Cui Tao, PhD - Semantic-WEb Based Annotation

8/7/2019 Cui Tao, PhD - Semantic-WEb Based Annotation

1/55

Semantic-Web based Annotationand Reasoning for Temporal

Information in Clinical Narratives

Cui Tao, PhD

Division of Biomedical Statistics and InformaticsDepartment of Health Sciences Research

Mayo Clinic, Rochester, MN


2/55

Acknowledgement

Christopher Chute, MD,DrPH

Kim Clark Piet de Groen, MD

Guergana Savova, PhD

Deepak Sharma

Harold Solbrig

This research issupported by theNational ScienceFoundation underGrant #0937060 tothe ComputingResearch

Association for theCIFellows Project


3/55

Introduction

Time is essential in clinical research

Uncover temporal pattern

Disease level Patient level

Explain past events

Predict future events


4/55

Introduction

Vast amount of data

Data embeded in narratives

Many temporal relations are notexplicitlystated in the clinical narratives,but rather needs to be inferred


5/55

Temporal Relation Reasoning

(Example) Patients INR value is below normal(Event 1)

today. (note date: 01/26/07)

He has had the chills

and body

aches (

Event 2)before the abnormal test. (Event 3) (note date:

01/26/07)

On Jan. 30, 2007, patient started Coumadindosing plan of 1.0 mg(Event 4).(note date:02/09/07)

Question: did the patient experience bodyachesbefore he started the Coumadin dosing plan?(was Event 2 before Event 4?)


6/55


(Example) Patients INR value is below normal

(Event 1) today. (note date:01/26/07)

He has had the chills and bodyaches (Event 2) before the abnormaltest. (Event 3) (note date: 01/26/07)

On Jan. 30, 2007, patient started

Coumadin dosing plan of 1.0 mg(Event 4).(note date: 02/09/07)

Question: did the patient experiencebodyaches before he started theCoumadin dosing plan?

Event1 = Event3


7/55








Event1 = Event3

+

Event2 before Event3



8/55








Event1 = Event3

+



Event1 01/26/07 +

Event4 01/30/07



9/55








Event1 = Event3

+



Event1 01/26/07 +

Event4 01/30/07




10/55

Introduction

The Semantic Web provides a suitableenvironment for temporal data

representation and reasoning: Standard mechanism with explicitand formal semantic definition

OWL DL

SWRL

Reasoning tools, querying andstorage mechanisms


11/55

System Overview

CNTRO


12/55

CNTRO Ontology


13/55

Clinical Narrative Temporal

Relation Ontology (CNTRO) Event

Time

Time Instant

Time Interval

Time Phase

Time Period

Temporal Relation


14/55

CNTRO Overview


15/55

Temporal Information

Representation (Instant) Example Sentence:

the second cycle ofchemotherapy was onJune 10, 2004


16/55


Representation (Interval) Example Sentence:

monitor patients heartrate for 72 hoursstarting from today(note date:2004-06-01)


17/55


Representation (Period & Phase) Example Sentence:

take antibiotics every8 hours for 10 daysstarting from today(note date:2004-06-01)


18/55


Representation (Uncertainty) Example Sentence:

see the patient backin approximately twoweeks prior to histhird cycle ofchemotherapy (notedate:2004-06-10)


19/55


Representation (Statement) Example Sentence:

patients bilirubin iselevated 2 weeksafter the second cycleof chemotherapy


20/55

CNTRO Evaluation

153 sentences with temporal information

178 events

98 time instants 10 time intervals

53 time phases

170 temporal relations 142/153 sentences were represented

without losing any time-relatedinformation


21/55

CNTRO 1.0

provides a formal semantic foundationto leverage the semantic-web

techniques still necessary to arrive at a shared set

of semantics and operational rules withcommonly used ontologies for the timedomain


22/55

CNTRO 2.0: A HarmonizedSemantic Web Ontology for

Temporal RelationInferencing in Clinical Narratives


23/55

Semantic Alignments

Initial Alignments

Time Ontology

SWRL Temporal Ontology Basic Formal Ontology

Measurement Units Ontology


24/55

Entity Definition

Time-related classes harmonization


25/55

Entity Definition

Event Class: any sort of occurrence, state,perception, procedure, symptom or situationthat occurs on a time line in clinical narratives

Subclasses:

Process (BFO)

Fiat_process_part (BFO)

Process_boundary (BFO) Process_observable

Process_repeat


26/55

Entity Definition

Process (BFO)

a clinical event that associates with a timeinterval

patients chemotherapy started from Jan.10 and ended on Mar. 19

Process_boundary (BFO)

Clinical events that associate with a timeinstant

the patient checked in at 3pm


27/55

Entity Definition

Fiat_process_part (BFO)

events with unknown or unspecified starting andending time

patient experienced headache during last week

Process_observable

observable states, conditions, or qualities

Process_repeat

clinical events recur periodically

take antibiotics every 8 hours for 10 days startingfrom today


28/55

Temporal Relation Harmonization

Allens time operators: between timeintervals

SWRL Temporal Ontology: betweentime intervals and instants

Time Ontology: between intervals,

additional general properties for instants CNTRO: between two events, or event

and time


29/55

Temporal Relation Harmonization

Define CNTRO properties using SWRLtemporal ontology properties:

Leverage the SWRL Temporal Built-Ins library to

perform temporal operations


30/55


31/55

Unit

Mapped the Measurement Units Ontology

muo:QualityValue muo:UnitOfMeasurementmuo:measuredIn

xsd:float

rdf:value

cntro:Duration cntro:Unit


32/55

Semantic Annotation


33/55

Semantic Annotation Tool

Knowtator extension

RDF exporter

Connects to the temporal reasoningframework API

Currently a manual process

Will connect with Mayos cTAKESpipeline


34/55


35/55


36/55

Temporal Reasoning Framework


37/55


Temporal Representation Normalization

OWL DL Reasoning

SWRL-based Reasoning


38/55

Temporal Representation

Normalization Normalize to xsd DateTime Data Type format

BYUOntos information extraction technique

Mayo Clinic cTAKES (clinical Text Analysisand Knowledge Extraction System)

Jan 23rd, 01/23/2010, 1/23/10, 1-23-10,etc

Today, yesterday, in 2 days, 2 hours ago, etc


39/55

Logical Characteristics of Properties

Transitive properties: i.e., before, after

Inverse properties, beforeafter

Symmetric properties, i.e., equal


40/55

Operators for Offsets

inverse operators and

Temporal relation between e1 and e4?

(3 days) (2days) = (1day)


41/55

Restriction Assertions

a cancer chemotherapy must happen after a cancer

diagnosis based on clinical evidence


42/55

Semantic Definition of Concepts With

OWL DL


43/55

SWRL Temporal Built-In Library

Works with Protg SWRL Tag

Implemented Allens temporal operators

Serves as basic functional blocks in ourtemporal relation reasoning framework


44/55

SWRL RuleML

Rule-based Definition forConsistency


45/55

SWRL RuleML

Rule-based Definition for properties


46/55

SWRL RuleML

Rule-based Definition for concepts

premature labor after 22 weeks but before 37

completed weeks of gestation without delivery


47/55

Implementation Status

findEvent(searchText)

returns a list of events that match thesearching criteria. Currently we look forevents based on text search.

GetEventFeature(event, featureflag)

returns a specific time feature for a givenevent.

Sample query: When was the patient diagnosed with

diabetes?

When did the patient start hischemotherapy?


48/55


getDurantionBetweenEvents(event1, event2)

returns the time interval between two events.

Sample query: How long after the patient was

diagnosed colon cancer did he start thechemotherapy?

getDuration(event)

returns the duration of a given event.

Sample query: How long did the symptoms ofrectal bleeding last?


49/55


getTemporalRelationType(event1, event2)

returns the temporal relations between twoevents if it can be retrieved or inferred.

Sample query: Was the CT scan after thecolonoscopy?

getTemporalRelationType(event1, time)

returns the temporal relations between an

event and a specific time if it can be inferredor retrieved.

Sample query: Is there any behavior changewithin a week of the test?


50/55


sortEventsByTemporalRelationsOrTimeline(events)

returns the order(timeline) of a set of events.

sample query:

What is the tumor status timeline asindicated in the patients radiology note?

What is the treatment timeline as recordedin oncology notes?

When was the first colonoscopy done? When was the most recent glucose test?


51/55

Pilot Study

Late stent thrombosis adverse event

Complaint files from Manufacturer and

User Facility Device Experience(MAUDE) database

Detect potential temporal patterns within

complaint files of similar adverse events


52/55

Pilot Study

MAUDE Database

Confirmed Late Stent

Thrombosis Reports

15 Reports

CNTRO Reasoning

System-generated

answersCompare

Human Expert

Human-generated

answers (gold

standard)

Temporal QuestionsTemporal Questions


53/55

Pilot Study

Evaluation

System answered 65 out of 73

questions correctly Accuracy: 89.04%

Errors caused by ambiguities in

annotations


54/55

Summary

A more robust and semantically crisp model

A semi-automatic annotation system

A temporal relation reasoning framework

A user friendly querying system

A GUI where users can view and query

information of interested in its original context

CNTRO.ORG


55/55

Documents

Cui Tao, PhD - Semantic-WEb Based Annotation