Dr. Shiyu Zhou, Univ. of Wisconsin-Madison
Dr. Patricia Flatley Brennan, Univ. of Wisconsin-Madison
Dr. Yong Chen, University of Iowa
Smart Asthma Management:Statistical modeling, prognostics, and
intervention decision making
Research Approaches
o Motivation
• Asthma is a common lung disease impacting on a largenumber of people
• Smart Asthma Management (SAM) system makes thepatient-centered chronic asthma care possible
o Objectives
• Patient condition modeling and estimation at populationlevel
• Patient condition model updating and individualizedprognosis
• Clinical decision support using partially observable Markovdecision process
Accomplishment & Future Work
Case Study
SCH:EXP:Collaborative Research: Smart Asthma Management: Statistical modeling, prognostics, and intervention decision making (SCH-1343969) e-mail: [email protected]
Rescue inhaler usage0 5 10 15 20 25 30
250
300
350
400
450
500
550
Day
Peak F
low
( L
/min
)
Longitudinal biomarker(time-dependent covariate)
Mixed Effects Model for the
Time-dependent Covariate
(Continuous)
Progresses� Datasets from Propeller Health and Project HealthDesign. Some
will be publicly available in ICPSR for use by other investigators.
� Son, J., Brennan, P., and Zhou, S., “Rescue inhaler usage
prediction in smart asthma management systems using joint
mixed effects logistic regression model,” IIE Transactions, 2014,
submitted
� Jia, H., Brennan, P., Zhou, S., Son J., and Hung, Y.,
“Personalizing statistical models for asthma prognosis and
therapeutics,” Artificial Intelligence in Medical Applications (AIMA)
Annual Conference, 2014 (poster, accepted)
� �,� �,� �,� �∗
Response: (logit transformed) Probability of rescue
inhaler use at time t for the ith patient
(individual probability)Individual covariate model
� �∗
��
� �
Separate model for the
time dependent covariate
II. Joint Mixed Effects Logistic Regression
III. Individualizing the Model
(1) Updating the time-dependent covariate model
�∗ = �,� � �,�
�∗ = �,� � �,�
Historical
Database
(Population)
Updated
Model
(Individual)
Learn Update
New observations
from patient p
�Population-wise
Distribution
Distribution
for patient p
�
(2) Updating the logistic regression model
Distribution
for patient pHistorical
Database
(Population)
Updated
Model
(Individual)
Learn Update
����Population-wise
Distribution
USE DAY PF
Y 1 350
⋮ ⋮ ⋮
Y 10 250New data from
patient p
Goal: Predict the number of expected inhaler usage in the
immediate next week (response range: 0 to 7)
Method: Joint Mixed Effects Logistic Regression
Criteria: Cross-validation and mean absolute error (MAE)
2 20 40
01
23
45
67
(a) M1
Number of observations (m)
Absolu
te e
rror
2 20 40
01
23
45
67
(b) M2
Number of observations (m)
Absolu
te e
rror
2 20 40
01
23
45
67
(c) M3
Number of observations (m)
Absolu
te e
rror
Criteria MethodNumber of observations (m)
2 20 40
MAE
(SD)
M1: the model which assumes
a population-wise behavior
2.098
(1.326)
2.105
(1.282)
1.957
(1.399)
M2: proposed individualized
prognostic model
1.690
(1.521)
1.070
(1.010)
0.861
(0.790)
M3: individually fitted model2.720
(2.762)
1.312
(1.440)
1.180
(1.599)
Future work:� Underlying asthma condition change detection method based
on individual rescue inhaler usage profile
� Optimal clinical intervention decision making algorithm based on
the patient-level statistical models
I. Joint Modeling Framework
Rescue Inhaler Usage Prediction using Joint Mixed Effects Logistic Regression Model
Mixed Effects Logistic
Regression for Rescue
Inhaler Use (Binary)
Joint Mixed Effects
Logistic Regression
Model
Both models consider random
effects that can be considered
as individual effects
Joining
models
IV. Individualized Prediction
Number of rescue inhaler use in a week is a key
indicator of the asthma control level of patients
(National Heart, Lung, and Blood Institute)
Historical
dataTime
Collect data from the
new patient up to �∗
∗ ∗+7New patient
enrolls Individualize (update) the model
Prediction with
individualized model
Smart Asthma Management (SAM)
Motivation and Objectives