AI framework spots psychosis risk in youth speech

Researchers at Nanyang Technological University (NTU Singapore) and the Institute of Mental Health (IMH) have engineered an artificial intelligence system that detects early signs of psychosis in young people through brief recordings of everyday conversation. The framework analyzes spontaneous speech to identify individuals at ultra-high risk (UHR), a clinical state marked by warning factors that significantly raise the likelihood of developing severe mental health conditions characterized by hallucinations, delusions, and disordered thinking.

Limitations of standard models

This study marks the first instance in Singapore focusing on early psychosis risk among youths via large language models (LLMs). Direct application of standard LLMs to predict clinical risk from raw transcripts proved ineffective, yielding results no better than random guessing and failing to identify most at-risk participants. The research team attributes this shortfall to the absence of suitable training datasets and the inherent nature of casual dialogue, which rarely includes explicit symptom descriptions or structured clinical data.

The challenge lies in "non-disclosing speech," where individuals do not verbally articulate their mental health struggles. For example, a person experiencing depression may not explicitly state they feel depressed. Consequently, assessing whether AI can discern potential risk solely from informal conversation addresses a critical barrier to early intervention.

Lemmon-AID performance metrics

To resolve these limitations, the team introduced LEMon-AID (Latent Emotion Proxy for Mental Health Detection with AI-integrated Decision-making). This approach utilizes LLMs to extract subtle emotional profiles—such as joy, trust, fear, and sadness—from text, then integrates these signals with traditional machine learning predictions. Published in npj Mental Health Research, the study evaluated transcribed samples from 429 participants aged 14 to 29 enrolled in the Longitudinal Youth-at-Risk Study (LYRIKS).

Initiated in 2008, LYRIKS is a landmark Singaporean study designed to assess social, clinical, and biological risk factors for youths at UHR for psychosis. Each participant provided a 45-second non-self-disclosing speech sample responding to neutral prompts, such as arguing why they should win an award for being the 'most interesting person'. LEMon-AID demonstrated superior performance compared to conventional methods, boosting overall predictive accuracy from 55 per cent under direct LLM prompting to 67 per cent. Furthermore, it increased detection sensitivity to identify up to 93 per cent of at-risk youths.

Transparent reasoning mechanisms

Wilson Goh, Assistant Professor of Biomedical Informatics at LKCMedicine and the NTU Centre of AI in Medicine, served as co-lead investigator. He noted that standard AI models struggle because people rarely discuss symptoms openly. Instead of requesting a direct diagnosis, LEMon-AID employs language models to interpret underlying emotional cues, connecting patterns with mental health risks while maintaining patient privacy and providing clear explanations.

Unlike traditional machine learning models that often function as 'black boxes', LEMon-AID generates natural-language explanations alongside its risk assessments. This transparency allows clinicians to understand the reasoning behind each prediction, positioning the technology as a safer tool for medical evaluation. Matheus Calvin Lokadjaja, the study’s first author and a Research Assistant at LKCMedicine, explained that shifting the AI's focus to mapping emotional signals like joy, sadness, or trust uncovered meaningful patterns previously hidden in everyday speech.

Clinical validation roadmap

Dr Jimmy Lee, Group Chief Research and Innovation Officer at NHG Health, Senior Consultant at IMH, and Associate Professor at NTU's LKCMedicine, led the LYRIKS study. He emphasized that early detection is vital for preventing long-term disability and restoring futures for young people with mental illness. The research aims to equip frontline community workers and healthcare providers with tools to identify at-risk youths before conditions worsen, encouraging timely care-seeking behavior.

Future work will explore various clinical deployments of the LEMon-AID framework. Researchers plan to incorporate multimodal approaches, integrating data types such as acoustic cues, vocal pitch, speech rhythm, and facial expressions to enhance system reliability. These additions aim to clarify how patterns across speech, behaviors, and acoustics inform mental health progression. Extensive real-world validation remains necessary before practical implementation can occur.

The study received support from the Singapore Ministry of Health through the National Medical Research Council (NMRC) Office and MOH Holdings Pte Ltd under the NMRC Translational and Clinical Research Flagship Programme (grant number NMRC/TCR/003/2008).