This document discusses using convolutional neural networks for medical text classification. It presents an approach using CNNs to classify sentences from clinical notes into categories. The model is trained on word embeddings from clinical papers and evaluated on labeled data from the Merck Manual. The CNN approach achieves better accuracy than baseline methods using doc2vec, mean word embeddings, and bag-of-words features with an SVM. Future work could include expanding the training data and applying the models to tasks like patient note classification and learning patient representations from their records.
Medical Text Classification using Convolutional Neural Network
1. Medical Text Classification using
Convolutional Neural Networks
Mark Hughes, Irene Li , Spyros Kotoulas and Toyotaro Suzumura
26, April, 2017
Informatics for Health
IBM Research Ireland
Japan Science and Technology Agency, Tokyo, Japan
IBM TJ Watson Research Center, New York, USA
2. Motivation: Medical Text Classification
( A 75-y-o woman) with sudden onset back pain last
night while lifting turkey from oven. The pain is worse
with movement or deep breath, better with rest. No
symptoms in legs, no fever or chills. No chest pain,
cough, wheezing, abdominal pain, headache… Married.
Two children. No smoking.
Unstructural
clinical notes:
Various Topics
Messy
Irrelevant
IBM Watson Smart Notes Project
Search info related to particular illnesses
--- sentence-level classification
3. State-of-the-art Representation of NLP
[1] Distributed Representations of Words and Phrases and their Compositionality, Mikolov et.al. 2013
[2] Distributed Representations of Sentences and Documents, Quoc V.Le et.al. 2014
[3] Gensim: https://radimrehurek.com/gensim/models/doc2vec.html
[4] Dai, Andrew M., Christopher Olah, and Quoc V. Le. "Document embedding with paragraph vectors." (2015).
Distributed Representations: dense vectors
• Embedding Models: Word2vec[1] , Doc2vec[2,3]
• Visualization Example:
– Semantically clusterred
– Unsupervised learning
– Large training corpus
4. Convolutional Neural Network Modeling Sentences
Figure from Kim, YoonConvolutional neural networks for sentence classification." arXiv preprint arXiv:1408.5882 (2014).
6. Datasets
[1]: US National Library of Medicine National Institutes of Health Search database http://www.ncbi.nlm.nih.gov/pubmed
[2]: Merck Manual Dataset http://www.merckmanuals.com/
Pre-trained Word2vec: 15,000 clinical research papers from PubMed[1].
Experiments: 26 Categories, 4000 sentences each, 1000 sentences validation
from Merck Manual[2].
7. Sentence embeddings + SVM
▪ Doc2vec, the distributed memory (PV-DM) model: represent each sentence
as a vector;
▪ Sentence vectors as inputs, supervised learning by SVM.
Mean Word embeddings + SVM
▪ Pair-wise mean sentence embeddings: each sentence is a vector, add zero
or eliminate if unseen;
▪ Sentence vectors as inputs, supervised learning by SVM.
Word embeddings with BOW(Bag-of-Word) Features
▪ K-means: word embeddings into 1000 clusters;
▪ BOW histogram: each sentence represented by a 1000-d vector;
▪ Sentence vectors as inputs, supervised learning by SVM.
Evaluation: Baselines
9. Conclusions & Discussions
Convolutional Neural Nets
• sentence-level classification in clinical domain;
• possible to be scaled up to paragraph/document level;
• the better ability to do classification compared with shallow
learning methods.
Representation Learning
• the ability to represent in a distributed way;
• pre-trained embeddings are useful for text
comparison/retrieval tasks.
10. Future Works
Dataset
• Extend in-domain knowledge: papers, books, relevant topics in
Wikipedia, etc;
• Test on fine graied set of clinical datasets.
Potential Applications
• Notes classification;
• Patient2vec (Use Case next page): representation learning on
individual patient, high level semantic representation of each
patient.
11. Patient2Vec: Every patient is a vector
Feature extraction from everything:
gender,age, body conditions, history
treatments, …