Formalizing guideline text into a computable model, and linking clinical terms and recommendations in clinical guidelines to concepts in the electronic patient record (EHR) is difficult as, typically, both the guideline text and EHR content may be ambiguous, inconsistent and make use of implicit and background medical knowledge. How can lexical-based IE approaches help to automate this task? In this presentation, various design patterns are discussed and some tools presented.
Automating the formalization of clinical guidelines using information extraction
1. Automating the formalization of clinical
guidelines using information extraction:
an overview of recent lexical approaches
05 August 2011
Phil Gooch
Centre for Health Informatics
City University, London UK
2. Clinical guidelines
• Contain recommendations for best practice based on systematic
reviews of clinical evidence, consensus statements and expert opinion.
• Goal is to reduce variation in medical care by promoting the most
effective treatments, and to provide a means of quality control in clinical
practice via audit
• Produced by a variety of organizations (e.g. NICE, RCP, SIGN) in a
variety of document formats usually not conducive to use at the point of
care.
3. Clinical decision support (CDS)
• Aims to provide diagnostic and treatment recommendations and
advice at the point of care, i.e. information tailored for the specific
patient under consideration by the clinician during a consultation
• CDS systems require a knowledge base (KB), usually derived from
guidelines, consisting of declarative knowledge (penicillin is-a
antibiotic) and procedural (if…then) rules, and some sort of electronic
patient record system (EPR)
4. Computer-interpretable guidelines
• Early systems ‘computerized’ guidelines by making them available ‘on
the computer’, e.g. as HTML or PDF
• Did not lead to improved guideline compliance or use!
• To standardize the format of the knowledge-base, ease development
of CDS, and to improve guideline use at the point of care, a number of
formalisms for representing guidelines have been developed
5. Computer-interpretable guidelines (CIGs)
Rule-based: ‘if ... then’, e.g. Arden Syntax for individual clinical decisions
LET Last_HgA1C BE READ LATEST {"HgA1C Value"};
LET Diabetic_Patient BE READ LATEST {"Problem: Diabetes"};
if Diabetic_Patient and Last_HgA1C Occurred not within past 6 months and Last_HgA1C is less
than or equal 7
then conclude true;
Document based, e.g. GEM, for complete guideline documents in XML
OO expression query languages e.g. GELLO:
observation.code == ‘SBP’ AND observation.value > 140 AND assessment.code ==‘LVF’
Task-network models (TNM), e.g. GLIF, Asbru, PROforma, for workflow-like
modelling of decisions over time
6. Formalization of guidelines into a CIG model
• Declarative: Mapping clinical concepts in the guideline to terms within a
controlled vocabulary (e.g. UMLS) or ‘virtual medical record’
• Procedural: Identification and extraction of eligibility criteria, clinical
actions (tests, treatment regimes, referrals), temporal constraints and
if…then decision rules
• Translation to a formal model, e.g. PROforma, GLIF, Asbru
• Time-consuming, iterative, manual process as the guideline text tends to
assume background knowledge, is incomplete or contains ambiguity and
vague terms
7. Example CIG fragment (Asbru)
<plan name="Doxycycline : 100 mg orally twice a day for 7 days"
plan_id="plan52769441">
<cyclical_plan plan_id="plan5675512">
<frequency value="12" unit="hour"/>
</cyclical_plan>
<duration>
<min value="7" unit="day"/>
<max value="7" unit="day"/>
</duration>
</plan>
8. Examples of vague guideline statements
Underspecification:
• Avoid the use of highly intensive management strategies to achieve
an HbA1c level less than 6.5% (48 mmol/mol)
• Monitor HbA1c every 2–6 months (according to individual need) until it
is stable on unchanging treatment
Qualitative terms requiring mapping to numeric values or ranges:
• The moderate use of alcohol may increase HDL-cholesterol
• If blood pressure remains uncontrolled on adequate doses of three
drugs, consider adding a fourth and/or seeking expert advice
9. Information extraction for guideline formalization
• Helpful to automate
• Knowledge base construction: text to formal model translation
• Identification of opportunities for decision support: mapping
guideline concepts and rules to concepts in the EPR
• Measurement of guideline compliance
10. Information extraction approaches
• Bottom-up: identification of individual clinical terms, temporal
expressions, units of measure
• Look-up lists, regular expressions
• Shallow parsing to identify noun phrases
• Terminology services: UMLS, MetaMap
• Co-reference resolution: WordNet
• Top-down: identification of guideline structure: preamble, eligibility,
recommendations, ‘action’ sentences and rules
• Shallow parsing to identify verb phrases
• Ontologies for semantic relations, e.g. UMLS Semantic Network
• Use of linguistic guideline patterns (see later)
11. Mapping text to UMLS concepts - problems
• Identification of clinical terms is dependent on context:
- family history of congestive heart failure
- probable diagnosis of congestive heart failure
- no evidence of congestive heart failure
- patient does not have established cardiovascular disease
• Clearly just identifying the raw concepts congestive heart failure and
cardiovascular disease and mapping them to UMLS terms is
inadequate.
12. Mapping guideline text to UMLS concepts - problems
• Guideline documents are typically large (100 pages), in PDF or XML
format
• Requires guideline text to be segmented to enable efficient processing
- How best to segment the text that maximizes contextual clinical concept
identification?
13. Solutions: Text segmentation
• Customised phrase chunker to identify candidate terms:
- Noun phrases (NP), prepositional phrases (PP), verb phrases (VP)
- Neoclassical combining forms phrases (Token groups containing
Latin/Greek prefixes, roots, suffixes)
- Past-participle and gerund NPs:
- 'results in increased blood pressure', 'fasting blood glucose'
- List expansion:
- 'mild, moderate and severe hypertension → mild hypertension,
moderate hypertension and severe hypertension'
- 'lowering of heart rate and blood pressure → lowering of heart
rate and lowering of blood pressure'
- Abbreviation expansion: 'waist circumference (WC)'
14. Solutions: GATE-MetaMap Server integration plugin
- Extracts clinical concepts, in context, from large guideline texts in
multiple formats and encodings (PDF, XML, RTF, ASCII, UTF-8)
- Exchanges data/annotations with a MetaMap server
- Implements Unicode Normalization Forms for UTF-8 → ASCII
- Provides flexible text chunking options
- Optimises input data to MetaMap for mapping to UMLS concepts
- Integrates with other information extraction pipelines
16. Guideline patterns
Serban et al. (2007), examples:
(med_context, target_group, recommendation_operator, med_action)
In the event of [pregnancy]med_context, [patients with diabetes]target_group
[should]recommendation_op be[prescribed calcium channel blocker]med_action
(target_group, med_context, med_goal)
For [diabetic patients]target_group with [kidney damage]med_context the [blood
pressure target is130/80]med_goal