SLIDE 1
Text Mining in Clinical Domain: Dealing with Noise
Author: Hoang Nguyen, Jon Patrick Source: KDD’16 Advisor: Jia-Ling Koh Speaker: Avon Yu
Date: 2018/12/4
- 1
Outline
- Introduc*on
- Method
- Experiment
- Conclusion
2
Text Mining in Clinical Domain: Dealing with Noise Author: Hoang - - PowerPoint PPT Presentation
Text Mining in Clinical Domain: Dealing with Noise Author: Hoang - - PowerPoint PPT Presentation
Text Mining in Clinical Domain: Dealing with Noise Author: Hoang Nguyen, Jon Patrick Source: KDD16 Advisor: Jia-Ling Koh Speaker: Avon Yu Date: 2018/12/4 1 Outline Introduc*on Method Experiment Conclusion 2
SLIDE 2
SLIDE 3
Introduction
- MoOvaOon
- High level of noise in clinical corpus.
- unknown word (ex. misspellings, acronym, abbreviaOons)
- non-word (clinical scores & measure ex. BP140/65, HR 72…)
- poor grammaOcal sentence
- Costly labelled data, which sOll o\en contain errors and
inconsistencies.
- Imbalanced data distribuOon.
3
SLIDE 4
Introduction
- Goal
- Introduces a general clinical data mining architecture
which is potenOal of addressing these challenges using:
- Pre-processing system (proof-reading)
- InteracOve model development
- AcOve learning
4
SLIDE 5
Introduction
- Framework
5
SLIDE 6
Outline
- IntroducOon
- Method
- Experiment
- Conclusion
6
SLIDE 7
Method
- StandardisaOon
- Ring-fencing tokeniser
- Finite State Recognizer (FSR) uses training examples to
recognize token paaerns consOtuOng a score or measurement that requires standardisaOon.
7
SLIDE 8
Method
- NormalisaOon & Clinical Concepts RecogniOon
- The Lexicon Management System (LMS) store the
accumulated lexical knowledge and contains categorizaOons of spelling errors, acronyms and non- word tokens.
dictionary for English and Medical terms
8
SLIDE 9
Method
- IteraOve Model Development
- The model is evaluated and the algorithm is revised in
a feedback process to produce a more accurate result.
9
SLIDE 10
Method
- IteraOve Model Development
- Feature selecOon:
- Bag of words(BOW)
- Proof reading
- Ring-fencing
- Lemma
- Medical term and gazeaer
- Bag of tags(BOT)
- Context feature
- NegaOon and modality
10
SLIDE 11
Method
- IteraOve Model Development
- New model is delivered to the Visual Annotator(VA) to
perform manual correcOon with the support of an annotaOon validaOon tool.
11
SLIDE 12
Method
- AcOve Learning
- The learner queries the most informaOve instances to
retrain the model instead of making a random selecOon.
12
SLIDE 13
Method
- AcOve Learning
- Pool-based acOve learning
13
SLIDE 14
Method
- AcOve Learning
- Simple AL
Data within the margin is less imbalanced than the enOre data.
14
SLIDE 15
Method
- AcOve Learning
- Self Confident
- Chooses the next example to be labeled so that, when
it is added to the training data, the future generalizaOon error probability is minimized
- log-loss funcOon:
15
SLIDE 16
Method
- AcOve Learning
- Kernel Farthest-First
- The most informaOve instance is the farthest instance
in the unseen pool from the current training set
16
SLIDE 17
Method
- AcOve Learning
- Balanced Explora*on and Exploita*on(Balance-EE)
- A combinaOon of Simple and KFF
- The probability p for exploraOon will be updated as:
17
SLIDE 18
Outline
- IntroducOon
- Method
- Experiment
- Conclusion
18
SLIDE 19
Experiment
- Dataset:
- All reports provided in a year’s data collecOon by the
imaging services in Australia.
- Sample of 16472 reports was drawn from Lake Imaging
and assigned to cancer (4784 reports) or non-cancer (11 688 reports) classes by the cancer registry
19
SLIDE 20
Experiment
- Descriptor (De)
- 形態學、地形學、細胞型態..
- EnOty (En)
- subject of the report
- LinguisOc (Li)
- lexical polarity, normality and modifier
- Radiologist’s coding (Ra)
- cancer stage , TNM
- Structure (St)
- heading tags
20
SLIDE 21
Experiment
21
SLIDE 22
Experiment
The evaluaOon of the reportability classifier presented here was executed independently at the Cancer Registry. The final version is implemented based on two ML algorithms, they are CondiOonal Random Fields(CRFs) and SVMs. ‘sensitivity’ is equal to ‘recall’ of the posiOve class (reportable) ‘specificity’ is the ‘recall’ of the negaOve class (non-reportable)
22
SLIDE 23
Outline
- IntroducOon
- Method
- Experiment
- Conclusion
23
SLIDE 24
Conclusion
- Presents a general system for text mining in clinical
domain with a focus on dealing with mulOple frequent kinds of noise.
- Can dramaOcally reduce human effort in idenOfying
relevant reports from the large imaging pool for further invesOgaOon of cancer.
- The classifier is built on a large real-world dataset and
can achieve high performance in filtering relevant reports.
24