Abstract:
Globally, depression is a leading cause of disability and impaired quality of life.
The COVID-19 global pandemic has seen an increase of mental health disorders
particularly depression, which often goes untreated. Several key barriers contribute
to this, including social stigma, di culty accessing mental health services as they
are overwhelmed, and lack of timely objective assessment approaches. A system
that helps monitor mood changes on a 24/7 basis will help identify when someone
may be starting to develop depression and therefore aid with earlier diagnosis and
relapse prediction.
The current literature has shown traditional machine learning techniques such
as regression and ensemble learning to be e ective in tackling the problem of mood
classi cation and prediction. In this work, we use deep learning and neural network
approaches to tackle this problem with the aim of improving we can improve the
accuracy and lowering the variance of existing methods.
We acquired a dataset from an existing study (of n = 14 participants) using
regression and ensemble learning techniques and developed our own Neural Network
model using multilayer perceptron models to tackle the classi cation task.
We tried a neural network approach across multiple subsets of their dataset with
varying success. Our best model performed on par with the existing Shah et al
model with 17 out of 32 total measurements improving on that of Shah et al and
providing consistently provided lower variance than their model.
Additionally, where we collected a similar dataset using technology and sensors
from smartphones and smartwatches for one month. A total of 15 healthy individuals
participated in the study. This data was then used in traditional monolithic
and compositional neural network models to perform mood classi cation using
smartphone-enhanced ecological momentary assessments and physiological data
collected from a Fitbit smartwatch.
The compositional and monolithic neural network approaches developed using
Keras provided promising results with accuracies ranging from roughly 60 - 90%. However, there were limitations to our collected dataset as the distribution
of labels was limited and may have resulted in over- tting. This was expected as
all participants were healthy controls, as it was challenging to recruit a depressed
group due to the COVID-19 pandemic. Future work can be done by adding extra
features and samples to improve the neural network models but overall this
work showed that the deep learning approach has the potential for accurate mood
prediction.