Transcript Prediction Models for Smart Home Based Health Care System

Vikramaditya Jakkula
Washington State University
[email protected]
IEEE Workshop of Data Mining in Medicine 2007 (DMMed '07)
In conjunction with IEEE International Conference in Data Mining 2007 (ICDM '07)
VJ
AI@WSU © 2007
1
Smart Environments
VJ
AI@WSU © 2007
2
MavHome: Smart Home Project

Project Unique
◦ Focus on entire home

House perceives and acts
◦ Sensors
◦ Controllers for devices
◦ Connections to the mobile user and Internet

Unified project incorporating varied AI
techniques, cross disciplinary with mobile
computing, databases, multimedia, and
others.
VJ
AI@WSU © 2007
3
MavHome: Core Technologies1

Minimal Sequential Patterns Using “ED”
Given an input stream S of event occurrences O, ED:
 Partitions S into Maximal Episodes, Pmax.
 Creates Itemsets, I, from the Maximal Episodes.
 Creates a Candidate Significant Episode, C, for each
Itemset I, and computes one or more Significance Values,
V, for each Candidate.
 Identifies
Significant
Episodes
by
evaluating
the
Significance Values of the candidates.
VJ
AI@WSU © 2007
4
MavHome: Core Technologies2

Decision Making using ProPHeT

ProPHeT is the main controlling component of the system.

It uses data filtered through Episode Discovery (ED) to
create a Hierarchical Hidden Markov Model (HHMM).

HHMM represents a user model that includes all of the
episodes (e.g., entering a room, watching TV, sitting in a
chair and listening to music, and so forth) that a person
performs in the environment.
VJ
AI@WSU © 2007
5
Experimentation Environment1
VJ
AI@WSU © 2007
6
Experimentation Environment2
MavHome Environment
MavLab
MavKitchen
MavPad
VJ
AI@WSU © 2007
7
Experimentation Environment3
MavHome Smart Apartment

The evaluation environment is a
student
apartment
with
a
deployed Argus and X-10 network

There are over 150 sensors
deployed in the MavPad that
include
light,
temperature,
humidity, and switches.
VJ
AI@WSU © 2007
8
Earlier Work
Identification of Lifestyle Behavior Patterns with
Prediction of the Happiness of an Inhabitant in a Smart
Home
•Identify correlations among everyday activity in smart home.
•we use the machine learning technique, the k-nearest neighbor
algorithm, to predict the state of wellness the inhabitant will
experience on the following day.
•Show that a simple sensor network in smart home can be used
to detect lifestyle patterns
Monitoring Health by Detecting Drifts and Outliers for a
Smart Environment Inhabitant
•help us gain information about different types of drifts and
outliers that are part of the inhabitant’s lifestyle
•anomalies in inhabitants health such as blood pressure, pulse
and temperature values.
•Gives information about sudden changes observed in inhabitants
health
VJ
AI@WSU © 2007
9
Experimentation Overview

Basic overall goal is to build a forecasting system for
healthcare system in smart home.

Used 90 days data for training and 61 days data for
testing.

Use Weka workbench for the experimentation process.

Experiment 1: Comparing different learning algorithms
prediction accuracy on health vital datasets collected in a
smart home.

Experiment 2: Learning to predict abnormal or unhealthy
days in a smart home residents life.
VJ
AI@WSU © 2007
10
Experiment I

Goal: Compare prediction accuracy of
classifiers to choose the best classifier to
predict the health vitals.

Challenges:
Health Vital value prediction is dependent
on many factors and major factors
including food intake, current health
condition/history/previous illnesses and
physical activity performed.
VJ
AI@WSU © 2007
11
Experiment I Results
Class
SMO
(Reg.)
NN
MLP
Lazy
LWL
KNN
Systolic
3.34%
18%
27.86%
47%
Diastolic
14%
8.20%
42%
53%
Pulse
2%
8.33% 16.66% 53.30%
Average
Accuracy
6%
VJ
12%
29%
AI@WSU © 2007
51%
12
Experiment II

Goal: Learn to predict abnormal days.

Challenges:
Unexpected events [Emotional/Physical]
Sudden health and environment changes
Food consumption and sleep
and so forth!
VJ
AI@WSU © 2007
13
Experiment II
Run#
KNN
10-Fold C.V.
Correctly
Classified
Instances
6
49
Incorrectly
Classified Instances
Acc (%)
Err. (%)
MAE
1
4
85.7
92.5
14.2
7.54
0.1788
0.0925
Abnormality to be the any value greater than 137/84 mm Hg (Myers
MG) and for pulse the normal range is 60 to 100 beats per minute
(Wikipedia) combined with physical activity.
Did not observe any significantly extreme values!
Future work includes observation on subjects from different age
groups and different genders.
VJ
AI@WSU © 2007
14
Conclusions and future work

K-NN outperforms other classifiers with an overall prediction
accuracy of 51% in experiment 1 and has an prediction
accuracy of 86% in experiment 2.

Predicting time series data is still a difficult challenge.

We observe that the prediction models act as useful
components to the health care system in smart homes.

Future work would include improving the prediction, collecting
more data over time and experimenting larger datasets.

Anomaly detection based prediction for health care system
and adaptive healthcare systems.
VJ
AI@WSU © 2007
15
Acknowledgements

I would like to thank my professor Dr. Diane J. Cook for her
encouragement and support.

I would also like to thank the Human subject who
participated in these trials.
VJ
AI@WSU © 2007
16
Questions!
VJ
AI@WSU © 2007
17