Integrating Wearable Sensor Data and Self-reported

Diaries for Personalized Affect Forecasting

Zhongqi Yang1 , Yuning Wang4 , Ken S. Yamashita2 , Elahe Khatibi1 , Iman Azimi1 ,
Nikil Dutt1 , Jessica L. Borelli2 , and Amir M. Rahmani1,3
1
Department of Computer Science, University of California, Irvine
2
Department of Psychological Science, University of California, Irvine
3
School of Nursing, University of California, Irvine
4
Department of Computing, University of Turku
arXiv:2403.13841v2 [cs.LG] 23 Mar 2024

{zhongqy4, ksyamash, ekhatibi, azimii, dutt, jessica.borelli, a.rahmani}@uci.edu, yuning.y.wang@utu.fi

Abstract—Emotional states, as indicators of affect, are pivotal wearable and mobile devices such as smartwatches, smart rings,
to overall health, making their accurate prediction before onset wristbands, and smartphones have been increasingly used for
crucial. Current studies are primarily centered on immediate short- real-time, remote monitoring of individuals’ affective states
term affect detection using data from wearable and mobile devices.
These studies typically focus on objective sensory measures, often in free-living settings [11]–[15]. Notably, [15] demonstrates
neglecting other forms of self-reported information like diaries andthe feasibility of using fully objective measurements from
notes. In this paper, we propose a multimodal deep learning model commercial wearable devices to predict affect status in everyday
for affect status forecasting. This model combines a transformer settings. This study employs smart rings, smartwatches, and
encoder with a pre-trained language model, facilitating the smartphones to continuously monitor physiological and environ-
integrated analysis of objective metrics and self-reported diaries.
To validate our model, we conduct a longitudinal study, enrolling mental information, as well as sleep, metabolic, and physical
college students and monitoring them over a year, to collect an activity patterns. Machine learning models are leveraged for
extensive dataset including physiological, environmental, sleep, data analysis to predict positive and negative affect statuses
metabolic, and physical activity parameters, alongside open- for the next day. This research highlights that utilizing data
ended textual diaries provided by the participants. Our results collected from wearable devices enables a purely objective
demonstrate that the proposed model achieves predictive accuracy
of 82.50% for positive affect and 82.76% for negative affect, a method to predict one’s emotional state the following day.
full week in advance. The effectiveness of our model is further Although the accuracy of this approach is not exceptionally
elevated by its explainability. high, it demonstrates the potential of wearable technology in
Index Terms—Affective Computing, Multimodal Machine forecasting affect status.
Learning, NLP, Wearable Sensor Data, Digital Mental Health However, there are two limitations of the existing studies.
Firstly, while current remote affect monitoring systems primar-
I. I NTRODUCTION ily rely on objective data gathered from wearable devices in
Affect is a broad term referring to one’s subjective feeling everyday settings, they often ignore the valuable insights that
states. Affective states can influence moment-to-moment emo- can be obtained from human-generated textual data in daily
tional experience as well as long-term mood and have impacts life, which are crucial for mental health analysis. Such textual
on overall well-being [1]. Prolonged negative affect has been data play a pivotal role in providing information, such as life
demonstrated to yield adverse consequences for both physical events or emotional narratives, which offer essential information
and mental health outcomes, including increased susceptibility for a comprehensive understanding of an individual’s mood
to mental health conditions such as depression [2]–[5], which trajectory [16], [17]. Incorporating such textual data (including
ranks among the leading causes of disability [6]. Accurate diaries, notes, etc.) complementing the objective measurements
identification, detection, and prediction of affect are crucial for still requires further exploration. Secondly, existing research
effective intervention and prevention strategies. concentrates on the instantaneous detection or short-term
In recent years, wearable and mobile devices, along with prediction of affective states. The reliance on fully collected
machine learning-based approaches, have been employed to data from the next day limits their predictive capabilities
track individuals’ affective states [7]. Existing works focus on to no earlier than the following day. This timeframe may
gathering and analyzing objective physiological and behavioral not be adequate for timely prevention strategies. There is a
data through the use of laboratory-based or wearable/portable pressing need for research focused on forecasting mental health
devices. Such research utilizing multimodal data, including conditions significantly in advance of their onset to enhance
electroencephalogram (EEG), photoplethysmogram (PPG), has preventive measures and overall mental well-being.
primarily been conducted in laboratory settings [8]–[10]. In this paper, we present a multimodal deep learning model
Expanding beyond the confines of laboratory environments, which is based on a transformer encoder and a pre-trained
language model, DistilBERT, to forecast affect status one week Affect Status

in advance. We accomplish this by fusing objective features

Linear Layer
from wearable and mobile devices and self-reported diaries. The
Linear Layer
objective features encompass a range of data, including sleep Predictive
Model
patterns, physiological metrics, physical activity, metabolic Transformer Encoder
Diary
rates, and environmental factors. For the diary data, we extract Submission
Diary
Content
Frequency
features from two primary aspects: the diary submission Sleep Pattern
Diary
Feature
frequency and the diary content. To evaluate the effectiveness Physiology Physical Activity Extraction
Physical Activity
of our proposed affect forecasting model, we collect a dataset Metabolism
Environment
Self-reported
Diaries
featuring multimodal information from wearable devices and
self-reported diaries detailing daily affective highs and lows.
This system also includes 20 unique discrete affect ratings Fig. 1: Affect Forecasting Model Workflow.
provided by participants daily and weekly. The affect ratings
are aggregated to establish labels for positive and negative affect is evaluated using a scale ranging from 0 (“Very Slightly")
statuses. Furthermore, we investigate model explainability by to 100 (“Extremely"), , where 0 meant “Very Slightly" and
analyzing the contribution of features as indicated by Shapley 100 “Extremely", including 20 affect words like “inspired" and
values, and the attention scores assigned to keywords in the “nervous," based on the Positive and Negative Affect Schedule
diaries. (PANAS) [19], [20]. Each affect word is scored individually
and as part of a composite, with Positive Affect (PA) and
II. DATA C OLLECTION Negative Affect (NA) calculated as the average of 10 positive
We collected data for assessing mental health and affect and negative items, respectively. Binary classifications for PA
status of college students including 25 undergraduate students and NA are assigned based on whether an emotion’s score
between the ages of 18 and 22 who could speak and write surpassed or did not meet the median for all participants,
English fluently. Participants could not be parents or married thereby maintaining an equitable binary label distribution. In
or returning to school after a 3 year period or longer to line with standard practice, we exclude the central 20% of
maintain a homogeneous sample that reflects the college student values from the overall distribution prior to assigning labels.
population. Participants must not have participated in any of III. A FFECT F ORECASTING M ODEL
our previous studies on this topic.
This section introduces the affect forecasting model designed
Over the course of the study, we tracked participants’
for the analysis of our multimodal data. The workflow of the
emotional states, physiological patterns, and behavioral habits
model is illustrated in Figure 1.
through smart devices and ecological momentary assessments
(EMAs). Participants were fitted with the Oura ring and Sam- A. Diary Feature Extraction
sung Gear Sport smartwatch and downloaded the corresponding Participants submitted daily self-reported diaries to describe
Oura and Samsung Android mobile apps in an effort to capture their daily feelings through the ZotCare application on smart-
an accurate depiction of their daily physical habits, sleep, phones. We introduce the self-reported diary feature extraction
and health. These devices were integrated into the ZotCare of the affect forecasting model for two key features from the
mHealth platform [18]. The Oura ring collects data on sleep and diaries: Diary Content and Diary Submission Frequency.
cardiovascular activity that are used in this study to assess sleep a) Diary Content: The analysis of self-reported diaries
quality by measuring sleep duration, average heart rate during poses a unique challenge in affective computing as they usually
sleep, and heart rate variability during sleep. The Samsung do not explicitly reveal emotional status. Moreover, these diaries
watch complements the physical activity feature set by adding rarely directly indicate physiological or behavioral factors
walk and run steps and captures the atmospheric pressure, which that could predict future affect status. Considering the model
enriches the environmental features available for analysis. The efficiency, we employ DistilBERT [21] complemented by a
detailed list of the objective features is the same as those linear layer to extract meaningful insights from the diaries.
described in [15]. DistilBERT functions to parse and interpret the underlying
Participants utilized the ZotCare application to submit self- sentiments of the diaries and offers a high-dimensional vector
reports functioning as daily diaries through EMAs. These representation of the input. Then the following linear layer
entries vary from expressions of personal sentiments and moods acts on the high-dimensional representation produced by
to narratives of daily experiences. Participants document key DistilBERT, transforming them into a streamlined, singular
moments of their day, detailing the high and low points of value vector as the Diary Content feature.
their day’s affective experiences, overall emotional experiences, b) Diary Submission Frequency: Another challenge in our
and notable daily events such as physical activities, academic study is the variability in the frequency of diary submissions
exams, or social interactions. by participants, serving as a factor that, as highlighted in recent
Participants were also asked to engage with the ZotCare app studies, reflects the mental health status of individuals [22].
to submit their daily and weekly affect status through EMAs We calculate the diary submission frequency based on how
that were used as labels with a one-week delay. The affect status many days a participant submits in a one-week window.
B. Predictive Model gradient propagation. Both DistilBERT and this linear layer
are then fine-tuned together on the training dataset.
The architecture of the predictive model is built on a Trans-
b) Step 2: The End-to-End Affect Forecasting Model
former Encoder [23] to aggregate the features from a wide range
Training: Following the initial fine-tuning of DistilBERT, Step
of modalities efficiently through its self-attention mechanism.
2 jointly trains the whole affect forecasting model encompass-
Following the transformer encoder, we integrate a 2-layer multi-
ing the fine-tuned DistilBERT, the transformer encoder, and
layer Perceptron to project the complex embeddings generated
the linear layers. Diaries are processed through DistilBERT
by the transformer into a singular, interpretable value that
to extract diary content feature, which are subsequently
directly corresponds to the affect status.
concatenated with objective features and diary submission
The input for the predictive model amalgamates diary
frequency. The combined inputs are then fed through the
features derived from Self-reported Diary Processing with
transformer encoder, followed by linear layers that map these
objective features, as detailed in Section II. To achieve this,
inputs to the corresponding affect status.
we concatenate both objective features and diary features to
create an extensive feature vector. D. Model Explainability

C. Model Training Strategy We demonstrate the explainability of the proposed model in

two aspects. We first highlight how individual features impact
The integration of diary features in our predictive model the forecasted affect status. We accomplish this by employing
is challenged by the initial ineffectiveness of DistilBERT. To path-dependent feature perturbation algorithms provided by
address this, we propose a sequential training process that the SHAP (SHapley Additive exPlanations) library [24] to
is executed in two steps. Initially, DistilBERT is fine-tuned calculate the Shapley values for each feature. Additionally, to
independently to adapt its pre-trained weights to our specific gain a deeper understanding of which keywords the model
affective context. We then train the entire model in a joint identifies as crucial for predicting affect status, we delve into
manner. This sequential approach ensures a balanced learning the significance of specific words in the diary entries. This
process to prevent any type of feature from disproportionately is achieved by analyzing the attention scores allocated by the
influencing the model’s development. This process is also fine-tuned DistilBERT model to these keywords.
illustrated in Figure 2.
a) Step 1: Fine-tuning DistilBERT: We observe that it is IV. R ESULTS
challenging when training DistilBERT and the predictive model
We evaluate the proposed model using leave-one-subject-out
jointly. Given that DistilBERT is a general-purpose model
cross-validation and a personalized cross-validation strategy.
trained on a diverse corpus, we must align DistilBERT to the
In the leave-one-subject-out cross-validation approach, each
affect forecasting task. However, jointly training DistilBERT
iteration involves training the model on the data from all but one
with the predictive model presents challenges. Initially, Distil-
subject and testing on the data from the excluded subject. For
BERT’s performance in identifying affect-related information
the personalized cross-validation, we utilize the latter half of a
from diary entries is suboptimal, leading the predictive model
participant’s data for testing, while the initial half, combined
to assign attention lower scores to the diary features. This leads
with data from other participants, is employed for training. This
to weak gradient signals being sent back to DistilBERT during
method ensures that each iteration of the model incorporates
training that may result in omission of the diary information.
historical data from an individual participant for training,
To address this challenge, it is essential to customize
fostering a personalized model. We refer to the models derived
DistilBERT specifically for the affect forecasting task. Our
from personalized cross-validation as ’Personalized,’ while
approach involves initially fine-tuning DistilBERT to better
those obtained from leave-one-subject-out cross-validation are
suit affect forecasting. We accomplish this by integrating a
designated as ’Non-personalized.’ The accuracy of our proposed
temporary linear layer on top of DistilBERT, which serves to
model for forecasting affect status are shown in Table I.
translate embeddings into predicted affect status and facilitates
We observe an enhancement in accuracy in the personalized
Step 2: cross-validation process that the personalized models, distinct
The End-to-End Affect Forecasting Model Training

Affect Status
for the training on the participant’s unique data set, outperform
Linear Layer
generic, non-personalized models. This result suggests that
Linear Layer
Step 1:
models tailored to individual behavioral and emotional patterns
Transformer Encoder
Finetuning DistilBERT
yield more precise and reliable predictions, emphasizing the
Affect Status
Diary
Objective Features Submission
Diary
Content
value of personalization in predictive analytics.
Linear Layer
Frequency

Fine-tuned
DistilBERT
TABLE I: Forecasting Accuracy Results
DistilBERT

Self-reported Weekly Shifted Personalized (%) Non-personalized (%)

Diaries EMAs Labels
Forecasting PA w/ Diary 82.50 81.20
Forecasting NA w/ Diary 87.76 80.22
Forecasting PA w/o Diary 81.08 79.03
Fig. 2: Training Process of the Affect Forecasting Model Forecasting NA w/o Diary 82.11 75.31
 Attention Scores of Keywords for Forecasting Negative Affect Status Target Calories - day 2
Ashamed
Fairly Heart Rate Std - day 2
Easily
Swing Deep Sleep Duration - day 7
Console
Interaction
Light Sleep Duration - day 7
Parts
Drove
Heart Rate Mean - day 6
Continuing Awake Time - day 3
Shitty
Rejected Met Min Inactive - day 2
Exercise
Dent Met Min Inactive - day 1
Rain
Sorted Light Sleep Duration - day 1
Kind
Upon
Target Calories - day 4
Lit Target Calories - day 6
Bark
Harder Diary Submission Frequency
Met Min Inactive - day 4
Average Attention Scores
Awake Time - day 4
Met Min Low - day 6
Attention Scores of Keywords for Forecasting Positive Affect Status Met Min Inactive - day 5
Ashamed Heart Rate Mean - day 4
Easily
Console Diary Content
Bark
Swing Move Every Hour - day 6
Fairly
Neglected Total Calories - day 3
Interaction
Exercise
Printing
Dent
Shitty
Confront
Fig. 5: Feature Shapley Values for Negative Affect Forecasting
Rain
Rejected
Recommendation duration emerge as top-ranked predictors, with higher duration
Enough
Sorted of deep sleep and lower light sleep duration contributing
Parts
Period strongly to the prediction of a positive affect status, and
Average Attention Scores lower duration is correlated to negative affect status. Diary
Submission Frequency also holds considerable weight as the
Fig. 3: Attention scores of keywords higher frequency of submitted diaries indicates an association
with positive affect status while a lower frequency correlates
Deep Sleep Duration - day 1
Deep Sleep Duration - day 7 with an increase in negative affect status. The physically active
Light Sleep Duration - day 5 time and the total movement contribute to our model’s positive
Deep Sleep Duration - day 3 affect predictions. Moreover, we observe that dietary metrics,
Diary Submission Frequency
such as calorie intake, along with activity metrics like MET and
Deep Sleep Duration - day 5
Move Every Hour - day 6
daily target calorie expansion, are influential in determining
Inactive - day 5 negative affect status, underscoring the multifaceted nature of
Light Sleep Duration - day 4 affect status predictions.
Light Sleep Duration - day 6
Light Sleep Duration - day 7 B. Diary Keywords Attention Scores
Meet Daily Target - day 4
Target Km - day 7
Recognizing the critical role of diary content in forecast-
Diary Content ing, we analyze the most influential keywords identified by
Cal Activity - day 7 their attention scores computed by DistillBERT. Figures 3
REM Sleep Duration - day 1
demonstrate the attention scores derived from the DistilBERT
Daily Movement - day 1
model for forecasting both positive and negative affect status.
Mean HRV - day 6
Light Sleep Duration - day 1 The most influential keywords are largely similar across both
Stay Active - day 5 positive and negative affect forecasting. It can be seen that the
words demonstrating emotional status, such as "ashamed", "eas-
ily", and "shitty" obtain higher attention scores. Additionally,
Fig. 4: Feature Shapley Values for Positive Affect Forecasting keywords pertaining to specific life events or activities, like
Additionally, to illustrate the impact of the diary features, "drove", "console", "exercise", and "rain", are also identified
we compare the proposed model against a purely objective as significant in affect forecasting.
model, which excluded the diaries and relies solely on objective
V. D ISCUSSION
features, designated as ’Forecasting w/o Diary’. Our findings
indicate that the inclusion of diary features leads to a boost in The results of this study not only validate the feasibility of
the model’s forecasting accuracy. one-week forecasting but also align with the accuracy of next-
day predictions reported in our previous study [15], highlighting
A. Features’ Shapley Value the effectiveness of our multimodal approach, especially
The Shapley values distribution of features are shown in incorporating self-reported diaries. Although forecasting a
Figure 4 and Figure 5. We observe that both light and deep sleep single loneliness level was previously shown to be feasible
using objective features [25], affect status forecasting, as severity of coronary artery disease. Kardiologia Polska (Polish Heart
a multi-facet case, is still challenging. We show that the Journal), 72(7):617–623, 2014.
[2] Nico H Frijda. Emotion, cognitive structure, and action tendency.
integration of diary data with objective modalities, such as Cognition and emotion, 1(2):115–143, 1987.
physiological, environmental, and physical activity information, [3] Katharine E Smidt and Michael K Suvak. A brief, but nuanced, review
gathered from wearable and mobile devices represents a novel of emotional granularity and emotion differentiation research. Current
Opinion in Psychology, 3:48–51, 2015.
and efficient approach to affective computing. [4] Nilani L Shankar and Crystal L Park. Effects of stress on students’
In terms of explainability, we find that levels of sleep had physical and mental health and academic success. International Journal
the greatest impact on the model output. This is consistent with of School & Educational Psychology, 4(1):5–9, 2016.
previous literature relating the quality of sleep to mood, where [5] Olivia Remes, João Francisco Mendes, and Peter Templeton. Biological,
psychological, and social determinants of depression: a review of recent
sleep deprivation is found to have a strong influence on the pos- literature. Brain sciences, 11(12):1633, 2021.
itive affective system [26]. Our findings also highlight certain [6] Rabie Karrouri, Zakaria Hammani, Roukaya Benjelloun, and Yassine
keywords with higher attention scores in affect forecasting. The Otheman. Major depressive disorder: Validated treatments and future
challenges. World journal of clinical cases, 9(31):9350, 2021.
significant terms are similar across both positive and negative [7] Yan Wang, Wei Song, Wei Tao, Antonio Liotta, Dawei Yang, Xinlei Li,
affect. Words demonstrating emotional status (i.e., ashamed) Shuyong Gao, Yixuan Sun, Weifeng Ge, Wei Zhang, et al. A systematic
have a higher attention score in general, which is consistent review on affective computing: Emotion models, databases, and recent
advances. Information Fusion, 83:19–52, 2022.
with previous research suggesting that emotionally intense [8] S. Deb et al. Emotion classification using segmentation of vowel-like
words can influence behavior and cognition [27]. Additionally, and non-vowel-like regions. IEEE Trans Affect Comput., 10(3):360–73,
studies have found that negative words are generally stronger 2017.
[9] Chiara Filippini, Adolfo Di Crosta, Rocco Palumbo, David Perpetuini,
in extremity in comparison to positive words [28]. This aligns Daniela Cardone, Irene Ceccato, Alberto Di Domenico, and Arcangelo
with our findings, as the highest score of emotion-tied terms Merla. Automated affective computing based on bio-signals analysis and
has a negative valence. We find keywords relating to day- deep learning approach. Sensors, 22(5):1789, 2022.
to-day activities (i.e., rain, console) are significant in affect [10] Cheng-Jie Yang, Nicolas Fahier, Chang-Yuan He, Wei-Chih Li, and Wai-
Chi Fang. An ai-edge platform with multimodal wearable physiological
forecasting. Additionally, this study indicates that a higher signals monitoring sensors for affective computing applications. In 2020
frequency of diary submissions is associated with an increased IEEE International Symposium on Circuits and Systems (ISCAS), pages
likelihood of positive affect status and a decreased likelihood of 1–5. IEEE, 2020.
[11] Eiman Kanjo, Eman MG Younis, and Chee Siang Ang. Deep learning
negative affect status. Some research has found links between analysis of mobile physiological, environmental and location sensor data
writing and mood fluctuations; for instance, a study found that for emotion detection. Information Fusion, 49:46–56, 2019.
completing daily diary entries reduced self-reported depression [12] Ziyu Wang, Zhongqi Yang, Iman Azimi, and Amir M Rahmani. Differ-
ential private federated transfer learning for mental health monitoring
levels as well as influenced emotion regulation strategies [29]. in everyday settings: A case study on stress detection. arXiv preprint
Although neutral terminology may not seem significant, simply arXiv:2402.10862, 2024.
writing and sharing daily activities may have implications on [13] Rui Wang, Fanglin Chen, Zhenyu Chen, Tianxing Li, Gabriella Harari,
Stefanie Tignor, Xia Zhou, Dror Ben-Zeev, and Andrew T Campbell.
mood [30]. Future work could expand on specific keyword Studentlife: assessing mental health, academic performance and behav-
frequency, studying if the increased presence of a specific ioral trends of college students using smartphones. In Proceedings of the
valence (positive or negative) can be associated with positive 2014 ACM international joint conference on pervasive and ubiquitous
computing, pages 3–14, 2014.
affect or negative affect forecasting, respectively. [14] Sara Taylor, Natasha Jaques, Ehimwenma Nosakhare, Akane Sano,
and Rosalind Picard. Personalized multitask learning for predicting
VI. C ONCLUSION tomorrow’s mood, stress, and health. IEEE Transactions on Affective
Computing, 11(2):200–213, 2017.
This study introduced a multimodal deep learning model [15] Salar Jafarlou, Jocelyn Lai, Iman Azimi, Zahra Mousavi, Sina Labbaf,
to predict affect status using a combination of data from Ramesh C Jain, Nikil Dutt, Jessica L Borelli, Amir Rahmani, et al.
wearable devices and self-reported diaries. The model exhibited Objective prediction of next-day’s affect using multimodal physiological
and behavioral data: Algorithm development and validation study. JMIR
satisfactory accuracy of 82.50% and 87.76% for positive Formative Research, 7(1):e39425, 2023.
and negative affect, respectively, affirming the practicality of [16] Matla Suhasini and Badugu Srinivasu. Emotion detection framework
forecasting affect status. A significant outcome of our study for twitter data using supervised classifiers. In Data Engineering and
Communication Technology: Proceedings of 3rd ICDECT-2K19, pages
is the improvement in accuracy upon incorporating diary data, 565–576. Springer, 2020.
underscoring the contribution of self-reported diaries to this [17] S Santhosh Baboo and M Amirthapriya. Sentiment analysis and automatic
process. Moreover, our results stressed the importance of emotion detection analysis of twitter using machine learning classifiers.
International Journal of Mechanical Engineering, 7(2), 2022.
personalized methods in monitoring mental health, suggesting
[18] Sina Labbaf, Mahyar Abbasian, Iman Azimi, Nikil Dutt, and Amir M.
that personalized approaches yield more accurate outcomes. Rahmani. Zotcare: a flexible, personalizable, and affordable mhealth
Future research could expand upon this study by applying the service provider. Frontiers in Digital Health, 5, 2023.
model to diverse populations and environments. This would aim [19] Akane Sano. Measuring college students’ sleep, stress, mental health
and wellbeing with wearable sensors and mobile phones. PhD thesis,
to enhance the integration of both objective and self-reported Massachusetts Institute of Technology, 2016.
diary data for more accurate predictions. [20] David Watson, Lee Anna Clark, and Auke Tellegen. Development and
validation of brief measures of positive and negative affect: the panas
R EFERENCES scales. Journal of personality and social psychology, 54(6):1063, 1988.
[21] Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf.
[1] Berkay Ekici, Ebru Akgul Ercan, Sengul Cehreli, and Hasan Fehmi Töre. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter.
The effect of emotional status and health-related quality of life on the arXiv preprint arXiv:1910.01108, 2019.
[22] Krishnapriya Vishnubhotla and Saif M Mohammad. Tweet emotion
dynamics: Emotion word usage in tweets from us and canada. arXiv
preprint arXiv:2204.04862, 2022.
[23] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion
Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention
is all you need. Advances in neural information processing systems, 30,
2017.
[24] Scott M Lundberg et al. A unified approach to interpreting model
predictions. NIPS’17, 30, 2017.
[25] Zhongqi Yang, Iman Azimi, Salar Jafarlou, Sina Labbaf, Jessica Borelli,
Nikil Dutt, and Amir Rahmani. Loneliness forecasting using multi-
modal wearable and mobile sensing in everyday settings. medRxiv, pages
2023–06, 2023.
[26] Patrick H Finan, Phillip J Quartana, Bethany Remeniuk, Eric L Garland,
Jamie L Rhudy, Matthew Hand, Michael R Irwin, and Michael T Smith.
Partial sleep deprivation attenuates the positive affective system: effects
across multiple measurement modalities. Sleep, 40(1):zsw017, 2017.
[27] Luis Carretié, José A Hinojosa, Jacobo Albert, Sara López-Martín,
Belén S De La Gándara, José M Igoa, and María Sotillo. Modulation of
ongoing cognitive processes by emotionally intense words. Psychophysi-
ology, 45(2):188–196, 2008.
[28] Jiemin Yang, Jing Zeng, Xianxin Meng, Liping Zhu, Jiajin Yuan, Hong
Li, and Nasir Yusoff. Positive words or negative words: Whose valence
strength are we more sensitive to? Brain research, 1533:91–104, 2013.
[29] M Suhr, AK Risch, and G Wilz. Maintaining mental health through
positive writing: Effects of a resource diary on depression and emotion
regulation. Journal of clinical psychology, 73(12):1586–1598, 2017.
[30] Joanna C Yau, Stephanie M Reich, and Tao-Yi Lee. Coping with stress
through texting: an experimental study. Journal of Adolescent Health,
68(3):565–571, 2021.