ADONIS: Automated Diagnosis System based on Sound and Precise Logical Descriptions

1
ADONIS Automated Diagnosis System based on Sound
and Precise Logical Descriptions

• Alejandro Rodríguez1, Jose Emilio Labra2,
• Giner Alor3, Juan Miguel Gómez1
• 1Universidad Carlos III, Madrid, Spain
• 2Universidad de Oviedo, Spain
• 3Instituto Tecnológico de Orizaba, Mexico

2
Motivation

• Initial Goal Build a system for students of
  medicine to aid diagnosis and self-study
• Develop a Support System which performs
  diagnostics
• In practice, it should be accompanied by the
  judgment of a medical professional
• Calculates probability of the diagnosis as a
  function of a number of parameters
• Symptoms, Laboratory Tests, Age, Sex, Visited
  Countries, etc.

3
ADONIS System

• A medical diagnosis system based on logical
  inference and probabilistic techniques
• Developed on top of an architecture named ODDX
• ODDX is based on Differential Diagnosis (DDX)
• Two versions Desktop and Online (Web)
• http//www.oddx.es

4
ODDx Features

• Based on semantic web technologies
• Ontologies and Reasoners
• Enables the integration with other systems
• Example Mashup which locates specialists
• User centered design
• Usability tests
• Multilingual (english/spanish/)
• Use of standards
• More clinical parameters
• Takes into account the Interaction of drugs a
  patient may have taken

5
Semantic Web

• Resources (URIs)
• Non Unique Name Assumption
• Two given URIs may refer to the same thing
• Open World Assumption (OWA)
• Decentralized
• Large datasets

6
Semantic Web Technologies

• RDF Graph based model
• Composability of different graphs
• Integration between heterogeneous sources
• Ontologies
• OWL (Web Ontology Language)
• Based on Description Logics (DLs)
• Tradeoff between expressivity vs efficiency
• Pellet (OWL Reasoner)
• Rules
• Combination between Description Logics and Rules
• Jena (RDF Library) supports Rules

7
ODDx - Desktop
8
ODDx - Desktop
Results
9
ODDx - Web
10
ODDx - Base Ontology

• Ontology for Medical Diagnosis
• Diseases, Symptoms, Laboratory Tests, etc.
• Implemented in OWL (Web Ontology Language)
• Development
• Meetings with medical students and professionals
• Medical literature World Health Association
  Disease Classification and Coding Systems
• ICD 10 International Statistical Classification
  of Disease and Related Health Problems (10th
  Revision, 2007)
• Chapter 1, Blocks A00 B99
• Chapter 1 contains blocks such as
• Intestinal infectious diseases (A00 A09)
• A00.0 Cholera due to Vibrio cholerae 01, biovar
  cholerae
• A00.1 Cholera due to Vibrio cholerae 01, biovar
  eltor
• A00.9, Cholera, unspecified
• Tuberculosis (A15 A19)
• Certain zoonotic bacterial diseases (A20 A28),
  etc.

11
ODDx - Base Ontology
12
Inference - 1st Approach

• ODDx was implemented in Java
• Uses JENA Library (Rule based reasoner)
• Supports a form of Negation as Failure using the
  NoValue builtin
• Example
• For each symptom
• rule_DIS_X_Sym_C
• (?i onthasDiagnosis ontDiseaseX) lt-
• (?i onthasSymptom ontSym_C)
• noValue(?i, hasNegSymptom,ontDIS_X_NOT_SYM)
  
• rule_DIS_X_NOT_REST_SYMPTOMS
• (?i onthasSymptom ?x)
• notEqual(?x, ontSYM_C)
• notEqual(?x, ontSYM_D)
• notEqual(?x, ontSYM_E) -gt (?i
  onthasNegSymptom ontDIS_X_NOT_SYM)
• NOTE Efficiency but non declarative semantics

13
Inference - 2nd Approach

• Using OWL (Description Logics)
• Pellet Reasoner SPARQL Queries
• Challenges
• Open World Assumption
• Composed Diseases

14
Open World Assumption

• Problem
• Given the knowledge
• patient1 hasSymptom Sym_A.
• patient1 hasSymptom Sym_B.
• what should be the answer to
• patient1 hasSymptom Sym_C ?
• Possible answers
• No (CWA), ?? (OWA)
• When we need the negative answer
• patient1 a a Restriction
• owlonProperty
  hasSymptom
• owlcardinality 2 .
• Sym_A owldifferentFrom Sym_B .
  

15
Composed Diseases
Sym_C
Sym_A
Sym_D
Sym_B
Disease X
DiseaseY
Sym_E
Disease X
SymptomsY owlequivalentClass a
owlClass owlunionOf ( a owlClass
owloneOf (symptomA
symptomB) SymptomsX ) .
SymptomsX owlequivalentClass a
owlClass owloneOf (symptomC
symptomD
symptomE ) .
16
Composed Diseases

• In MEDBOLI it was possible to infer
• Sym_C, Sym_D, Sym_E ? DiseaseX
• Sym_A, Sym_B, DiseaseX ? DiseaseY
• but with
• Sym_A, Sym_B, Sym_C, Sym_D, Sym_E
• the system did not infer DiseaseY

17
Solution in Description Logics

• HasDiagnosisX ? hasSymptom SymptomsX ?
• ? hasSymptom SymptomsX ?
• hasSymptom 3
• HasDiagnosisY ? hasSymptom SymptomsY ?
• ? hasSymptom SymptomsY ?
• hasSymptom 5

18
Evaluation

• Description Logics times were very high (gt 8
  hours for very simple tests)
• Using Rules (Jena) ? 10s per inference
• Another approach procedurally computing the
  uniques diseases that can be formed

19
Conclusions Future Work

• Applying Description logics to Medical Diagnosis
  offers a new challenge for reasoners
• Efficiency vs Expressivity
• In the mean time
• We found a new solution using OWL 2 constructs
• Qualified cardinality
• Property chains
• Future work
• Systematic evaluation of performance
• Other formalisms Probabilistic Description
  Logics

20
End of Presentation