1 of 56

Machine Learning for Medical Text and EHR (word2vec, RNNs)

Joseph Paul Cohen, PhD

Montreal Institute for Learning Algorithms

Topics:�

  1. Medical Concept Representation
  2. Clinical Event Prediction

Tutorial

2 of 56

Where is the Deep Learning research?

[Shickel, Deep EHR : A Survey of Recent Advances in Deep Learning Techniques for Electronic Health Record, 2018]

(~word2vec)

(word2vec)

Topic of today!

3 of 56

Concept Representation

Mr. Smith is a 63-year-old gentleman with coronary artery disease, hypertension, hypercholesterolemia, COPD and tobacco abuse. He reports doing well. He did have some more knee pain for a few weeks, but this has resolved. He is having more trouble with his sinuses. I had started him on Flonase back in December. He says this has not really helped. Over the past couple weeks he has had significant congestion and thick discharge. No fevers or headaches but does have diffuse upper right-sided teeth pain. He denies any chest pains, palpitations, PND, orthopnea, edema or syncope. His breathing is doing fine. No cough. He continues to smoke about half-a-pack per day. He plans on trying the patches again.

Clinical Note

Clinical Publication

Representations for:

Patient

Doctor

Visit

Disease

Drug

Symptom

Helpful to understand similarity or make semi-supervised predictions!

4 of 56

Word Embeddings for Biomedical Language

Word Embeddings for Biomedical Language

Extract relationships between words and produce a latent representation

5 of 56

What to do with word embeddings?

  • We can compose them to create paragraph embeddings (bag of embeddings).
  • Use in place of words for an RNNs (More on this later!)
  • Augment learned representations on small datasets

  • `

[Cultural Shift or Linguistic Drift, Hamilton, 2016]

Study how the meaning between two texts varies (or hospitals, or doctors)?

[Pennington, 2014]

[Mikolov, 2013]

Study the compositionality of the

learned latent space

6 of 56

Token representations

One-hot encoding: binary vector per token

Example: �cat = [0 0 1 0 0 0 0 0 0 0 0 0 0 0 … 0]�dog = [0 0 0 0 1 0 0 0 0 0 0 0 0 0 … 0]�house = [1 0 0 0 0 0 0 0 0 0 0 0 0 0 … 0]

Note!

If x is one hot

The dot product of Wx

= a Single column of W

M x N

N x 1

=

M x 1

7 of 56

word2vec

involving respiratory system and other chest symptoms

Target word

Context word

Context word

Context word

Context word

involving

respiratory

doctor

chest

Mikolov, Efficient Estimation of Word Representations in Vector Space, 2013

  1. Each word is a training example
  2. Each word is used in many contexts
  3. The context defines each word

system

1

1

0

1

0

0

0

0

0

1

-1

5.1

involving

respiratory

doctor

chest

system

Context window = 2

8 of 56

Learning in progress

9 of 56

king + (woman - man) = ?

The point that is closest is queen!

10 of 56

Follow along online!

11 of 56

window_size = 2�idx_pairs = []�for sentence in tokenized_corpus:� indices = [word2idx[word] for word in sentence]�� for center_word_pos in range(len(indices)):�� for w in range(-window_size, window_size + 1):� context_word_pos = center_word_pos + w

if context_word_pos < 0 or

context_word_pos >= len(indices) or

center_word_pos == context_word_pos:� Continue

� context_word_idx = indices[context_word_pos]� idx_pairs.append((indices[center_word_pos], context_word_idx))

Code by github user: mbednarski

Output:

[[16, 12],

[16, 1],

...

[ 1, 16],

[ 1, 12]]

for each word, treated as center word

center_word_pos = 0, 1, 2, ...

for each window position

w = -2, -1, 0, 1, 2

make sure not to jump out of the sentence

sentence = ['paris', 'france', 'capital']

indices = [2, 5, 11]

12 of 56

class SkipGram(nn.Module):� def __init__(self, vocab_size, embd_size):� super(SkipGram, self).__init__()� � self.W1 = Variable(torch.randn(embd_size, vocab_size).float(), requires_grad=True)� self.W2 = Variable(torch.randn(vocab_size, embd_size).float(), requires_grad=True)� � def forward(self, focus):� � z1 = torch.matmul(self.W1, focus)�

z2 = torch.matmul(self.W2, z1)�

softmax = F.log_softmax(z2, dim=0)�

return softmax

�model = SkipGram(vocabulary_size, 2)

Initialize two matrices

|E| x |V|

|V| x |E|

Encoder:

|E| x |V| dot |V| x 1 = |E| x 1

Softmax over context prediction

Decoder:

|V| x |E| dot |E| x 1 = |V| x 1

Code by github user: mbednarski

13 of 56

A note on the softmax function

To predict multiple classes we project to a probability distribution

word2

word1

word3

3.2

5.1

-1.7

24.5

164.0

0.18

exp

normalize

0.13

0.87

0.0

14 of 56

Simplex

Because it is on a simplex; the correction of one term impacts all

Tumor

word3

word2

word1

word3

Image credits: http://gureckislab.org/

word1

word2

word2

word1

word3

0.13

0.87

0.0

x

15 of 56

Infinite ways to generate the same output.

A correction of one sends gradients to others

We can learn unseen classes through a process of elimination.

https://github.com/ieee8023/NeuralNetwork-Examples/blob/master/general/simplex-softmax.ipynb

word2

word3

word1

Word1

Word3

Word2

16 of 56

Softmax and Cross-entropy loss

To predict multiple class we can project the output onto a simplex and compute the loss there.

word2

word1

word3

3.2

5.1

-1.7

24.5

164.0

0.18

exp

normalize

0.13

0.87

0.0

loss

0.0

0.13

0.0

17 of 56

Open Access Subset of PubMedCentral

Subset of files that are available open access

  • Papers available in PDF or XML
  • 1.25 million biomedical articles and 2 million distinct words
  • Available over FTP for bulk download (CompSci Friendly!)
  • Metadata includes journal name and year�

Breast Cancer Res. �Genome Biol.�Arthritis Res.�BMC Cell Biol.�…

Journal names

Example XML data

18 of 56

Word Embeddings for Biomedical Language

Word Embeddings for Biomedical Language

Representations are biased by the data.

We can use this to our advantage to control the domain.

Wang, A Comparison of Word Embeddings for the Biomedical Natural Language Processing, 2018

(Mayo Clinic)

(c) Wikipedia + Gigaword

19 of 56

(Mayo Clinic)

Wikipedia + Gigaword

Specific

General

Example medical words and their five post similar words based on each training corpus of text

The full name of diabetes is "diabetes mellitus"

Articles say diabetics are at increased risk of hypertension (high blood pressure)

A type of peptic ulcer

One symptom is an ulcer

Colon cancer is associated with breast cancer!

20 of 56

Indiana University Hospital Reports

Chest X-ray images from the Indiana University hospital network

1000 reports available in XML format!

21 of 56

['heart size normal lungs ...',

'the heart size and ...',

'the heart is normal in ...',

'the lungs are clear the ...',

'heart size normal lungs ...',

'heart is mildly heart ...',

'the lungs are clear there ...',

'cardiac and mediastinal ...',

def clean(s):� for c in [".",",",":",";","\"","/","[","]","<",">","?"]:� s = s.replace(c, " ").lower()� return s��corpus = []��for f in os.listdir("ecgen-radiology"):� tree = xml.etree.ElementTree.parse("ecgen-radiology/" + f)� root = tree.getroot()� node = root.findall(".//AbstractText/[@Label='FINDINGS']")[0]� corpus.append(clean(str(text)))

<Abstract><AbstractText Label="COMPARISON">None.</AbstractText><AbstractText Label="INDICATION">Positive TB test</AbstractText><AbstractText Label="FINDINGS">The cardiac silhouette and mediastinum size are within normal limits. There is no pulmonary edema. There is no focal consolidation. There are no XXXX of a pleural effusion. There is no evidence of pneumothorax.</AbstractText><AbstractText Label="IMPRESSION">Normal chest x-XXXX.</AbstractText></Abstract>

Python code to parse the XML

Output!:

Input:

22 of 56

Hyperparameters!

Depending on the configuration of the model the embeddings can vary:�

No information

Compression

We can vary: the dimension of the embedding, learning rate, token words, window size, etc..

23 of 56

Size of embedding space matters

Dim = 2

Dim = 100

*via t-sne

Distance

Distance

24 of 56

25 of 56

Time series medical records

Tasks:

  • Predict probability of event in future
  • Predict duration until next visit
  • Find similar patients/events/visits

We need to define what events are!

26 of 56

ICD (International Classification of Diseases)

ICD is the foundation for the identification of health trends and statistics globally, and the international standard for reporting diseases and health conditions. It is the diagnostic classification standard for all clinical and research purposes. ICD defines the universe of diseases, disorders, injuries and other related health conditions, listed in a comprehensive, hierarchical fashion (http://www.who.int/)

1893 - Causes of Death (International Statistical Institute)

1975 - ICD-9 - (WHO)

1990 - ICD-10 - (WHO)

2022 - ICD-11 - (WHO)

There are alternative standards but they can require a fee

27 of 56

ICD (International Classification of Diseases)

Example ICD-9 codes:

786 Symptoms involving respiratory system and other chest symptoms

786.0 Dyspnea and respiratory abnormalities

786.1 Stridor

786.2 Cough

786.3 Hemoptysis

786.4 Abnormal sputum

786.5 Chest pain

786.6 Swelling, mass or lump in chest

786.7 Abnormal chest sounds

786.8 Hiccough

786.9 Other

ftp://ftp.cdc.gov/pub/Health_Statistics/NCHS/Publications/ICD9-CM/2011

28 of 56

ICD (International Classification of Diseases)

Example ICD-9 codes (786.5)

Cardialgia (see also Pain, precordial) 786.51

Diaphragmalgia 786.52

chest 786.59

anginoid (see also Pain, precordial) 786.51

chest (central) 786.50

atypical 786.59

midsternal 786.51

musculoskeletal 786.59

noncardiac 786.59

substernal 786.51

wall (anterior) 786.52

costochondral 786.52

diaphragm 786.52

heart (see also Pain, precordial) 786.51

intercostal 786.59

over heart (see also Pain, precordial) 786.51

pericardial (see also Pain, precordial) 786.51

pleura, pleural, pleuritic 786.52

precordial (region) 786.51

respiration 786.52

retrosternal 786.51

rib 786.50

substernal 786.51

respiration 786.52

Pleuralgia 786.52

Pleurodynia 786.52

Precordial pain 786.51

chest 786.59

Prinzmetal-Massumi syndrome (anterior chest wall) 786.52

painful 786.52

Lots of grouping!

29 of 56

ICD-10 is very detailed

V97.33XD: Sucked into jet engine, subsequent encounter

V00.15: Heelies Accident �Applicable To Rolling shoe, Wheeled shoe, Wheelies accident.

Supertypes: �V00-Y99 External causes of morbidity�V95-V97 Air and space transport accidents�V00 Pedestrian conveyance accident�V00.1 Rolling-type pedestrian conveyance accident

30 of 56

Time series medical records

Sequence of medical codes over time

Tasks:

  • Predict probability of code in future
  • Predict duration until next visit
  • Find similar patients/codes/visits

787.2

787.2�358.2

682.1

31 of 56

MLPs on time series data

787.2

682.1

Before

After

-5 years

+1 year

0

1

0

0

0

1

0

0

0

.

.

.

787.2

358.2

MLP

1

682.1

787.2�358.2

Multi-hot

vectors!

32 of 56

Med2Vec

Word2Vec for time series patient visits with ICD codes.

Embeddings learned for codes and demographics.

Visits

(ICD Codes)

Visits + Demographics

ICD Codes over time

Choi, Multi-layer Representation Learning for Medical Concepts, 2016

Visit embedding

Visit embedding conditioned on demographics

33 of 56

Baseline methods

One-hot: In order to compare with the raw input data, the binary vector for the visit is used.

Stacked autoencoder (SA): Using the binary vector concatenated with patient demographic information as the input the SA is trained to minimize the reconstruction error. Then will be used to generate visit representations.

Sum of Skip-gram vectors (word2vec): First learn the code-level representations with Skip-gram only. Then for the visit-level representation, add the representations of the codes within the visit.

Choi, Multi-layer Representation Learning for Medical Concepts, 2016

34 of 56

Med2Vec

Evaluation: Predicting codes of the next visit

Choi, Multi-layer Representation Learning for Medical Concepts, 2016

Private

) :

Dataset

CHOA

35 of 56

36 of 56

Example clinical note

Mr. Smith is a 63-year-old gentleman with coronary artery disease, hypertension, hypercholesterolemia, COPD and tobacco abuse. He reports doing well. He did have some more knee pain for a few weeks, but this has resolved. He is having more trouble with his sinuses. I had started him on Flonase back in December. He says this has not really helped. Over the past couple weeks he has had significant congestion and thick discharge. No fevers or headaches but does have diffuse upper right-sided teeth pain. He denies any chest pains, palpitations, PND, orthopnea, edema or syncope. His breathing is doing fine. No cough. He continues to smoke about half-a-pack per day. He plans on trying the patches again.

  • Notes are still written together with codes selected for each visit.
  • Often explaining details which do not have equivalent codes.
  • Natural language is very difficult to parse with certainty.

37 of 56

Predicting codes from notes

Converting text to codes can

  • Adapt old databases
  • Correct errors
  • Upgrade ICD versions

Jagannatha, Bidirectional RNN for Medical Event Detection in Electronic Health Records, 2016

... with upset stomach <done>

<ICD-9 787.0

Nausea and vomiting>

38 of 56

RNNs, Different types of sequential prediction tasks

one to one one to many many to one many to many many to many

Input

Output

State

Taken from http://vision.stanford.edu/teaching/cs231n/slides/2016/winter1516_lecture10.pdf and Francis Dutil

Cat

"This is a cat"

“It's hairy and I'm allergic to it”

Cat

“Ceci est un chat”

“This is a cat”

“Meow Meow”

39 of 56

RNNs

An RNN applies a function over a sequence of inputs [x1, x2, …, xT]

which produces a sequence of outputs [y1, y2, …, yT]

and each input produces a internal state [h1, h2, …, hT].

Sequence of outputs

Sequence of internal states

Sequence of inputs

xt

yt

ht

Image du blog de Christopher Olah, slide from l’École d’automne 2018, César Laurent

40 of 56

RNNs

  • A simple RNN

  • W, U and V are the parameters of the network.
  • The weights are shared over time.

Image du blog de Christopher Olah, slide from l’École d’automne 2018, César Laurent

xt

yt

ht

U

V

W

41 of 56

Unrolling the RNN over time

The weights are shared over time.

x0

xt

x1

x2

xT

y2

yT

y1

y0

yt

h2

hT

h1

h0

ht

U

U

U

U

W

W

W

W

V

V

V

V

V

U

Image du blog de Christopher Olah, slide from l’École d’automne 2018, César Laurent

42 of 56

Unrolling the RNN over time

The weights are shared over time.

x0

xt

x1

x2

xT

y2

yT

yt

h2

hT

h1

h0

ht

U

U

U

U

W

V

V

V

V

V

U

Image du blog de Christopher Olah

W

W

W

43 of 56

Unrolling the RNN over time

The weights are shared over time.

x0

xt

x1

x2

xT

y2

yT

y1

y0

yt

h2

hT

h1

h0

ht

U

U

U

U

W

W

W

W

V

V

V

V

V

U

Image du blog de Christopher Olah

44 of 56

Predicting future events

785.1

345.1

xt

782.2

358.2

682.1

782.2

787.2

yt

h2

hT

h1

h0

ht

U

U

U

U

W

W

W

W

V

V

V

V

V

U

785.1

-

785.1

A multi-hot vector

Predicting next time step

45 of 56

Vanishing gradients

46 of 56

Problem with the basic RNN

Slide from l’École d’automne 2018, César Laurent

The shade of gray shows the influence of the input of the RNN at time 1. It decreases over time, as the RNN gradually forgets its first input.

Issue addressed by:

LSTM

GRU

Attention

Recurrent batch norm

Weight regularization

Layer norm

More reading: [Pascanu 2013]

47 of 56

We can get creative with RNN designs

Slide from l’École d’automne 2018, César Laurent

x0

x1

x2

z2

z1

z0

h22

h21

h20

x0

x1

x2

y2

y1

y0

h12

h11

h10

h0

h’0

h’1

h’2

h’i

hi

h2

h1

Stacked RNNs

Bi-directional RNN

48 of 56

RNNs for next visit prediction (Doctor AI)

  • Treat codes equally: ICD diagnosis codes, procedure codes, and medication codes
  • Grouped codes into higher-order categories

Medical codes R40000

y = High level medical codes R1778

d = time since last visit

Choi, Doctor AI: Predicting Clinical Event via Recurrent Neural Networks, 2016

Patients from Sutter Health Palo Alto Medical Foundation

49 of 56

Pretraining RNNs (Doctor AI)

MIMIC II has 2,695 patients with 2+ visits

Pretrained using a larger Sutter Health dataset�~300,000 patients

Choi, Doctor AI: Predicting Clinical Events via Recurrent Neural Networks, 2016

50 of 56

Related work

Lipton, Zachary C et al. “Learning to Diagnose with LSTM Recurrent Neural Networks.” International Conference on Learning Representations. 2016

Che, Zhengping et al. “Recurrent Neural Networks for Multivariate Time Series with Missing Values.” Nature Scientific Reports. 2018

51 of 56

Medical Natural Language Inference

Romanov, Lessons from Natural Language Inference in the Clinical Domain, 2018

MedNLI dataset derived from the MIMIC-III dataset

52 of 56

Medical Natural Language Inference

Romanov, Lessons from Natural Language Inference in the Clinical Domain, 2018

Studying the errors we can see the limits of the model.

53 of 56

Discussion

"While code-based representations of clinical concepts and patient encounters are a tractable first step towards working with heterogeneous EHR data, they ignore many important real-valued measurements associated with items such as laboratory tests, intravenous medication infusions, vital signs, and more." [Shickel et al,, 2018]

54 of 56

Discussion

"While some researchers downplay the importance of interpretability in favor of significant improvements in model performance, we feel advances in deep learning transparency will hasten the widespread adoption of such methods in clinical practice." [Shickel et al, 2018]

55 of 56

Discussion

"Many studies claim state-of-the-art results, but few can be verified by external parties. This is a barrier for future model development and one cause of the slow pace of advancement." [Shickel et al, 2018]

56 of 56

References

Ching, T., et al. Opportunities And Obstacles For Deep Learning In Biology And Medicine. Journal of The Royal Society Interface. 2018

Shickel, B. et al. Deep EHR : A Survey of Recent Advances in Deep Learning Techniques for Electronic Health Record. IEEE Journal of Biomedical and Health Informatics, 2018