AI 如何帮助精神分裂症的早期诊断

September 12, 2026 6 min read Add Us On GoogleAdd SciAm Listening for schizophrenia: how AI can help with early diagnosis The speech of people with schizophrenia sounds subtly different from their healthy counterparts. Can AI hear those differences better than human psychiatrists? By Astrid Landon & Knowable Magazine Sebastian Kaulitzki/Science Photo Library via Getty Images Schizophrenia is tough to diagnose. Patients may present with hallucinations (sometimes), social withdrawal (maybe) or delusions (not always). More generally, they just sound unlike themselves. Clinicians rely on their expertise and subtle cues to determine how different patients’ speech is, along with which symptoms appear over time, to justify leaning toward schizophrenia rather than another mental illness. This leads to delays in diagnosis. Americans with psychotic disorders — more than 3 million of whom have schizophrenia — receive a diagnosis a year and a half, on average, after their first symptoms appear. Researchers are now investigating whether artificial intelligence could improve diagnosis and care by listening to and analyzing what clinicians can’t hear or quantify, even if the software is working off just a few minutes of conversation. AI won’t make its grand entrance into the clinic tomorrow. But it’s being hailed as the new frontier in psychiatric care, one that could enable early, accurate detection and personalized monitoring of illnesses based on indistinct symptoms. “We have the tools to do that with the kind of precision that we have never had before,” says Thomas Insel, a psychiatrist and neuroscientist who led the US National Institute of Mental Health for 13 years. On supporting science journalism If you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today. Schizophrenia is a disorder that interferes with people’s perception of reality, their thinking and their emotional regulation. It affects about 23 million people worldwide and is usually diagnosed between the late teens and early 30s. No one knows what causes it, but research suggests it could be a combination of genetics, environment, brain chemistry and substance use. Clinicians stress the importance of detecting the disorder as early as possible, because the longer it is left untreated, the poorer the response to treatment and the greater the risk of brain tissue loss, worsening symptoms and suicide. But psychiatrists often make errors. Part of the problem is that diagnosis currently depends on subjective assessments. Based on what a patient says and how they say it, clinicians fill out one of several different rating scales to rank the severity of symptoms and arrive at a diagnosis. But the process is difficult to standardize, and clinicians’ scores for a given patient can differ by 30 to 50 percent. Artificial intelligence could help to automate the process, making diagnosis both speedier and more accurate. “For the first time, we could have a way of saying objectively, how delusional is this? How loose are these associations? How incoherent is it?” says Insel. Psychiatrists view speech as a solid marker to evaluate someone’s mental state since it can indicate disordered thinking, a hallmark of schizophrenia. People with disordered thought tend to move erratically from idea to idea — a chaotic path that AI could help to identify. Clinicians also know that schizophrenic patients sound different. Their tone can be more monotonous or robotic, they can take longer pauses. And the volume of their voice can lack contrast compared to that of healthy people who speak louder or quieter depending on the context and content of the conversation. But these differences are hard to use for diagnosis, especially in the early stages. A team of researchers in the Netherlands wondered whether AI could help. They selected audio recordings of people who had previously been diagnosed with schizophrenia by psychiatrists, then used software to measure 88 features, such as loudness, length of pauses, vowel pronunciation and intonation. They then used these measures to train an AI program to distinguish people with and without schizophrenia. When the team tested their AI using audio files from new patients it had not heard before, it differentiated schizophrenic patients from healthy controls with 86.2 percent accuracy and could even distinguish between different subtypes of the illness. People with schizophrenia often have distinctive speech patterns. Dutch researchers found that the speech of people with the mental illness tended to be more fragmented, with bursts of syllables separated by longer pauses. Their speech is also somewhat more variable in loudness than that of healthy people. AI may one day offer a convenient way to assess these differences. Adapted from J.N. De Boer et al/Psychological Medicine 2021/Knowable Magazine The approach isn’t especially useful for the clearest cases, says study coauthor Alban Voppel, a researcher now at McGill University in Montreal who studies the use of AI in psychiatry. But AI could help to detect patients in early stages or at high risk of becoming schizophrenic, and to predict relapse. “There are hopeful signs that we can pick up on very subtle things and things that psychiatrists find much harder to detect or harder to predict,” he says. Sunny Tang, a psychiatrist and researcher at the Feinstein Institutes for Medical Research near New York City, took a different approach. Instead of studying the sounds of speech, her team looked at its content. They built a type of AI called a machine learning model to assign a mathematical address, similar to the address of a house, to each word in a transcript of a patient’s conversation. By looking at the constellation of addresses, Tang’s program can tell whether a sentence stays in the same general neighborhood or leaves town and returns again, indicating disorganized thinking. In this way, Tang says, she can measure “how the meaning flows in the course of what somebody says.” When Tang provided transcripts of speech by people with and without schizophrenia, she found that the AI could distinguish the two groups with 87 percent accuracy. Clinical raters, who assessed patients without the AI, were only 68 percent accurate. People with schizophrenia are known to wander from topic to topic as they speak. To quantify this, researchers used artificial intelligence to map the conversational “neighborhood” explored by those diagnosed with the condition and healthy people in response to a prompt. The two groups differ in measurable ways, they found. Adapted from S.X. Tang et al/Schizophrenia 2021/Knowable Magazine Beyond diagnosis, clinicians may eventually be able to use AI for routine monitoring of the severity and progression of a patient’s symptoms over time. Such tracking helps clinicians provide more precise and responsive care, by evaluating whether a treatment is producing the desired effects and detecting early signs of relapse. Today, such monitoring requires clinicians to meet with patients regularly, at a huge cost of time and money. AI could potentially provide the same care for a fraction of the cost, and enable it to be done remotely. “You can have someone speak into an app or a device for a couple of minutes and give a fairly competent, confident rating on this person’s psychosis,” says Tang, who is hoping to have a tool ready for clinical trials in 2030. Much work remains before AI diagnosis is ready for the clinic, however. For one thing, artificial intelligence is only as good as the data it’s trained on — and there are serious deficiencies in those data. Many studies use small groups of people who are not representative of the population as a whole, according to a survey of studies using AI to make psychiatric diagnoses. “We need bigger samples, more diverse samples,” says Jeffrey Girard, a psychologist at the University of Kansas who led the analysis, published in the 2026 Annual Review of Clinical Psychology. Another problem is that people can speak more slowly not because they are mentally ill, but because they are older or speaking a second language. People may also sound different in stressful contexts such as an emergency room, or under the influence of medications or physical illness, notes Sandra Just, a clinical psychologist at UiT the Arctic University of Norway. This creates a “research-to-practice gap” where voice markers can do well in studies but poorly in the real world, says John Torous, a psychiatrist and informaticist at Harvard’s Beth Israel Deaconess Medical Center. Ideally, doctors would know what a patient usually sounds like to see if their speech has changed. But collecting such baseline data would involve identifying people at high risk of schizophrenia, then recording their calls, asking them to engage with an app daily or keeping track of clinical interviews — even if they currently show no symptoms. Doing that raises sticky ethical issues around privacy and security. “This is a really complicated problem,” says Brita Elvevåg, a cognitive neuroscientist at UiT the Arctic University of Norway who has studied speech patterns for nearly 30 years. “How can we preserve personal privacy now that we’re starting to use technology? It worries me that people don’t necessarily take that into consideration.” Because of all these limitations, researchers don’t see AI becoming a schizophrenia diagnostician anytime soon. “It’s a tool, it’s not a panacea,” says Vijay Mittal, a researcher and clinical psychologist at Northwestern University in Illinois. “It’s early days and there’s a lot of excitement [but] there needs to be a lot of caution.” This article originally appeared in Knowable Magazine, an independent journalistic endeavor from Annual Reviews. Sign up for the newsletter. Subscribe to Support Independent Journalism Great science journalism requires human expertise, time, effort and creativity. And it costs money. That’s why I and the journalists here at Scientific American hope you’ll join our community. When you subscribe, you are supporting staff and freelance journalists who are passionate about telling science stories that are true, important and compelling. Our editors and reporters are often experts in their fields, which means they understand the nuances of big discoveries and can untangle the breakthroughs from the hype. With a subscription, you are also supporting rigorous fact-checking to ensure the words we publish are precise and accurate. And you’re supporting original illustrations, graphics and photos that bring you closer to an advanced laboratory, an ice sheet in Antarctica or a space mission in orbit. You’re helping us craft other types of high-quality journalism as well: Our newsletters are carefully written, edited and curated by staffers you have or will come to know and love. Our Science Quickly podcast is based on original reporting, collaboration with editors and scientists and exacting production. Subscriptions keep this engine running so we can continue to deliver thoughtful, rigorous and independent science journalism to you. In an era of viral misinformation, this work is crucial. If you value what we do, I hope you’ll consider joining us as a subscriber Thank you, Jeanna Bryner, Editor in Chief, Scientific American Subscribe