Medical knowledge representation within HEARTFAID platform

1
Medical knowledge representation within HEARTFAID
platform

• Dragan Gamberger
• Rudjer Boskovic Institute, Croatia

2
HEARTFAID project title

• A KNOWLEDGE BASED PLATFORM of services for
  supporting medical-clinical management of the
  heart failure within the elderly population
• knowledge representation and its usage (decision
  support) has the central role

3
HEARTFAID central problem

• How to present knowledge in a general and
  systematic way so that it can be used by
  different services (some that are not even known
  in advance)
• This is one of central issues in AI today

4
Content

• Requirements and related work in medical domains
• Our results
• - HF ontology
• - procedural knowledge
• - medical plans
• - reasoning
• Main achievements, future research topics

5
Requirements and related work

• Knowledge representation is really a difficult
  task !

6
Knowledge base requirments I

• Building the knowledge base consists of
• - collecting the relevant medical knowledge,
• - its systematization,
• - and technical formalization.
• Building it presents the challenge of
  transferring all aspects of relevant medical
  knowledge into the platform. The success in this
  work significantly determines the overall
  performance of the system.

7
Knowledge base requirements II

• On the other side
• DSS completely determines the requirements set on
  formal aspects of the knowledge representation
  task.

8
Knowledge base requirements

• It is essential to satisfy both requirements
• Present all relevant knowledge (systematically
  and in the user friendly form)
• 2) Enable effective reasoning on so represented
  knowledge
• Some systems are better in one aspect, others in
  the other, but it is necessary to satisfy both.
• This requirement is important for all
  applications (not only medical).

9
Existing approaches I

• Arden syntax (1993), part of HL7 standard, each
  rule is stored in a single file and is called a
  Medical Logic Module (MLM), it does not refer to
  any kind of domain description (ontology), the
  system needs to interact directly with a clinical
  database, execution components are not freely
  available, it has found a practical usage in real
  clinical environments. (good for reasoning but
  expressiveness very low)
• Asbru (1998) is a guideline modeling tool for
  medical plans, highly aware of the time
  dimension, each plan is decomposed into more
  sub-plans performed sequentially, concurrent
  (parallel execution) or cyclical, good for
  situations where actions are taken in a
  predefined order, no freely available execution
  engines. (appropriate for time related and order
  related knowledge)

10
Existing approaches II

• GLIF (2000) provides a framework for developing
  medical guidelines in the form of a flowchart,
  suitable for describing logic sequence of
  actions, appropriate for presenting sequence of
  tests for diagnosing disease or prescribing
  therapy. (guideline modeling tool without
  execution engine GLEE?)
• PROforma (2003) is a knowledge composition
  language for workflow management, uses
  first-order logic, represents guidelines as a
  directed graph, major focus point is on guideline
  safety. (complex commercial system with specific
  knowledge representation form)
• - for all of them integration with real patient
  data is a serious problem!

11
Heartfaid approach

• Already in the project definition it has been
  determined
• Ontological form of knowledge representation is
  the most appropriate for the concrete task.
• In the early project phase it has been decided
• semantic web OWL ontology form with integrated
  rules in the SWRL form is the optimal solution.
• Why ?
• Semantic web is modern and promising methodology
  for knowledge representation
• we do not need explicit time component
• can be effectively handled by available open
  source interpreters and reasoning systems

12
Our results

• HF ontology - procedural knowledge - medical
  plans - reasoning

13
HF ontology

• In its current form the ontology presents the
  detailed taxonomic overview of the HF domain with
  around 200 classes describing HF related
  concepts.
• Examples are "Cardiac_hypertrophy",
  "Blood_pressure_signs" or "Heart_murmurs". These
  concepts are interconnected with super-class and
  sub-class properties into a hierarchical
  tree-like structure.
• When possible, classes are specified with their
  CUI number (Concept Unique Identifier according
  to UMLS) and with a list of synonyms. For
  example, for the class "Heart_diseases" its CUI
  is C0018799 and its synonyms are
  "Disorder_of_heart", "Cardiac_diseases",
  "Cardiopathy".

14
HF ontology
Basic elements are classes HF_concept,
Patient_characteristics, Testing,
Treatment
Protégé tool used to display a part of the HF
ontology
15
HF ontology

• Individuals or instances are members of the
  classes and typically present exhaustive list of
  concrete concepts relevant for the class. The
  ontology includes more than 2000 individuals.
• For example, the "Cardiac_hypertrophy" class has
  following six instances "Cardiomegaly",
  "Combined ventricular hypertrophy",
  "Left_atrial_hypertrophy", "Left_ventricular-hyper
  trophy", "Right_atrial_hypertrophy", and
  "Right_ventricular_hypertrophy".

16
HF ontology
Patient_characteristics class hierarchy with some
instances
17
HF ontology

• The ontology contains properties that connect
  individuals in different classes. These
  properties are relevant because they enable
  introduction of relations among concepts. The HF
  ontology includes definitions of more than 100
  properties.
• For example, individual "Valvular_heart_disease"
  from the class "Heart_valve_diseases" is
  indicated by the individual "Dyspnea" from the
  class of "Signs_and_symptoms".

18
HF ontology
19
Procedural knowledge

• The HF ontology presents the detailed taxonomic
  overview of complete heart failure domain
  including relevant relations among concepts. It
  represents descriptive knowledge about the
  domain.
• Platform should also be able to perform some
  actions, typically in the form of suggestions for
  patients and medical personnel. The knowledge
  representing sufficient and necessary conditions
  that some actions can be done is the so called
  procedural knowledge.
• Descriptive and procedural knowledge together
  present the knowledge base of the HF platform.

20
Procedural knowledge
An example Diagnosis Systolic heart
failure IF Patient has either heart failure signs
or heart failure symptoms AND ECG abnormal (left
bundle branch block AND anterior Q
waves) AND patient has (ischemic heart
disease) AND Chest X-ray abnormal (cardiothoracic
ratio gt 0.5) AND natriuretic peptides abnormal
(BNP gt 100 pg/ml))
21
Procedural knowledge

• For the integration of descriptive and procedural
  knowledge it is important that production rules
  can use only the concepts defined in the HF
  ontology.
• Rules are formal enough to present knowledge and
  at the same time medical experts can easily
  control the expected performance of the platform

22
Procedural knowledge

• HF procedural knowledge has been divided into 10
  functional subtasks in order to enable easier
  human control of the completeness and consistency
  conditions
• HF diagnosis
• Alternative or additional diagnosis
• Heart failure severity assessment
• HF general treatment process based on severity
  assessment
• HF medications contraindications, adverse effects
  additional treatment rules
• Prognosis estimation for HF patients
• Non-pharmacological management and
  recommendations
• Specific medication prescription and dosage
• Acute decompensation of congestive heart failure
• Heart failure cause and CAD risk factors

23
Procedural knowledge integrated into HF ontology
24
Medical plans

• Medical plans are textual and visual
  presentations of procedures that take place after
  detection of some events.
• Events can be any type of health disorder
  including signs, symptoms, and diagnosis.
• The main characteristic of medical plans is that
  they are event driven and because of that they
  present actionable view of the medical knowledge.

25
Medical plans
Medical plan for heart failure patients with
increased body temperature.
26
Medical plans

• Heart attack 5
• Cerebral stroke 5
• Ankle edema 3
• Dizziness 2
• Coughing 1
• Dyspnea 3
• Paroxysmal nocturnal dyspnea 3
• Orthopnea 3
• Oliguria 2
• Involuntary weight gain 3
• ......
• HF system has
• 38 interconnected plans for signs, symptoms and
  diagnosis assessment and treatment
• 15 plans for medications prescription and dosage

urgency levels
27
Medical plans
Plan 23 Pulmonary edema Severity level
5 Information for this plan has been obtained
from the Acute Heart Failure Guidelines and from
Congestive Heart failure Guidelines, 2005 Assume
congestive heart failure - verify by X-ray - it
is severe left heart backward failure, presenting
with shortness of breath, dry cough (sometimes
frothy sputum), pallor or even cyanosis, cold and
clammy skin, normal or elevated blood pressure,
systolic function is often preserved, also more
than 50 of patients have LVEFgt45, severe
respiratory distress, lung crackles (rales) and
orthopnea, O2 saturation less than 90 - can be
caused by long-term hypertension, especially in
older women (65) - possible causes are
myocardial ischemia or infarction, aortic or
mitral valve disease, severe arrhythmias, left
heart tumour, severe hypertension, anaemia,
thyrotoxicosis, brain tumour or trauma -
treatment order is O2, then CPAP or NIPPV, then
if necessary mechanical intubation, intravenous
administration of antihypertensive agents
(vasodilators(nitrates), nitrates are more
effective than diuretics, however diuretics can
also be used, do not use beta-blockers) -
consider coronary angiography if patient does not
respond to treatment - consider ultrafiltration
for temporary relief End plan 23
28
Medical plans

• The syntax of the medical plans highly resembles
  the traditional workflow management. The
  difference is that the medical plans will not be
  executed by machines they are written in an
  almost free graph/text form with main purpose to
  be fully understandable by humans. Their main
  characteristic is that they are event driven and
  their main advantage is a clear systematization
  of the medical procedures and interconnections
  among them.
• In the Heartfaid platform medical plans are only
  a middle step between experts and the guideline
  modelling tools which persuade the experts to
  clearly state the procedure they would normally
  perform when facing a specific problem.

29
Lessons learnt

• Semantic web is great for knowledge
  representation excellent choice
• OWL SWRL not a good combination (open world
  assumption!)
• Integration of procedural knowledge into OWL
  successful combination

30
Open versus closed world

• Patient X receives ACE inhibitors. But there is
  no information about diuretics.
• Open world we cannot conclude anything about
  diuretics
• Closed world we assume that X does not receive
  diuretics !!

31
Open versus closed world

• Open world we can not conclude X does not
  receive diuretics
• Closed world we assume that X does not receive
  diuretics !!

fine for web applications
medical applications need this the problem is
that reasoning process is different
32
Conceptualization of procedural knowledge

COSI closed world OWL interpreter
33
Descriptive ontology
individual "Valvular_heart_disease" from the
class "Heart_valve_diseases" is indicated by
the individual "Dyspnea" from the class of
"Signs_and_symptoms".
34
Procedural knowledge conceptualization
35
COSI Closed world OWL interpreter
36
COSI Closed world OWL interpreter
37
Main achievements and future work

• Integration of descriptive and procedural
  knowledge !

38
Achievements

• We have developed HF ontology
• - it is public on http//lis.irb.hr/hear
  tfaid/ontology/
• the ontology has been requested by Stanford
  University Medical Center for research in the
  study on risk factors for congestive heart
  failure
• it means a long term result valuable also outside
  Heartfaid project

39
Achievements

• We have developed procedural HF related knowledge
  (10 sets of rules)
• Novelty of this project we have integrated
  procedural knowledge into the ontological form !
• Brand new we have developed COSI (Closed world
  OWL syntax interpreter) for decision support
  purposes which is now used for the platform DSS !
• - we can use the same methodology on any domain

40
Achievements

• We have developed a complete set of medical plans
  for HF domain.
• - it is public on http//lis.irb.hr/hear
  tfaid/plans/
• they are not used for reasoning
• topic of the future work

41
Future work

• For the Heartfaid project
• Improvement of the HF knowledge base (iteratively
  by doing experiments with real patient data)
• Improvement of the COSI reasoning speed
• Long term scientific interests
• Application of the developed HF knowledge base
  for other medical applications
• Application and testing of the developed
  knowledge representation and reasoning
  methodology on other domains
• Experiments with medical plans and their direct
  usage for decision support purposes

42
Laboratory for Information Systems RBI Zagreb
Croatia

• team who prepared the HF knowledge base (summer
  2007.)

• 3 weeks ago

43
Thank you for your attention!

• Questions?