Title: Temporal Sequence Analysis of Clinical Laboratory Results for Patient Follow-up and Effective Data Display
1Temporal Sequence Analysis of Clinical Laboratory
Results for Patient Follow-up and Effective Data
Display
- James Harrison, M.D., Ph.D.
- Associate Professor of Pathology
- Faculty-in-residence, Center for Biomedical
Informatics - University of Pittsburgh
- jhrsn_at_pitt.edu
- Dept. of Electrical Engineering, Univ. of
Pittsburgh - February 25, 2004
2Standard Clinical Data Display
3Standard Clinical Data Display
4Phenytoin levels
Discharge
5How does this happen?
- Data reduction
- Classification
- Cross-sectional view
- Loss of information
6Outline for Today
- Development and application of a simple temporal
event monitor (LabScanner) - Extension of the event monitor lightweight
general-purpose temporal abstraction system
(PROTEMPA) - Applications of temporal abstraction
- Patient identification and monitoring in the
clinical laboratory - Clinical data presentation and decision support
7Simple Temporal Pattern Recognition
- Goal Process monitor with simple rule
maintenance - Case identification for training, Q/A, clinical
reporting, consultation - Clinical lab (TDM) perspective
- Data access limited to text file export from LIS
- No cross-correlation of primary data with other
results - Constraints
- Need to recognize short patterns, but data sparse
and irregular - Underlying physiological model may be changing
- Pre-analytic errors, time and procedural errors
8A Pragmatic Pattern Detection Strategy
- Define features of short problem sequences from
inspection and annotation of real data
(heuristics) - Model-independent, based on observable sequence
features only - Derive a minimal set of sequence templates
(Discovery Rules) designed to detect problem
sequences - Implement Discovery Rules in data-scanning
software
9Simple Patterns in Patient DataSustained high
and low values
Phenytoin
10Simple Patterns in Patient DataIncreasing and
decreasing trends
Phenytoin
11Simple Patterns in Patient DataVariability and
Frequency
Phenytoin
12General Features of Discovery Rules
- Pattern features
- Value (state), trend, variability, frequency
- -- plus --
- Patient features
- Age, gender, hospital location
- Interval features
- Number of events, max/min overall time span,
max/min time between neighboring events
13Example Discovery RulesIdentification of Basic
Temporal Sequences
Discovery Rule Categories for Phenytoin (10-20 mg/l) Discovery Rule Categories for Phenytoin (10-20 mg/l) Discovery Rule Categories for Phenytoin (10-20 mg/l) Discovery Rule Categories for Phenytoin (10-20 mg/l)
Values Trends Variability Frequency
3 in up to 5 sequential values gt 20 (lt 28 days) 3 in up to 5 sequential values lt 10 (lt 28 days) Increasing values 3 values with an increase gt 0.7 per day, with a projected value gt 20 within 2 days, and the last value gt 15 (lt 28 days) Decreasing values 3 values with a decrease gt 0.7 per day with a projected value lt 10 within 2 days and the last value lt 15 (lt 28 days) 3 values with an average deviation gt 3.5 (lt 28 days) gt 3 values in 4 days gt 1 value in 8 hours
Target (therapeutic) range. Total pattern
length lt 28 days.
JH Harrison and P Rainey. Am J Clin Pathol
1995103710-717
14The Sliding Window Detection Algorithm
- Window size defined in rule
- Does the sequence of values within this window
completely satisfy a rule?
Example Increasing values?
15The Sliding Window Detection Algorithm
- Window size defined in rule
- Does the sequence of values within this window
completely satisfy a rule?
Example Increasing values?
16The Sliding Window Detection Algorithm
- Window size defined in rule
- Does the sequence of values within this window
completely satisfy a rule?
Example Increasing values?
17The Sliding Window Detection Algorithm
- Window size defined in rule
- Does the sequence of values within this window
completely satisfy a rule? - Four parts of a positive match
- Initiation phase values that contribute to the
pattern but do not completely satisfy it by
themselves
Example Increasing values?
18The Sliding Window Detection Algorithm
- Window size defined in rule
- Does the sequence of values within this window
completely satisfy a rule? - Four parts of a positive match
- Initiation phase values that contribute to the
pattern but do not completely satisfy it by
themselves - Trigger and initiation phase together completely
satisfy the rule
Example Increasing values?
19The Sliding Window Detection Algorithm
- Window size defined in rule
- Does the sequence of values within this window
completely satisfy a rule? - Four parts of a positive match
- Initiation phase values that contribute to the
pattern but do not completely satisfy it by
themselves - Trigger and initiation phase together completely
satisfy the rule
Example Increasing values?
20The Sliding Window Detection Algorithm
- Window size defined in rule
- Does the sequence of values within this window
completely satisfy a rule? - Four parts of a positive match
- Initiation phase values that contribute to the
pattern but do not completely satisfy it by
themselves - Trigger and initiation phase together completely
satisfy the rule - Persistence phase If immediately subsequent
windows also completely satisfy the same rule,
they are concatenated to form a single pattern
(rule is greedy).
Example Increasing values?
21The Sliding Window Detection Algorithm
- Window size defined in rule
- Does the sequence of values within this window
completely satisfy a rule? - Four parts of a positive match
- Initiation phase values that contribute to the
pattern but do not completely satisfy it by
themselves - Trigger and initiation phase together completely
satisfy the rule - Persistence phase If immediately subsequent
windows also completely satisfy the same rule,
they are concatenated to form a single pattern
(rule is greedy).
Example Increasing values?
22The Sliding Window Detection Algorithm
- Window size defined in rule
- Does the sequence of values within this window
completely satisfy a rule? - Four parts of a positive match
- Initiation phase values that contribute to the
pattern but do not completely satisfy it by
themselves - Trigger and initiation phase together completely
satisfy the rule - Persistence phase If immediately subsequent
windows also completely satisfy the same rule,
they are concatenated to form a single pattern
(rule is greedy). - End last point that satisfies the rule.
Example Increasing values?
23The Sliding Window Detection Algorithm
- Window size defined in rule
- Does the sequence of values within this window
completely satisfy a rule? - Four parts of a positive match
- Initiation phase values that contribute to the
pattern but do not completely satisfy it by
themselves - Trigger and initiation phase together completely
satisfy the rule - Persistence phase If immediately subsequent
windows also completely satisfy the same rule,
they are concatenated to form a single pattern
(rule is greedy). - End last point that satisfies the rule.
Example Increasing values?
24The Sliding Window Detection Algorithm
- Window size defined in rule
- Does the sequence of values within this window
completely satisfy a rule? - Four parts of a positive match
- Initiation phase values that contribute to the
pattern but do not completely satisfy it by
themselves - Trigger and initiation phase together completely
satisfy the rule - Persistence phase If immediately subsequent
windows also completely satisfy the same rule,
they are concatenated to form a single pattern
(rule is greedy). - End last point that satisfies the rule.
Example Increasing values?
25Initial Implementation Data Input
File downloaded from LIS
Flat file import
26Rule Management
- Pattern Types
- Value comparison
- Trend
- Variability
- Frequency
27Patient List and Display
28Incidence of Patterns in Clinical Laboratory Data
29Printed Reports
30LabScanner Downloadhttp//jhh.opi.upmc.edu/labsca
nner/
31Pattern Incidence in Ten Monitored DrugsJune -
September, 2000
- 8311 patients, 35813 TDM values imported
- Limit to patients with three or more values
- 2927 patients, 27512 values scanned
- 9 of values out-of-range high
- 16 out-of-range low
- 1071 patients with patterns (37)
- 673 patients with persistent patterns (23)
- Average persistence 5.9 days, 5.5 specimens
- Average 37 new patterns per day
32Resource Utilization Associated with Patterns
52 Pediatric PTN patients with patterns
(excluding freq.), Jan - Mar 2001 43 Ped.
patients with at least 3 PTN values but no
patterns (same period)
Hospital Days
Hospital Costs
Resource utilization from the day of the first
level to 4 weeks after the last level
33Summary of Simple Temporal Abstraction
- Patterns exist in drug level data that are not
clearly recognized by clinicians or laboratories
but may identify patients at risk for increased
cost of care, suboptimal care, or medical error - Discovery rules can be developed and applied in
software that identify patients showing such
patterns - We have developed a software tool that can
analyze data retrospectively or prospectively to
detect such patterns - Large datasets can be processed (ca. 10,000
patients) - Identifying these patients may be useful for
education, QA/QC, clinical process improvement,
and consultative follow-up - Characteristics of the data patterns, such as
persistence, may provide new QA indicators - Tool is limited to single instances of simple
patterns cannot correlate patterns within or
across types of events
34Complex Temporal AbstractionsAggregating Simple
Patterns Through Relationships
- The temporal abstraction task has the goal of
abstracting high level concepts from time-stamped
data.
Digoxin-Quinidine interaction
High digoxin
Rising digoxin
Rising digoxin
Declining digoxin
On quinidine
On digoxin
time
35PROTEMPAA Temporal Abstraction Engine
- PRoblem-Oriented TEMPoral Analysis
- A symbolic rule-based software framework for
specifying, detecting, and visualizing temporal
patterns in time series data - Supports simple and complex temporal abstraction
- Complex abstractions may include results from
multiple clinical laboratory tests or other
events - Modular, Java-based may be implemented
stand-alone or in a server environment
36Temporal Pattern RelationshipsSimple
Abstractions Combine to Form Complex Abstractions
Disjoint
Equals
Meets
Overlaps
Coincident
Starts
- Patterns have
- Their own basic characteristics
- Relationships to other patterns
Ends
During
Contains
37Features of PROTEMPA Rules
- Patient features
- Age, gender, location
- Interval features
- Max/min overall span, max/min point-point
distance - Optional fixed min/max start and min/max finish
times - Simple pattern features
- Value (state), trend, variability, frequency
(others possible) - Rule Chain (complex abstraction) features
- Listing of contributing simple patterns and their
relationships - Specification of at least one of the following
defines a relationship - Max/min time between starts and/or ends
- Max/min times between start1-end2 and/or
end1-start2 - Represented as 8-tuple
- Negated patterns
38Temporal Constraint NetworksA framework for
storage of rule chain data
- Nodes are start and end points of intervals
- Constraints define spans and relationships of
intervals - Constraint values are maintained in matrices
- Efficient algorithms are available to test
network consistency
39PROTEMPA Processing Sequence
- Locate simple pattern intervals (sliding window)
- Aggregate intervals into complex abstractions via
constraint network consistency checking - Continue aggregation until no new rule chains are
identified
Simple Abstraction Modules (Math or stat
functions across sequences)
Value
Complex Abstraction Module (Temporal Constraint
Networks)
Trend
Raw Data
Identified Intervals
Report Patterns
Rule Chain Matching
Output
Variability
Frequency
Others
Re-analysis
40PROTEMPA Framework
Dataset
Input
Scanner Engine
Temporal Pattern Detector
Output
Pattern Matches
Rules Database
Input
Encode as
Visualize matches as
Patterns of Interest
Selection Criteria
41PROTEMPA Stand-alone Interface
42Application PatientPatterns
- Designed to allow the clinical laboratory to
specify temporal patterns for monitoring and
follow-up - The PROTEMPA framework is implemented as a Java
servlet running in a JBoss/MySQL environment - LIS transactions transferred and scanned every
six hours - Rule chains identify situations appropriate for
follow-up - Laboratory personnel log on via the web to view
found patterns
43PatientPatterns Interface
Prototype implementation Pattern detection
engine runs as a servlet accesses laboratory
data in MySQL DB via JDBC user interaction
through the Web.
44Rule Entry and Editing
45Interval Relationship Specification
46Prototype Web Display
47Application Adaptive Clinical Displays
- Lab data displays are usually tabular and static
- Sequence, form and proximity of displayed data
affects decision-making - Graphical data display is best for decisions
requiring quantitative comparisons or sequence
assessment under time pressure - Clinically significant situations may be
recognizable by processing data at display time - Automated aggregation and optimal display of
data based on the content of the data may improve
clinical decision-making (a form of decision
support) - Project funded by NLM October, 2003
48Adaptive Display Evaluation Protocol
- Set of validated clinical cases for diagnostic
training - Extend with one week hospital course including
diagnostic cues - Physician subjects review and write orders under
time pressure using alternative clinical displays - Cognitive analysis think-aloud protocol,
videotape - Transcriptions/tapes coded for searching vs.
integration/evaluation activities - Evaluate orders for completeness and accuracy
49Alternative Displays for Investigation
- Control Cerners new web display, flowsheet
component - Tabular text, static
- Organized by laboratory
- Test Semigraphical display
- PROTEMPA processing on query
- Display organization prioritized by rule chains
based on data content (decision cluster) - Default display format for each test (graphics or
text) controlled by rule chains based on data
content - Both offer similar data access after default
display
50Adaptive Display Hypotheses
- The adaptive display will provide aggregated,
rapidly-interpretable data for decision-making (a
decision cluster) - Data viewing and decision-making will occur
primarily in the adaptive portion of the display - Cognitive load will be decreased (fewer steps)
and the ratio of searching/integrative steps will
decrease - Orders will be more complete and accurate
- Diagnostic cues will be recognized more frequently
51Summary
- We have developed a flexible software tool
- Analyzes medical events retrospectively or
prospectively - Detects simple temporal patterns alone or in
arbitrary combinations - May be implemented as a standalone tool or a
server - Supports a hypothesis-testing form of data
mining - Likely to be useful for clinical laboratory and
other QA - Possibly helpful for identifying suboptimal
therapy or medical errors, either in retrospect
or as a process monitor - Possibly useful in data display and decision
support - Possibly useful for annotating temporal
relationships in data sets to allow improved
clustering/machine learning
52Acknowledgements
- Pete Rainey, MD, PhD
- Thira Choojitarom, MD
- Juan Lertora, MD, PhD
- Andrew Post, MD
- Valerie Monaco, PhD, MS