Study reveals AI’s potential to detect loneliness by deciphering speech patterns

In a new study published in Psychiatry Research, scientists have discovered that artificial intelligence (AI) can detect loneliness by analyzing unstructured speech. This research offers promising new methods for identifying and addressing loneliness, particularly in older adults, through the nuanced analysis of how people communicate.

Loneliness is a pervasive issue affecting people of all ages, with older adults being particularly vulnerable. Defined as the distress caused by a gap between desired and actual social relationships, loneliness can significantly impact both mental and physical health. The problem is exacerbated by age-related factors such as the loss of loved ones, decreased mobility, and health challenges.

Traditional assessments of loneliness rely on self-report scales like the UCLA Loneliness Scale and the DeJong Giervald Scale, which can be time-consuming and subject to biases. Recognizing these issues, the researchers aimed to develop an AI model that could analyze speech data to detect loneliness, thereby offering a more scalable and less intrusive method of assessment.

“The impact of loneliness and social isolation can be devastating as we age. This has been something my patients have been reporting for a while and I wanted to know what we can do to combat this growing problem,” said study author Ellen E. Lee, an associate professor at UC San Diego and staff psychiatrist at the San Diego VA Healthcare System.

The study included 97 older adults, with ages ranging from 66 to 101 years, living independently in a senior housing community in Southern California.

Free daily newsletter

To gather sociodemographic data, trained study staff conducted clinical interviews. Data collected included age, sex at birth, racial background, years of education, and marital status.

The primary assessment tool for loneliness was the UCLA Loneliness Scale (Version 3), a validated 20-item self-report survey that measures various aspects of social functioning without explicitly using the word “lonely.” Participants were categorized as lonely or not lonely based on their scores on this scale.

Qualitative interviews covered six main topics: social relationships, loneliness, successful aging, meaning and purpose in life, wisdom, and technology use. These semi-structured interviews were conducted by a single interviewer and were audio-taped and transcribed.

Using the linguistic features extracted from the interview transcripts, the researchers developed an AI model based on transformer neural networks. Explainable AI (XAI) techniques were employed to identify which aspects of the speech data were most indicative of loneliness. The analysis revealed that both semantic and non-semantic elements of speech were significant indicators.

The semantic elements of speech, which relate to the meaning and content of the words, revealed that lonely individuals frequently referenced social status, religion, and expressed more negative emotions. For instance, in discussions about the meaning and purpose of life, lonely participants often mentioned social status and religion more prominently. This suggests that these individuals might be seeking validation or comfort in these areas.

Conversely, non-lonely individuals often talked about family and lifestyle, indicating a focus on social connections and activities that likely contribute to their sense of fulfillment and community.

The use of personal pronouns also varied significantly. Lonely individuals used first-person singular pronouns like “I” and “me” more frequently, reflecting a more self-focused perspective. In contrast, non-lonely individuals used first-person plural pronouns such as “we” and “ours,” suggesting a greater sense of inclusion and connection with others. This aligns with the understanding that loneliness is often accompanied by a heightened sense of isolation and self-focus.

“We found that non-social themed words were also reflective of loneliness depending on the different types of interview questions and prompts,” Lee told PsyPost. “For example, lonely individuals use more feeling adjectives when they are asked to describe wisdom. This may reflect how we tend to respond to different questions when we feel more lonely or highlight other qualities or experiences that are related to loneliness.”

Non-semantic elements, which include aspects of speech that convey how something is said rather than what is said, were also critical indicators of loneliness. The study found that conversational fillers (e.g., “uh,” “um”), non-fluencies (e.g., repetitions, false starts), and internet slang (e.g., “lol”) were more prevalent in the speech of lonely individuals.

Moreover, the use of causation conjunctions (e.g., “because,” “therefore”) was more common among lonely individuals, indicating a tendency to provide more detailed and perhaps rationalized explanations of their experiences.

The findings provide evidence that “the way we communicate may reflect our feelings about social relationships,” Lee said. “Newer AI approaches that allow us to open the ‘black box’ and how language can be linked to social functioning.”

But the study, like all research, has some limitations. “These findings are from a relatively small sample of older adults in San Diego, who were generally well-educated and primarily White,” Lee noted. “Our models can help generate new hypotheses about different concepts, however, we will need more inclusive and diverse participants to build unbiased and informative models.”

“We hope to expand our work to more diverse populations, looking at older adults with serious mental illnesses , and understand how to better assess loneliness to capture the large heterogeneity and personal nature of loneliness.”

The study, “Decoding loneliness: Can explainable AI help in understanding language differences in lonely older adults?“, was authored by Ning Wang, Sanchit Goel, Stephanie Ibrahim, Varsha D. Badal, Colin Depp, Erhan Bilal, Koduvayur Subbalakshmi, and Ellen Lee.