AI & Clinical Informatics in Psychiatry

1. Why a practicing psychiatrist builds software

A research finding that never reaches the exam room is like a recall notice that never gets mailed. The defect is documented, the fix is known, and none of it protects the person actually driving the car.

That's the gap I kept running into. My first decade of research asked whether ADHD was being treated the way the evidence said it should be, then traced its connection to adolescent cannabis use. The prescribing findings and the cannabis findings were solid. They were published in JAMA Network Open, Pediatrics, and JAMA Psychiatry. But sitting with a patient, I couldn't reach for any of it. The knowledge lived in PDFs and my own memory, not in the workflow of the visit.

Clinical informatics is the discipline that closes that gap: it moves evidence from the literature to the point of care. I didn't come to it as an engineer who wandered into medicine. I came to it as a clinician who got tired of watching good evidence stall one step short of the patient.

3. This isn't new for me — it started in 2016

The informatics thread goes back further than the current AI wave. In 2016 I analyzed Veterans Affairs electronic-health-record data to show improved clozapine safety monitoring in patients with schizophrenia. Clozapine is the most effective antipsychotic we have and one of the most dangerous to monitor, so getting its safety checks right in near real time was an early, concrete case of using EHR data to make care safer.

Everything since has been the same move at larger scale: take a real clinical dataset, find where care and evidence diverge, and build something that narrows the gap. First it was VA records and clozapine. Then national claims and prescribing. Now it's natural-language processing over the clinical note and the point-of-care decision itself.

4. CEBA-ADHD: a knowledge graph for the diagnosis itself

The deployed system is the foundation for three federally submitted grants. The first is CEBA-ADHD — Clinical Evidence-Based Assistant for ADHD — submitted as an NIMH R01 with me as principal investigator and Xuhai "Orson" Xu, PhD, of Columbia's Data Science Institute as co-investigator handling the underlying architecture.

A large language model on its own is a confident generalist. It will answer an ADHD diagnostic question fluently whether or not it's right. CEBA-ADHD grounds that fluency in a clinical knowledge graph, so the model's reasoning is anchored to structured clinical evidence rather than to whatever it absorbed from the open internet. Think of it as the difference between a smart resident working from memory and the same resident working with the guidelines open on the desk.

In a 315-case pilot, adding this structure improved diagnostic accuracy across five frontier language models, with up to a 12-point gain on ADHD cases. That's the part that matters: the improvement held across models, which means the value is in the clinical scaffolding, not in one vendor's model. The next step is a prospective evaluation embedded in routine care, with Dr. Xu's group on the knowledge-graph architecture while I oversee clinical validity and deployment.

5. PAWS: a digital therapeutic built safety-first

The second grant is PAWS (Pawsitive Companion), submitted as a NIDA UG3/UH3 with me as a multiple principal investigator alongside Dr. Frances Levin. It's a multi-modal AI digital therapeutic for youth cannabis use disorder that pairs a large language model with wearable data.

Building an AI that talks to teenagers about substance use raises an obvious question before any other: what happens when a kid in crisis says something dangerous? We answered that first. The safety architecture was tested against 50 expert-designed crisis scenarios and handled all 50. Safety wasn't a feature we added at the end; it was the spec we designed to from the start, the way you engineer a bridge to its worst-case load rather than its average day.

The project has drawn letters of support from eight academic medical centers and an FDA pre-submission. It grows directly out of my cannabis research: my 2023 JAMA Network Open and 2026 Pediatrics studies showed that even low-frequency adolescent cannabis use carries measurable risk, which is exactly the under-served population a scalable digital therapeutic can reach. A related patent, CAPTI (Cannabis Assessment Psychoeducation Tool and Intervention), is in progress.

6. Measuring it like medicine, across sites

Deploying a tool in one practice proves it can work. Proving it works means measuring it somewhere else, too. I'm extending the decision-support program as a minimal-risk quality-improvement initiative beyond Integrative Psych to ColumbiaDoctors, the New York State Psychiatric Institute, and a Union Square practice.

The endpoint isn't "do clinicians like it." It's concordance with the actual standards of care: whether point-of-care suggestions move practice closer to AAP, AACAP, and NICE guidelines. That's a stricter bar than user satisfaction, and it's the right one. A tool that clinicians enjoy but that nudges care away from guidelines is worse than no tool. I'm also working with Dr. Mark Olfson toward a population-scale extension using national claims data, currently a submitted R01.

7. The bigger problem: records that don't travel

Every piece of this work keeps running into the same wall. Mental health records aren't standardized, and they don't follow the patient. When someone moves between a private practice, a hospital system, and a clinic, their psychiatric history mostly doesn't come with them in any usable form. It gets reconstructed from scratch, from memory and free text, at every handoff.

Compare it to banking. Your financial history follows you between institutions in a common, machine-readable format; that's why you can switch banks without losing your credit history. Psychiatry has no equivalent, so we rebuild the patient's story over and over and lose most of it to the free-text note in between.

The next step in this program is to build an open-access, standardized mental health record format — a structured file that captures diagnosis, treatment history, and outcomes in a common schema any clinician or EHR system can read. The point is twofold: make prospective mental-health data usable for research from the outset instead of excavating it retroactively, and let a patient carry a real clinical history between institutions instead of starting over. It's the same line of work that began with the VA project in 2016, scaled up from one better tool inside one practice to a shared standard that makes care and research easier across all of them.

8. Frequently asked questions

What is Dr. Sultan's work in AI and clinical informatics?

He designs and deploys AI tools that run inside real psychiatric care. His deployed clinical decision-support system was evaluated over six months across 16 clinicians, significantly improving decision support (p=.004) and reducing chart-review time (p<.001) with large effect sizes (0.73–0.85). Its extraction reached 93% precision and 98% recall across 395 records, and its intake-summary module reached 100% clinician adoption.

What is CEBA-ADHD?

A decision-support tool built on a clinical knowledge graph, submitted as an NIMH R01 (Dr. Sultan as PI; Xuhai "Orson" Xu, PhD, co-investigator). In a 315-case pilot it improved diagnostic accuracy across five frontier language models, with up to a 12-point gain on ADHD cases.

What is PAWS?

A multi-modal AI digital therapeutic for youth cannabis use disorder that pairs a language model with wearable data, submitted as a NIDA UG3/UH3 (Dr. Sultan as multiple PI with Dr. Frances Levin). Its safety architecture handled all 50 expert-designed crisis scenarios and it has drawn letters of support from eight academic medical centers and an FDA pre-submission.

Does the AI replace the psychiatrist?

No. The deployed system is a minimal-risk, read-only, non-EHR-altering adjunct. It surfaces knowledge and drafts summaries; the clinician makes every decision. The point is to give the physician more time and better information, not to automate judgment.

Why does a practicing psychiatrist build software?

Because a finding that never reaches the exam room changes nothing. After years documenting how ADHD and cannabis are actually treated, Dr. Sultan found the evidence wasn't reaching clinicians during a visit. Building the tools that carry it there is the next step.

What is the portable mental health record?

An open-access, structured, portable file format capturing diagnosis, treatment history, and outcomes in a common schema any clinician or EHR can read, so mental-health data is usable for research from the outset and follows the patient between institutions.

About Dr. Ryan Sultan

Dr. Ryan Sultan is Assistant Professor of Clinical Psychiatry at Columbia University Irving Medical Center, Director of the Sultan Lab for Mental Health Informatics, and a double board-certified Adult and Child/Adolescent psychiatrist. He is also Medical Director of Integrative Psych in Manhattan, where his AI clinical decision-support system runs in daily care. His research has moved from documenting how ADHD is treated, to understanding its connection to adolescent cannabis use, to building the clinical informatics tools that put that evidence to work at the point of care.

He is principal investigator of an NIDA K12 award and principal or multiple-PI on federally submitted grants including CEBA-ADHD (NIMH R01) and PAWS (NIDA UG3/UH3). His 2019 JAMA Network Open study on antipsychotic prescribing in youth with ADHD has been cited 440+ times.

Read full bio → | Publications → | Sultan Lab →