Upload
hanh
View
19
Download
1
Tags:
Embed Size (px)
DESCRIPTION
Technical Solutions Underlying Wireless Health Systems. John A. Stankovic BP America Professor Department of Computer Science University of Virginia. http://www.cs.virginia.edu/wsn/medical/ http://wirelesshealth.virginia.edu/. What’s Wrong With Wires. - PowerPoint PPT Presentation
Citation preview
Technical Solutions
Underlying Wireless Health
Systems
John A. StankovicBP America Professor
Department of Computer ScienceUniversity of Virginia
http://www.cs.virginia.edu/wsn/medical/http://wirelesshealth.virginia.edu/
What’s Wrong With Wires
And we don’t want a patient tethered to a bed or fixed medical device
Realisms and Main Points
• Realisms– Humans and their behaviors are not simple
• Example: Sleep
– Environments are not simple• Example: Acoustics
Realities – Sounds Encountered
Physiological: Sneezing, nose blowing, sniffling, clearing throat, hiccup, eating, burp, humming, laughter, drinking, snoring
Objects: phone vibrating or ringing, typing, mouse wheel, unwrapping food, papers rustling, clothes rustling, television, piano, moving furniture, doors opening and closing, objects dropping or moving, footsteps, pouring liquid, coffee percolation, dishwasher, cleaning sounds
Ambient: truck backing up, siren, birds chirping, passing airplane, traffic, motorized tools (lawnmower, etc)
Main Point
• Many current solutions work ONLY when humans and environments are (assumed to be) very constrained
Outline
• Motivation
• Architecture
• ADLs and Semantic Anomaly Detection
• Control Heart Rate
• Vision: Aggression in Dementia Patients (not today)
• Acoustics: Reverberation (not today)
Motivation - The Problems
• Aging Populations• High Cost of Medical Care• Lack of Facilities• Quality of Life Issues
• Solution: Home Health Care CCRC Assisted Living
Vision - Smart Living Space• Humans-in-Loop• Heterogeneous• Evolution• Open• Privacy
Reality: NOT SIMPLE
Research Q: Architecture
• Alarmnet • Web based (Empath)
InferenceDB
Med Records
WebServer
Our Solution: Empath
• Realisms– Off-the-shelf
– Non-expert user
– Easy installation and maintenance
– Remote monitoring
– Extensible (HW)
– Evolve (Behavior)
– Patients are very different than grad students
• Main Points– Web-like architecture
is quite flexible– Single hop
Cloud Controller
Inference Modules
Monitoring Modules
MongoDB
Web Server
CDMA Modem
4G
MQTT Broker
Home Controller
System Repository
Monitoring Modules
Collectors Sensors
Base Station
Architecture
Applications – Versions of Architecture
• Depression and General Anomalies
• Dementia and Incontinence
• Exercise and Epilepsy
Depression Detection andGeneral Anomalies
• Multi-modal
• Passive
• Combines Objective and Subjective Measures
Empath: Depression Monitoring
Motion and Contact Sleep Data PHQ-9 Acoustic Weight
DB
Eating Sleep Quality Movement Mood Weight Gain/Loss
Patient DisplayCaregivers
Display
Depression Inference
Cloud
Caregivers Display
Dementia and Incontinence
• Bedwetting and sleep agitation relationship
• Interventions on sleep
• Sleep labs impractical
• Goal: 50 patients
Dementia and Incontinence(subsystem of Empath)
• Passive Sleep system uva built
• UVA Tempo device uva built
• Wetness sensors – Dry Buddy off-shelf
• Acoustic sensors – Screaming off-shelf
ID Age Sex No. of Nights
No. of Nights Wetness Detected
No. of Nights Wetness Reported
Who Deoplyed
Deployment Time (Minutes)
No. of Intermediate Visits
1 65 Female 6 2 5 Grad. Student
20 0
2 87 Male 8 5 5 Grad. Student
20 0
3 71 Male 5 4 5 Grad. Student
20 0
4 73 Female 5 5 5 Nurse 30 0
5 79 Male 5 0 0 Nurse 30 1
6 88 Male 5 4 5 Nurse 30 1
7 89 Female 5 5 5 Nurse 30 0
Example Correlations
Epilepsy Study(subsystem of Empath)
• Linkage: between exercise, stress, sleep and epilepsy seizures
• Measure: HR, BP (CareTaker), sleep quality, frequency of seizures
• Intervention: qigong (Chinese healing art)
Many Seizures at Night
Status
• Depression – 2 homes
• Dementia – 11 real patients (on-going)
• Epilepsy – 2 real patients
Research Q: (Robust) Activities of Daily Living
• Monitor eating, sleeping, exercising, …
• What’s normal
• Detect anomalies
• Intervene: send alarms, inform someone, prevent, …
Realisms
• Activity Recognition (AR) of ADLs– Higher accuracy required– Overlapped activities– Across room activities– Many realities (missing data) ….
• Anomaly detection– Accurate AR and what’s normal support good
anomaly detection
Main Point
• Normal behavior is very complex– Per day– On Wednesdays– Two times per week– Every other month– In summer when condition X exists– Grouping of activities– Context dependent– …
Big Picture
28
Motion Sensors
Contact Sensors
Bed Sensors
Body Sensors
Pre-Processing
Activity Recognition
Sleep Eat Toilet Shower Exercise
Learn Regular Behavior
Entertainment
Behavioral Anomaly Detection
AALO (Active Learning based Activity recognition in the presence of Overlapped activities)
29
Motion Sensors
Contact Sensors
Bed Sensors
Body Sensors
Pre-Processing
Activity Recognition
Sleep Eat Toilet Shower Exercise
Learn Regular Behavior
Entertainment
Behavioral Anomaly Detection
AA
LO
Overla
pped
Acti
vities
Holmes: A comprehensive anomaly detection system
30
Motion Sensors
Contact Sensors
Bed Sensors
Body Sensors
Pre-Processing
Activity Recognition
Sleep Eat Toilet Shower Exercise
Learn Regular Behavior
Entertainment
Behavioral Anomaly Detection
Hol
mes
Context-basedMultiple Models
Temporal & CausalCorrelations
Semantic RulesReduce False
Positives & Negatives
31
Robust Activity Recognition
• Accurate detection and summary of daily activities are essential
• Proper training and periodic retraining are necessary
• Necessary to ensure user comfort
31PervasiveHealth 2012
Overlapped Activities
• Interrupted by room changes
• Concurrent in the same room
• Across rooms (not done yet)
PervasiveHealth 2012 32
33
Room Level Segmentation
• First consider spatial regularity
• In real life:– Within one episode, multiple activities may take place
• We use Itemset identification
– An activity may not complete within one episode
PervasiveHealth 201233
Segmentation into Occupancy Episodes
(<timestamp, sensor-firing> pairs)(<roomID, entranceTime, duration, usedSensors>)
34
Room Level Segmentation
• Successive Occupancy Episode Merging (SOEM)– to identify overlapped activities interleaved across
multiple rooms
• We construct a new occupancy episode by merging two occupancy episodes in the same room that are separated by less than a time interval threshold
PervasiveHealth 201234
35
Example Evaluation
Public Dataset
• Single resident home• Activity annotation by the
resident using Bluetooth headset
• 26 days of data• 4 rooms (Bedroom,
Kitchen, Toilet, Shower)• ‘Out of house’ considered
as another room
List of Sensors
PervasiveHealth 201235
kitchen_microwave kitchen_groceries
kitchen_cups kitchen_fridge
kitchen_plates toilet_door
kitchen_dishwasher toilet_flush
kitchen_freezer shower_door
kitchen_pans bedroom_door
kitchen_washingmach front_door
36
Effectiveness of SOEM
PervasiveHealth 201236
• Accuracy of ‘Sleep’ & ‘Prepare Dinner’ detection improved• SOEM does not reduce the accuracy for any activity
37
Comparison with Supervised Algorithms
PervasiveHealth 201237
Semantic Anomaly Detection
• Sensor Level Anomalies
• Activity Level Anomalies– Point– Context– Collective
• Semantics
Sensor Level Anomalies
• May be due to different types of sensor failures– Fail-stop failure (may be detected by Ping)– Stuck-at failure– Transient– Obstructed field-of-view– Sensor movement
Sensor Level Anomalies
• How to detect:– Use temporal and spatial correlation among
different sensors
– Use multiple classifiers
– Many systems assume no sensor failures!
K. Kapitanova, E. Hoque, D. Alessandrelli, J. Stankovic, S, Son and K. Whitehouse, Being SMART About Failures: Assessing Repairs in Activity Detection, Ubicomp, Sept. 2012.
Activity Level Anomalies
• Point Anomalies– For each activity, one or more models based on
training data
– Different set of features for different activities• Sleep: sleep onset time, duration, number of
interruptions, movement level• Showering: Start time, duration, water usage
– Use statistical anomaly detection
– Too many false alarms if used alone!
Context Anomalies
• For each activity, one or more models based on training data– Daily, specific day / days of week, specific
time periods in a day
42
Collective Anomalies
• Monitor instances of same activity for N days collectively– Example: exercise twice per week
• Monitor whether other activities happen before/after/during that activity in each time period– Example: have coffee before breakfast
• Monitor instances of all activities for N days collectively– Example: every week shop, exercise, see doctor, …
Semantics
• Watching a pet for a few days
• Entertaining visitors
• Power outage
• Recovering from major medical operation
• Sensors died
• Extreme weather: cold/heat wave
• Social security check did not arrive: cannot purchase food or medication
• Major medication change
• Life changing events: sibling etc died, new grandchild etc.
Holmes Framework
45
Training Data
Semantics: Expert Knowledge & User Feedback
Explainable Scenarios
Learn Regular Behavior
Calculate Anomaly Score and Initial Filter
Testing Data
Regular Behavior Models
True Anomalies
SemanticAnomaly Filters
MusicalHeart
• Music non-invasive recommendation system to control heart rate– Bio-feedback (human-in-the-loop)– Context aware
• Sitting, lying, walking, jogging, exercising• Home, traveling
– Personalized– Learns
Contexts: jogging, relaxing, traveling, etc.
Wellness - Musical Heart
Realisms
• Movement (rustling) of ear buds
• Music reaction is personal
• Max HR per person• Context dependent• Impact changes over time
Main Points
• Human in the Loop• PI Control matches well
Calm Your Heart
Exercise Your Heart
3 Parts to Musical Heart
Sensor EquippedEarphones
Smartphone AppCloud
Musical Heart: Hardware (SEPTIMU)
• SEPTIMU v2 • Sensors: IMU Microphone IR Reflective Sensor Thermometer
• Computation: Tiny OS
• Communication: Audio Jack Bluetooth
• Power: Li-Polymer battery
Beijing, China
Musical Heart: Software Architecture
In-phone SensorsIn-phone Sensors
Heart RateDetector
Heart RateDetector
Activity LevelDetector
Activity LevelDetector
ContextDetectorContextDetector
Sensor Data DispatcherSensor Data Dispatcher
Septimu SensorsSeptimu Sensors
SeptimuMic.
SeptimuMic.
SeptimuAcc.
SeptimuAcc.
In-phoneGPS
In-phoneGPS
In-phoneWiFi
In-phoneWiFi
CombinerCombiner
MusicPlayerMusicPlayer ProxyProxy
Music Recommender
Sensing
Heart Rate and Activity Detector
MusicRecommender
Algorithm 1: Heart Rate Detection• Filtering: A low pass filter to remove non-
heart beat signals.
Signals from Ear(music + heart beats)
After filtering(heart beats)
music
Detect Heart Rate
• V1: Microphone– Movement of ear phone, ambient sounds, music,
…– Dynamic programming solution– Accuracy: (1-7 bpm error)
• V2: LED– Accuracy: 1-2 bpm
Algorithm 2: Activity Level Inference
Threshold1
Threshold2
We use k-means clustering to learn the thresholds
Result: Activity Level Inference
L1 L2 L3
L1 0.9998 0.0002 0
L2 0 0.9997 0.0003
L3 0 0.0280 0.9720
Predicted level
Act
ual le
vel L1 L2 L3
L1 0.989 0.011 0
L2 0 0.951 0.049
L3 0 0.037 0.963
Predicted level
Act
ual le
vel
• Single Activity • Sequence of Activities
99.1% 96.8%
Algorithm 3: Music Recommendation
I
PProcessSet Point
Proportional
Integral
Algorithm 3: Music Recommendation
I
PSet Point
Proportional
Integral
Process: Human Heart
CurrentHeart Rate
Algorithm 3: Music Recommendation
I
P
Desired Heart Rate
Chart
CurrentHeart Rate
Set-point: Desired Heart Rate
Proportional
Integral
Algorithm 3: Music Recommendation
I
P
Desired Heart Rate
Chart
CurrentHeart Rate
Find a matching song
Reflex
History
u
Deployment: Goal directed jogging • Typical Cardio Exercise Plan:
Time Desired Intensity (%)
5 min 60% - 70%
3 min 70% - 80%
2 min 80% - 90%
3 min 70% - 80%
2 min 80% - 90%
5 min 60% - 70%
Result: Achieving Desired Heart Rate
Time Desired Intensity (%)
5 min 65%
3 min 75%
2 min 85%
3 min 75%
2 min 85%
5 min 65%
Result: Achieving Desired Heart Rate
Time Desired Intensity (%)
5 min 65%
3 min 75%
2 min 85%
3 min 75%
2 min 85%
5 min 65%
Desired Intensity
Achieved Intensity
Control Input(music feature, u)
u
Result: Achieving Desired Heart Rate
Time Desired Intensity (%)
5 min 65%
3 min 75%
2 min 85%
3 min 75%
2 min 85%
5 min 65%
Desired Intensity
Achieved Intensity
Control Input(music feature, u)
Result: Achieving Desired Heart Rate
Time Desired Intensity (%)
5 min 65%
3 min 75%
2 min 85%
3 min 75%
2 min 85%
5 min 65%
Desired Intensity
Achieved Intensity
Control Input(music feature, u)
Result: Achieving Desired Heart Rate
Time Desired Intensity (%)
5 min 65%
3 min 75%
2 min 85%
3 min 75%
2 min 85%
5 min 65%
Desired Intensity
Achieved Intensity
Control Input(music feature, u)
Result: Achieving Desired Heart Rate
Time Desired Intensity (%)
5 min 65%
3 min 75%
2 min 85%
3 min 75%
2 min 85%
5 min 65%
Desired Intensity
Achieved Intensity
Control Input(music feature, u)
Result: Achieving Desired Heart Rate
Time Desired Intensity (%)
5 min 65%
3 min 75%
2 min 85%
3 min 75%
2 min 85%
5 min 65%
Desired Intensity
Achieved Intensity
Control Input(music feature, u)
Result: Achieving Desired Heart Rate
Time Desired Intensity (%)
5 min 65%
3 min 75%
2 min 85%
3 min 75%
2 min 85%
5 min 65%
Desired Intensity
Achieved Intensity
Control Input(music feature, u)
Result: Achieving Desired Heart Rate
Time Desired Intensity (%)
5 min 65%
3 min 75%
2 min 85%
3 min 75%
2 min 85%
5 min 65%
Desired Intensity
Achieved Intensity
Avg. Deviation: 11.4%Control Input
(music feature, u)
Summary
• 1000s of deployments for WH
• Technical Solutions– Architecture– AR and Anomaly Detection– MusicalHeart– Vision: Kintense– Acoustics: in the home (reverberation)
Summary
• (Personalized) Wireless Health– Body Sensor Networks– Environmental and AR Networks
• Easy to Modify over Time– Incorporate new technology as it becomes
available– Adapt as medical conditions change
Kintense – Detecting Aggressive Behavior
• Main Point– Cohen-Mansfield metrics for dementia patients
• Hitting• Kicking• Throwing• Punching• …
– Usable in other contexts, e.g., reception areas, …
Our Solution: Kintense
• Kinect
• Array of Supervised Learners
• Unsupervised Learner• Human input
• Robust – People– Orientation– Speed– In-home obstacles
• Performance– DTW vs Gesture
Recognition vs Kintense
Kintensehttp://www.youtube.com/watch?v=LUJLXcPFMpg
Resonate• Get 1 clean speech sample in each room for
each person• Use reverberation impulse response simulator
– Generate many training samples at different locations
• Create suite of classifiers• Tested
– 6 different types of rooms and long term deployments in home and office
– Close to ideal performance
PervasiveHealth 2012 78
Reducing User Labeled Data
• User needs to label only 19 clusters
• Any supervised algorithm requires labeling of 291 activity instances
• With supervised algorithms, amount of user labeling proportionally increases with number of training days