Basim Majeed, Ben Azvine, Steve Brown (BT) Trevor Martin (University of Bristol) Intelligent Systems for Telecare
Mar 28, 2015
Basim Majeed, Ben Azvine, Steve Brown (BT)
Trevor Martin (University of Bristol)
Intelligent Systems for Telecare
Presentation Outline
• Introduction
• Centre for Care in the Community
• Well-being Monitoring
• Intelligent Data Analysis
• Questions
IntroductionEffects of the demographic shift
2.0
2.5
3.0
3.5
4.0
4.5
1995 2005 2015 2025 2035 2045 2055Year
Rat
io P
erso
ns
Ag
ed 1
6-64
to
65+
10
15
20
25
30
35
40
45
50
UK
Lo
ng
Ter
m H
ealt
hca
re C
ost
(£B
)
Support Ratio 1 UK Long TermHealthcare Cost 2
1. Office for National Statistics, 2002.2. Royal Commission Report into Long Term Care, 1999.
Introduction
• Intelligent Telecare provides new ways of enabling elderly or vulnerable people to maintain their independence and live in their own homes for longer.
• BT is leading a DTI sponsored group of academic partners to develop an Intelligent monitoring system for telecare.
• System uses a range of low cost sensors placed in the home to monitor a person's activity and build up a picture of their behaviour.
• Target users are care professionals within social services (initially at Liverpool).
Centre for Care in the Community
Three Research Teams
• Domain Specific Modelling
• Sensor Network
• Intelligent Data Analysis
DTI/BT Funded
Well-being monitoringIdentifying the activities
Identifying the activities to monitor
• Older population split into ‘priority groups’ identified by the DoH
• Relevant activities identified for each priority group
• A core set of activities relevant to physical, mental and social elements of well-being has been identified
• Specific questions for each activity have been identified
Core activity monitoring set for the physically frail priority group
• Leaving & returning home (Social interaction)
• Visitors (Social interaction)
• Preparing food & eating appropriately (*ADL’s)
• Sleeping (*ADL’s)
• Leisure activities (Personal goals)
• Personal appearance (Personal goals)
*Activities of Daily Living
Client home
Kitchen
Lounge
cupboard
Back door
TV
Coffeetable
‘Radio’chair
Fire
plac
e
Fridge/freezer RMU
Gas
Oven/hob
Sofa &
armchairs
Window sill
sink drainer
‘TV’’chair
Spare bedroom
Master bedroom
Bath
W.C.
Basin
BathroomLand
ing
Double bed
War
drob
es
War
drob
es
Draws
Pile of various objects
Intelligent Data AnalysisConverting sensor data into Activity information
Target Group
Well-being concepts
Sleep monitoring
Visitor detection
…….
Domain Knowledge
Provide Answers to Core Activity Questions
1 4 7
Well-Being
Time Horizon
2.00
2.50
3.00
3.50
Wak
e In
terv
als
per
Nig
ht
Data from a Sensor Network
System Overview
The Challenge• People are different, fickle, unpredictable, and unlike physical systems
that have known responses to external influences
• The challenge is to provide a system that can adapt to changes
• These characteristics are very important in well-being monitoring, more so than e.g. consumer behaviour analysis
• Cause No Harm!
• Our Approach: Assist the carer in decision making by providing an easy-to-configure system that hides the complexity of the analysis
System Overview
Inte
rfac
e to
Se
ns
or N
etw
ork
User Interface
Sensor Object Management
AnalysisAlgorithms
Raw Data Pre-processing
Pre-Processed Data ReportData
Sensor Information
Report Generator
Date Time Sensor ID Value
07-May-04 00:00:13 70 FALSE07-May-04 00:00:14 70 TRUE07-May-04 00:00:22 70 FALSE07-May-04 00:00:23 70 TRUE07-May-04 00:00:24 72 TRUE07-May-04 00:00:47 70 FALSE07-May-04 00:00:48 70 TRUE07-May-04 00:47:25 72 TRUE07-May-04 00:47:41 72 TRUE07-May-04 00:47:42 70 FALSE07-May-04 00:47:46 72 TRUE07-May-04 00:47:54 72 TRUE07-May-04 00:47:57 100 TRUE07-May-04 00:48:06 71 TRUE07-May-04 00:48:09 100 TRUE
Raw Data:• Contains logical inconsistencies
• Subject to intermittent errors
• Huge amount of events captured
• Decision making needs abstract data:
when, where and what
Pre-processing Level - I
• Grouping of similar events into blocks
• Categorising sensors into location and activity types
• Capturing additional sensor information e.g. activity levels, silence
• Abnormal toggle sensor detection
• Sensor location consistency checkIn
terface to S
enso
r Netw
ork
User Interface
Sensor Object Management
AnalysisAlgorithms
Raw Data Pre-processing
Pre-Processed Data ReportData
Sensor Information
Report Generator
Date Name SensorID Start Time Duration Comment06-May-04 BEDROOM_PIR 72 22:51:33 01:56:21 location07-May-04 BED_OCCUPANCY 70 00:00:14 00:00:08 activity_T07-May-04 BED_OCCUPANCY 70 00:00:23 00:00:24 activity_T07-May-04 BED_OCCUPANCY 70 00:00:48 00:46:54 activity_T
07-May-04 SILENCE 110 00:00:49 00:46:35BED_OCCUPANCY-
BEDROOM_PIR07-May-04 LANDING_PIR 100 00:47:57 00:00:00 location07-May-04 BATHROOM_PIR 71 00:48:06 00:00:00 location07-May-04 LANDING_PIR 100 00:48:09 00:00:00 location
Pre-processing Level - II
Interface to
Sen
sor N
etwo
rk
User Interface
Sensor Object Management
AnalysisAlgorithms
Raw Data Pre-processing
Pre-Processed Data ReportData
Sensor Information
Report Generator
• Using Fuzzy values for start time and duration of sensor event blocks
• Converts time axis into a more meaningful and manageable number of regions
• Allows reasoning with Fuzzy rules
• Duration membership functions are learnt for each sensor and each client
Bed Occupancy
Kitchen Occupancy
Analysis Techniques
• Provide answers to core activity questions
– Abstraction of sensor events into daily activities
– Identify key points in the data sequence (e.g. data silences)
– Analyse surrounding sensor data to classify activities (e.g. bed in use - asleep)
– Trend analysis
Interface to
Sen
sor N
etwo
rk
User Interface
Sensor Object Management
AnalysisAlgorithms
Raw Data Pre-processing
Pre-Processed Data ReportData
Sensor Information
Report Generator
Visitor Activity:
Challenge: No identification sensors (non-intrusive
sensing)
To infer the existence of visitors:
• Various metrics are used to describe regions of
activity
– Activity Levels
– Delay in activity level changes
– Non-adjacent rooms activity
• Regions between entrance events are then
compared
• Changes are used to accumulate evidence of visitor
activity
Activity Specific Analysis
• Use of Activity Levels for visitor evidence accumulation:
• Average level of activity between door events
• The delay after a door event before behaviours show change
• Region B clearly has a higher rate of activity
• Sliding a window W across region B shows that the change in activity level occurs at the beginning of region B
VisitorActivity
• A Changing Rooms measure comprises:
• The rate at which the location of activity changes
• The proportion of changes between non-adjacent rooms
• The longest sequence of consecutive non-adjacent changes
All Room Changes Non-Adjacent Changes
Non-Adjacent Bursts
• Fuzzy rules are used to accumulate the evidence of a visit
VisitorActivity
Training to obtain fuzzy membership functions
• At each door event in the training set the changes in activity level are recorded
• The results are split into ordered negative and positive changes
• Each half is split into three, using their 1st & 3rd quartiles to define fuzzy sets
• The fuzzy sets of each half are combined, then averaged with initial sets
Big Rise
Big Rise
Q1Q3Q1 Q1 Q1Q3 Q3 Q3
0 0 -1 1 1 -1
RiseSteadyBig Fall Fall
Act. Level Change Act. Level Change
A AB C D EF CD E F
Activity Level Change
1 -1
Big RiseRiseSteadyBig Fall Fall
Activity Level Change 1 -1
RiseSteadyBig Fall Fall
Activity Level Change
Initial Sets (Equally Sized)
Negative and Positive Changes Combined Sets (Equal Data Share)
Final Sets (Averaged)
VisitorActivity
Important Notes:
• Data splits are {A: 40%, B: 40%, C: 20%}, {D: 20%, E: 40%, F: 40%}
• Equally sized sets are not sensitive to an individual’s data spread
• Sets using equal data can be over-sensitive when data spread is uneven
• Averaged sets used are a compromise, avoiding both problems
• The same style of training is applied to the changing rooms measure
VisitorActivity
User GUI
User GUI
Deployment
Client- Web browser based - Thin client- Multi-user- Pure HTTP/HTTPS communication
Back End- Extensive calculation on server - SQL database driven
SQL
App. Server
DBThin client
High Level Application objects over HTTP