Towards Patient-Specific Treatment:: Medical Applications of Machine Learning

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 22

Towards Patient-Specific Treatment:

Medical Applications of Machine Learning

Russ Greiner
Alberta Innovates Centre for Machine Learning & Department of Computing Science University of Alberta

Personalized Medical Treatment


Often, many treatment options for a disease
breast cancer, leukemia, Crohns disease,

Which is best for Patient#73? Dunno. Try the first, or the one that works on previous patient, or latest-&-greatest drug, or Better: identify treatment best for specific patient Not just Stage 4 melanoma but Stage 4 melanoma, 60yo male, BMI=20, + histology + genetics +

Approach
Requires knowing the connections:
Patient Features DiseaseX Patient Features Effectiveness of Treatment7

Unfortunately, NOT known Fortunately, there is often data from previous patients (with known outcomes) LEARN the connections from that historical data

Supervised Learning Framework: Learning a Classifier


-catenin m
4

-catenin n
0

E-cadherin n
4

m
2

p120 c
0

age n
0 60

size

pten

recur?

c
3

m
2

c
0

m
0

c
0

n
2

Y : Y N : N

:
0 1

:
1 4

:
4 3

:
3 0

:
0 2

:
2 0

:
0 4

:
4 0

:
0 0

:
0 2

:
2 1

:
1 2

:
2 0

:
70 62

:
2 6

:
4 1

:
1

:
2

:
0

:
1

:
4

:
3

:
0

:
2

:
0

:
4

:
0

:
0

:
2

:
71

:
2

:
2

Learner
Novel patient
-catenin m
0

-catenin n
4

E-cadherin n
2

m
0 2

p120 c
1

age n
2 70

size

pten

recur?

c
1

m
3

c
0

m
0

c
4

n
0

Classifier
2 4

Topics Seeking Students


Applications fMRI [functional Magnetic Resonance Imaging] IDM [Intelligent Diabetes Management] Cancer Heterogeneity [LDA?] Predict Metabolites Foundational PSSP [Patient Specific Survival Prediction] Experimental Design

fMRI for Psychiatric Disorders


Many psychiatric disorders look similar
depression and bipolar disorder often have similar presentations
BOLD

but different treatments Variability in population of patients, with same disorder


What works for one patient, might not work for another

Use fMRI (functional Magnetic Resonance Imaging) to distinguish disorders to better identify best treatment Our results to date:
Intl competition: Distinguish ADHD vs control: best performance, 2 pubs

Next step:
First episode psychosis; Depression vs BiPolar; w/ Dept of Psychiatry (M Brown, A Greenshaw, S Dursun, R Ramasubbu,)
Skill: Imaging, Signal Processing + MachineLearning (Neurophysiology is useful)

Intelligent Diabetes Management


Each patient with TypeI diabetes must regulate his/her own insulin:
Give self a dose of insulin (4x / day), based on
Current blood glucose level Anticipated carbohydrate consumption Stress

Dose is based on formula


specific to patient varies over time

Diabetes MD can adjust formula based on patients Diabetes Diary


but only on visits every 3-6 months? none in 3rd word countries

Automate this policy adaption process


Reinforcement Learning! w/ Alberta Diabetes Institute
(E Ryan, P Senior, )

Skill: Reinforcement Learning, Implementation, (Endocrinology is useful)

Cancer Heterogeneity
Proper treatment depends on specific (sub)disease Many diseases are mixtures of subtypes
Single patient really has multiple subtypes

For each individual with BreastCancer


Disease = mixture of Strains Strain = distn over mutations

Challenge:
Given set of patients, each with set of mutations Compute set of (latent) strains, each w/ distn over mutations

Topic Modeling LDA


Specific document = mixture of Topics Topic = distn over words LDA (Latent Dirichlet Allocation) Given set of Documents, Can learn
topics, and associated distributions
Strain Mutations Mutation

Specific patient = mixture of Strains Strain = distn over mutations


Patient Strains Strain Patient

Then use this to characterize new document

Skill: Graphical models, (Oncology is useful)

LDA for Cancer Heterogeneity


Colleague in BC Cancer has ~2000 breast cancer samples
including mutations for each

Learn strain model For new patient:

(using LDA?)

Given mutations, predict strain mixture

Later: consider EVOLUTION of cancer subtypes (within single patient)

Skill: Kernel Methods, (BioChemistry is useful)

Predict Metabolites
When you eat something (food, drink, drug, ):
Your body transforms it into products metabolized into metabolites The body then uses those metabolites

Food Producers / Drug companies / REALLY want to understand this process!


What will DrugX metabolize into?? No one knows this now But

We have ~1000 [X, metabolites[X] ] pairs

Topics Seeking Students


Applications fMRI [functional Magnetic Resonance Imaging] IDM [Intelligent Diabetes Management] Cancer Heterogeneity [LDA?] Predict Metabolites Foundational PSSP [Patient Specific Survival Prediction] Experimental Design

Stage 4 Stomach Cancer


Based on 128 patients Median survival time:
11 months

0.9

80% confidence interval


0.5

10% to 90% 2 51 months

Based ONLY on cancer location/stage


0.1

2 months 11 months

51 months

Stomach Stage 4 Cancer


#1314 Cancer Type, Stage
HGB CREATININE_SERUM ALBUMIN Predicted Survival Time

#1523

Median: 11 80% CI: 2- 51

Stomach, Stomach, Stage4 Stage4 143 108 43 21


[6.2 - 76]

60 305 35 2.2
[0.8 4.8]

Skill: Foundational machine learning,

PSSP Status
time time time time to to to to

Next steps: * Multiple time probes * More robust * High dimensional data *

Survival Prediction Survival Analysis Useful for predicting


death for patient relapse for patient failure for machine part re-injury for athlete

Extends Cox Regression, Kaplan Meier curve, Risk Assess,


Involves ALL patient information, Deals effectively with censored data Is well calibrated: statistics mean something!

Website: http://pssp.srv.ualberta.ca/

Experimental Design

Skill: Foundational machine learning, experimental design

Two standard treatments for Leukemia Several tests available + 5 news ones Want a good policy:
Which tests to run, to best identify which patient should get which treatment

OUR GOAL:
Efficiently find this good policy: training using minimum #patients, #tests
Part of national project, funded from Terry Fox Research Institute

Other Tasks
2300 Chemicals

Application Pull
Metabolomic Tasks (small molecule) Brain Tumor Analysis Microarray tasks [transplant, cancer, diseases] Medical imaging histologically stained slides

3200

Enzymes

30,000 Genes

Technology Push
large p, small n tasks
Microarray, SNP, MRI (brain tumor), fMRI

Covariate Shift Explaining Gene Signature Anomaly

you tell me

SubCell Location

SNP Analysis

Microarray
30,000

Metabolomics Proteomics

26

Cadherin-catenin: M Pasdar, J Mackey, SNP: S Damaraju, Microarray (ER+/-): K Graham, J Mackey, PolyomX team Brain Tumor: A Murtha, J Sander, P Boulanger, BTAP, fMRI: M Brown, A Greenshaw, Metabolic (Cachexia): V Baracos, D Wishart, Metabolic (Colon Cancer): R Fedorak, MTI, Crohns Disease: K Madsen, Diabetes: E Ryan, P Senior, Pathology slides: J Hugh, N Ray, AICML: N Asgarian, R Eisner, B Hoehn, My students, colleagues in Dept of Computing Science

Collaborators

Results
Given the subcell location of a BreastCancer patients junctional proteins a subjects genetics [Single Nucleotide Polymorphisms] the expression level of the genes in a womens breast cancer tumor biopsy a patients metabolic profile fMRI scan of subject MRI scan of brain tumor patient summary of Crohn patients gut flora Predict her 5 year disease free survival whether she will develop breast cancer whether she is ER+ or ERwhether s/he will lose muscle mass (cachexia) whether s/he should be screened for colon cancer whether subject has ADHD whether subject has psychosis long vs short range survival 1 year relapse, or not s
Pub ?

y y s y y y

See http://tinyurl.com/MedInfo-Papers

Recent Pubs.. all with students


Multiple Pubs in
PLoS One Nucleic Acids Research Bioinformatics BMC Bioinformatics Metabolomics Frontiers in Systems Neuroscience PLoS Biology Journal of Nutrition Theoretical Biology and Medical Modelling Human Genetics Breast Cancer Research & Treatment Current Oncology Radiation Research Clinical Cancer Research Analytic Chemistry Journal of Chromatography B

as well as
Computerized Medical Imaging and Graphics Medical Image Computing and Computer-Assisted Intervention (MICCAI) NIPS ICML AAAI IJCAI UAI EMBC ISVC

Topics Seeking Students


fMRI IDM
[functional Magnetic Resonance Imaging] [Intelligent Diabetes Management]
Imaging, Signal Processing + MachineLearning (Neurophysiology is useful)

Reinforcement Learning, Implementation (Endocrinology is useful )

Cancer Heterogeneity
Graphical Models, (Oncology is useful )

Useful, but NOT necessary you can learn what you need!

Predict Metabolites
Skill: Kernel Methods, (BioChemistry is useful)

PSSP

[Patient Specific Survival Prediction]

Foundational machine learning,

Experimental Design
Foundational machine learning, experimental design

You might also like