Signal Processing - September 2017 - 196

PERSPECTIVES
Daniel Bone, Chi-Chun Lee, Theodora Chaspari,
James Gibson, and Shrikanth Narayanan

Signal Processing and Machine Learning for
Mental Health Research and Clinical Applications

H

uman behavior offers a window into
the mind. When we observe someone's actions, we are constantly inferring his or her mental states-their
beliefs, intents, and knowledge-a concept
known as theory of mind. For example:
■ "Is a person growing impatient?"
■ "Are they feeling down?"
■ "Are they being truthful?"
■ "Will they buy my product?"
We can perceive changes in emotion
and even if someone is intoxicated. The
basis for our "mind reading" abilities is
rooted in the influence (conscious and
subconscious) that our mental state has
on the cognitive and motor processes that
control the coordinated production of
rich, complex behaviors. Our behaviors
are multimodal and encoded at different temporal scales, ranging from rapid
facial microexpressions or a quick rise
of vocal pitch to indicate a question, to
slower body gestures like waving hello.
Behavioral differences across individuals and changes within a person are
also critical indicators of behavioral and
mental health. For instance, softer, slower, and less articulate speech is a progressive symptom of Parkinson's disease
(which can improve with treatment) [1],
and three primary indicators of autism
spectrum disorder are poor eye contact,
lack of conversational skills, and an atypical speech prosody (the rate, rhythm,
and intonation of speech) [2]. Behavioral
markers are also relevant to relationship
Digital Object Identifier 10.1109/MSP.2017.2718581
Date of publication: 6 September 2017

196

Table 1. The prevalence of selected health conditions in the United States [6]-[9].
Condition

Ages

Prevalence

Autism spectrum disorder

Children*

1.5% (lifetime)

Posttraumatic stress disorder

Adults

3.5% (one year)

Mood disorders (e.g., depression)

Adults

9.5% (one year)

Alcohol addiction/abuse

All

6.6% (one year)

Illicit drug use (nonmarijuana)

All

2.5% (one year)

Parkinson's disease

≥80 years old

1.9% (lifetime)

Dementia (e.g., Alzheimer's disease)

≥60 years old

6.5% (lifetime)

*Typically diagnosed in children but symptoms persist over the life span.

quality; for example, extended use of
second-person singular pronouns (e.g.
you) during conversational speech has
been linked to increased levels of blame
in romantic relationships [3].
The span of mental and behavioral
health conditions is vast (Table 1) and so
is the accompanying cost to individuals and society at large. The National
Institute of Mental Health (NIMH) has
estimated that the total costs of serious mental illness, affecting 6% of the
population, is in excess of US$300 billion a year [4]; this doesn't even include
developmental or substance abuse disorders. Aside from financial costs,
these disorders reduce life expectancy;
unlike many other health conditions,
years of life lost due to neurological,
mental, and behavioral disorders has
increased recently, representing a rising burden that will impose new challenges on the health system [5]. Thus,
translational research that improves
aspects of health awareness, access,
IEEE SIGNAL PROCESSING MAGAZINE

|

September 2017

|

treatment quality, and cost will have a
profound impact.
The core means of behavioral and
mental health assessment have changed
little over the past decades: human evaluation (by direct observation, interview, or
self-report) is the primary tool. Humans
are excellent signal processors, having the
ability to transfer knowledge from related
experiences and being able to account for,
and adapt to, context with relative ease.
But humans also have limitations in their
perceptual abilities. First, to label "large"
data sets, an army of human workers may
spend hundreds of hours meticulously
annotating videos; this solution does
not scale with the pace of collected realworld data. Second, human judgment is
subjective and idiosyncratic; reaching
an agreement on qualitative judgments
such as the level of rapport between individuals is challenging, and it is unlikely a
person can precisely estimate quantitative
(continued on page 189)
1053-5888/17©2017IEEE
Table of Contents for the Digital Edition of Signal Processing - September 2017
