AI and the Future of Psychiatry: Guardrails & Clinical Support

Artificial intelligence may offer the most immediate value in psychiatry by replacing burdensome tasks rather than replacing clinicians, shared Christoph U. Correll, MD, in a conversation with John J. Miller, MD. Avatar-based interviews completed at home before a visit, combined with self-report scoring, could standardize the history and free clinicians to focus on the chief complaint and the relevant history. Clinical decision support systems linked to the electronic health record could integrate intake data and ambient speech to update diagnostic and treatment predictions, but only if guardrails prevent hallucination, such as requiring the system to generate questions rather than recommendations when information is insufficient.¹ Correll reported that work with the Mayo Clinic involving 40,000 patients suggested better outcomes among clinicians whose prior choices matched the system's suggestions, although clinician uptake and the value of off-label options remained open questions. Precision psychiatry, in which phenotype-based models predict outcomes such as hospitalization and suicidality after different treatment choices, was framed as the future, contingent on standardized measurement feeding these systems.²

John J. Miller, MD: With all the buzz around AI, I have been captivated, and I still have a lot to learn. I feel like I am back in kindergarten with AI. I know it is something you are interested in and one of your passions. Can you help us understand how you see AI fitting into the future of the delivery of psychiatric treatment to help our patients?

Christoph U. Correll, MD: Thanks for the opportunity. I see AI as a tool that we obviously need to harness but also be careful with and have good guardrails around. It has various options to improve care.

One is task replacing, not people replacing, of tasks that are onerous and cause people to be worn out and burned out. Examples are repeating all the same questions and doing the interviews with patients about information they can easily deliver to an avatar. There could be avatar-based discussions before the patient ever comes to the hospital, while they are at home in their comfort. They are told, "Before you come, you have to do this, and we want to understand the prior history: how many times have you been hospitalized, what medication have you been given," and so forth. The patient would then see what the program has put out as the output, and this can be adjusted to an intake form that the hospital or the clinician may want. The patient can make some final changes, and it then goes to the clinician.

There could also be self-reports that generate certain T scores, so the clinician can say, "Oh, this symptom is more akin to a patient having depression or not, maybe PTSD," whatever the patient reported. That saves a lot of time and already goes into depth. When the patient is then seen by the clinician, the clinician can focus on the positives, meaning the things that the patient adores. It is not that you are not doing an interview anymore, but if the patient has never had substance use or a trauma, you may not want to go there. It allows you to focus on the real chief complaint and the historical information that is relevant.

These systems can also say whether the patient was consistent or inconsistent, whether they just answered a certain way, or whether the symptoms they reported do not hang together. That could flag that you may need to actually do the interview because the patient might have been too thought disordered. That is one way to prepare the clinician and also unload some of the tasks that are not why we studied medicine or went into nurse practitioner school.

The idea then is that a clinician decision support system (CDSS) linked with the electronic health record could gather all of the information that we cannot keep online and that is available, and then take the information from the patient's intake or from ambient speech and update its prediction of what the patient has in terms of diagnosis and treatment plan. At the end of the visit, the clinician could see what the CDSS would say: "Well, this patient has depression; augmentation might be the best here. Here are 2 or 3 options." It is important, though, that these systems do not hallucinate. They should not just please the patient or the doctor or give an answer because they want to give an answer.

Miller: How do you control for that?

Correll: We are developing one of those CDSSs right now. We are now at least cleared to sell it, but we are looking into developing it with a claim that we are better than usual care. For that, you need to show that there is clearly no hallucination.

What we developed there is guardrails so that when there is not enough information, the system is disallowed from making a recommendation and instead generates questions. These questions are needed so that I can make a recommendation. The guidelines recommend certain treatments, but I need to know whether the patient has hepatic problems or has been suicidal or manic before. If that is not spoken about or in the history, the system fills in the gap.

I think these systems can help clinicians have the whole 360 view of the past, but also what is needed in the present, and draw on evidence-based treatment. We need evidence-based and also measurement-based treatment, which AI can really deliver. In the future, there should also be modeling of what patients did after they got certain treatments. I think that is much more black box. But recommending and opening the view of the maybe expert clinician by telling you these are the options is easier than making good treatment decisions.

Miller: So you are starting on a solid foot with good evidence-based information. Or even more than solid. You are starting on a very comprehensive and evidence-based foot.

Correll: Now, we will look at whether that holds, because we are working with the Mayo Clinic on 40,000 patients. It showed that clinicians who did in the past what our decision system now suggests had better outcomes for their patients. But the question is, will clinicians use it? We need to educate them and show that they are benefiting from this and can then deselect or select treatments.

Miller: That is where you make hay with a solid basis. Is there an advantage to off-label treatments? The CDSS, the clinical decision support system, would not necessarily suggest that, because it is not evidence-based.

Correll: We will look into whether the outcomes are actually better by dialing in something that is not evidence-based or not approved but has at least clinical data behind it. AI will make all of this available to give you at least suggestions. The AI should not make the treatment decision ultimately, but it gives you at least a 360 view of the patient's data and also the available data in the field, so that you can make the most informed decisions.

Miller: Would it also have a database of hundreds of thousands, maybe millions, of patients with a specific diagnosis, so that it is able to compare the initial data they enter with what a large population of people with diagnosis X likely have? It would have this large data set of lots of patients that it is able to refine and see similarities and phenotypes in.

Correll: Yes, that is another possibility. We are getting more precision psychiatry, with certain phenotypes, but they need to be measured; otherwise the AI has nothing to deal with. We can now use natural language processing, which we are also doing, on the notes that the clinician writes. Obviously, we are dependent on what is in the patients' notes, whether certain symptoms are written down or, if they are not written down, were really absent. But yes, you could standardize what clinicians need to filter out, or use the avatar interview before each visit. That is possible in 5 to 10 minutes, and they then get these symptoms standardized, which the system could use to make predictions: if an SSRI is started in this patient rather than an SNRI, or augmentation with an atypical antipsychotic, what are the predictive values of having less hospitalization or less suicidality? That is the future, but we need the measurement of these systems to feed into the system.

Dr Miller is Medical Director, Brain Health, Exeter, New Hampshire; Editor in Chief, Psychiatric Times; Volunteer Consulting Psychiatrist, Seacoast Mental Health Center, Exeter; Consulting Psychiatrist, Insight Meditation Society, Barre, Massachusetts.

Dr Correll is professor at the Institute of Behavioral Science, Feinstein Institutes for Medical Research; medical director of the Recognition and Prevention Program in the Department of Psychiatry at Zucker Hillside Hospital; and professor of Psychiatry and Molecular Medicine at the Donald and Barbara Zucker School of Medicine at Hofstra/Northwell.

References

1. Kleine AK, Kokje E, Hummelsberger P, et al. AI-enabled clinical decision support tools for mental healthcare: a product review. Artif Intell Med. 2025;160:103052.

2. Salazar De Pablo G, Studerus E, Vaquerizo-Serrano J, et al. Implementing precision psychiatry: a systematic review of individualized prediction models for clinical practice. Schizophr Bull. 2021;47(2):284-297.