Automatic Detection of Psychological Distress Indicators and Severity Assessment from Online Forum Posts

Proceedings of COLING 2012: Technical Papers, pages 2375–2388, COLING 2012, Mumbai, December 2012. Automatic Detection of Psychological Distress Indicators and Severity Assessment from Online Forum Posts Shirin Saleem1, Rohit Prasad1, Shiv Vitaladevuni1, Maciej Pacula1, Michael Crystal1, Brian Marx2,3, Denise Sloan2,3, Jennifer Vasterling2,3 and Theodore Speroff4,5 (1) Raytheon BBN Technologies, 10 Moulton St, Cambridge, MA, U.S.A. (2) National Center for PTSD at VA Boston Healthcare Sytem, Boston, MA, U.S.A. (3) Boston University School of Medicine, Boston, MA, U.S.A. (4) VA Tennessee Valley Healthcare System, Nashville, TN, U.S.A. (5) Vanderbilt University School of Medicine, Nashville, TN, U.S.A. {ssaleem,rprasad,svitalad,mpacula,mcrystal}@bbn.com {brian.marx,denise.sloan,jennifer.vasterling}@va.gov ted.speroff@vanderbilt.edu ABSTRACT Psychological disorders are frequently under-diagnosed and consequently have an irreversible impact on individuals and society. The stigma associated with such disorders makes face-to-face discussions with family members and clinicians difficult for many individuals. In contrast, people openly relate experiences on Internet forums. This paper describes a novel system that analyses forum posts to: (1) detect distress indicators that directly map to the Diagnostic and Statistical Manual of Mental Disorders (DSM) IV constructs, and (2) assess the severity of distress for prioritizing individuals who should seek clinical help (i.e. triage). For distress indicator detection, we use support vector machines (SVMs) trained on a suite of innovative intra- and inter-message features. We show significant improvements in multi-label classification accuracy using human- generated rationales in support of annotated distress labels. For triage assessment, we demonstrate the effectiveness of Markov Logic Networks (MLNs) in dealing with noisy distress label detections and encoding expert rules. KEYWORDS: Psychological Distress, Web forums, Text classification, Annotator rationales, Support Vector Machines, Probabilistic Logic, Markov Logic Networks. 2375 1 Introduction Psychological health disorders pose a growing threat to individuals, their family members and to society. Disorders such as Depression, Post-Traumatic Stress Disorder (PTSD), and mild Traumatic Brain Injury (mTBI), are often under-diagnosed and under-treated (Kessler et. al, 1999). Failure to intervene early and effectively impacts individuals and their family members adversely and results in profound long-term costs to society. The standard approach to diagnosing psychological health disorders is through a series of clinically administered diagnostic interviews and tests (Weathers et. al, 2001). However, assessment of patients using these tests is expensive and time-consuming. Furthermore, the stigma associated with mental illnesses motivates inaccurate self-reporting by affected individuals and their family members, thus making the tests unreliable. In recent years, there has been a tremendous growth in social interactions on the Internet via social networking sites and online discussion forums. In contrast to clinical tests, the Internet is an ideal, anonymous medium for distressed individuals to relate their experiences, seek knowledge, and reach out for help. Web-forum discussions of symptoms, thoughts and experiences are open, descriptive, and honest, making them an ideal source for observing communications of individuals for assessing psychological status. In this paper, we present a multi-stage text classification system for assessing psychological status of individuals based on their text postings on online web forums. Specifically, our system combines state-of-the-art NLP and machine learning techniques to: (1) extract fine-grained psychological distress …

Content truncated to 500 words