Detecting and Decoding Loneliness, Using AI
Stevens teams up with IBM and two universities to spot signs of alienation from transcribed interviews
A Stevens artificial intelligence expert has teamed up with a major corporate AI partner and two universities to develop an AI-based system that predicts loneliness in the elderly. The technology may help us better understand the roots of people's alienation, too.
Professor K.P. “Suba” Subbalakshmi — who has previously developed AI-powered systems to detect depression, misinformation, deception and even orbiting space junk — collaborated on the project with the University of California-San Diego, Jiangnan University and IBM.
For the study, the team surveyed about 100 residents of a California continuing-care senior housing community ranging in age from 66 to 101. Subjects averaged 83 years old, while nearly two-thirds were women. Their states of mind were assessed in two ways: through personal interviews with a trained psychiatrist, as well as with the UCLA Loneliness scale (UCLA-3) — a commonly used 20-point written assessment tool that participants complete themselves to self-report moods.
About 40% of the interviewees scored as “lonely” on the UCLA-3 scale, based on those evaluations.
Next the team transcribed recorded audio of all the in-person interviews and fed transcripts, along with the UCLA-3 survey data, into a specially designed model that quickly ‘learned’ to tell the linguistic differences in responses between lonely and non-lonely people and offer predictions.
Presented with the interview transcripts and no other context, the AI proved 89% to be accurate at correctly classifying whether people did indeed report feeling lonely or not.
Internal communications: AI that reports on itself
But that’s not all.
Certain component of the AI the research team designed helped the group better understand the diversity and nuances of the responses of the participants’ answers.
These layers and portions of the algorithmic system are known as explainable AI (XAI), and can provide powerful insights into the conclusions and predictions of AI systems that are normally opaque in their operation.
“XAI is a very exciting field,” explains Subbalakshmi. “Here we used a component of the network known as additive attention to process and analyze each section of the interviews.
“Incorporating the additive attention module not only gave our model strong predictions; it also enabled us to learn more afterward about the most useful details of the interview responses themselves.”
In other words, as they worked on their analyses the algorithms also reported back on which sections of each interview were most important in the final prediction about whether a person felt lonely or not.
Using increased numbers of emotional adjectives and verbs in answers to the human interviewers, for instance, tended to point toward a person that felt lonely, as did increased references to religion and highly analytical responses. The increased use of personal pronouns also seemed to point to loneliness.
“XAI can play a crucial role not only in identifying individuals at risk of loneliness but also in understanding loneliness itself,” says Subbalakshmi.
That’s important, she says, since it’s now believed there are various subtypes of loneliness that probably each require different, more personalized and precisely targeted interventions to improve well-being
“Loneliness caused by, for example, actual social isolation might be remedied by building more and stronger social connections,” Subbalakshmi explains. “But a feeling of loneliness experienced by a person who is already surrounded by friends and family could stem from a different cause requiring a different intervention.
"This proof-of-concept study shows how the new techniques of XAI may hold the potential to help professionals separate out and elevate important clues and cues in conversations with the elderly, in order to help them improve or recover emotional wellness."
The research, which was supported by both IBM Research and the National Institutes of Health, was reported in Psychiatry Research [339, 16078].