Healthwear: Medical Technology Becomes Wearable

3479 users shared this document! click

0018-9162/04/$20.00 © 2004 IEEE 34 Computer Healthwear:
Medical Technology
Becomes Wearable T he concept of computing is rapidly
expanding from simply using a desktop
PC, where people sit and type for a small
part of the day. Every day, more than one
billion people carry around portable com- putation devices that have sensors and Internet-
capable connections—but we call them cell phones
rather than computers. The most recent cell phones go far beyond tele- phony: They are truly wearable computers. These
location-aware devices have sensors for detecting
sounds, images, body motion, and ambient light
level, have a secure Internet connection, and can
download and upload programs as well as audio
and image ?les. They also can serve as a situation-
aware intelligent assistant, whether as personal
agents that use the digital equivalent of 3M’s
Post-it notes to augment reality or as a means of
forming tight-knit intellectual collectives in which
people can supercharge their social networks. As part of this change in the way we use com- puters, my research group at the MIT Media Lab
(http://hd.media.mit.edu) has been developing
healthwear, wearable systems with sensors that can
continuously monitor the user’s vital signs, motor
activity, social interactions, sleep patterns, and other
health indicators. The system’s software can use the
data from these sensors to build a personalized pro-
?le of the user’s physical performance and nervous
system activation throughout the entire day—pro-
viding a truly personal medical record that can, we
believe, revolutionize healthcare. HEALTHWEAR OVERVIEW Until recently, researchers have had little success in extending healthcare into the home environment,
yet there clearly is a huge demand for this service.
Americans currently spend $27 billion on healthcare
outside the formal medical establishment because
they ?nd it dif?cult to access, expensive, and painful
(www.rwjf.org). A clear demand for better integrat-
ing the home into the healthcare environment exists.
Not only that, but a dramatic shift in the composi-
tion of the US population makes it absolutely neces-
sary to develop such distributed systems. Caregiver shortage Although the US had 25 caregivers for each dis- abled person in 1970, the success of our healthcare
system will lower the ratio of caregivers to at-home
disabled to 6 to 1 by 2030 (www.agingstats.gov).
How will those six people care for a disabled per-
son? Certainly, a centralized system of visiting
nurses is not an option for providing this care—
such a system would leave too few individuals
working at other jobs in the economy to support it.
Thus, a more highly distributed system is not only
desirable, but absolutely necessary. These statistics provide the driving force behind the development of healthwear. This concept offers
an unobtrusive method for acquiring in-depth
knowledge about the body that could help manage
chronic medical conditions such as cancer, diabetes,
degenerative disorders of the nervous system, or
chronic pain. Perhaps just as importantly, the deploy- Widespread adoption of sensors that monitor the wearer’s vital signs and
other indicators promises to improve care for the aged and chronically ill
while amassing a database that can enhance treatment and reduce
medical costs. Alex (Sandy) Pentland Massachusetts
Institute of
Technology P u b l i s h e d b y t h e I E E E C o m p u t e r S o c i e t y C O V E R F E A T U R E ment of continuous monitoring devices provides an
excellent opportunity to fully inform medical
providers about a patient’s condition, thus helping
the patient obtain the best treatment possible. Already, health-conscious individuals are wear- ing small digital pedometers and exercise monitors.
Indeed, some companies such as Nissan in Japan
give such devices to employees to heighten health
awareness and decrease medical insurance costs. In
the future, people who dress for success may also
wear a healthwear personal trainer that helps keep
them active, knowledgeable, and involved. Opportunities and concerns As new sensor, computing, and communication technology becomes available, healthcare profes-
sionals will be able to organize huge medical data-
bases for use in tracking every test taken and
medicine prescribed over an individual’s lifetime. In
addition to helping drive down healthcare costs, this
data can provide powerful epidemiological infor-
mation for use in improving our knowledge about
keeping society healthy. For example, today because
the huge expense of clinical trials limits the size and
sensitivity of drug testing, harmful interactions are
often detected only months or years after a drug is
introduced to the general populace. Continuous,
quantitative behavior logging has the potential to
generate enough data so that researchers could dis-
cover these interactions more quickly. Another application that is potentially even more important is the early detection of epidemics like
SARS or biological weapons attacks. Today, reports
of the treatment of an unusual number of patients
with similar symptomatology at a medical facility
often provide the ?rst warning of a potential epi-
demic. Widespread continuous monitoring could
detect such outbreaks much sooner by noticing
when unusual numbers of people are behaving
lethargically or staying home from work. However, creating such an information archi- tecture requires safeguards to maintain individual
privacy. Indeed, we believe that this issue demands
immediate, thoughtful attention and public debate,
perhaps beginning with the current concern about
using cell phone signals to track people. The current
forces for creating huge databases and big medi-
cine are powerful and all too successful. The poten-
tial solution is to place control and ownership of
as much personal information as possible in the
hands of the individual user, sharing only infor-
mation cleansed of identifying features. This power-
to-the-people approach favors using wearable
sensing devices rather than sensors in the sur- rounding environment because the information
starts out in the control of the individual, and the
legal tradition in the US is that individuals own the
data collected from their bodies. MITHRIL In J.R.R. Tolkien’s Middle Earth stories, mithril is a precious metal used to craft armor with prop-
erties that protect its wearer from evil. The term
thus seems an apt name for the technology that pro-
vides the basis for healthwear. Highly ?exible, the
MIThril architecture provides a modular system
tied together by wireless networking protocols and
a uni?ed multiwired-protocol power and data bus
for sensors and peripherals. 1 Hardware components Figure 1 shows the MIThril system. Designed for use with either a modern programmable cell phone
or a wireless personal digital assistant (PDA),
MIThril offers input, output, and general compu-
tation functions and can support a wide range of
physiological measurements. 1,2 The MIThril hard- ware architecture is designed to be modular and
easily con?gurable so that it can handle a variety of
sensors and tasks. The software architecture sup-
ports using the ad hoc, on-the-?y combination of
sensor signals from multiple users to control sig-
naling and outputs. A sensor hub interfaces with the MIThril body bus, which combines the Philips I2C multiple-
device serial protocol and power lines. The sensor
hub provides a bridge to the sensor data, enabling
data acquisition, buffering, and sequencing, and it
can be used as a stand-alone data-acquisition sys-
tem. 2 This is particularly useful for large-group applications that do not require real-time process-
ing, wireless communication between users, or May 2004 35 Figure 1. MIThril
system. Plugging
the biosensor hub
into a cell phone
or wireless PDA
provides a system
that offers input,
output, and general
computation
functions and can
support a wide
range of
physiological
measurements. 36 Computer complex user interaction and thus do not
require a cell phone or wireless PDA to be
part of the system. Currently supported devices include accelerometers for motion detection, IR
active-tag readers for location and proxim-
ity detection, audio input and output devices,
battery monitors, GPS, analog two-channel
EKG/EMG, two-channel galvanic skin
response sensors, and skin-temperature sen-
sors. MIThril uses an RS-232 interface to
communicate with a wide range of commer-
cially available sensors for monitoring pulse
oximetry, respiration, blood pressure, EEG,
blood sugar, and CO 2 levels. Software architecture The core MIThril software components include the Enchantment Whiteboard, the Enchantment
Signal system, and the MIThril Real-Time Context
Engine. These tools provide the foundation for
developing modular, distributed, context-aware
wearable and ubiquitous computing applications. The Enchantment Whiteboard implements an interprocess communications system suitable for
distributed, lightweight, embedded applications.
Unlike traditional interprocess communications
systems such as RMI and Unix/BSD sockets—
which are based on point-to-point communica-
tions—the Enchantment Whiteboard uses a
client-server model in which clients post and read
structured information on a whiteboard server. This
lets any client exchange information with any other
client without the attendant complexity in negoti-
ating direct client-to-client communication. These
exchanges can take place without the client know-
ing anything at all about the other clients. Clients can subscribe to portions of the Enchantment Whiteboard, automatically receiving
updates when changes occur. Further, clients can lock
a portion of the whiteboard so that only the locking
client can post updates. It also supports symbolic
links across servers, letting whiteboards transpar-
ently refer to other whiteboards across a network. Intended to act as a streaming database, the Enchantment Whiteboard captures the current
state of some system, person, or group. On modest
embedded hardware, the board can support many
simultaneous clients distributed across a network
while making hundreds of updates a second. We
have used the Enchantment Whiteboard with the
Enchantment Signal system for bandwidth-inten-
sive voice-over-IP-style audio communications
between teams of up to 50 users. LIFE PATTERNS The MIThril system provides a modular frame- work for real-time understanding of sensor data.
The results of this process can be used locally for
reminders and wearer feedback, or they can be
broadcast to other users to enable smart-group
communications and increased awareness of other
members’ health and activity levels. Pattern recog-
nition techniques are the basis for modeling and
interpreting the output of the wearable sensors. The
standard pattern-recognition approach breaks this
process into four stages: • Sensing. A digital sensing device measures something in the real world, resulting in a dig-
ital signal of sampled values. For example, a
microphone sensor converts continuous ?uc-
tuations in air pressure—sound—into discrete
sampled values with a specified resolution,
encoding, and sampling rate. • Feature extraction. A raw sensor signal is transformed into a feature signal more suitable
for a particular modeling task. For example,
the feature extraction stage for a speaker-iden-
ti?cation-classi?cation task might involve con-
verting a sound signal into a power-spectrum
feature signal. • Modeling. A generative or discriminative sta- tistical model—such as a Gaussian mixture
model, Support Vector Machine hyperplane
classi?er, or hidden Markov model—classi?es
a feature signal in real time. For example, a
Gaussian mixture model could be used to clas-
sify accelerometer spectral features as walking,
running, sitting, and so on. • Inference. The results of the modeling stage, possibly combined with other information, are
fed into a Bayesian inference system for com-
plex interpretation and decision making. We use machine-learning techniques to record raw sensor measurements and create statistical
models of users’ behavior and the surrounding con-
text. Most commonly, we use hidden Markov mod-
els—which are also the basis of speech recognition
systems—for behavior modeling. We have used this
approach to build systems that use sensor mea-
surements of hand motions to perform real-time
recognition of American Sign Language and even to
teach simple T’ai Chi movements. 3 Typically, these systems have vocabularies of 25 to 50 gestures and
a recognition accuracy greater than 95 percent. We have applied this same basic approach to audio and video to accurately identify the setting The core MIThril software components provide the foundation for developing modular, distributed, context- aware wearable and ubiquitous computing applications. in which conversations take place—in a restaurant,
in a vehicle, and so on—and even to classify the
type of conversations a user engages in during the
day. 4,5 Once we model the behavior and situation, we can classify incoming sensor data to build a
model of the user’s normal behavior. We can then
use this model to monitor health, trigger reminders,
or even notify caregivers. Information about the wearer’s social interac- tions is particularly interesting. Understanding face-
to-face encounters is critical to developing inter-
faces that respect and support the wearer’s social
life. Social interactions are also very sensitive indi-
cators of mental health. Thus, an important chal-
lenge for our behavior modeling technology is to
build computational models that we can use to pre-
dict the dynamics of individuals and their interac-
tions. The number of parameters is a significant
factor in a model’s learnability and interpretabil-
ity. The requirement for minimal parameterization motivated our development of coupled hidden
Markov models (CHMM) to describe interactions
between two people, where the interaction para-
meters are limited to the inner products of the indi-
vidual Markov chains. 6 As a practical matter, a CHMM is limited to the interactions between two
people. We have therefore begun using a general-
ization of this idea, called the “in?uence model,”
which describes the connections between many
Markov chains as a network of convex combina-
tions of the chains. 7 This allows a simple parame- terization in terms of the “in?uence” each chain
has on the others, and we can use it to analyze com-
plex phenomena involving interactions between
large numbers of chains. To apply the influence model to human net- works, we have extended the original formulation
to include hidden states and to develop a mecha-
nism for learning the model’s parameters from
observations. 8 Modeling human behavior this way allows a simple parameterization of group dynam-
ics in terms of the in?uence each person has on the
others, and we have found that it provides a sensi-
tive measure of social interactions. HEALTHWEAR APPLICATIONS Several ongoing projects hint at the capabilities healthwear will offer. These applications include
medical monitoring and feedback systems for those
with chronic medical conditions, monitoring social
networking to reinforce healthy behavior, and men-
tal monitoring to detect the symptoms of depres-
sion or dementia. Medical monitoring and feedback Healthwear promises to be especially effec- tive for monitoring medical treatments.
Currently, doctors prescribe medications
based on population averages rather than
individual characteristics, and they check the
appropriateness of the medication levels only
occasionally—and expensively. With such a
data-poor system, it is not surprising that
medication doses are frequently over- or
underestimated and that unforeseen drug
interactions occur. Stratifying the population
into phenotypes using genetic typing can
improve the problem, but only to a degree.
Continuous monitoring of motor activity,
metabolism, and so on can be extremely effective
in tailoring medications to the individual. For example, consider Parkinson’s patients. For them to function at their best, their medications
must be optimally adjusted to the diurnal variation
of symptoms. For this to occur, the managing clin-
ician must have an accurate picture of how the
patient’s combined lack of normal movement
(hypokinesia) and disruptive movements (dyskine-
sia) ?uctuates throughout a typical day’s activities. To achieve this, we combined the MIThril sys- tem’s wearable accelerometers with standard sta-
tistical algorithms to classify the movement states
of Parkinson’s patients and provide a timeline of
how those movements ?uctuate throughout the day. Two pilot studies were performed, consisting of seven patients, with the goal of assessing the abil-
ity to classify hypokinesia, dyskinesia, and bradyki-
nesia (slow movement) based on accelerometer
data, clinical observation, and videotaping. Using
the patient’s diary as the gold standard, the result
was highly accurate identi?cation of bradykinesia
and hypokinesia. In addition, the studies classi-
?ed the two most important clinical problems—
predicting when the patient “feels off” or is about
to experience troublesome dyskinesia—perfectly. 9 Memory glasses Regardless of age, we’ve all had our moments of forgetfulness. We accept such memory lapses as
human fallibility, but we would be grateful if
researchers could find a way to cue our natural
memory and help us overcome these lapses.
Perhaps such a device also could, for example, help
improve an elderly person’s memory or provide
critical cues for emergency medical technicians,
doctors, or ?re?ghters in a nondistracting way. Toward this end, we are developing memory glasses that might someday help people with chal- May 2004 37 Understanding face-to-face encounters is critical to developing interfaces that respect and support the wearer’s social life. 38 Computer lenges ranging from complex memory loss to sim-
ple absent-mindedness. Figure 2 shows a prototype
of this wearable, proactive, context-aware mem-
ory aid based on the MIThril platform and wear-
able sensors. 10 Memory glasses function like a reliable human assistant, storing reminder requests and delivering
them under appropriate circumstances. Such a sys-
tem differs qualitatively from a passive reminder
system such as a paper organizer, or a context-blind
reminder system such as a modern PDA, which
records and structures reminder requests but which
cannot know the user’s context. Perhaps the major obstacle to this vision is that people resist being reminded to exercise, take their
medicine, or skip that extra helping of dessert.
Subliminal memory aids—visual and audio
reminders that lie just below the user’s threshold of
perception—may offer one way around this prob-
lem. Our research shows that under the right con-
ditions, subliminal text or audio cues can jog the
memory much like overt cues even though the per-
son receiving the cues is not aware of them. In one
experiment, for example, subliminal text cues
improved performance on a name-recall task by 50
percent compared to the uncued control. 11 Perhaps more important than this positive effect, our
research suggests that incorrect or misleading sub-
liminal cues do not interfere with memory recall.
This contrasts starkly with the effect of overt mis-
cues, which have a signi?cant misleading effect. A practical system might use a Bluetooth con- nection between cell phones to obtain the names of
nearby friends. Similarly, a combination of infor-
mation about location, proximity to others, time,
and surrounding sounds could assist in situation recognition. The system could then use this con-
text information to trigger the appropriate prompt,
which would ?ash across the user’s glasses or be
communicated through an earpiece. If the system
presented the prompt subliminally, users would not
consciously process the reminder and so would be
unaware that the prompt was jogging their mem-
ory. Thus, the subliminal prompts that the mem-
ory glasses provide would not interrupt a user’s
daily routines. SOCIAL NETWORKING Reinforcing an individual’s social support sys- tem may be the most effective way to encourage
adopting more healthy behavior patterns. Thus,
one aspect of healthwear’s core functionality is
interpersonal communications supported by con-
tinuous biomedical sensing. 12 Embedded social networking Healthwear’s social networking capabilities answer broad and immediate needs. For example,
aging parents now commonly live far away from
their families. Healthwear can help in such a situ-
ation by promoting communication between fam-
ily members when it senses a suspicious change in
an elder member’s behavior. In one version, healthwear occasionally but con- tinuously leaves phone messages reminding grown
children to call their parents and vice versa. How-
ever, when a marked change in behavior occurs—
such as decreased food consumption, socializing,
or sleeping—healthwear increases the frequency of
these reminders. The system would not tell people
something is speci?cally wrong or describe why it
left a particular message, nor would it call the doc-
tor except in extreme circumstances, because doing
so could violate people’s privacy and might actu-
ally interfere with proper medical support. Instead,
healthwear strengthens the social support network
when the need is likely to be most signi?cant. DiaBetNet Children also need social support networks, and they tend to be extremely sensitive to social con-
text. We focused on this tendency when we created
DiaBetNet, a computer game for young diabetics
that uses belt-worn motion sensors, a wireless
Internet connection, and a standard PDA for an
interface. 13 DiaBetNet capitalizes on their passion for social games to encourage children with dia-
betes to keep track of their food intake, activity,
and blood sugar level. A typical day in the life of a diabetic child using Figure 2. Memory
glasses. A
wearable,
proactive, context-
aware memory aid,
the memory glasses
system combines
the MIThril platform
with wearable
sensors to provide
a device that
functions like a
human assistant,
storing reminder
requests and
delivering them
under appropriate
circumstances. DiaBetNet would unfold as follows. In the morn-
ing, the child clips his wireless accelerometer and
DiaBetNet case—with wireless Internet connection,
PDA, glucose meter, and wireless receiver for the
accelerometer—onto his belt and goes off to school.
Throughout the day, the PDA records his activity
from the accelerometer, data from measuring glu-
cose and injecting insulin from the glucose meter,
and user-entered information about food con-
sumption. At any time, the user can see a graph on
the PDA that summarizes the day’s activity, carbo-
hydrate consumption, and glucose data. From time
to time, a wireless Internet connection sends this
data to a secure central server. DiaBetNet is a group gaming environment that requires guessing blood-sugar levels based on infor-
mation that wearable sensors collect: The more
accurate the answers, the higher the score. For
example, imagine that a user named Tom begins to
play DiaBetNet with others on the wireless net-
work. Transformed into his cherished alias, Dr. T,
Tom ?nds that his fellow players were all within 30
milligrams per deciliter of guessing their blood
sugar levels correctly, but his guess was closer than
anyone else’s. Tom challenges a DiaBetNet player called Wizard and looks through Wizard’s data. Although
Wizard was euglycemic in the morning, he ate a
late lunch. Therefore, Tom decides that Wizard’s
glucose level would be high and guesses 150 mg per
dl. Wizard guesses his glucose to be 180 mg per dl.
Tom wins again and grabs five more points. He
shoots a brief conciliatory message to his van-
quished foe and signs off. In clinical trials, 93 percent of DiaBetNet partic- ipants successfully transmitted their data wirelessly
to the server. The Game Group transmitted signif-
icantly more glucose values than the Control
Group. The Game Group also had signi?cantly less
hyperglycemia—glucose 250 mg per dl—than the
Control Group. Youth in the Game Group dis-
played a signi?cant increase in diabetes knowledge
over the four-week trial. Finally, more youth in the
Game Group monitored their hemoglobin levels. 14 Mental monitoring Healthwear technology also can assist in the early detection of psychological disorders such as
depression. Even though they are quite treatable,
mental diseases rank among the top health prob-
lems worldwide in terms of cost to society. Major
depression, for instance, is the leading cause of dis-
ability in established market economies. 15 Researchers have long known that speech activ- ity can be affected in pathological states such as
depression or mania. Thus, they have used audio
features such as fundamental frequency, amplitude
modulation, formant structure, and power distri-
bution to distinguish between the speech of nor-
mal, depressed, and schizophrenic subjects. 16 Similarly, movement velocity, range, and frequency
have been shown to correlate with depressed
mood. 17 In the past, performing such measurements out- side the laboratory was dif?cult given the required
equipment’s size and ambient noise. However,
today even common cell phones have the compu-
tational power needed to monitor these correlates
of mental state. We also can use the same method-
ology for more sophisticated inferences, such as the
quantitative characterization of social interactions.
The ability to use inexpensive, pervasive computa-
tional platforms such as cell phones to monitor
these sensitive indicators of psychological state
offers the dramatic possibility of early detection of
mental problems. Perhaps the most sensitive measure of mental function is social interaction, which clearly reveals
attitudes, emotions, and cognitive function. 18 To investigate this idea, we are using a MIThril-based
device dubbed the sociometer to collect data about
daily interactions with family, friends, and strangers
such as: • How frequent are the interactions?
• Are the interactions energetic or lethargic?
• Are the interactions appropriate without long gaps or frequent interruptions? Figure 3 shows an example of the sociometers that we used to collect almost 1,700 hours of inter-
action data from 23 subjects. Participants in this
study also ?lled out a daily survey that provided a
list of their interactions with others. The sociometer and conversation-detection algo- rithms classi?ed 87.5 percent of the conversations
as greater or equal to one minute, a far greater
accuracy than achieved using the survey method. May 2004 39 Figure 3. MIThril-
based sociometer.
A biosensor hub in
the badge-like
device, which is
worn over the
shoulder, collects
data about the
wearer’s daily
interactions. 40 Computer The few conversations that the automatic sociome-
ter method missed typically took place in high-
noise, multiple-speaker situations. 19 Once collected, researchers can use the in?uence model, a statistical framework that is a generaliza-
tion of the hidden Markov models commonly used
in speech recognition, to model the interaction
data. Modeling spoken behavior this way allows a
simple parameterization of group dynamics in
terms of the in?uence each person has on the oth-
ers. Our initial experiments show that these in?u-
ence parameters are effective indicators of status
within a social network and the degree of coupling
to the social network. 20 J udging from the adoption rates of advanced
cell phones and wearable health tools such as
pedometers, within this decade much of the US population will likely have access to continuous,
quantitative monitoring of its behavioral health sta-
tus, coupled with easily accessible biosignals. How
will this change our lives and our society? An exciting possibility is that with the wide- spread adoption of healthwear, researchers could,
for the first time, obtain enough data to really
understand health at a societal level. For example,
correlating a continuous, rich source of medication
data from millions of people could make drug ther-
apies more effective and help medical professionals
detect drug interactions more quickly. If correlated
with medical conditions, the data could illuminate
the etiology and preconditions of disease far more
powerfully than is possible today and, further, serve
as an early warning system for epidemic diseases
like SARS. Comparing the medical data with
genomic and protonomic data from different pop-
ulation samples could provide a powerful method
for understanding complex gene and environment
interactions. However, when considering the effects of health- wear systems, we would be wise to recall Marshall
McLuhan’s dictum that “the medium is the mes-
sage.” The way in which a new technology changes
our lifestyle may well be more important than the
information it conveys. Healthwear will likely be considered more per- sonal and intimate than traditional health tools
because it will form a constant part of a user’s phys-
ical presence. Psychological studies have shown
that clothes do indeed make the man. Thus, health-
wear will not only be part of what the user wears
but part of who that user is. Body-worn technol-
ogy will likely change our self-perception and self- con?dence in ways that are today unpredictable. While it could be more effective at promoting healthy behavior than traditional approaches,
healthwear also could be more seriously abused.
However, with more than one billion cell phones
already being worn every day, there is no escape
from being absorbed into this far more intimately
connected new world. Our goal now should be to
design this technology to make that world a very
human place to live. ? References 1. R. DeVaul et al., “MIThril 2003: Applications and Architecture,” Proc. 7th Int’l Symp. Wearable Com-
puters, IEEE Press, 2003, pp. 4-11; www.media.mit.
edu/wearables. 2. V. Gerasimov, Every Sign of Life, doctoral disserta- tion, Dept. Media Arts and Sciences, MIT, 2003. 3. T. Starner, J. Weaver, and A. Pentland, “Real-Time American Sign Language Recognition Using Desk
and Wearable Computer-Based Video, Hidden
Markov Models, IEEE Trans. Pattern Analysis and
Machine Vision, Dec. 1998, pp. 1371-1375. 4. A. Pentland, “Smart Rooms, Smart Clothes,” Scien- ti?c Am., Apr. 1996, pp. 68-76. 5. S. Basu, Conversational Scene Analysis, doctoral dis- sertation, Dept. of Electrical Engineering and Com-
puter Science, MIT, 2002. 6. N. Oliver, B. Rosario, and A. Pentland, “A Bayesian Computer Vision System for Modeling Human Inter-
actions,” IEEE Trans. Pattern Analysis and Machine
Intelligence, Aug. 2000, pp. 831-843. 7. C. Asavathiratham, The In?uence Model: A Tractable Representation for the Dynamics of Networked
Markov Chains, doctoral dissertation, Dept. of Elec-
trical Eng. and Computer Science, MIT, 2000. 8. T. Choudhury et al., “Learning Communities: Con- nectivity and Dynamics of Interactive Agents,’’ Proc.
Int’l Joint Conf. Neural Networks, Special Session
on Autonomus Mental Development, IEEE Press,
2003, pp. 2797-2802; http://hd.media.mit.edu. 9. D. Klapper, Use of a Wearable Ambulatory Monitor in the Classi?cation of Movement States in Parkin-
son’s Disease, master’s thesis, Harvard-MIT Health
Sciences and Technology Program, 2003. 10. R. DeVaul, Memory Glasses: Wearable Computing for Just-In-Time Memory Support, doctoral disser-
tation, Dept. of Media Arts and Sciences, MIT, 2004. 11. R. DeVaul, V. Corey, and A. Pentland, “The Mem- ory Glasses: Subliminal vs. Overt Memory Support
with Imperfect Information,’’ Proc. 7th Int’l Symp.
Wearable Computers, IEEE Press, 2003, pp. 146-
153; www.media.mit.edu/wearables. 12. M. Sung and A. Pentland, “LiveNet: Health and Lifestyle Networking through Distributed Mobile
Devices,” tech. report TR 575, MIT Media Lab,
2003; http://hd.media.mit.edu. 13. V. Kumar et al., “DiaBetNet: Learning and Predict- ing Blood Glucose Results to Optimize Glycemic Con-
trol,” poster exhibit, 4th Ann. Diabetes Technology
Meeting, Atlanta, 2002; www.diabetestechnology.
org. 14. V. Kumar, The Design and Testing of a Personal Health System to Motivate Adherence to Intensive
Diabetes Management, master’s thesis, Harvard-MIT
Health Sciences and Technology Program, 2004. 15. C.L.J. Murray and A.D. Lopez, The Global Burden of Disease, Harvard Univ. Press, 1996. 16. D.J. France et al., “Acoustical Properties of Speech as Indicators of Depression and Suicidal Risk,” IEEE
Trans. Biomedical Eng., July 2000, pp. 829-837. 17. M.H. Teicher, “Actigraphy and Motion Analysis: New Tools for Psychiatry,” Harvard Rev. Psychiatry,
1995, vol. 3, pp. 18-35. 18. P. Franks, T.L. Campbell, and C.G. Shields, “Social Relationships and Health: The Relative Roles of Fam- ily Functioning and Social Support,” Social Science
& Medicine, Apr. 1992, pp. 779-788. 19. T. Choudhury and A. Pentland, “Modeling Face-to- Face Communication Using the Sociometer,” W9
Workshop, Proc. Int’l Conf. Ubiquitous Computing,
IEEE Press, 2003, pp. 3-8; http://hd.media.mit.edu. 20. T. Choudhury, Sensing and Modeling Human Net- works, doctoral dissertation, Dept. of Media Arts
and Sciences, MIT, 2003. Alex (Sandy) Pentland is the Toshiba Professor of
Media Arts and Sciences at MIT and heads the MIT
Media Laboratory’s Human Dynamics research
group. His research interests include wearable com-
puting, human-machine interfaces, computer
graphics, arti?cial intelligence, and machine and
human vision. Pentland is a cofounder of the IEEE
Computer Society’s Technical Committee on Wear-
able Information System and the IEEE NNS
Autonomous Mental Development Technical
Committee. He received a PhD from MIT. Contact
him at pentland@media.mit.edu. May 2004 41

Comments