Philip Resnik's Research on Mental Health

Brief Overview

Language is a crucial window into people's mental state. In my work on mental health, colleagues, students, and I are applying natural language processing and machine learning methods to help uncover evidence in language use connected with mental health issues with a particular focus on suicidality, schizophrenia, and depression. In earlier work, I explored combining unsupervised and supervised machine learning methods; this actually emerged as an outgrowth of work I was doing in computational political science -- the key observation being that , just as evidence in language can be connected to ideological distinctions like liberal versus conservative, so too it can be connected using similar kinds of computational models to other distinctions related to someone's mental life, like whether or not they are suffering from depression.

In 2014, with clinical psychologist Rebecca Resnik and AI researcher Meg Mitchell (who later founded Google’s Ethical AI team), I co-founded the Workshop on Computational Linguistics and Clinical Psychology (CLPsych), which brings together technologists and clinicians and has helped to build critical mass and attention in the computational linguistics research community to mental health as a problem space. In order to help solve the critical problem of shared community-level access to relevant data, my group at UMD created the UMD Reddit Suicidality Dataset, which (with ethical governance in collaboration with the American Association of Suicidology) has been shared with more than thirty research teams internationally, and with the help of an Amazon Machine Learning Research Award I've been collaborating with NORC at the University of Chicago to create the UMD/NORC Mental Health Data Enclave, a secure infrastructure backed on AWS that helps solve the problem of shared access to sensitive datasets by bringing researchers to the data rather than vice-vera.

Over the past several years I have been focusing on laying the underpinnings for application of NLP and machine learning technology in the real-world mental health ecosystem. This work begins with a paradox: if technology gets better at surfacing people whose mental health problems aren't receiving adequate attention, that's going to make an already overburdened mental healthcare system worse. Colleagues and students and I have been tackling that problem via a shift in focus, from the problem of classification to the problem of prioritization. Obviously technology should not be deciding who should or should not receive care, but how can we use it to help clinicians better understand what's going on with their patients, particularly in between healthcare visits, in order to inform their treatment and be aware of more severe problems more quickly? To that end I've been working with colleague Doug Oard and our student Han-Chin Shing on clinician-centered model design and evaluation, and I've been working with colleague John Dickerson and students Samuel Dooley and Christine Herlihy on a multi-stage machine learning framework that optimizes available resources in order for individuals with more severe problems to progress through to higher degrees of intervention. These collaborative efforts also include work with Deanna Kelly in the UMD School of Medicine, focusing on data collection and assessment for schizophrenia and depression, and Carol Espy-Wilson, looking at ways to integrate language, speech, and facial emotion data in mental health.


Relevant Recent Publications