1. Identifying Condition-Action Statements in Medical Guidelines using Domain-Independent FeaturesBy
Hossein Hematialam and Wlodek Zadrozny
Presented by
Seethalakshmi Gopalakrishnan

2. Contents Introduction Motivation Proposed Method
Condition-Action extraction
Classification
Evaluation & Model performance
Conclusion & Future work

3. Abstract
Clinical Practice Guidelines to Computer interpretable guideline
Report Experiments with several machine learning techniques to classify sentences as to whether they express conditions and actions.

4. Introduction
Statements to assist with practitioners and patients decisions.
Establish criteria regarding diagnosis, management and treatment.
Example:
If the A1C is 7.0% and a repeat result of 6.8%, the diagnosis of diabetes is confirmed- “What is true”
Topical and oral decongestants and antihistamines should be avoided in patients with ABRS” –”what to do or not to do”

5. CDSS
Clinical decision-support system- decision support for health care professionals, and using clinical data or knowledge.
Deciding which questions to ask, tests to order, procedures to perform, treatment to indicate or which alternative medical care to try.
CDSS Fall into two categories
Determining “what is true” about a patient
Determining “what to do” for the patient.

6. Motivation
Condition-action statements provide information about expected process flow.
If a guideline-based CDSS could extract and formalize condition-action statements, it could help practitioners in the decision-making process.
Help automatically asses the relationship between therapies, guidelines and outcomes and help the impact of changing guidelines.

7. Proposed Method
Automated process to find and extract condition-action statements from medical guidelines.
Supervised machine learning model for classifying sentences as to whether they express a condition or not.
Natural language processing and information extraction tools to extract conditions and resulting activities.

8. Condition-Action Extraction
Given an input statement classify it into
NC( no Condition)
CA(condition-action sentence)
CC ( condition-Consequence)
Combine CC and CA to determine both what is true and what to do.

9. Limitations of condition-action extraction
Guidelines may contain statements with a condition referring to a consequence in another statement.
Condition and action in two different sentences. In this case separate evidence statements were developed.

10. Classification POS tag as features in the classification models.
Condition-action sentences have a modifier in sentences.
Example:
“In the population aged 18 years or older with CKD and hypertension, initial antihypertensive treatment should include an ACEI or ARB to improve kidney outcomes”
Condition: the population aged 18 years or older with CKD and hypertension
Modifier : “In”
If, in , for, to, which and when – frequent modifiers

11. Classification-contd.
Core NLP Shift-Reduce Constituency Parser sentences
If, in , for  IN
To  TO
When , which  WHADV
Regular expressions are used to find the parses which are promising candidates for extraction of condition-action pairs.

12. Classification -contd
POS tags are extracted as features for the model.
Example: “In adults with hypertension, does initiating antihypertensive pharmacologic therapy at specific BP thresholds improve health outcomes”
”(ROOT (S (PP (IN In) (NP (NP (NNS adults)) (PP (IN with) (NP (NN hypertension))))) (, ,) (VP (VBZ does) (S (VP (VBG initiating) (S (NP (NP (JJ anti- hypertensive) (JJ pharmacologic) (NN therapy)) (PP (IN at) (NP (JJ specific) (NN BP) (NNS thresholds)))) (VP (VBP improve) (NP (NN health) (NNS outcomes))))))) (. ?)))”

13. Created features for the model based on POS tags and their combinations.
Sets of features and the combinations are learned automatically from annotated examples.

14. Evaluation-Data preparation
Converting guidelines from PDF or html to text format, editing sentences only to manage conversion errors.
Use regular expressions to select candidate sentences.
Sentences do not include modifiers. Eg. “Beta-blockers, including eye drops, are contraindicated in patients with asthma”
Annotators – paraphrase candidate sentences as “if condition, then consequence”. “If patients have asthma, then beta blockers, including eye drops, are contraindicated”.

15. Model Performance
Hypertension, asthma and rhinosinusitis guidelines and gold standard guidelines were applied to evaluate the model.
278,172 and 761 candidate statements respectively.
Get rid of 38,116,5 no condition statement from guidelines.
Weka, ZeroR, Naïve Bayes classifier, J48 and random forest classifiers.
Results based on 10 fold cross validation on respective dataset.

16. Condition- Action
Condition- Effect
Action
No Con- dition
Asthma 38 7 8
224
Rhinosinusitis 39 15
726
Hypertension 63 14 1
238
Asthma
Condition-Action
Total
Precision
Recall
F-measure
ZeroR
0.69
NaiveBayes
0.455
0.263
0.333
J48
0.444
0.211
0.286
0.67
RandomForest
0.5
0.079
0.136

17. Rhinosinusitis
Condition-Action
Total
Precision
Recall
F-measure
ZeroR
0.80
Naïve Bayes
0.5
0.412
0.452
J48
0.581
0.258
0.357
0.81
Random Forest
0.844
0.392
0.535
0.84
Hypertension
Condition-Action
Total
Precision
Recall
F-measure
ZeroR
0.72
Naïve Bayes
0.581
0.397
0.472
0.74
J48
0.619
0.413
0.495
Random Forest
0.931
0.429
0.587
0.81

18. Conclusion
Random Forest classifier- good for condition-Action statement extraction.
Wenzina and Kaiser(Asthma guidelines)- used information extraction rules and semantic pattern rules.
Problem of automated extraction of condition-action from clinical guidelines based on an annotated corpus.

19. Future Work
Performance is low compared to a collection of manually created extraction rules.
Experiment with discourse relations- important for understanding lists and tables.
Annotating with standard annotation tools like BRAT.

20. THANK YOU