kugami
commented
6 years ago According to this paper: "Identifying risk factors for heart disease over time: Overview of 2014 i2b2/UTHealth shared task Track 2"
general information this task focused on identifying medical risk factors related to Coronary Artery Disease in the longitudinal medical records of diabetic patients a note about the corpus: the training data consisted of 60% of the total corpus (790 records) and the testing data consisted of the remaining 40% (514 records)
submissions -- what did other people do? note: this will be listed based on the past ranking, so 1. equals first place in the past competition
1st place: approach of mention level classification task,
- using the spans highlighted by the annotators
- they reannotated two thirds of the training corpus
- standardized both the mention spans and annotated both positive and negative mentions
this data was used for training, and then preprocessing: identification of section headers, negation words, modality words and output from ConText, rules were used for locating trigger words, medication and measurement, SVM classifiers were used to identify the validity and polarity of each mention, smoking status was identified by using a single 5-way classifier and a separate rule-based classifier for family history
2nd place: Divided risk factors into three categories
- Phrase-based factors (finding relevant phrases in the text)
- Logic-based factors (form of analysis needed after identifying the relevant phrase)
- Discourse-based factors (require parsing a sentence, e.g. identifying smoking status or family history)
After preprocessing the texts with MedEx (=medication information extraction system for clinical narratives)
- Using Conditional Random Fields (CRF) and
- Structural Support Vector Machines (SSVMs) >>> identifying risk based factors
- NegEx used for negation
- SVMs (normal) used for identifying discourse based risk factors
- Multi label classification approach to assign temporal attributes to risk factors
3rd place: Approach as a multiple text categorization task
- Combination of tag and attribute value pairs = independent target category
- Feature sets were created for these pairs
- centered around “hot-spot keywords”
- fed into Weka’s JRip classifier
- a second classifier built for smoker classification
- SVM utilized, could be overruled by regular expression
- A third classifier supplemented more information to the output
- this classifier based on Stanford’s Named Entity Recognition Tool
Important Takeaway’s Between the top-performing approaches, there were some similarities Pre-processing tools to gain syntactic information and only one (3rd place) added temporal attributes Nearly all the systems used medical lexicons, either as UMLS, Drugs.com and Wikipedia Only one of the teams did not mention using a lexicon of medical terms Hypertension and Family History had the best performance over all risk factors (result partly due to the collection of files – mostly indicated no family history at all) Among the top few – mostly all of them had similar performance in the system tests
michelole
Take a look on what other people did in the past in similar tasks.
See https://dbmi.hms.harvard.edu/programs/healthcare-data-science-program/clinical-nlp-challenges/7-2014-deid-heartdisease