Classifying Severity Level of Psychiatric Symptoms on Twitter Data

Abstract

Internet-based social media sites such as Twitter represent a growing level of modern experience. These sites claim a large number of users and their influence is increasingly being experienced in clinical practice. Psychiatric disorders currently are affecting many people from different cultures, ages, and geographic locations. Although, the majority of individuals who experience symptom of psychiatric disorder practice the desire to be isolated, which drives them to use online channels to share their feelings. Hence, this sites provide a way to detect undiagnosed psychiatric disorders. In order to address this issue, we propose a model to classify the severity level of psychiatric symptoms (i.e. depression, anxiety, and bipolar) based on a data extracted from Twitter. The model is employed by fusing the linguistic features of Term Frequency Inverse Document Frequency (TFIDF) weighed by N-gram (unigram, bigram, and trigram), and word2vce, with Pattern of Life Feature (PLF) that take polarity, subjectivity, and gender. The experiment, is conducted by incorporating with machine learning classifiers of Support Vector Machine (SVM), Logistic Regression (LR), Random Forest (RF), and Naïve Bayes (NB). Experimental results show that SVM with features of TFIDF weighed by unigram combined with PLF outperforms with an accuracy score of 97.3%. In future, the proposed model could be employed to include lexicon-based features for classifying various psychiatric symptoms with a combined approach of machine learning and lexicon-based classification.