HealthSense: Classification of Health-related Sensor Data through User-Assisted Machine Learning

1
HealthSense Classification of Health-related
Sensor Data through User-Assisted Machine
Learning
From Prof. Gregory D. Abowd
Presenter Mi Zhang Feb.
23rd, 2009
2
Outline

• Part I Background
• Part II HealthSense Platform
• Part III Experiments
• Part IV Future Work

3
Part I Background

• Problem Overview
• Research Methodology
• System Features Contributions

4
Problem Overview

• How to automatically detect Directly Undetectable
  health-related conditions?
• What is Directly Undetectable condition?
• Can NOT be detected by direct sensing technology
• Example Pain, Depression, Itching

5
Research Methodology

• Assumption
• Assume the occurrence of Directly Undetectable
  Conditions is correlated with events that can be
  observed based on patient feedback.
• Adopt techniques from Activity Gesture
  Recognition
• Model Design Methodology
• Build up an initial model by a preliminary
  supervised learning procedure
• Use user-query feedback and machine learning
  techniques to continuously improve the model

6
System Features Contributions

• System Features
• Update the system with Online Supervised Learning
• Use patient inputs/feedback to assist with
  classification
• Both techniques contribute to classification
  accuracy
• Contributions
• A seminal work on detecting directly undetectable
  health-related events from on-body sensor streams

7
Part II HealthSense Platform

• Platform Architecture
• Feature Extraction Classification Strategy
• User Query Process

8
Platform Architecture

• Classic 3-tier Architecture
• Tier 1 Sensor Tier
• Witilt 3-Axis Accelerometer Bluetooth interface
• Communicate with Tier 2 via Bluetooth
• Tier 2 Mobile Device
• Nokia N800 PDA Bluetooth, 802.11, running Linux
• Communicate with Tier 3 via 802.11
• Tier 3 Back-end Server
• Web Server Apache Tomcat
• Database Apache Derby
• Machine Learning Engine Weka
• GUI Berkeley PtPlot

9
Feature Extraction Classification Strategy

• Extracted features
• Frequency-Domain Energy
• Product- Moment Correlation Coefficient
• Standard Deviation
• Root Mean Square (RMS)
• Classification Strategy
• Each Event Window is a classification unit
• Two Categories Occurrence, Non-occurrence

10
User Query Process

• Strategy
• Server queries the user ONLY when positive
  classification occurs
• Does NOT handle negative classification
• How
• Server detects a positive classification
• Sends a SMS to Mobile hub
• GUI of Yes/No questions
• Mobile hub sends a SMS back to server

11
Part III Experiments

• Case Overview What Why?
• Step 1 Choosing the right sensors
• Step 2 Choosing the right features
• Step 3 Choosing the right window size
• Step 4 Choosing the right classifier
• 4 Experiments Results Analysis

12
Case Study Detecting An Itch

• What
• Detect a scratching
• Differentiate normal daily scratching from
  medically relevant scratching
• Why
• Detecting pain or depression is infeasible at
  this stage
• Itch is also Directly Undetectable
• Critique NOT a direct proxy for Pain Depression

13
Step 1 4

• Step 1 Choosing the right sensors
• 2 wrist-mounted 3-Axis Accelerometers
• Step 2 Choosing the right features
• Frequency-Domain Energy, Product- Moment
  Correlation Coefficient, Standard Deviation, Root
  Mean Square (RMS)
• May NOT be the most appropriate features for this
  case
• Step 3 Choosing the right window size
• In this case 128 samples _at_ 60Hz (Case by Case)
• Tradeoff Accuracy vs. Decision Period
• Step 4 Choosing the right classifier
• Neural Network, Decision Tree, K-Nearest
  Neighbors, Bayesian Network (Naive Bayes)
• In this case Decision Tree performs the BEST

14
Experiment Results

• Test 1 No feedback
• Accuracy 63 73
• Accuracy remains fairly constant throughout the
  test
• Test 2 No scratching, Has users Feedback
• Accuracy 62 93
• Feedback helps a lot
• Test 3 Has scratching, Has users Feedback
• Accuracy 62 93
• Feedback helps a lot
• Test 4 With Feedback, differentiate normal daily
  scratching from medically relevant scratching
• Require a priori knowledge of the locations of
  medically relevant scratches
• Accuracy 81 100
• Feedback helps a lot

15
Part IV Future Work

• Expand feature pool Indentify important
  features
• Goal Improve Classification Accuracy
• Add more sensors (Different types, numbers,
  places)
• Goal Improve Classification Accuracy
• Improve their model Handle false negatives
  properly
• Goal Improve Classification Accuracy
• Method I Allow patients to voluntarily notify
  the system when system fails to detect scratch
  (Not good timing issue)
• Method II Manually validate of both positive and
  negative classification (Not good Annoy and
  obtrusive)

16
Questions

• Any Questions ? No

17

• Thank You Very Much !