IEEE Computational Intelligence Magazine - May 2021 - 50

both a trait (self-control) and cognitive ability
(working memory capacity) [116] and has been
predicted by EEG signals and trans-dermal devices [114], [115]. Physical activity (IPAQ) is commonly predicted by using mobile sensing, e.g.,
[119], or specialized devices such as accelerometers [120]. Physical activity patterns appear to be
stable, as patterns earlier in life can predict some
activity patterns later in life [121].
Sleep quality has been estimated using wearables by
[122], [123]. More invasive techniques can use chest sensors
and polysomnography (to measure body acceleration and
position) [124]. Specialized work to estimate sleep for
patients with schizophrenia also exists (see a comparative
analysis in [125]).
Tobacco consumption has been monitored using air sensors
as in [117]. Also, smoker group membership (never, established,
former, non-daily, and daily) has been predicted using family
history, depression, consumption of other substances, and
demographics [118]. To our knowledge, tobacco consumption
has not been monitored using wearable sensors.
Job performance is usually measured through either subjective
rating scales [48], [137] or objective performance outcomes,
such as sales amounts and production numbers. [48]. Using
wearable sensor data to estimate job performance has been
explored by [17], who demonstrate that wearables can be used
to detect when a person is focused on their work via physiological features. Another approach estimates job performance
based on personality and individual traits [11], [13]-[15] with
varying degrees of success. This includes conscientiousness [12],
[15], [84], and extraversion [13]. Links have been found
between personality, cognitive ability, and other traits with job
performance [86]. However, as mentioned, estimation of job
performance more commonly relies on various types of questionnaires including self-reports, and supervisory and peer
evaluations [11], [48], [137]. Job performance varies depending on demographic information (e.g., age, gender) [85] and
individual traits (personality, emotional intelligence) [9], [10].
Cognitive ability, personality, affect, and anxiety are all not
only related to each other as discussed, but can both affect and
be affected by job performance, and physical and health variables. The mutual effects of personality traits and job activities
are well documented [87].
Mobile and wearable sensor data are powerful sources of
information about human behavior which could help us identify patterns of activities, human mobility, [25]-[27], well-being
[28], and job and academic performance [17]. Machine learning models have been demonstrated to achieve high accuracy
on very specific tasks on very small samples, e.g., estimating
work load category using wearables on a cohort of twenty academic participants [17]. Subjective perceptions of job performance are, however, harder to predict using wearable data even
in small samples and specific work locations and environments
[18], as opposed to various work locations and occupations as
in the present case. To our knowledge, ours is the first model

We recruited 757 individuals working in knowledge
fields in several institutions in the US as part of a
large-scale longitudinal research study. We collected
data from wearables, phones, social media, and
location patterns as predictors.
personality modeling has focused on predicting the Big 5 personality traits from sensing and system adaptability. For a
recent review of the most popular computational approaches
for automated personality detection (including datasets, applications, and machine learning methods), see [95] and most
recently [126]-[128] that improved the results on the Facebook, Essays, and Kaggle and Essays datasets, respectively. Some
of these techniques combine various types of complex features
using Deep Residual Networks and sophisticated techniques
requiring large amounts of data.
Affect detection has been an active area of research in the
field of affective computing [94], [103]-[105], [129]. Positive
and negative affect, as well as perception and satisfaction with
health and life, have been predicted using wearable sensors
[106]. Affective modeling also involves sentiment analysis,
including predicting sentiment intensity [108] using a
stacked ensemble that contains neural networks as part of the
architecture. In the case of sentiment analysis, however, the
number of instances used for training in [108] is in the order
of thousands, while in our case we collected posts from social
media for 392 participants. The interested reader is referred
to the survey in [107], [129], which uses common tasks in
affective computing and sentiment analysis (emotion recognition, polarity detection, multimodal fusion) to present a
taxonomy of platforms: knowledge based, statistical methods,
and hybrid. Notice that sentiment analysis is often done in
the context of specific scenarios, e.g., with respect to social
media posts and news. In contrast, our paper is devoted more
to sensor-based modeling for understanding individual
traits. Emotion categorization involves psychological models, e.g., [130]-[134] and sensing algorithms [91], [107],
[135], [136] at the intersection of psychology linguistics,
computer science, engineering, and other fields. Most emotion categorization models differ on the number and the
list of emotions. However, emotions can be roughly
grouped into three types: positive, neutral, and negative, as
discussed in [109]. In our paper, we model generalized affect
(positive and negative) [35] as a fairly stable trait that
describes the experience of emotions because affect is associated with well-being and productivity.
Beyond positive and negative affective states we consider
anxiety. Previous work has focused on predicting anxiety from
sensors, e.g., [104], [110]. Methods that predict stress and anxiety based on ECG monitoring have been proposed [111].
Physical and health variables have been predicted by a variety
of methods. Alcohol consumption is found to be related to

50

IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | MAY 2021
IEEE Computational Intelligence Magazine - May 2021

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - May 2021
