Page 148 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

Chapter 5 Depression discovery in cancer communities using deep learning 137

The positive or negative sentiments in conversation can help
in finding the mental disorders in cancer patients specifically
depression [65,66].
Zainuddin et al. [67] have done aspect-based SA on twitter data
and found the “hate” sentiment inside the data using ML algo-
rithms. A systematic review has been provided by Giuntini et al.
[68] on recognizing depression disorder in narrative provided by
the patients on social network using techniques of emotion anal-
ysis and SA. In Refs. [64] and [66] also, the researchers propose to
detect depression using social media posts.
The linguistic inquiry word count (LIWC) lexicon, which com-
prises of over 32 types of psychosomatic concepts [69], has been
extensively used for detection of mental disorders. The lexicon has
been used for recognizing insomnia [70], distress [71], postpartum
depression [72], depression [73], and posttraumatic stress disor-
der (PTSD) [92]. For the accurate categorization of these mental
ailments, investigators looked for features that were exclusive to
a certain ailment as well as those that overlaid with one other.
For instance, larger usage of first-person pronouns [75] than
second- and third-person pronouns [76] is being used to discover
users suffering from distress and depression. Similarly, age is
recognized as a differentiator between depression and PTSD.

3. Proposed system architecture

For the task of depression detection on Twitter, the training
data set that we propose to use is a fragment of the database given
by Ref. [77], which comprises of tweets from a total of 1145 Twitter
users. These tweets have been pigeonholed as control, depressed,
and PTSD. Furthermore, the data set is labeled based on two fea-
tures: sex and age. The data set has a total of 742,793 tweets in
the depressed category. This set of tweets in the depressed category
forms the labeled training data set used in this work.
The overall design of the proposed scheme for depression dis-
covery in cancer communities is outlined in Fig. 5.1. We propose
a technique for establishing which users may be slipping into
depression or are depressed from their tweets in cancer support
groups.
To this end, an efficient NN-based model is proposed that en-
hances the word embedding. In our model, to derive enhanced
embedding, the two frequently used word embeddings as shown
in Fig. 5.1, skip gram and CBOW, are used.