Researchers from the University of Pennsylvania, New York University and Penn's Linguistic Data Consortium have received a two-year, $4 million grant from the Wellcome Trust to develop a scalable, AI-powered platform for training mental health clinicians.
STELLAR - short for Steering-Vector Enhanced LLM Agents for Realistic Digital Twins in Mental Health - will create virtual patients known as "digital twins" in the form of AI-driven simulations that allow trainees to practice clinical interviews with patient profiles whose psychiatric symptoms can be precisely adjusted.
One challenge in educating mental health clinicians is to prepare trainees for the complexity of real clinical conversations, where symptoms can overlap, shift over time and be expressed differently from person to person. STELLAR's goal is to provide an ethical, repeatable way for trainees to simulate interviewing patients across a wide range of symptom presentations, backgrounds and clinical scenarios.
"STELLAR brings together behavioral data, clinical expertise and AI to ask a very practical question," says Sharath Chandra Guntuku, Research Associate Professor in Computer and Information Science (CIS) within Penn Engineering and one of the project leads. "Can we build training tools that better prepare clinicians for how varied and complex patients are?"
"The promise of this approach is that we can move beyond stylized and potentially biased simulations," adds João Sedoc, Assistant Professor of Technology, Operations and Statistics within NYU's Stern School of Business and another project lead. "If we can create digital patients that simulate controllable plausible symptom expression and responsibly evaluate, we can augment current clinician training practices with the kinds of conversations that are essential to better mental health care."
The power of patient simulations
STELLAR's simulations will not be copies of individual patients, but rather composites based on real-world data that can help clinicians practice realistic conversations in a controlled setting.
For mental health training, being able to precisely control the symptoms students encounter is especially important. A trainee might need to practice interviewing a patient with mild anxiety, then a patient whose anxiety overlaps with depression or psychosis.
STELLAR is designed to make those variations possible, allowing researchers and trainers to adjust how symptoms appear, how strongly they are expressed and how they interact with one another.
In psychiatry, the details of symptom experience matter: how someone describes distress, how symptoms overlap, how severity changes over time and how context shapes the clinical interaction."
Raquel Gur, Karl and Linda Rickels Professor of Psychiatry with secondary appointments in Neurology and Radiology in Penn's Perelman School of Medicine
Turning data into patient simulations
To build those simulations, the researchers will draw on clinical data from the Philadelphia Neurodevelopmental Cohort, a research resource founded by Penn Medicine and the Children's Hospital of Philadelphia (CHOP) that includes psychiatric assessments and clinical interviews from thousands of young people.
The project will also draw on data from social media platforms, where mental health symptoms can appear in everyday language. "Many mental health symptoms do not appear only in formal clinical settings; they also come through in the way people talk day to day, including online," notes Guntuku.
Patient simulations will only be useful for clinician training if they are grounded in real clinical speech and evaluated as clinical interactions, not just plausible AI dialogue."
Neville Ryant, researcher with the Linguistic Data Consortium (LDC) at Penn and project lead on STELLAR
"LDC's role is to bring speech and language science into the core of the project: adapting speech-recognition tools to clinical interviews, creating high-quality transcripts and annotations, and helping evaluate both what the simulations say and how they say it," adds Ryant.
"That includes assessing the language generated by the models, the naturalness of synthetic voices, how well those voices reflect target speech patterns and the behavior of the avatar during real trainee interactions."
Centering human patients
Although STELLAR will use AI to create digital patients, the project cannot succeed on technical performance alone. The simulations must also be evaluated by the people most likely to understand what a realistic, respectful and useful interaction should feel like.
To that end, the team will involve people with lived experience of mental health conditions throughout the project, along with family members and caregivers who bring valuable perspectives to the evaluation process.
Their feedback will help the researchers assess whether the simulations reflect real patterns of symptom experience, avoid flattening or stereotyping patients, and prepare trainees for the complexity, diversity and nuance of real clinical conversations.
"By involving individuals with lived experience throughout the project, STELLAR can help us ask not only whether a digital patient is clinically accurate, but whether the interaction feels respectful, realistic and attentive to experiences that are too often missed," says Gur.
Enhancing training for mental health clinicians
Ultimately, STELLAR could give training programs a scalable way to prepare clinicians for more varied and complex patient encounters, including combinations of symptoms that may be difficult to encounter consistently during traditional training.
"By connecting real-world symptom expression with controllable digital simulations, we hope to make clinician training more scalable, rigorous and representative," says Guntuku.
Social media linked to worse mental health in Canadian seniors