Stanford neuroscientists: Teen mental health problems can be predicted by speech patterns. Here’s what to watch for

Computer language models, which analyze what children say during stressful events, are better at predicting their future mental health than a panel of experts in the field. That’s the finding from a new study published in Nature Mental Health. Researchers ran recorded audio interviews of more than 200 children—ages 9 to 13 (on average 11 years old)—through four natural language processing models, asking the kids to speak about stressful events in their personal lives. The result: Those models were “very accurate” at predicting whether these same children developed mental health conditions six years later in adolescence, according to the study. Researchers drilled down to the actual words the children used—and what they found was fascinating. Across the models, the style of the kids’ speech mattered more than the content . For example, “small connector words,” such as “and,” “to,” and “but,” were better predictors of stress than the actual descriptions of the events. The language models showed the predictive power of linguistic style in indicating mental health, consistent with previous research that linked patterns in the use of certain words, such as “I” and first-person pronouns, and frequent use of prepositions and conjunctions, to mental health. Although less predictive, the actual content of speech did indicate some future problems and resilience as well. In short, “the study identifies markers of risk before individuals are diagnosed,” Chase Antonacci, the study’s lead author and a neuroscience doctoral student in Stanford’s School of Humanities and Sciences, said in a news release . The team utilized in-depth interviews from the Stanford lab of the study’s senior author, Ian Gotlib, originally conducted as part of a larger project that is following a group of young people over many years. It used what’s known as a Traumatic Events Screening Inventory (TESI) to ask about topics such as financial insecurity, parental divorce, abuse, and experiencing a natural disaster. The experts originally rated the child’s cumulative experience of stress with a single number. However, “we realized that there was probably so much richness, variability, and nuance that we were losing by reducing these clinical interviews to a single number, so we thought about other ways we could leverage those audio recordings,” Antonacci said. The findings show “great potential for a language-based assessment,” according to Gotlib, who said the next step involves running a larger dataset through the models. “If these findings hold, it means we may be able to just take smartphone recordings of children talking, analyze that speech, and identify which children are at risk, years before they might develop a disorder,” Gotlib said.