AI Framework Developed to Identify Early Signs of Psychosis in Youths

Psychosis often surfaces during adolescence or early adulthood, characterized by hallucinations, delusions, and disordered thinking. This study is notable as it is the first in Singapore to examine early psychosis risk among youths using large language models (LLMs), a category of AI that processes and generates human language.

The team asserts that this technology could assist primary care providers and community workers in recognizing at-risk youths before symptoms escalate. Identifying individuals at ultra-high risk (UHR) of psychosis can be challenging since youths often do not openly express mental health issues, a phenomenon termed "non-disclosing speech."

To explore this, the researchers published their findings in the journal npj Mental Health Research. They analyzed transcribed speech samples from 429 participants aged 14 to 29 involved in Singapore's Longitudinal Youth-at-Risk Study (LYRIKS), launched in 2008 to investigate various risk factors for psychosis in youths.

Participants provided 45-second speech samples in response to a neutral prompt, such as why they should be awarded the title of "most interesting person." The study revealed that standard LLMs performed poorly when asked to predict clinical risk from these raw transcripts, yielding results no better than random guessing and failing to identify many at-risk individuals.

The research team developed a new model called LEMon-AID (Latent Emotion Proxy for Mental Health Detection with AI-integrated Decision-making), which utilizes LLMs to extract subtle emotional indicators like joy, trust, fear, and sadness. This model improved overall predictive accuracy from 55% with conventional methods to 67%, enabling the identification of up to 93% of at-risk youths.

Wilson Goh, co-lead investigator of the study and assistant professor at LKCMedicine, emphasized that standard AI models often couldn't detect early mental health risks in everyday conversations, where symptoms are not explicitly communicated. LEMon-AID, however, analyzes emotional cues rather than seeking direct diagnoses, thus protecting patient data while providing clear explanations.

Dr. Jimmy Lee, group chief research and innovation officer at NHG Health and senior consultant at IMH, reinforced the importance of early detection in preventing long-term disability among young people with mental health issues. He highlighted that this research makes strides towards equipping frontline workers and healthcare providers with better tools for early identification of at-risk youths.

Matheus Calvin Lokadjaja, the study's first author, noted that standard AI models largely failed to detect risk from casual speech. By shifting focus to emotional signals throughout the conversation, the researchers uncovered significant patterns previously overlooked.

Future development of the LEMon-AID framework will include various clinical applications. Researchers plan to integrate multiple data types such as vocal pitch, speech rhythm, and facial expressions to enhance system reliability and understand mental health progression better. Further extensive research and real-world validation will be needed before practical applications can be realized.