Using AI to Write Psychiatric Progress Notes

COMMENTARY

It is 6 PM on any given winter evening, and in Baltimore the sun is set. My office, bright during the day, is now dark. I am tired and eager to go home. What stands in the way is the final task of my workday, and the part I dread most: writing progress notes.

My notes are computerized, with a template for each patient that carries through from session to session. If I stay focused, they do not take very long—perhaps 5 to 10 minutes per note unless there is something complicated or worrisome going on—and yet I still dread writing them. I check my voicemail and email, respond to text messages, scroll through social media, and play Wordle, either to decompress or to procrastinate, and that too makes my day longer than it has to be. I have started to wonder if AI could offer a streamlined way of doing things.

I am in solo private practice, and I see patients for medication management in conjunction with psychotherapy. Most patients are seen for an hour, so I have a low-volume practice. I do not use an EMR, not because I am averse to the technology, but because I have not found a reason to do so. I print off notes and file them in paper files—quaint, but effective. In many ways, my desire to have AI write my notes is a solution in search of a problem.

I have a very specific idea in mind: I would like to be able to paste in my last progress note, then verbally tell the program what happened in the session and have it generate a new and timely note. If the time for the new vs the old is equal, it is still a win because the process is novel and perhaps I will not dread doing it.

I began by using ChatGPT. It is not HIPAA-compliant, so I created a fictional patient. It was great—fast, accurate, and able to do exactly what I asked. If there was a mistake, I corrected it and had a new version in seconds. I liked that I could ramble about the session, and it would organize my thoughts into a coherent, articulate note—more polished than my own. But it could not save the notes for future use, nor could I legally use it for real patients.

Mark S. Komrad, MD, a psychiatrist in Towson, Maryland, wrote about his experiences with MDHub, “It has many different templates to organize what it hears, or you can customize your own template. It also generates a summary for the patient.” Komrad tried it with 30 patients. “I found it pretty remarkable, but not a time saver. Though it made relatively few errors, it left out some things that I thought were important, and by the time I was done proofing it, editing it and adding to it, cool as it was, it did not save me any time compared to just writing my own note.”

Paul Nestadt, MD, a Baltimore psychiatrist, also shared his experience. “I've used the ambient listening scribe that my university provides, and although it is a neat party trick, I haven't found it to save any time.”

I was not looking for a scribe—AI that listens to the full session and distills it into a progress note. I cannot imagine what it might do with the nuances of a psychotherapy session, or how my patients would feel about having their sessions recorded. Wouldn’t this influence what people might bring to treatment? Or how I would respond? Am I ready to be constantly judged by a computer?

After trying several programs, I settled on a new one offered by OpenEvidence, the AI app I use for clinical consultation. “OpenEvidence Visits” is HIPAA-compliant and it is free. It creates a file for each patient and saves information for future visits. It lets me input a prior note, update it verbally, and generate a new one. The platform does a good job of creating organized, professional notes—it even articulates therapeutic interventions better than I do—and it offers to add references (especially gratifying when they validate my approach).

I started using OE Visits just days after its release, and since there are no formal instructions, I am not sure I am using it optimally. After I finish dictating, it takes a few minutes to generate the note. If I dictate a change, it takes longer. I can type edits directly, which is faster, but at that point there is no real improvement over my own word-processed notes. Sometimes, if I pull up an old note and ask it to make changes, it just adds new material to the existing text instead of replacing it. Copying forward a note also takes time, and it clears fields I would prefer to keep consistent from session to session.

For every note, I tell the program to use paragraph form instead of bullet points, since I try to fit notes on one page—but it does not seem to “learn” this. If I forget to specify, it reverts to bullet points for medications, conditions, and the mental status exam. Occasionally, fonts appear in odd sizes—oversized headings, for example—and sometimes it adds phrasing I never use (“affect was congruent,” being a favorite). Weeks into this project, we are still getting to know each other. Even though it is HIPAA-compliant, I still avoid using patients’ last names or birthdates—there is an odd sense of not knowing exactly where my data is going.

It would be wonderful if I could pull up the notes on my phone and print them later, but I have not figured that out yet. It would also help if notes could print directly from the platform; so far, every system I have tried requires copy-pasting into another program first.

I have enjoyed the novelty of using AI to write progress notes. I stopped procrastinating and no longer dreaded the task—and without that resistance, I sometimes get out of the office a bit sooner. But I have also realized that my fun new project often takes longer than my old method, and it is oddly refreshing now to open a Word document and write a note myself—one not as polished as the computer’s version, but one that still gets the job done.

Dr Miller is a clinical psychiatrist and writer in Baltimore. She is on the faculty at the Johns Hopkins School of Medicine.