S Social Sensing for i l S i f Epidemiological p de o og ca - - PowerPoint PPT Presentation

S i l S i f Social Sensing for Epidemiological p de

• og ca

Behavior Change

Presented by: Michal Dobosz Michal Dobosz

CS525M: Mobile and CS525M: Mobile and Ubiquitous Computing Spring 2011

Definition of Epidemiology Definition of Epidemiology

• “Epidemiology is the study of

Epidemiology is the study of patterns of health and illness and associated factors at the population associated factors at the population level.”

Outbreak investigation – Outbreak investigation – Biology Biostatistics

2

– Biostatistics – Social Science disciplines

Worcester Polytechnic Institute 2

Introduction Introduction

• How is individual behavior affected

How is individual behavior affected by illness and stress?

• Measure characteristic behavior
• Measure characteristic behavior

change in symptomatic individuals

M bil h li ti – Mobile phone application

• Co-location
• Communication

3

• Communication
• Predict health status of an individual

Worcester Polytechnic Institute 3

Benefits Benefits

• Understanding how people behave

Understanding how people behave when they are infected

Lack of realistic social interaction data – Lack of realistic social interaction data and spatio-temporal data

• Modeling can be made more accurate
• Modeling can be made more accurate

– Results can be used in the SIR model

Number and frequency of contacts on

4

• Number and frequency of contacts on

Susceptible -> Infected transition

– Face-to-face interaction in contagion

Worcester Polytechnic Institute 4

Face to face interaction in contagion process

The Experiment The Experiment

• Two months of data from an

Two months of data from an undergraduate residence hall

– Individuals surveyed daily for symptoms y y y p – Behavioral changes when individuals are sick

• Total communication, communication

patterns, network diversity, entropy of movement

5

movement

Worcester Polytechnic Institute 5

Related Work Related Work

• Mobile Phones as Social Sensors

Mobile Phones as Social Sensors

– Eagle and Pentland

• Reality Mining – social network structure,

y g and recognition of patterns in daily user activity

Gonzalez et al – Gonzalez et. al

• Call detail records used to characterize

spatio-temporal regularity

6

• Google Flu Trends

Worcester Polytechnic Institute 6

Related Work Cont. Related Work Cont.

• Sociometric

Sociometric Badge

– Identify human y activity patterns and analyze con ersational conversational prosody features – Vocal features

7

– Vocal features, body motion, relative location

Worcester Polytechnic Institute 7

Data Source Data Source

• Undergraduate Dormitory

Undergraduate Dormitory

– 80% participated in the study, most of the remaining 20% were spatially isolated – Pro-technology orientation – Even distribution among academic years – 54% males and most were Engineering, Mathematics, and Science majors

8

• Incentives

– Windows Mobile Phones and $1 a survey

Worcester Polytechnic Institute 8

Data Sets Data Sets

• Social Interaction Data from Mobile

Social Interaction Data from Mobile Phones

– Call data records – SMS logs – Bluetooth proximity and WLAN location p y sensing (every 6 minutes)

• Symptom Data via Daily Self-Report

9

y p y p

– Physical and Emotional Symptoms – 20/69 participants FLU immunized

Worcester Polytechnic Institute 9

p p

Survey Questions Survey Questions

• Do you have a sore throat or cough?

y g

• Do you have a runny nose, congestion or

sneezing?

• Do you have a fever?
• Have you had any vomiting, nausea or

diarrhea? diarrhea?

• Have you been feeling sad, lonely or

depressed

10

p

• lately?
• Have you been feeling stressed out lately?

Worcester Polytechnic Institute 10

y g y

Survey Data Survey Data

• Immunized Participants not

Immunized Participants not considered

• Survey Data

Survey Data

– 63% survey completion rate – Grouped into 48-hour periods Grouped into 48 hour periods – Symptoms labeled as FLU by medically trained epidemiologist

11

p g

• 12 cases identified, lasting 5-7 days

Worcester Polytechnic Institute 11

Sensor Data Sensor Data

• Total Communication – Phone Calls and

SMS

• Communication (10PM – 9AM on weekdays)
• Communication Diversity
• Physical Bluetooth Proximity day and night

(10PM – 9AM on weekdays)

• Physical Bluetooth Proximity excluding

t d t

12

students

• University WLANs and non-University

WLANs

Worcester Polytechnic Institute 12

WLANs

Behavioral Effects of Low Intensity Symptoms (Runny y p ( y Nose, Sore Throat and Cough)

13 Worcester Polytechnic Institute 13

Behavior Effects of Higher-Intensity Symptoms (Fever and Influenza) Symptoms (Fever and Influenza)

14 Worcester Polytechnic Institute 14

Behavior Effects of Fever Behavior Effects of Fever

15 Worcester Polytechnic Institute 15

Behavior effects of CDC- d fi d i fl defined influenza

16 Worcester Polytechnic Institute 16

Symptom Classification Using B h i l F t Behavioral Features

• Cell phones can predict illness

p p

• K-nearest-neighbor-clustering

– stress + depression p – runny nose + sore throat – fever + influenza – runny nose + sore throat + fever + influenza

17

• Bayesian-network classifier with MetaCost

– Accuracy between 60% - 80%

Worcester Polytechnic Institute 17

Temporal Flux Between Behavior, St d Ph i l S t Stress and Physical Symptoms

• Granger causality

Granger causality test

– Poor noise immunity

• Phase Slope

p Index (PSI) Method

18 Worcester Polytechnic Institute 18

Conclusions Conclusions

• Limitations

Limitations

– Bluetooth signal strength – Statistical tests assume independent Statistical tests assume independent samples

• Doctors and nurses can use

diagnostic information

– Early detection of conditions

19

y – Better healthcare – Lower costs

Worcester Polytechnic Institute 19

References References

• Social Sensing for Epidemiological

g p g Behavior Change, Anmol Madan, Manuel Cebrian, David Lazert and Alex Pentland, MIT Media Lab and Harvard University MIT Media Lab and Harvard University, Cambridge MA

• http://hd media mit edu/badges/
• http://hd.media.mit.edu/badges/
• http://www.google.org/flutrends/
• http://en wikipedia org/wiki/Epidemiology

20

• http://en.wikipedia.org/wiki/Epidemiology

Worcester Polytechnic Institute 20