The Case for Automated Clinical Documentation: Can Somebody Please Make This?

Imagine for a moment that you are a busy clinician who runs her own practice of psychiatry. You spend your days seeing patients — perhaps one after the other — and attending to their needs.

This could mean evaluating new patients, following up with existing patients, prescribing new medications, refilling old ones, providing psychotherapy, ordering tests, making referrals, filling out forms and generating documentation for third parties (e.g. disability insurers). Optimally, you obtain collateral information when necessary by speaking to the patient’s other treating clinicians (e.g. a psychiatrist speaking to a psychotherapist who is treating the same patient) and family members.

Each patient encounter, therefore, could potentially mean a lot more time than the time that you spend face-to-face with your patient in the office. If you allow for it — and, in this day and age, it’s become more and more difficult to not do so — you might also have to field emails from your patients, their family members and other treating clinicians. These can come at any time of the day and night, of course; you have to prioritize to which ones to attend most urgently. All the while, there are voicemail messages to which you must listen (or read, in the case of transcribed messages), phone calls to return from prospective patients, current patients, and psychotherapists who call with concerns about mutual patients. And, finally, in order to generate income from all these activities, you have to either bill for your service by invoicing patients directly or submitting claims to their health insurers.

If you are a private practitioner, you may be engaged in all these activities by yourself. If you practice in a group practice setting or a clinic, you likely have some of these functions handled by administrative assistants, receptionists or billers. In settings such as Kaiser Permanente, your work may be even more circumscribed, in that you deal with patient encounters, treatment planning, treatment initiation, and care coordination — while all other functions of patient care are rendered by the support staff.

Oh, and there’s one thing that we haven’t even broached yet: clinical documentation.

If you are a mental health clinician, you likely chose your profession because you enjoy working with patients, sitting with their suffering and doing your best to relieve it. In order to do so, you need to spend adequate time with your patients to ask them about their symptoms, perform an examination, address their concerns and questions, and obtain relevant information from collateral sources. Moreover, in situations that are not clear cut, you might need to consult with colleagues and familiarize yourself with the latest information in the professional literature to answer a particular question. Thus, the more time you have for these activities, the better your chances of rendering optimal care to your patients.

So, how can digital health technologies help facilitate this scenario, where clinicians have more face-to-face time with their patients, more time for care coordination, and more time to survey the research literature to optimize care?

* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *

As someone who has been enthusiastic about digital health technology in the mental health space for some time now, I frequently wonder about how novel technologies could actually help the practice of psychiatry and clinical psychology. Certainly, the most widely adopted innovations have been electronic health records (EHR) software, electronic medication prescription software and practice management software. EHR software, in essence, helps generate health records of patient encounters that are professional requirements of practice. They replace the paper chart. Practice management software, on the other hand, deals with all other issues pertaining to running a practice of psychiatry or psychology: 1) patient scheduling 2) billing 3) claims generation and 4) analytics.

It’s possible to run a practice of psychiatry or psychology with very little in the way of overhead expenses. Scheduling, invoicing/billing and claims generation can now be automated through a complete EHR/practice management solution. Receptionists are not necessary anymore to schedule patients and manage calls, as patients can now directly message their care providers through email or through a patient portal function on the clinician’s EHR software. For people who still feel more comfortable with voicemail, automated transcription services can, in essence, turn their voice messages into text emails for clinicians to review. By and large, patients can now routinely reach their clinicians through secured electronic mail and even in large HMOs, such as Kaiser Permanente, this seems to be the preferred mode of communication.

While these innovations have helped reduce clunky and outdated methods of practicing the business of healthcare, they have come with their own pain points, unfortunately. So, now, clinicians no longer need to have a file cabinet (or many) with aging patient charts. But, they may not have the same satisfaction of being able to quickly review a number of notes by flipping through a patient’s chart. And while electronic prescribing solutions have made prescribing medications a cinch, it’s not clear just how much time they actually save relative to handwriting a prescription. Certainly, they don’t generate paper records that have to be re-entered into the chart (i.e. most e-prescribing solutions integrate with the clinician’s EHR, so that prescribed medications are automatically integrated into the patient’s medical records), thereby potentially saving some time and paper waste.

Many EHR solutions now offer patient portals, through which patients can schedule follow-up appointments by themselves, review some of their medical records, and message the provider directly and securely. This, of course, requires patients to generate yet another online account and come up with yet another password to remember.

But, in the course of my day-to-day work, the task that I find least appealing is the generation of clinical notes. Certainly, it’s an important activity and one that is required professionally. For psychiatrists — particularly those who practice medication management — this involves following a template that corresponds to a billing code such as CPT Code 99214, a Level 4 office visit with an established patient. Current documentation requirements for this code include having:

1. Patient identifier information (e.g. name, date of birth, medical records number)
2. Visit identifier information (e.g. date of service, CPT code)
3. A chief complaint (e.g. follow-up for Major Depressive Disorder)
4. A history of the present illness detailing current symptoms, duration, etc.
5. A review of systems
6. A medications list
7. Pertinent family and social history relating to the history of the present illness
8. An examination, including mental status examination
9. Diagnoses
10. An assessment and plan to address the relevant clinical issues (e.g. ongoing depression, ongoing anxiety, ongoing substance abuse, chronic pain, social isolation, etc.)

Recognizing that there is room for improvement for physicians in the process of generating a clinical note, new strategies have been employed by health systems to mitigate the burden of documentation. In some emergency departments, for example, human scribes produce the clinical note, following physicians along from patient to patient, room to room. This relieves the physician of this burden, allowing her to focus singularly on patient care.

Augmedix, a digital health company co-founded by Ian Shakil and Pelu Tran in 2012 introduced a new twist to this formula. Employing remote human scribes, the company harnessed Google Glass technology to, essentially, live stream patient-doctor encounters for the purpose of generating clinical documentation. No longer did you need to have human scribes on location to document patient encounters. The cost of live transcription was optimized by migrating this function to a low-cost labor market.

However, it struck me that certain fields of medicine are probably not well-suited to this method of clinical documentation. Psychiatric patients — who talk about very private matters with their clinicians — would probably squirm at the idea of having intermediary scribes eavesdropping into their sessions — even if they were half a world away. And women presenting for gynecological examinations would likely similarly balk at the idea of having their examination live streamed to a complete stranger.

Because the psychiatric examination — known as the mental status examination — does not require touching patients physically, it would be relatively easy for human scribes to document — video live-streamed or otherwise. But, since privacy is such an important consideration among psychiatric patients, it would not be ideal to employ human intermediaries for the purposes of clinical documentation generation.

So, where does this leave us, then?

When I first became interested in digital mental health technologies, I was interested in two specific solutions: 1) the development of artificial intelligence technologies that could essentially replace psychiatrists in the work they do (i.e. evaluating patients, rendering diagnoses, developing treatment plans, and administering treatments) and 2) automated clinical documentation technology that would not require a human intermediary.

At various networking meetings for people interested in health technology, I would ask software engineers about the feasibility of the latter. Most of the time, I was told that the idea, while interesting, was not yet feasible, primarily on account of limitations of natural language understanding, an artificial intelligence problem. Automatic speech recognition, on the other hand, was not so much of a limitation.

In my mind, the schematic for this solution — automated clinical note generation — seemed simple enough. It would require audio recording of the clinical encounter between psychiatrist and patient (easy enough), transcription of the two-person audio conversation to text (probably doable) and then interpretation of the dialogue into clinically relevant units of information (the tough part) that would comprise the clinical note.

Doing a little Google research and speaking with some engineers, I learned that it was possible to record the conversation between psychiatrist and patient into two different channels using a device such as a Tascam digital recorder. Separating the audio into two channels corresponding to two speakers would presumably result in more accurate transcripts. At the same time, I learned about parsing and entity extraction and learned about the current limitations of entity extraction as a way of understanding human language. (Entity extraction examines text and labels named entities into various categories (e.g. person, organization, time, place, etc.), resulting in a very rudimentary analysis of text).

At this point, I was beginning to understand the complexity of the problem: how could you train a computer to understand transcribed human conversation well enough to summarize it, in effect, into a clinically meaningful note? I wondered whether it were even necessary, then, for said computer to “understand” the transcribed conversation. What if it could be trained, instead, to recognize patterns within unedited transcripts that resulted in specific information in a progress note? In this type of solution, a computer would have to be trained using machine learning to examine thousands of free text transcripts of psychiatrist-patient visits and their corresponding progress notes to develop an algorithm to go from one to the other.

To give you a sense of what this technology might look like in action, consider the following exchange between patient and psychiatrist, an actual transcript of the first five minutes of a psychotherapy session I had with one of my patients, anonymized for his protection:

MD: so, how’ve you been since we last met?

PT: doing pretty good. I’ve been, umm, I actually have been feeling a little more in control this week

MD: um hum

PT: So, I, umm, for some reason I feel like at work I’ve made my choices given the circumstances

MD: um hum

PT: Like I can decide how I want to react to it

MD: um hum

PT: And I think that’s opposite of the past where I’d like fight to change, fight to change so much

MD: yeah

PT: so, I feel more in control that way

MD: um hum

PT: My house is kind of slowly, slowly plodding along

MD: um hum

PT: But, I feel like that’s…so some circumstances have been helping

MD: ah hah

PT: and I finished all the documents with my lawyer, so that’s

MD: For the complaint?

PT: Yeah

MD: uh huh

PT: So, I have just have yet to sign it, but everything is kind of, so I feel like things are, um, a little more in control

MD: um hum, well that’s good

PT: Yeah, and I guess that equates, at some level, to my being a little happier

MD: um hum

PT: so, you know, that’s kind of where I’ve been

MD: Mood is a little bit happier?

PT: Yeah

MD: Do you feel like you still have some level of anxiety and depression that you are living with?

PT: Yeah, um, it’s not so on the sleeves so much…I do get impatient, especially in the car, and at work, I keep having to coach myself, “Let it go. Let it go.” Because I keep thinking of ideas to make things better. [laughs].

MD: Um hmm.

PT: Then, I just keep thinking, “Ehh! Not my problem. Not my problem” So…[laughs]

MD: No one is going to appreciate it and….

PT: Yeah, why bother?

MD: Yeah. Yeah.

PT: I….My value system is detached from the way…. it’s such a mismatch of, like, manager to the staff.

MD: Umm hmm.

PT: And Kelly and I have talked about that…. like there is…. was no consideration whatsoever of matching a manager to the people that work there

MD: Umm hmm.

PT: There was no one who tried to say, “What, you know, what are people doing there?”

MD: Mmm. Hmm.

PT: “Who are the people? Maybe we should have them meet to see if it’s, you know, get feedback, like none of that. It was all about, like, “I like having lunch with her.” And, um, I heard Theresa telling someone, she said, “Yeah…. I mean…. you know it’s just great managing people who are so independent and work so well on their own.” And she was, you know, trying to pray…. praise us, but what she was really saying was, “I don’t have to do anything.”

MD: Mmm. Hmm. That’s what she’s grateful for.

PT: Yeah.

MD: Yeah. So, have you had any more of those, like, uh, road rage episodes?

PT: I don’t think I did this week.

MD: Well, that’s good.

PT: I’ve been impatient.

MD: Uh huh.

PT: But, I don’t think I’ve actually gotten out of the car. [Laughs] Or

MD: Mmm hmm.

PT: I might have tapped my horn, but I have not yelled at anyone, or…

MD: umm hmm.

PT: So

MD: So, you’re, you’ve noticed some impatience, but you’re, like, handling it okay. Like you’re not having outbursts.

PT: Right, right.

MD: And, with the house, is there some movement? Is like, Is there….

PT: Yeah, Gordon and I got another guy over to look at the back of the house

MD: Mmm hmm.

PT: And he had some more, um, he actually painted our building like ten years ago

MD: Mmm hmm.

PT: And, so he knew the building and was, like, eight years ago. So, he kind of was more clear about what needed to be done. And he kind of actually verified more of the middle of the road advice we got before….

MD: Which is you could wait?

PT: We could wait and

MD: But, it has to be done.

PT: And he basically said that the person you need to fix what you need to fix is a carpenter. We didn’t know that. Like we didn’t know if we needed a contractor, a siding company. So, he’s like no, he goes you just take care of this area you need a carpenter to do it.

MD: Mmm hmm.

PT: So, that was helpful.

MD: Yeah.

PT: Because, you know, people like us — I’m assuming you, too — don’t know

MD: Much about these kind of things, yeah….

PT: [laughs]

MD: So, they’d cut out, like, an exterior, uh, part of the….

PT: Yeah, the part where they replaced the window and now there’s a crack above the window

MD: Yeah

PT: Not a crack, but the siding doesn’t meet correctly

MD: Right, right, yeah. How about your eating, George? How has that been this week?

PT: It’s been okay. I haven’t been as disciplined as before.

MD: Mmm hmm.

PT: And, in fact, I would say that for the first time since I got real serious about my kidney stuff, um, I kind of binged a little bit this week.

MD: Hmm.

PT: And, um, I don’t know, I just got off track.

MD: Was that on, like your comfort foods, or, or

PT: Yeah, comfort foods and the amount of food.

MD: Umm hmm.

PT: And, I, I mean, it’s far less than I used to, but…

MD: Umm hmm.

PT: One day, I ate a bunch of chocolate. I just couldn’t stop myself

MD: Umm hmm.

PT: Which is high in potassium…

MD: Oh, right.

PT: And, then, two nights in a row, I had…. I had big dinners, in addition to having lunch.

MD: Mmm hmm.

PT: And, then, because I did that, I had a half pint of ice-cream afterwards, because, you know [laughs], so…anyway, I don’t feel that it is totally out of control, it was just a matter of few days.

MD: Umm hmm.

PT: And, then I keep thinking, “Oh, my poor kidneys, having to process all this stuff”

MD: Yeah…. that helps you kind of get back on track?

PT: Yeah

MD: Yeah….and, then, going back to Lexapro, you’ve kept the dose at 15?

PT: Mmm hmm.

MD: Are you having side effects?

PT: No. I’m mean, not anything new.

MD: So, no, the appetite, you feel, is not too dysregulated.

PT: No

MD: Okay

Now, consider how this dialogue would need to be interpreted in order to generate a meaningful and succinct clinical progress note such as this one:

HPI: Patient reported that he’s been “doing pretty good” since our last meeting. He stated that he feels a “little more in control”. Although he continues to have some level of anxiety and depression, it does not feel as intense at this time. He agreed that his mood is a bit happier. He denied having road rage episodes in the past week, but reports that he feels impatient still. He has not gotten out of his car to confront other drivers.

Regarding his eating habits, he stated that he binged for a few days in the past week, but that it was not as out of control as it had been in the past. He stated that on one occasion, he ate “a bunch of chocolate” and then a “half pint of ice cream afterwards.”

He continues on Lexapro 15 mg/day and denied having any new side effects. He does not believe that Lexapro is dysregulating his appetite.

Thus, the transcript-to-note transformation algorithm would have to discard large chunks of conversation that have nothing to do with psychiatric symptoms per se (e.g. in this case, the patient’s description of the type of repair work that needs to be done on his home, his dissatisfaction with his work, etc.). Moreover, it would have to detect when I asked the patient questions that pertained to important clinical information (e.g. about his mood, about his impulse control with respect to his regulation for anger, about his impulse control with respect to his eating behavior, about his medication dose and side-effects). And it would have to “understand”, in a way, when the patient was answering these questions and what his responses are.

Thus, the AI at the heart of this technology would need to:

1. Understand who was speaking and when

2. Understand when the physician asks important clinical questions (e.g. about depressive symptoms such as mood, energy level, sleep, motivation, etc.)

3. Understand when the patient responds to these questions and what he says

4. Understand when the patient brings up symptoms on his own and what these are

5. Group reporting of similar symptoms together, to make the final note more comprehensible

6. Recognize names and dose ranges for all psychiatric medications

7. Recognize descriptors of side-effects to medications

The psychiatric mental status examination, too, could be extrapolated from this transcribed dialogue. The typical “fields” or “variables” in the mental status examination include:

1. General appearance
2. Demeanor/attitude
3. Psychomotor functioning
4. Speech
5. Mood
6. Affect
7. Thought processing
8. Thought content
9. Insight
10. Judgment

While it might be a bit ambitious to have AI technologies weigh in on a patient’s demeanor and attitude or psychomotor functioning simply by examining the para-verbal features of the patient’s speech and by analyzing the text of the dialogue between patient and clinician, it doesn’t seem too far-fetched, however, to imagine that this technology could populate certain of these variables or fields: speech (volume, rate, rhythm, emotional prosody, latency of response, etc.), subjective mood, and thought content (e.g. presence of suicidal ideations, homicidal ideations, auditory hallucination, visual hallucination, paranoid ideations, ideas of reference, delusions, nightmares, flashback experiences, obsessions, obsessive ruminations, excessive worries, etc.).

The final document produced by this algorithm could, therefore, potentially have detailed information about:

1. The Chief Complaint

2. The History of the Present Illness

3. Medications (discussed in the session and extrapolated from electronic medical records)

4. Mental Status Examination

5. Diagnoses (extrapolated from electronic medical records)

6. Treatment plan

Then, after the patient encounter, the physician could simply fill in the gaps or modify, as needed, the records created by the note-generating algorithm to finalize an accurate encounter/progress note.

So, what effect would this have on the practice of psychiatry and clinical psychology?

My hunch is that it would significantly free up clinicians’ time so that they could focus more on clinical care and less on documentation. This might result in improved care coordination with other providers (i.e. calls to therapists, primary care physicians) or increased time spent performing research into treatment planning (i.e. reading recent research pertaining to a clinical question or situation).

At this time, the main impediments to creating such an application appear to be:

1. Generating a large enough volume of accurate transcripts of patient-physician encounters

2. Training a computer to analyze thousands of such transcripts to understand how a final clinical document is generated from its corresponding transcript

I’m confident, however, that it’s just a matter of time — and likely not too far off into the future — before this technology is ready to get up and running.

So, could somebody please start to make this? Please?