Machine Learning for Healthcare

HST.956, 6.S897

Lecture 1: What makes healthcare unique?

Prof. David Sontag & Pete Szolovits

Need 2 scribes for today’s lecture
 The Problem
• Cost of health care expenditures in the US are
over $3 trillion, and rising
• Despite having some of the best clinicians in
the world, chronic conditions are
– Often diagnosed late
– Often inappropriately managed
• Medical errors are pervasive
 Outline for today’s class
1. Brief history of AI and ML in healthcare
2. Why now?
3. Examples of how ML will transform
healthcare
4. What is unique about ML in healthcare?
5. Overview of class syllabus
 1970’s: MYCIN expert system
• 1970’s (Stanford): MYCIN expert Related Workand Goal

Dialogue interface
system for identifying bacteria I amready
causing severe infections ** THISIS A 26 YEAR
OLDMALE
PATIENT
Myunderstanding
is:
• Proposed a good therapy in Theageof thepatient
is 26
Thesexof thepatient
is male

~69% of cases. Better than ** FIVE DAYSAGO,HE HADRESPIRATORY-TRACT

Whatis his name?
SYMPTOMS

infectious
The Contextof the MYCIN
disease experts
Experiments ** JO
Myunderstanding
is:
Thename
of thepatientis Jo
EXPERTSYSTEM Respiratory-tract
is oneof thesymptoms
that thepatienthad
** A COUPLE OFDAYS BEFORE THEADMISSION,
HEHADA MALAISE
Description User Please
give methedateof admission

I
=~ inter-
of newcase qp~ Inference ** MARCH12, 1979
Engine
face
USER Myunderstanding
is:

Advice& t Thepatient
Malaise
wasadmittedat thehospital
is oneof thesymptoms
3 daysago
thatthepatient
had5 daysago
Explanation qp_~ Knowledge
Base [

FIGURE33-1 Short sample dialogue. The physician’s inputs

appearin capital letters after the double asterisks.
FIGURE1-1 Majorparts of an expert system. Arrowsindicate
informationflow.
some of the frames might rule out others, thus enabling the s
 1980’s: INTERNIST-1/QMR model
• 1980’s (Univ. of Pittsburgh): Probabilistic model relating:
INTERNIST-1/Quick Medical 570 binary disease variables
Reference 4,075 binary symptom variables
45,470 directed edges
• Diagnosis for internal medicine
Elicited from doctors:
flu pneumonia diabetes
15 person-years of work
Diseases

Led to advances in ML & AI

(Bayesian networks, approximate
Symptoms
inference)
fatigue cough chest high
pain A1C
Problems: 1. Clinicians entered symptoms manually
2. Difficult to maintain, difficult to generalize
[Miller et al., ‘86, Shwe et al., ‘91]
 1980’s: automating medical discovery

Discovers that prednisone

elevates cholesterol
(Annals of Internal Medicine, ‘86)

[Robert Blum, “Discovery, Confirmation and Incorporation of Causal Relationships

from a Large Time-Oriented Clinical Data Base: The RX Project”. Dept. of Computer
Science, Stanford. 1981]
 pings

which restricts the output to be between 0 and 1. If

the incoming weighted activity is larger than the

1990’s: neural networks in medicine

(negative) bias weight, the activation is positive. Pos-
itive activations cause node outputs that tendj to 1.
Negative activations cause outputs that tend to 0.
Thus, the bias weight acts as a threshold above
which the node is active. For small activation levels,
the sigmoidal function is approximately linear.
• Neural networks with
Perceptrons are the basic processing element in

clinical data took off in

most neural network models. A feed-forward neural
network, called the multilayer perceptron (MLP), is

1990, with 88 new studies

illustrated in figure 2. The network consists of sen-
sory units that make up the input layer, one or more
that year
hidden layers of processing units (perceptrons), and
one output layer of processing units (perceptrons).

• Small number of features

Every unit is connected to every unit in the layer
below. The input signal propagates through the net-

(inputs)
work a layer at a time. Because MLPs are trained
with an algorithm called error back-propagation,

• Data often collected by

they are also known as &dquo;backprop&dquo; networks.
There are many other types of networks, varying

chart review
in node models and patterns of connectivity,34
3’,‘~’4
FIGURE 2. A multilayer perceptron. This is a two-layer percep-
but the MLP is the network used in nearly all med- tron with four inputs, four hidden units, and one output unit.

Problems: 1. Did not fit well into clinical workflow

2. Hard to get enough training data
3. Poor generalization to new places

[Penny & Frost, Neural Networks in Clinical Medicine. Med Decis Making, 1996]
 395

Table 1 9 25 Neural Network Studies in Medical Decision Making*

*For reference citations, see the reference list

tP = pnor probability of most prevalent category.
$D ratio of tramng examples to weights per output
=

§A single integer in the accuracy column denotes percentage overall classification rate and a single real number between 0 and 1 indicates the
AUROCC value Neural = accuracy of neural net, Other accuracy of best other method
=
 Outline for today’s class
1. Brief history of AI and ML in healthcare
2. Why now?
3. Examples of how ML will transform
healthcare
4. What is unique about ML in healthcare?
5. Overview of class syllabus
 The Opportunity:
Adoption of Electronic Health Records
(EHR) has increased 9x in US since 2008
00000
85.2%* 96.9%*
96%
Certi ed EHR 94%*
75.5%* 83.8%*
71.9%
Basic EHR
Percentage 59.4%*
of hospitals
44.4%*
in the US
27.6%*

15.6%
12.2%
9.4%

2008 2009 2010 2011 2012 2013 2014 2015

[Henry et al., ONC Data Brief, May 2016]

Large datasets
Laboratory for
Computational
Physiology

De-identified
health data from
~40K critical care
patients

Demographics,
vital signs,
laboratory tests,
medications,
notes, …
Large datasets

“Data on nearly
230 million
unique patients
since 1995”

$$$
 Large datasets
President Obama’s initiative to create a 1 million
person research cohort Core data set:
• Baseline health exam
THE PRECISION MEDICINE INITIATIVE • Clinical data derived
from electronic health
records (EHRs)
• Healthcare claims
• Laboratory data
WHAT IS IT?
[Precision Medicine Initiative (PMI) working Group Report, Sept. 17 2015]
Precision medicine is an emerging approach for disease
prevention and treatment that takes into account people’s
individual variations in genes, environment, and lifestyle.
The Precision Medicine Initiative will generate the scientific
 Diversity of digital health data
proteomics

lab tests
imaging

social media
phone

vital signs devices genomics

 Standardization
• Diagnosis codes: ICD-9 and
ICD-10 (International
Classification of Diseases)
…

…
… [https://blog.curemd.com/the-most-bizarre-
[https://en.wikipedia.org/wiki/Lis
t_of_ICD-9_codes] icd-10-codes-infographic/]
 Standardization
• Diagnosis codes: ICD-9 and
ICD-10 (International
Classification of Diseases)
• Laboratory tests: LOINC
codes
• Pharmacy: National Drug
Codes (NDCs)
• Unified Medical Language
System (UMLS): millions of
medical concepts
[http://oplinc.com/newsletter/index_May08.htm]
Standardization
 Standardization

OMOP
Common
Data
Model v5.0
Breakthroughs in machine learning

Why now?
• Big data
• Algorithmic advances
• Open-source software
Breakthroughs in machine learning
• Major advances in ML & AI
– Learning with high-dimensional features (e.g., l1-
regularization)
– Semi-supervised and unsupervised learning
– Modern deep learning techniques (e.g. convnets,
variants of SGD)
• Democratization of machine learning
– High quality open-source software, such as
Python’s scikit-learn, TensorFlow, Torch, Theano
Industry interest in ML & healthcare
 Industry interest in ML & healthcare
• Major acquisitions to get big data for ML:
– Merge ($1 billion purchase by IBM, 2015)
medical imaging
– Truven Health Analytics ($2.6 billion purchase by
IBM, 2016)
health insurance claims
– Flatiron Health ($1.9 billion purchase by Roche,
2018)
electronic health records (oncology)
 Outline for today’s class
1. Brief history of AI and ML in healthcare
2. Why now?
3. Examples of how ML will transform
healthcare
4. What is unique about ML in healthcare?
5. Overview of class syllabus
ML will transform every aspect of healthcare

The stakeholders:

Source for figure:

http://www.mahesh-vc.com/blog/understanding-whos-paying-for-what-in-the-healthcare-industry
Emergency Department:
• Limited resources
• Time sensitive
• Critical decisions
What will the ER of the future be like?
Behind-the-scenes reasoning about the patient’s
conditions (current and future)

flu pneumonia diabetes • Better triage

Diseases
• Faster diagnosis
Drives
• Early detection of
Symptoms adverse events
fatigue cough chest high • Prevent medical
pain A1C
errors

Automatically extracted from

electronic health record
What will the ER of the future be like?

Propagating best practices

What will the ER of the future be like?

Anticipating the clinicians’

needs
 What will the ER of the future be like?
Pneumonia Detection on Chest X-Rays
h Deep Learning

n * 1 Kaylie Zhu 1 Brandon Yang 1 Hershel Mehta 1

Reducing the need for specialist consults
ti Bagul 1 Curtis Langlotz 2 Katie Shpanskaya 2
. Lungren 2 Andrew Y. Ng 1

ect
ex-
go-
nal
ur-
X-
tal-
our
e a Input
Chest X-Ray Image
for-
sts.
adi- CheXNet
121-layer CNN
ion
end
tX- Output
on Pneumonia Positive (85%)

Arrhythmia?

Figure sources: Rajpurkar et al., arXiv:1711.05225 ’17

h pneu- Rajpurkar et al., arXiv:1707.01836, '17
se every
X-rays
 creenshots of the system now running at BIDMC hospital on note: 69 y/o M patient with sev
RUQ pain. Began soon after eating bucket of ice cream and cupcake. Also is a heavy drinker. Left:
ectly proposes both ‘RUQ abdominal pain’ and ‘Allergic reaction’ as possible chief complaints. Rig
What will the ER of the future be like?
does not see the label they want, they can start typing and see a list of suggested auto-completes. Ag
t likely labels describe ‘RUQ abdominal pain’ and ‘Allergic reaction’.

Automated documentation and billing

Triage note

Predicted
chief
complaints Contextual
auto-
complete
ML will transform every aspect of healthcare

The stakeholders:

Source for figure:

http://www.mahesh-vc.com/blog/understanding-whos-paying-for-what-in-the-healthcare-industry
 What is the future of how we treat
chronic disease?
• Predicting a patient’s future disease progression

Time
Disease burden

Undiagnosed
condition
Time

Figure credit: https://www.cdc.gov/kidneydisease/prevention-risk.html

 What is the future of how we treat
chronic disease?
• Predicting a patient’s future disease progression
• Precision medicine
Choosing first line therapy in multiple myeloma
A) KRd: carfilzomib-lenalidomide-dexamethasone, B) VRd: bortezomib-lenalidomide-dexamethasone
Progression on KRd
Diagnosis and first-line treatment

Treatment A
Progression on VRd

Response to treatment A

Patient w.
condition X Treatment B
Response to treatment B

Time
 What is the future of how we treat
chronic disease?
• Early diagnosis, e.g. of diabetes, Alzheimer's,
cancer

Liquid biopsy

Figure sources: NIH,

https://www.roche.com/research_and_development/what_we_are_working_on/oncology
/liquid-biopsy.htm
 What is the future of how we treat
chronic disease?
• Continuous monitoring and coaching, e.g. for the
elderly, diabetes, psychiatric disease

Figure source (left): http://www.emeraldforhome.com/

 What is the future of how we treat
chronic disease?
• Discovery of new disease subtypes; design of
new drugs; better targeted clinical trials

Figure sources: Haldar et al., Am J Respir Crit Care Med, 2008

http://news.mit.edu/2018/automating-molecule-design-speed-drug-development-0706
 Outline for today’s class
1. Brief history of AI and ML in healthcare
2. Why now?
3. Examples of how ML will transform
healthcare
4. What is unique about ML in healthcare?
5. Overview of class syllabus
 What makes healthcare different?
• Life or death decisions
– Need robust algorithms
– Checks and balances built into ML deployment
– (Also arises in other applications of AI such as autonomous
driving)
– Need fair and accountable algorithms
• Many questions are about unsupervised learning
– Discovering disease subtypes, or answering question such
as “characterize the types of people that are highly likely to
be readmitted to the hospital”?
• Many of the questions we want to answer are causal
– Naïve use of supervised machine learning is insufficient
What makes healthcare different?
• Very little labeled data
Recent breakthroughs in AI
depended on lots of labeled data!
What makes healthcare different?
• Very little labeled data
– Motivates semi-supervised learning algorithms
• Sometimes small numbers of samples (e.g., a
rare disease)
– Learn as much as possible from other data (e.g.
healthy patients)
– Model the problem carefully
• Lots of missing data, varying time intervals,
censored labels
What makes healthcare different?
• Difficulty of de-identifying data
– Need for data sharing agreements and sensitivity
• Difficulty of deploying ML
– Commercial electronic health record software is
difficult to modify
– Data is often in silos; everyone recognizes need for
interoperability, but slow progress
– Careful testing and iteration is needed
 Goals for the semester
• Intuition for working with healthcare data
• How to set up as machine learning problems
• Understand which learning algorithms are
likely to be useful and when
• Appreciate subtleties in safely & robustly
applying ML in healthcare
• Set the research agenda for the next decade
 Outline for today’s class
1. Brief history of AI and ML in healthcare
2. Why now?
3. Examples of how ML will transform
healthcare
4. What is unique about ML in healthcare?
5. Overview of class syllabus
 Course staff
• David Sontag (instructor)
– Associate professor in EECS, joint IMES &
CSAIL
– PhD MIT, then 5.5 years as professor at NYU
– Leads clinical machine learning group

• Peter Szolovits (instructor)

– Professor in EECS, associate faculty in IMES
– Researching AI in medicine since 1975 (!)
– Leads clinical decision making group in CSAIL
 Course staff
• Willie Boag (teaching assistant)
– PhD student with Pete Szolovits
– Research in clinical NLP
– Master’s thesis on quantifying racial disparities in
end-of-life care
• Irene Chen (teaching assistant)
– PhD student with David Sontag
– Research in fairness in ML, and modeling disease
progression
– Before PhD, worked for 2 years at Dropbox
• Office hours Monday 1pm, 32-G 9th floor lobby
 Prerequisites & Enrollment
• Must submit pre-req quiz (on course website) by 11:59PM
EST today

• We assume previous undergraduate-level ML, and comfort

with:
– Machine learning methodology (e.g. generalization, cross-
validation)
– Supervised machine learning techniques (e.g. support vector
machines, neural networks)
– Optimization for ML (e.g. stochastic gradient descent)
– Statistical modeling (e.g. Gaussian mixture models)
– Python

• Because of space limitations, no listeners or auditors will

be permitted
 Logistics
• Course website:
https://mlhc19mit.github.io/

• All announcements made via Piazza – make sure you are

signed up for it!
• Recitation (optional): Fridays, starting next week (details
TBD)
• Grading:
– 40% homework (6 problem sets)
– 40% course project
– 20% participation (scribing, MLHC community consulting,
reading responses, and in-class discussion)
 Homework (tentative)
• PS0 (due Monday): human subjects training & data use
agreements
• PS1: Predicting mortality in ICUs using labs and clinical text
• PS2: Risk stratification using health insurance claims
• PS3: Clinical natural language processing
• PS4: Physiological time-series
• PS5: Causal inference (theory)
• PS6: Inferring the effect opioid prescription on addiction
 6.S897/HST.956 vs 6.874
• Our class will focus on clinical data and its use
to improve health care
• For reasons of time & scope, we will not go
deep into applications in the life sciences
– For this, we recommend taking 6.874
Computational Systems Biology: Deep Learning
in the Life Sciences
https://mit6874.github.io/