Temporal Data Management: Semantic Web Engineering

Discussion Leader: Cui Tao

Assistant Professor in Medical InformaticsMayo Clinic

Temporal Reasoning Journal Club

December 1, 2011

Articles to Discuss

• Time-Oriented Question Answering from Clinical Narratives Using Semantic-Web Techniques. Tao C, Solbrig HR, Deepak S, Wei W-Q, Savova G, Chute CG. International Semantic Web Conference, Lecture Note of Computer Science. 2010; 6497:241-56. http://www.springerlink.com/content/67623p256743wv4u/

• Representing Complex Temporal Phenomena for the Semantic Web and Natural Language. Feng Pan, 2007 PhD thesis University of Southern California http://www.isi.edu/~hobbs/time/pub/pan-phdthesis.pdf

Why Semantic Web?

The Semantic Web provides a suitable environment for temporal data representation and reasoning:• Standard mechanism with explicit

and formal semantic definition• OWL DL• SWRL • Reasoning tools, querying and

storage mechanisms

Ontologies Available Online

• Clinical Narrative Temporal Relation Ontology (CNTRO): http://cntro.org/index.html

• Time-OWL: http://www.w3.org/TR/owl-time/

Goals

CNTRO• For representing events,

time information, and temporal relations in clinical narratives

• Can also represent structured data

• Presumably can be generalizable

Time-OWL• For describing the

temporal content of Web pages and the temporal properties of Web services

• Can also represent structured data

Classes

CNTRO• ValidTime

• TimeInstant• TimeInterval• TimePeriod• TimePhase

• Duration

• DurationUnit

• Granularity

• Event

Time-OWL• TemporalEntity

• Instant• Interval

• ProperInterval• DateTimeInterval

• DurationDescription

• DateTimeDescription

• TemporalUnit

• DayOfWeek

Properties (Time-OWL)

Temporal Relations

Duration Description

DateTimeDescription

Properties (CNTRO)Property Domain Range

After Event Event/ValidTime

Before Event Event/ValidTime

During Event Event/ValidTime

Equal Event Event/ValidTime

Contain Event Event/ValidTime

Finish Event Event/ValidTime

Start Event Event/ValidTime

hasDuration TimeInterval Duration

hasDurationUnit Duration DurationUnit

hasDurationValue Duration decimal

hasStartTime TimeInterval TimeInstant

hasEndTime TimeInterval TimeInstant

hasValidTime Event ValidTime

hasPeriod TimePhase TimePeriod

hasModality ValidTime boolean

hasNormalizedTime TimeInstant dateTime

hasOrigTime TimeInstant string

CNTRO Overview

Temporal Relations

Both adapted Allen's interval algebra

After Event Event/ValidTime

Before Event Event/ValidTime

During Event Event/ValidTime

Equal Event Event/ValidTime

Contain Event Event/ValidTime

Finish Event Event/ValidTime

Start Event Event/ValidTime

Time_OWL

CNTRO

Duration

Time_OWL

CNTRO

Time Description

Time_OWL

CNTRO

Time-OWLFeatures

• Relations for intervals/instants

• Time Zones

• Day of Week, Day of Year

• Specific definition of months, weekdays, etc

• Temporal sequence

Time-OWL

Time-OWL

Time-OWL

Every other week on Monday, Wednesday and Friday until December 24, 1997, but starting on Tuesday, September 2, 1997.

EveryOtherWeek:

hasTemporalUnit = unitWeek

hasGap = 2

MWFEveryOtherWeek:

hasStart = 09/02/1997

hasEnd = 12/24/2007

hasithTemporalUnit =1,3,5

hasTemporalUnit=unitDay

hasContextTemporalUnit=unitWeek

CNTRO Features

• Periodic Time Interval

• Relation between Two Events

• Time Offset

• Relative Time

• Uncertainty

CNTRORepresentation (Period & Phase)

• Example Sentence: take antibiotics every 8 hours for 10 days starting from today (note date:2004-06-01)

Operators for Offsets

Was e1 before e4?inverse operatorsαand ββ(3 days) α(2days) = β (1day)

SWRL RuleML

• Rule-based Definition for Consistency

SWRL RuleML

• Rule-based Definition for properties

SWRL RuleML

• Rule-based Definition for concepts

“premature labor after 22 weeks but before 37 completed weeks of gestation without delivery”

CNTRO API

• findEvent(searchText) • returns a list of events that match the

searching criteria. Currently we look for events based on text search.

• GetEventFeature(event, featureflag)• returns a specific time feature for a given

event.• Sample query:

• When was the patient diagnosed with diabetes?

• When was the patient started his chemotherapy?

CNTRO API

• getDurantionBetweenEvents(event1, event2) • returns the time interval between two events.• Sample query: How long after the patient was

diagnosed colon cancer did he start the chemotherapy?

• getDuration(event) • returns the duration of a given event. • Sample query: How long did the symptoms of

rectal bleeding last?

CNTRO API

• getTemporalRelationType(event1, event2) • returns the temporal relations between two

events if it can be retrieved or inferred. • Sample query: Was the PT scan after the

colonoscopy?

• getTemporalRelationType(event1, time) • returns the temporal relations between an

event and a specific time if it can be inferred or retrieved.

• Sample query: Is there any behavior change within a week of the test?

CNTRO API

• sortEventsByTemporalRelationsOrTimeline(events) • returns the order (timeline) of a set of events. • sample query:

• What is the tumor status timeline as indicated in the patient’s radiology note?

• What is the treatment timeline as recorded in oncology notes?

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

DiscussionSemantic harmonization

Time-related classes harmonization

Relation harmonization

DiscussionInstances vs. Intervals

• Time instant: a time interval with a very short duration

• Time interval: a time instant on a coarse level of granularity

• “patient’s last cycle of chemotherapy was on Jan. 19”• Process or occurrence?• Interval or instant?

DiscussionTemporal Uncertainties

• Approximated temporal expression:• In approximately 2 weeks• About 3 hours• In the AM• Late last year

• Insufficient level of granularity:• Duration between Jan. and Jun.?• Event A: Jan. and Event B: Jan. 16;

what was the temporal relation?

DiscussionTemporal Relation on Granularity

• “PT WENT INTO CARDIAC ARREST AND THEY WERE UNABLE TO KEEP HEART BEATING FOR MORE THAN A COUPLE HOURS. PT PASSED AWAY THAT NIGHT.”

• Cardiac arrest before death (granularity: hour)

• Cardiac arrest equal death (granularity: day)

DiscussionVague Event Duration

• Missing durations is one of the most common sources of incomplete information for temporal reasoning in natural language applications

• Empirical approach?• Long, short• Coarse-grained duration information

• Set up ranges of durations

DiscussionTemporal Uncertainties

Coarse temporal notion:• Early next week, middle of last year,

short after 11PM, before breakfast• Short after 11:30PM on the 16th,

before or on the 17th?

DiscussionTemporal Uncertainties

Ambiguities:

Last cycle of chemotherapy was on Jan. 16.

1) patient’s last cycle of chemotherapy STARTED on Jan. 19;

2) patient’s last cycle of chemotherapy ENDED on Jan. 19;

3) patient’s last cycle of chemotherapy STARTED and ENDED on Jan. 19.

Questions