Monitoring Health by Detecting Drifts

1

• Monitoring Health by Detecting Drifts
• and Outliers for a Smart Environment
• Inhabitant
• Gaurav Jain, Diane J. Cook, Vikramaditya Jakkula

2
MavHome

• UTA 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

3
MavHome

• Goals
• The goals of an intelligent environment control
  system should be to
• 1. Maximize the safety and security of the
  inhabitant(s)
• 2. Maximize the comfort of the inhabitant(s) by
  automating their environment the fullest and most
  desirable extent possible
• 3. Minimize the consumption of natural
  resources in an effort to reduce costs and
  maximize environment efficiency.

4
Environment

• MavHome Environment
• MavDen
• MavKitchen
• MavPad

5
Environment-Contd
6
Overview
7
Core Technologies

• Minimal Sequential Patterns Using ED
• Given an input stream S of event occurrences
  O, ED
• 1. Partitions S into Maximal Episodes, Pmax.
• 2. Creates Itemsets, I, from the Maximal
  Episodes.
• 3. Creates a Candidate Significant Episode, C,
  for each
• Itemset I, and computes one or more
  Significance
• Values, V, for each Candidate.
• 4. Identifies Significant Episodes by evaluating
  the
• Significance Values of the candidates.

8
Core Technologies

• 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.

9
Core Technologies
10
Need for Health Monitoring
Problem Elderly, disabilities and the chronic ill
need health care. Personal preference Increased
care cost Inadequate infrastructure Solution Low-
cost automated health monitoring system at
home Lanspery Hyde state For most of us, the
word home evokes powerful emotions and is a
refuge
11
Drift Detection Algorithm

• Diurnal algorithm
• Uses autocorrelation plots
• Three Steps
• Update history
• Detect drifts
• Report Generation

• Input history h, frequency sets, action list and
  their criticalities
• OutFile report file
• update h with the frequency sets
• for each action a loop
• find the drift type d in action as history
• send the drift d for action a to the report
  manager
• the report manager generates the final report
  based on the criticality of each action, the
  current drift parameters and previous drift
  parameters.

12
Update History

• Maintains six-hourly, daily, weekly history
  queues.
• Input is four six-hourly frequency sets.
• Different window sizes are posible
• Large window vs. small window

13
Detecting Drifts

• Input action a, history h, reporter r
• OutFile drift type d and its parameters p
• check if action a has drift type d no drift
• if yes then
• send the drift type and its parameters to the
  reporter
• return to the calling function
• check if action a has drift type d cyclic or
  increasing
• if yes then
• send the drift type and its parameters to the
  reporter
• return to the calling function
• send the drift type as chaotic to the reporter

14
Test for no-drift

• No-drift?
• constant for a significant period of time, and
• may have random noise.
• Only the top half of the autocorrelation plot is
  used. Why?
• Test
• autocorrelation plot values lt threshold. Why?
• Less than 10 of these values should lie outside
  the (m 2s, m 2s) range. Why?

15
Test for Cyclic

• cyclic trend shows high upward peaks in
  autocorrelation graph

16
Test for Sloping

• High degree of autocorrelation is between
  adjacent and near-adjacent observations.
• High value at lag one
• Value decreases with increase in lag
• Slope length is the smallest lag at which the
  values stops decreasing.
• Note Random noise is suppressed by the
  autocorrelation plot

17
Test for Chaotic

• No test for chaotic
• Anything not yet classified ends up any chaotic
• Causes
• large number of irregular changes
• heavy non-random noise in data in all the windows
• sudden large changes in the distribution
• seen for a short period of time when drift type
  changes
• Reporting of drifts will be discussed after
  presentation of the outlier detection algorithms

18
Outlier Detection Algorithms

• Two types of outliers
• Extremely high or low value in periodic frequency
• Occurrence unexpected action in an ordered
  sequence of actions.
• Separate algorithms for each
• Autocorrelation-based outlier detection
• Uses drift detection method
• Outlier if last data point lies outside (m 3s,
  m 3s),
• Tested for all window sizes.
• If found outlier, then drift detection is not
  done.
• Prediction-based outlier detection
• Why a two methods?

19
Prediction-based Outlier Detection

• Live-monitoring method
• Uses Active LeZi (ALZ) 2 to find the expected
  pattern in the data.
• ALZ uses data compression to predict the next
  action in a sequence. It determines the
  probability distribution for each action at any
  point of time.
• When an action occurs this probability
  distribution is used to determine if the action
  is an outlier or not.

20
Prediction-based Outlier Detection

• To determine if an action x is an outlier we
  calculate the anomaly measure n(x).
• Two methods are used to calculate anomaly
  measure. Why?
• To determine the importance an outlier we
  calculate the urgency factor u(x)

urgency factor, u (x) n(x) c(x) report if u
(x) gt 0.1
21
Report generation for Autocorrelation-based
algorithms

• Which drift or outlier is important to report?
• Uses
• current classification,
• the previous classification,
• the criticality of the action, and
• other parameters (confidence, length of drifts
  etc.)
• Three levels
• Level one Critical drifts and outliers
• Level two Important drifts and outliers
• Level three All drifts and outliers

22
Report generation

• Level one
• If action criticality is above medium, and either
  the classification changed from the previous or
  cycle period changes.
• Level Two
• All outliers
• criticality is above medium
• Previous classification changes
• cycle period changes
• confidence changes by some amount
• Level three
• Classification of each action.

23
Experiments

• HMS was tested using both synthetic and real data
  (activity and health).
• Five sets
• Synthetic set one
• Synthetic set two
• Real set one
• Real set two
• Health set
• Step1 verify algorithms using synthetic sets
• Step 2 analyze how the algorithm work on real
  and health sets

24
Nature of data
Synthetic set one

• Synthetic set one
• To test autocorrelation-based algorithm
• Hundred days nine action
• 10639 data points
• Random criticalities
• Synthetic set two
• To test prediction-based algorithm
• 100 data points
• Four actions

25
Nature of data

• Real data
• Activity data from MavPad
• Seven weeks
• Real set one tested on prediction-based
  algorithm
• Electrical outlets usage, light usage and
  overhead fan usage.
• 2163 data points 79 actions
• Real set two tested on autocorrelation-based
  algorithm
• Real set one data plus motion sensor data
• 334935 data points 157 actions
• Health Data tested on autocorrelation-based
  algorithm
• Systolic, diastolic and hear rate are taken as
  action
• 2 months one value each per day
• each action is associated with its value instead
  of frequency
• Most missing values were added manually

26
Experiments using Autocorrelation-based method

• For Health Data
• sensitive to sudden large changes
• could detect drifts due to long term trends even
  with small amounts of noise.

27
Figure Line graph for graph confidence
diastolic vs. number of days for health set.
Figure Line graph for graph confidence heart
rate vs. number of days for health set.
28
Reminder Assistance System

• Automation assistance is beneficial when
  activities are difficult to perform.
• Such reminder service would benefit individuals
  suffering from dementia.
• Reminders Triggered in two situations
• when user queries for next routine activity
• Critical anomaly is detected.

29
Conclusion

• HMS help us gain information about different
  types of drifts and outliers that are part of the
  inhabitants lifestyle.
• Detect anomalies in inhabitants health.
• Gives information about sudden changes observed
  in inhabitants health.
• Successful demonstration of MavHome software
  Architecture can monitor and provide automated
  assistance for inhabitants.

30
Future Work
We are currently collecting health-specific data
in the MavHome sites. We will be testing in the
living environments of recruited residents at the
C.C. Young Retirement Community in Dallas,
Texas. Lifestyle Trends and patterns of
inhabitants would be analyzed over period of
time.
31

• Thank You