Survivors of cancer are at-risk for the lifelong e ects of disease and treatment. A signi cant number of them face Post-Traumatic Stress Disorder (PTSD) that may adversely a ect their mental health. Twitter is a social networking site that allows users to interact with others by posting short messages (tweets). These tweets, which to a certain extent re ect the users psychological state, are convenient for data collection. However, Twitter also contains a mix of noisy and genuine tweets. The process of manually identifying the genuine tweets is expensive and time-consuming. Thus, we stream the data using cancer as a keyword and lter the tweets with cancer-free and PTSD related keywords without having to label each tweet manually. Convolutional Neural Network (CNN) learns the representations of the input to identify cancer survivors with PTSD. The experiments on real-world datasets show that the model outperforms the baselines and correctly classi es the new tweets.
PTSD is an anxiety disorder that occurs in some people after experiencing or
witnessing life-threatening events, may severely a ect daily life activities. Being
diagnosed with cancer often causes psychological distress due to painful
treatments and traumatic experiences of cancer survival [
Recent work in psychology aims to analyze the manner of self-declared
mentally ill users from their interactions and behavior based on written posts [
Most of the mental health studies focus only on capturing the psychological problem in society. Besides, the applied analysis approaches such as manual labeling, crowd-sourcing, Twitter Firehouse, and Linguistic Inquiry Word Count (LIWC) are time-consuming and expensive. Thus, data preparation and analysis procedure are challenging. To tackle these challenges, we propose an algorithm to create a labeled dataset using cancer-free related keywords and PTSD features. Also, without having to manually check whether a tweet was written by a cancer survivor with PTSD or not, DNN is able to captures important features from the input dataset. In this work, we used the DNN approach that learns to extract a di erent level of meaningful representations of texts.
Experts from various elds strive to propose reliable detection models for
mental health problems. Despite having the same mission, they touched the
issues from di erent angles. Human emotions can be expressed in many forms of
physiological states such as nervous system responses, blood ow, facial
expression, and vocal acoustics [
Treated cancer patients may nd themselves at risk of getting cancer
recurrence [
The Proposed Framework for Cancer Survivors Living with PTSD Diagnosis on Social Media
Feature Extraction: The top part of Figure 1 shows the overview of previous
work related to depression detection using social media data [
CNN Architecture: We adopted only one convolutional layer with embedding layer in the CNN network setting to produce results for tweets classi cation. It requires the speci cation of the vocabulary size, the size of the real-valued vector space, and the maximum length of words in input tweets. For convolutional feature maps, we used word embedding with 100-dimensional for text representation. 32 lters were applied with a kernel size of 8 and a recti ed linear (ReLu) activation function. Followed by a pooling layer, the lters will generate feature maps and reduces the output by half. The end layer uses a sigmoid activation function to output a value between two categories of positive and negative in the tweets based on the concatenated the previous vectors. 3
In this experiments, the dataset with PTSD positive represents the diagnosed group, while PTSD negative represents the control group. For the diagnosed group, we retrieved tweets from a user who publicly stated that they survived cancer and had PTSD. To construct the PTSD negative group, we make uses of tweets from `Twitter User Gender Classi cation' dataset from Kaggle website 3. Both groups have the same total tweets of 10k to create balanced datasets. The data preparation phase includes three tasks: (1) splitting the dataset into 80% for training and 20% for testing, (2) cleaning the dataset to remove punctuation, stop words, and numbers; (3) de ning a vocabulary of preferred words from a training dataset by stepping through words and keeping only tokens with minimum occurrences of ve. We used Keras API running on Tensor ow to train deep learning models. All the models were trained with ten epochs through the training data. The e cient Adam implementation of stochastic gradient descent was used.
The three baselines that are capable of handling text dataset used for evaluating our proposed algorithm are Multiple Layer Perceptron (MLP), CNN n-gram, and Recurrent Neural Network (RNN). Our results indicate that CNN can effectively identify cancer survivor with PTSD. Experimental results in Table 1 show the accuracy of CNN of 98.5% is higher than CNN n-gram by 0.2%. We ran the experiments multiple times due to the stochastic nature of DNN to get the reasonably accurate result. 3 https://www.kaggle.com/crowd ower/twitter-user-gender-classi cationgenderclassi er-DFE-791531.csv
PTSD is a severe anxiety disorder that a ects individuals who are exposed to traumatic events such as cancer. Cancer survivors are at risk of short-term or long-term e ects on physical and psychosocial well being. Therefore, the evaluation and treatment of PTSD are essential parts of cancer survivorship care. We propose a prediction system with a CNN model that can produce promising results. Experimental results demonstrated that CNN was able to capture important signals from texts in determining PTSD among cancer survivors. The social media users with cancer history who su er from depression will bene t from the prediction system. It will act as an alarming system by detecting the depression presence based on users' postings.