Workshop on Computing for Physical and Mental Health

The first CLPsych workshop was held in June 2014. Although it was hosted at the annual conference of the Association for Computational Linguistics, it was unusual for an ACL workshop, because its main focus wasn't advances in language science and technology per se: the primary goal of the workshop was to create a new conversation about how language technology might have an *impact* in actual clinical practice. Here we are five years later -- the fifth anniversary CLPsych workshop was held in June 2018. What kind of progress have we seen at the intersection of CL and Psych? What's the progress like, and have the core challenges changed or remained the same? What is the place of this research within the broader range of related research? And where do we go next?

Suicide is a top 10 cause of death worldwide, claims the lives of more than 42000 Americans each year (WHO, 2014; CDC 2016). Part of the problem is that experts are no better than chance at predicting who will attempt to take their life (Franklin et al., 2016). We present work detailing the use of natural language processing on social media language data to screen for suicide risk, with apparent utility far exceeding clinicians. In many ways, the technical challenges here are the easiest ones, we will discuss the largest remaining challenges -- pragmatic and ethical integration of this technology into the system of care.

The fields of psychology and psychiatry have wrestled with creating accurate, reliable diagnostic nosology with acceptable predictive value to inform treatment. Despite advances in neuroscience, many practices in mental health continue to be guided by descriptive, not explanatory, theories. Difficulties such as guild issues, 'silos' of expertise, low inter-rater reliability in diagnosis, and human biases impact not only the efficacy of clinical practice but the progress of research. The field of computational linguistics, specifically natural language processing, has potential to advance human ability to detect signal and enhance clinical diagnosis and treatment planning, however, clinicians and scientists must proceed in a state of mindful collaboration in their search for 'ground truth.'

All Machine learning and data science projects have one thing in common: data. The relatively low occurrence rate of deaths from suicide in the US, combined with the complicated nature of the topic created a challenging situation for research dedicated to suicide. This talk covers the lessons learned from the partnership between a social media community, The American Association of Suicidology and Qntfy a startup software company. It encourages others who hope to make scientific progress amidst challenging circumstances.

Recent years has demonstrated the tremendous value of social media data in understanding mental health. Various efforts have demonstrated the ability to diagnose mental health disorders from online media, track population level trends, and support online mental health settings. An open question remains how social media data and related analytics can be used to shape the delivery of mental health care. In this talk, I'll present some of the issues with using such data in care and discuss a path forward.

A key focus of the Georgetown Information Retrieval Lab in the recent years has been the searching and mining of health-related data, from clinical notes, scientific literature, Web, and social media. Mental health is a significant and growing public health concern. As language usage can be leveraged to obtain crucial insights into mental health conditions, there is a need for large-scale, labeled, mental health-related datasets of users who have been diagnosed with one or more of such conditions. The focus of this talk is on the construction of the publicly available self-diagnosed mental health datasets (RSDD and SMHD) from Reddit posts, by creating high precision patterns, and without the need for manual labeling.

Classification and prediction of suicide ideation in high-risk groups is crucial in preventing suicide. In 2015, suicide rates among United States military veterans were 2.1 times higher compared to non-veteran adults. The purpose of this work is to automate detection of suicide ideation in veterans from acoustic and semantic features of speech. A total of 188 audio interviews were recorded by veterans (N=94) in addition to self-report psychiatric scales and questionnaires. We then built a classifier that differentiates between suicidal and non-suicidal veterans based on acoustic features of speech and sentiment analysis of the transcribed narratives. Support Vector Machine (SVM) classifier correctly identified veterans with suicidal ideation with an overall accuracy of 86.4% and area under the receiver operating characteristic curve (AUC) equal to 79.2%. Our findings indicate that speech analysis can be useful in suicide ideation detection. This machine learning approach could help clinicians identify high-risk veterans in a clinical setting.

Mental illnesses are predicted to become the leading disease burden globally in the next decade, according to the World Health Organization. Suicide is already a major public health problem as a leading cause of death in the United States. The effects of suicide go beyond the person who acts to take his or her life: it can have a lasting effect on family, friends, and communities. Health care providers understand the risk factors and they can help prevent suicide using evidence-based treatments and therapies. Advances in technology create opportunities for close collaboration between data analysis of published literature and mental health researchers.

In this work, we used our text mining and natural language processing tools to analyze the PubMed literature that contained the term of interest “suicide” and documents that were annotated with a Mesh Term relevant to the term “suicide”. We applied unsupervised clustering algorithms to analyze the landscape of the suicide-related literature and we identified groups of documents discussing specific topics and groups of closely related terms associated with each topic. We further identified gene, protein, drug, and disease name mentions in this dataset and associated them with computed clusters. In addition, we examined PubMed queries and present statistics on queries related to mental health. Based on the queries and documents clicked, we identified journals as well as PubMed articles that PubMed users have found to be most insightful.

In many NLP applications, temporal information is either not explicitly modeled or downright ignored (e.g., through the removal of stop words). In this talk, I argue that temporal information is vital for a complete understanding of text, especially in the health domain. We’ll look at clinical and radiology notes to show how temporal information can be modeled in these domains to construct a patient timeline from the unstructured text. Then we’ll explore how temporality affects statements of mental health diagnoses in online text.

Research Domain Criteria (RDoC), is a framework for studying mental disorders. Its aim is to support a better understanding of the basic dimensions of human behavior, from normal to abnormal. In support of this goal, the RDoC framework is organized into five domains, each capturing a different and cross-diagnostic perspective on psychiatric illness and health. In the CEGS N-GRID challenge, we focused on just one domain: positive valence. Abnormalities of positive valence may be observed in disorders as diverse as substance abuse and dependence, mania, gambling, obsessive-compulsive disorder, and depression. The goal of the CEGS N-GRID challenge was to determine lifetime maximum symptom severity of patients in positive valence domain, based on their psychiatric intake interview reports. 1000 reports were de-identified and annotated by experts for positive valence symptom severity. This talk will review this data set and give an overview of other related datasets generated by National NLP Clinical Challenges (n2c2, formerly known as i2b2) team since 2006.

Intelligent natural language processing techniques can be very helpful for radiology interpretation and education. Our primary focus has been on determining substantive changes between preliminary trainee radiology reports and faculty final reports, but applications for report summarization, radiology-pathology correlation, and pathology adequacy determination are all future directions where these techniques could be quite fruitful.

The Agency for Healthcare Research and Quality notes that, "In a learning health system (LHS), internal data and experience are systematically integrated with external evidence and that knowledge is put into practice. As a result, patients get higher quality, safer, more efficient care, and health care delivery organizations become better places to work." One important prototype LHS in pediatric medicine, the learning collaborative, is often aimed at improving care and outcomes for children living with chronic diseases. In this talk we describe a conceptual model of learning collaboratives and discuss opportunities for machine methods in three general areas of research concerning learning collaboratives: (1) learning from diverse data mined from EHRs, patient registries, and social media; (2) measurement and evaluation of improvement of patient health and care center performance through time; and (3) campaigning to increase patient activation and engagement in care improvement. Examples from preliminary work in each area are shown to illustrate.

This talk will provide an overview of various efforts developing and integrating NLP models in three healthcare applications: patient safety event reports, emergency department chief complaints, outpatient physician notes. Supervised and unsupervised modeling approaches have been applied to patient safety event reports to identify themes as well we classify event types. The free-text in chief complaints of emergency department patient visits were used to classify patients at higher risk for spinal cord compression. Lastly classification models were built to identify patients with completed diabetic retinal eye exams from physician notes.

Smartphones are powerful tools for obtaining information and connecting with others. But does being constantly connected to the internet have hidden costs for the fabric of social life? We explore the effects of smartphones both on fundamental social behaviors and on the benefits of face-to-face interactions.