Title: Incorporating Data Mining Applications into Clinical Guidelines
1Incorporating Data Mining Applications into
Clinical Guidelines
- Presented by
- Reza Sherafat
- sherafr_at_mcmaster.ca
- March 28, 2006
- McMaster University
2Agenda
- Current trends in Healthcare
- Clinical decision support systems
- Data Interoperability problem
- Data mining results as the source of knowledge
- Knowledge interoperability
- Integration of data mining results with clinical
guidelines (plus some case studies) - Summary
- References
3Current trends in Healthcare
- The Healthcare professionals are overwhelmed with
information. - Preventable medical errors cause thousands of
deaths each year and loss of billions of dollars. - Healthcare information systems are deployed for
various purposes, telemedicine, patient care,
Electronic Health Records and decision support. - A good start by many standards organizations to
define and maintain healthcare standards. - HL-7 (most popular healthcare data standard) 4
4Clinical decision support systems (CDSS)
- Effectiveness of clinical decision support
systems (a question to be answered) - Even a recommendation system should NOT to flood
the practitioner with so many irrelevant cases
and also should NOT ignore possible important
cases. - Many different approaches to provide decision
making support. - We focus on guideline-based CDSS that try to
support clinical best practices at the point of
care decision making. - Arden Syntax by HL-7 5
- Guideline Interchange Format (GLIF) 3
5Arden Syntax 5
- The idea behind Arden Syntax is to have a simple,
yet powerful enough procedural language that can
encode the necessary logic for deciding upon a
single problem. - Decision making knowledge is encoded as IF-THEN
rules in separate Medical Logic Modules (MLM). - Each module is responsible for making a single
decision and is run on an engine that can access
the EHR systems. Based on the result of
evaluation of the rules an action (an alert or
reminder) is taken. - A library of modules
- Modules have data sections that should be mapped
to institution specific data repositories the
rest of the module is already ready to use.
6Arden Syntax (Contd)
- knowledge
- type data_driven
- data
- last_creat read last "Creatinine level"
- last_BUN read last "BUN level"
-
- evoke ct_contrast_order
- logic
- if last_creat is null and last_BUN is null then
- alert_text "No recent serum creatinine
available. Consider - patient's kidney function before ordering
contrast studies." - conclude true
- elseif last_creat gt 1.5 or last_BUN gt 30 then
- alert_text "Consider impaired kidney function
when - ordering contrast studies for this patient."
- conclude true
3
Source 9
7Guideline Interchange Format (GLIF) 3
- Three different types are models are mentioned
- Guideline Models
- Data models
- Data mining models
- A guidelines specification standard
- Flowchart-like diagrams (Guideline Models)
- 3 levels of abstraction 3
- Conceptual modeling
- Computable level
- Implementation details
8Conceptual Modeling (Level 1)
- The conceptual models have simple building
elements (steps) - action step,
- patient state step,
- decision step,
- branch step and
- synchronization step
- It is easy to build and understand models
- Some steps may involve user interaction,access
to a data source or triggering an event.
9Second and third levels in GLIF
- Computable level deals with encoding the decision
making logic (expressions) - Implementation level is concerned with how to map
and bind the variable to local (institution
specific) medical records.
10Data Interoperability
The semantics of the communication The semantics
convey the actual "meaning" of the message. The
semantics is conveyed via a set of symbols
contained within the communication. An external
"dictionary", thesaurus, or terminology explains
the meaning of the symbols as they occur.
A syntax for communication The syntax defines the
structure and layout of the communication. Common
syntax representations include ASN.1, XML, X.12,
HL7, IDL, etc.
Services to accomplish the communication Examples
include the post office, a telephone
switchboard, SMTP, FTP, Telnet, RPC, ORB
services, etc.
Source 7
11Data Interoperability (Contd)
- Three different types are models are mentioned
- Guideline Models
- Data models
- Data mining models
- THE KEY IDEA Through standardization
- HL-7 has built a standard Reference Information
Model (RIM) - RIM is in the form of a large class diagram that
model the healthcare domain. - Some other XML based standards like Clinical
Document Architecture (CDA) use RIM as their main
data model.
12Data Interoperability (Contd) RIM
Services
Stake holder
Organization
Person
Patient
ClinicalObservation
13Data mining results as the source of knowledge
- Three different types are models are mentioned
- Guideline Models
- Data models
- Data mining models
- Data mining research has been active in building
models that can describe or predict. - Applications of data mining studies
- Likelihood of coincidence of particular diseases
- Adverse drug usage
- Diagnosis
- Patient clustering based on risk factors
- Verification of known medical knowledge
14Knowledge interoperability
- THE KEY IDEA Through standards
- Use standards for knowledge sharing and exchange
- The mined knowledge should be incorporated into
the guideline model to be used for decision
making at the decision steps. - PMML (Predictive Markup Language) 6 data
mining knowledge is encoded using the PMML
standard. - GLIF3 Medical knowledge is encoded in guideline
models
15Framework for interoperability of mined knowledge
Source 2
16Framework for interoperability of mined knowledge
(Contd)
- Three phases
- Knowledge preparation
- Mining the patients data
- Interoperation
- To make both data and mined knowledge available
at the point of care through use of standard - Interpretation
- Access the knowledge base with the patient data
that needs decision making
17Integration of data mining results with clinical
guidelines
Guideline Execution
Guideline modeling
Knowledge Extraction
18Integration of data mining results with clinical
guidelines (Contd)
- Knowledge extraction
- Building data mining models on usually large
data warehouses - Guideline modeling
- Building guideline models
- PMML encoding
- Institution specific data bounding
- Guideline Execution
- Execution engine will follow the flow defined in
the guideline model - Accessing patient data from EMR systems
- Interact with the healthcare personnel
- Alert, recommend or remind
19Integration of data mining results with clinical
guidelines (Contd)
Source 2
20Implementation
- Extending Guideline Interchange Format3 (GLIF3)
constructs - To support the new functionality needed for the
data mining models - Guideline Execution Environment (GEE)
- To execute the guideline models, and
access/interpret the data mining models - Provision of the mined knowledge as webservices
when the knowledge base is not available locally - Very helpful for small devices e.g. handheld
computers
21Case study
- A decision tree classifier
- For melanoma skin cancer diagnosis 6
Source 2
22Case study (Contd)
Source 2
23Guideline Execution Environment
The guideline execution environment widget
Guideline selection list
Different flows within a guideline in execution
Source 1
24Guidelines Meta Model
Data mining decision nodes as an ontology class
Data mining decision nodes slots
Source 1
25Guideline modeling
Slot widget to specify the new attributes of
a data mining decision node
Source 1
26Summary
- We described
- A knowledge management framework to for data
mining results - The environment in which the framework can be
deployed - How to integrate data mining results in clinical
guidelines - How knowledge interoperability is achieved.
27References
- Incorporating Data Mining Applications into
Clinical Guidelines, R. Sherafat, K. Sartipi, The
19th IEEE International Symposium on
Computer-Based Medical Systems, 2006 - Data and Knowledge Interoperability in
Distributed Healthcare Systems, R. Sherafat, K.
Sartipi, The 13th Annual International Workshop
on Software Technology and Engineering Practice,
2005 - Guideline Interchange Format 3 (GLIF3),
www.glif.org - Health Level-7 (HL-7), www.hl7.org
- Arden Syntax, http//hl7.org/library/standards_non
1.htmArden20Syntax - Data Management Group (DMG), www.dmg.org
28References (Contd)
- Rules for melanoma skin cancer diagnosis,
http//www.phys.uni.torun.pl/publications/kmk/ - Version 3 Intermediate Tutorial - Working the HL7
Version 3 Methodology, George W. Beeler,
http//hl7.org/library/data-model/V3_Tutorials/V3_
Intermediate_May00.ppt - An Arden Syntax MLM for CT study with contrast in
patient with renal failure, http//cslxinfmtcs.csm
c.edu/hl7/arden/mlm/astm_ct_contrast.txt
29Questions