Transcript KBS和KM

Temporal Abstractions for Interpreting
Diabetic Patients Monitoring Data
Advisor： Dr. Hsu
Graduate：Min-Hong Lin
IDSL seminar 2002/1/30
1
Outline

Motivation
Objective
Background

IDDM(Insulin Dependent Diabetes Mellitus)
 Temporal Abstractions
TA for Interpreting Diabetic Patients Monitoring Data






A Case Study
Conclusions
Comments
2
Motivation

Several medical domains require:




Analyzing and interpreting a large number of
longitudinal data.
Data coming from long-term monitoring of chronic
patients.
Physician must interpret information on the basis
of a comprehensive analysis.
In the interpretation task, the data pre-processing
phase is a crucial step.
3
Objective



To show that the data analysis and interpretation
tasks can be carried out by transforming the raw
data into a abstract view of the patient’s history.
Propose a novel approach based on the
combination of a temporal abstraction method
with statistical and probabilistic techniques.
Apply TA to the long-term monitoring of Insulin
Dependent Diabetes Mellitus(IDDM) patients.
4
Insulin Dependent Diabetes Mellitus



Diabetes Mellitus is a major chronic disease in
developed countries (about 3%).
Diabetes Mellitus is characterized by an alteration
of the glucose metabolism due to a decreased
endogenous production of insulin.
In IDDM the patients must take exogenous insulin
in order to prevent extremely high blood glucose
levels(hyperglycemia).
5
Insulin Dependent Diabetes Mellitus(cont’d)



The patients perform self-monitoring of the Blood
Glucose Levels (BGL) and glycosuria at home,
and report the monitoring and therapeutic data in a
diary.
The accuracy of the patients’ self-care is very
important, since the onset and development of
diabetic complications is strictly related to the
degree of metabolic control.
The physicians revise the therapy during
periodical visits, every two/six months.
6
IT for Diabetes Care


To improve the quality of the therapy revision
process and the patients’ management, several
computer-based systems have been proposed since
the early 80’s.
Different systems categories:




Day-by-day advisory systems V.S. visit-by visit
advisory systems
Model-based systems V.S. Data-driven systems
Statistical analysis and graphical representations
Time-series analysis and temporal abstractions
7
Temporal Abstractions




TA is an AI methodology.
TAs are used in data interpretation to solve the
temporal abstraction task whose goal is to abstract
high level concepts from time-stamped data.
In the medical domain, TAs can be used to
describe patients states holding over time periods.
The principle of the TA method is to move from a
time-point to an completely interval-based
representation of the monitoring data.
8
Temporal Abstractions(cont’d)

TA task is decomposed into two main type of TA
subtasks:

Basic TA: solved by mechanisms that abstract timestamped data into intervals(input data are events and
outputs are episodes)



State TAs: detect episodes associated to qualitative levels of
time-varying variables, like normal or abnormal states.
Trend TAs: detect patterns like increase, decrease and
stationarity in a numerical time series.
Complex TA: solved by mechanism that abstract
intervals into other intervals(input and output data are
episodes)
9
TA method ontology
10
Basic TA task

For each basic TA it defines two parameters:


Granularity: represents the maximum allowed
temporal gap between two measurements that
can be aggregated into the same episode.
Minimum extent: represents the minimum time
span of an episode to be considered relevant.
11
An example of the basic TA task
12
Complex TA task


The mechanism solving the complex TA task
searches for temporal relationships between
episodes.
The temporal relationships investigated can be
expressed through temporal operators defined in
the Allen algebra(before, after, meets, overlaps,
starts, finishes, equals, during)
13
An example of the complex TA task
14
An example of the complex TA task(cont’d)


If the input episodes refer to patterns extracted
from different time series, the method can detect
patterns in multi-dimensional data.
For example, the problem of investigating whether
a persistent cough and high fever occur
simultaneously in a patient’s history.
PERSISTENT COUGH OVERLAPS HIGH FEVER
15
Temporal Abstraction for the Interpretation of
Diabetic Patients’ Monitoring Data

Data Pre-Processing



Analyze the original time-series using TAs
Derive a collection of new time-series by
computing the TAs that are true in each time
point of the original time scale.
Data Interpretation
16
Data Pre-Processing

Subdivide the 24-hours daily period into a set of
consecutive non-overlapping time slices

Perform the analysis on the time series of three
variables ( BGL, glycosuria and insulin dosages)
17
Data Pre-Processing(cont’d)

Define a set of basic and complex abstractions for
each time slice.
18
Data Pre-Processing(cont’d)


Characterize the patients’ behavior through the
concept of Abstract State(ABST), that corresponds
to the combination of the TAs that are true in that
period.
The general form of the abstract state in the i-th
day for the j-th time-slice(ABSTij) is:
19
An example of data pre-processing
20
An example of data pre-processing(cont’d)
21
An example of data pre-processing(cont’d)
22
Data Interpretation

The Blood Glucose Modal Day(BG-MD)

The Blood Glucose Modal Daily Pattern(BG-MDP)

Exploiting TA Time-Spans
Exploiting Complex TAs

23
The Blood Glucose Modal Day(BG-MD)



The BG-MD is a characteristic daily BGL pattern that summarizes the
typical patient’s response to the therapy in a specific monitoring period
It is usually derived as the collection of the most probable blood
glucose qualitative levels in each time slice.
K:states, N: monitoring days, D:collect measurements, M=N-D:
missing data, dl is the number of occurrence of the l-th state in the
monitoring period.

Ignorance (IG) in the monitoring period:

The modal day can be extracted by taking the BGL states with the
highest pinf in each time slice.
24
The Blood Glucose Modal Daily Pattern(BG-MDP)




BG-MDP is the most frequent sequence of abstract states
of the BGL variable in the different time slices of one day.
Search parameters: the maximum allowed ignorance (MIG)
in one time slice (I.e. the maximum number of allowed
missing data) and the minimum probability bound for the
joint probability distribution Pinf.
Time slices selection: only the time slices that have
ignorance level lower than MIG will include in the BGMDP
Search: perform an exhaustive search of the daily patterns
that have a lower probability bound higher than Pinf.
25
Exploiting TA Time-Spans




If the time span of the BGL normoglycemic episodes have
an exponential distribution, it is clear that the patient is not
able to control his/her glucose metabolism for a long
period.
Perform a non-linear least-squares fitting of the model
Once we have estimated λ, we test the hypothesis that the
data follows an exponential law with parameter through
the x2 statistics.
The exponential distribution hypothesis is rejected with
degree α if x2h-1 > x2h-1(1- α)
26
Exploiting Complex TAs

In clinical practice, the physician usually tries to
combine the information coming from the
different variables under monitoring.



Somogyi effect : is detected by looking for
“hyperglycemia at Breakfast with absence of glycosuria”
Dawn effect : is detected by searching for
“hyperglycemia at Breakfast with presence of
glycosuria”
Metabolic Instability : in which a “BGL increase” is
immediately followed by a “BGL decrease ” or viceversa.
27
An example of Exploiting Complex TAs
28
A Case Study

A 14 years-old female patient, monitored for a period of
165 days. The BGL data are show below:
29
A Case Study (cont’d)
30
A Case Study (cont’d)
31
A Case Study (cont’d)
32
A Case Study (cont’d)
33
A Case Study (cont’d)
34
A Case Study (cont’d)
35
A Case Study (cont’d)
36
Conclusions


TA method requires an intensive knowledge
acquisition effort for the TA definition, both in
terms of types of TAs to be applied to the
particular problem and in terms of the parameters
that some TAs require.
TA method will enable the assessment of a new
class of data mining systems.
37
Comments

How to prevent from diabetic:



Diet
Physical exercise
Check the BGL regularly
38