Reduction
• Pre-processing
– Discretization algorithms
– Filtering, cleaning
• Compression
– Principal components analysis
– Partial least squares
• Variable/Model Selection
– Multivariate strategies
– Interpretation
Stochastic Search
• Approximate solution
strategies
– Greedy
– Annealing
– Genetic algorithms
– Ant colony optimization
– Other evolutionary approaches
Evaluation
• How good is the
prediction?
• Strategies for evaluation
when number of cases is
small
– Cross-validation
– Jackknife
– Bootstrap
0.90
0.80
0.70
Sensitivity
– Calibration
– Discrimination
– Bias and variance
1.00
0.60
LR
0.50
Score
aNN
0.40
0.30
0.20
0.10
0.00
0.00
0.20
0.40
0.60
1 - Specificity
0.80
1.00
Improving Performance
Combining Models/Ensembles
• Boosting
• Bagging
• Stacking
Bioinformatics
• Phylogenetic trees
• Haplotype tagging (SNP patterns)
ACTGCAT
ACGCTAG
ACTCCAA
2
GTTGCAA
CCTGCTT
Figures by MIT OCW.
Sugested General Books
• Duda R, Hart P, Stork
D.
Pattern Classification
Wiley Interscience
($103)
• Hastie T, Tibshirani
R, Friedman J.
The Elements of
Statistical Learning
Springer ($67)
Duda, Richard O., Peter E. Hart, and David G. Stork.
Pattern Classification. 2nd ed. New York, NY: Wiley, 2001.
ISBN: 0471056693.
Hastie, Trevor, Robert Tibshirani, and Jerome Friedman.
The Elements of Statistical Learning: Data Mining,
Inference, and Prediction. New York, NY: Springer, 2001.
ISBN: 0387952845.
Decision Analysis Module
• Chernoff and Moses
Elementary Decision Theory. Dover ($12)
• Hunink et al
Decision Making in
Health and Medicine:
Integrating Evidence
and Values. ($65)
Chernoff, Herman and Lincoln E. Moses.
Elementary Decision Theory. New York, NY:
Dover Publications, 1986, c1959. ISBN: 0486652181.��
Hunink, M.G. Myriam and et. al.
Decision Making in Health and Medicine: Integrating
Evidence and Values. Cambridge, UK: Cambridge
University Press, 2001. ISBN: 0521770297.