Transcript Temporal Reasoning and Planning in Biomedicine Maintenance

Temporal Databases and
Maintenance of
Time-Oriented Clinical Data
Yuval Shahar, M.D., Ph.D.
A Clinical Scenario
Ms. Jones was seen in the diabetes clinic on January 14 1997 at 11 A.M. Her
blood-glucose value at that time was measured in the clinic as 220 mg/100ml.
She complained of vomiting and dizziness for the past 2 or 3 days.
She was eventually hospitalized on the same day. A more accurate bloodglucose test that was taken at the same time as the one performed in the clinic
returned from the laboratory on January 15 1997, with a value of 380
mg/100ml.
Ms. Jones was discharged on January 17 1997, and was seen again in the clinic
on January 24 1997. At that time, several renal-function serum and urine tests
were performed in addition to measuring blood-glucose values. A complete
neurological assessment was carried out as well.
Uses of Clinical Data
• Clinical decision making
– monitoring
– diagnosis
– therapy
• Clinical research
• Administration and other tasks
– Quality assessment
– Billing
• Legal records
Clinical Database Features
• Clinical data is time oriented
– different temporal aspects, such as when was the
data valid, versus when was the data recorded
• Often, there is inherent uncertainty regarding
the time, value, or both aspects of the data
• Data are often incorrect, incomplete, or
inconsistent
• Might require a specialized database
management system (DBMS)
Data Quality Issues
• Correctness
– Validation during data entry
– Validation by global data analysis
• Completeness
– Missing observations
• Possible bias due to hidden contexts
• Possible completion from neighboring values
• Possible completion from related data types
• Consistency
– Consistent semantics over patients and time
A Temporal Query
• “Determine if the oncology patient (currently
under therapy by a chemotherapy protocol) had
within the past 6 months at least two
episodes that lasted for more than 3 weeks, of
Grade II bone-marrow toxicity (due to a
specific chemotherapy drug)”
• Responding to such queries is necessary to
support clinical management, such as when
using a clinical guideline
The Time-Oriented Database (TOD)
• Developed at Stanford during the 1970s
• A cubic, three-dimensional structure
– patients X visits X clinical parameters --> value
• Microcomputer version: MEDLOG
• Two file structures:
– One indexed by patients, for individual information
– One indexed by parameter type, for statistical analysis
The ARAMIS Database
• The American Rheumatism Association
Medical Information System (ARAMIS)
• Developed at Stanford during the 1970s and
maintained since that time in multiple sites
• Contains longitudinal data concerning
multiple patients who have rheumatic
diseases or arthritis
• Originally used TOD, then MEDLOG and
other tools for analysis of chronic diseases
Types of Temporal Dimensions
(Snodgrass and Ahn, 1986)
• Transaction time: The time in which (or during which) data
are stored in the database (e.g., in which “the patient has
mild anemia” was recorded)
• Valid time: The time during which the data were true (e.g.,
the period during which the patient did, in fact, have mild
anemia)
• User-defined time: A time stamp or interval that is specific
to the application (e.g., the time in which the anemia level
was determined in the laboratory)
=> Transaction time and valid time define the database type
Database Types, A Temporal View
• Snapshot databases: Have no time aspect (flat records)
• Rollback databases: Have only transaction time (e.g., a
series of time-stamped updates to the patient’s current
address and phone number)
• Historical databases: Have only valid time (e.g., a series
of updates of the patient’s state of anemia during
January 1997, deleting previous values that refer to that
time period, keeping only the latest updates)
• Bitemporal databases: Have both transaction time and
valid time (e.g., on February 12 1997, it was recorded
that, during January 1997, the patient had mild anemia)
A Tale of Two Data Types
Parameter name
(event name)
Blood-glucose level
Insulin administration
Blood-glucose level
Insulin administration
Blood-glucose level
Value/unit
(attribute:value)
120 gr/100cc
dose:2 units
130 gr/100cc
dose:3 units
190 gr/100cc
Valid
start time
1/17/96:9:00a.m
1/17/96:8:00a.m.
1/18/96:9:00a.m.
1/18/96:8:00a.m.
1/18/96:9:00a.m.
Valid
stop time
1/17/96:9:00a.m
1/17/96:8:02a.m.
1/18/96:9:00a.m.
1/18/96:8:01a.m.
1/18/96:9:00a.m.
Transaction
time
1/17/96:10:00a.m.
1/17/96:10:05a.m.
1/18/96:6:00p.m.
1/18/96:7:00p.m.
1/20/96:8:00a.m.
Time and Uncertainty
• There is often uncertainty as to when the clinical
episode started or ended, and what its duration
was
• One way of representing such uncertainty is by
using a Variable Time Interval (sometimes
augmented by min/max duration constraints)
Beginning
Body
Time
End
Temporal Reasoning and Temporal Maintenance
• Temporal reasoning supports inference tasks
involving time-oriented data; often connected with
artificial-intelligence methods
• Temporal data maintenance deals with storage and
retrieval of data that has multiple temporal
dimensions; often connected with database systems
• Both require temporal data modelling
DB
TM
Clinical
decision-support
application
TR
Examples of TemporalMaintenance Systems
• The TNET system and the TQuery query
language (Kahn, Stanford/UCSF)
• TSQL2, a bitemporal-database query
language (Snodgrass et al., Arizona)
• The Chronus/Chronus2 projects (Stanford)
The TQuery Language
•
•
•
•
(Kahn, 1991)
Used within the TNET temporal network system, which
was used by the Stanford ONCOCIN oncology-therapy
automated protocol-based system during the 1980s
Each TNODE represents a time interval during which a
clinical event happened
TQuery allows users to store and retrieve data using
clinical contexts rather than dates
Query::= <Function Attribute-Name When>
– (that is, perform Function on Attribute-Name during When)
• When::= <Interval-Name Range <When> Pname
Pcondition> (a recursive temporal specification)
Tquery Examples
• (Visit (1 4))
– The first to fourth TNODES with type label = Visit
• (Visit FIRST (Cycle (-4 –1)))
– The first of each of the TNODES with label = Visit from the
last four TNODES with type label = Cycle
• ((Visit Tx) All
((CmTx POCC) ALL)
WBC (NCOMPARE > $ 4.5)
-Select from all (type chemotherapy, subtype POCC) all
the nodes with (type visit, subtype Tx) in which the
value of attribute WBC exists and is greater than 4.5
TSQL2
(Snodgrass, 1995)
• Designed by a committee of researchers, headed by
Snodgrass at Arizona University
• Consolidates existing approaches
• Inherits from SQL-92 temporal types such as DATE
– Adds the PERIOD data type
• A linear, bounded at both ends, time line
• No commitment to discrete, dense, or continuous temporal
ontologies: Queries must include granularity to be
meaningful
• A bitemporal conceptual data model, timestamps tuples by a
set of bitemporal chronons; each chronon (t, v) is a rectangle
in valid time/transaction time space
The Bitemporal Conceptual
Data Model
Valid
Time
17/3/95
23/2//95
Hospitalized(Jane)
5/1/95
27/11/94
23/2/95
1/4/95
21/6/95 3/7/95
Transaction Time
TSQL2: Examples
• What drugs were prescribed to Jane in 1996?
SELECT Drug
VALID INTERSECT (VALID (Prescription),
PERIOD ‘[1996]’ DAY)
FROM Prescription
WHERE Name = ‘Jane’
• Insert a prescription with a known period of validity
INSERT INTO Prescription
VALUES (‘Jane’, ‘Dr. Max’, ‘Lasix’, ’50mg’,
INTERVAL ‘4:00’ MINUTE)
VALID PERIOD ‘[2000-07-23 – 2000-8-14]’
Chronus II
(O’Connor et al., 1999)
• A Stanford model, influenced by TSQL2 and the
previous Das Chronus system, which it considerably
enhances
• Designed to support queries in the EON guidelinebased therapy system and the Tzolkin temporal
mediator to patient data
• Supports most of SQL-92 as well as extensions such as
valid time, indeterminacy, multiple calendars,
hierarchical types, temporal joins, etc.
• Temporal indeterminacy uses the Snodgrass model of
lower support, upper support, and a probability mass
function to denote the event’s temporal distribution
Chronus II: Example
• Select employees that have worked as a
mechanic for longer than two months:
TEMPORAL SELECT Name
FROM Occupation
WHERE Title = ‘Mechanic’
WHEN DURATION(Occupation, ‘Months’) >2
Summary
• Clinical databases require representations that
include a strong emphasis on time and
uncertainty
• Bitemporal databases are necessary to support
clinical, research, administrative and legal
requirements