When Language Reveals What’s Coming: AI-Enabled Early Detection of Psychosis
The warning signs of psychosis can appear years—sometimes decades—before a first episode. Kids show up at community mental health clinics with diffuse symptoms. Anxious. Withdrawn. Something’s off, but nothing’s definitive. They get bounced between providers, misdiagnosed, treated for the wrong thing. Meanwhile, subtle changes in how they organize their thoughts are already visible, hiding in plain conversation.
The problem isn’t that clinicians don’t care. It’s that the gold-standard test to detect these early signs takes over five hours to administer and score. Nobody has that kind of time.
Amar Mandavia, a clinical psychologist at VA Boston and Boston University, and Enrique “Kike” Gutiérrez, an AI researcher at the Polytechnic University of Madrid, are working to change this. Through MIT’s linQ Catalyst program—a “needs-based innovation” fellowship that starts with problems, not solutions—they built CHiRP, a tool that extracts formal thought disorder signals from ordinary clinical conversations.
Four Signals in Speech
CHiRP measures four markers that clinicians have long recognized but rarely had time to formally assess:
- Looseness of association: abrupt topic shifts without preparing the listener
- Incoherence: thoughts that don’t connect logically
- Poverty of content: speech that sounds full but says little
- Illogical thinking: conclusions that don’t follow from premises
Gutiérrez demonstrated loose association during the episode: “My name is Kike. I love listening to music. My favorite character in the Simpsons is Patrick. I shifted topics three times without preparing you.”
The manual test that measures these patterns requires an hour to administer and four more to score. CHiRP runs in the background during a routine intake conversation.
The Ethical Knot
Early detection sounds straightforwardly good. But Mandavia raised a hard question that the field hasn’t fully resolved: does labeling someone as “at risk for psychosis” help them, or does the stress of that label trigger the very condition you’re trying to prevent?
“There is this hypothesis that actually adding this idea of being labeled as prodromal could be a stressor that could then lead to the very condition that you’re trying to prevent,” Mandavia explained. The team is taking cues from genetic counseling—where similar dilemmas exist—and building transparency into every step of the process.
The Payment Problem
Even if the technology works, who pays for prevention? Mandavia described approaching private insurers with what seemed like a compelling pitch: help us identify at-risk patients earlier, and you’ll save on long-term care. The response was sobering.
“A lot of folks who end up developing schizophrenia down the road become a problem for the government system—Medicaid or Medicare. They don’t tend to be taken care of by private insurers. So their incentive for early identification was not as loud as we thought.”
This is the structural challenge that makes so much preventive care hard to fund: the entity that pays for intervention isn’t the one that reaps the savings.
What to Take Away
Mandavia’s closing thought stuck with me: “What we’re coming up with is a proxy of a better tool that listens better than a provider can.”
That’s a humble framing for important work. The real insight may not be the algorithm itself, but the reminder that clinical conversations already contain information we’ve been too busy to hear.
Listen to the full episode with Amar Mandavia and Enrique Gutiérrez on Practical AI in Healthcare.