Automation Complacency, The Stepladder of AI in EHR’s, “Writing a note is how I think”. WHAT NOW?

A navel-gazing reflection on GPT, human cognitive effort, and the stepladder to the future. Where do YOU stand?

The image above generated by DALL-E embedded in the new BING, with the prompt “Doctors using a computer to treat patients, optimistic futuristic impressionistic image”. Wow. Not sure what the VR doctor coming out of the screen is doing.

Thanks to Dr. Brian Montague for prompting this post with his quote during a recent Large PIG meeting:

I find that I do a lot of my thinking when I write my progress note. If/when ChatGPT starts to write my note, when will I do that thinking?  — Brian Montague MD

That stopped me in my tracks.

We are so hell-bent on simplifying our work, reducing our EHR burden, we sometimes forget that this work is MORE than just pointing, clicking and typing.

It is also about THINKING. It is about assembling the data, carefully coaxing information and patterns out of our patients through skillful interview, parsimonious lab testing, and careful physical examination. It is how we, as physicians and APP’s, use our bodies and minds to craft an image of the syndrome, the disease: our hidden opponent.

Just like inserting a PC into the exam room changed dynamics, inserting GPT assistants into the EHR causes us to rethink … everything.

Pause to reflect

First, I think we should recall the technology adoption curve.

I fully acknowledge that I am currently dancing on the VERY PEAK of the peak of over-inflated expectations. Yes. That’s me right at the top.

Of concern, viewing the announcements this week from Google, Microsoft, and many others gives me chills (sometimes good, sometimes not) of what is coming: automated, deep-fake videos? Deep-fake images? Patients able to use GPT to write “more convincing” requests for … benzodiazepines? opiates? other controlled meds?

AND YET, think of the great things coming: GPT writing a first draft of the unending Patient Advice Requests coming to doctors. GPT writing a discharge summary based on events in a hospital stay. GPT gathering data relating to a particular disease process out of the terabytes of available data.

And where do we think physician/APP thinking might be impacted by excessive automation?

Automation Complacency

I refer you back to my book review of the book “The Glass Cage” by Nicholas Carr. As I said before, although this was written to critique the aircraft industry, I took it very personally as an attack on my whole career. I encourage you to read it.

In particular, I found the term “automation complacency” a fascinating and terrifying concept: that a user, who benefits from automation, will start to attribute MORE SKILL to the automation tool than it actually possesses, a COMPLACENCY that “don’t worry, I’m sure the automation will catch me if I make a mistake.”

We have already seen this among our clinicians, one of whom complained: “Why didn’t you warn me about the interaction between birth control pills and muscle relaxants? I expected the system to warn me of all relevant interactions. My patient had an adverse reaction because you did not warn me.”

Now, we have this problem. We have for years been turning off and reducing the number of interaction alerts we show to prescribers precisely because of alert fatigue. And now, we have complaints that “I want what I want when I want it. And you don’t have it right.” Seems like an impossible task. It IS an impossible task.

Thank you to all my fellow informaticists out there trying to make it right.

GPT and automation: helping or making worse?

Inserting a Large Language Model like GPT, that understands NOTHING, but just seems really fluent and sounding like an expert, could be helpful, but could also lull us into worse “automation complacency.” Even though we are supposed to (for now) read everything the GPT engine drafts, and we take full ownership of the output, how long will that last? Even today, I admit, as do most docs, that I use Dragon speech recognition and don’t read the output as carefully as I might.

Debating the steps in clinician thinking

So, here is where Dr. Montague and I had a discussion. We both believe it is true that a thoughtful, effective physician/APP will, after interviewing the patient and examining them, sit with the (formerly paper) chart, inhale all the relevant data, assemble it in their head. In the old days, we would suffer paper cuts and inky fingertips in this process of flipping pages. Now we just get carpal tunnel and dry eyes from the clicking, scrolling, scanning and typing.

Then when we’ve hunted and gathered the data, we slowly, carefully write an H/P or SOAP note (ok, an APSO-formatted SOAP note). It will include the Subjective (including a timeline of events), Objective (including relevant exam, lab findings), Assessment (assembly of symptoms into syndromes or diseases) and Plan (next steps to take).

During this laborious note-writing, we often come up with new ideas, new linkages, new insights. It is THIS PIECE we worry most about. If GPT can automate many of these pieces, WHERE WILL THE THINKING GO!?! I do not trust that GPT is truly thinking. I worry that the physician will instead STOP THINKING.

Then THERE IS NO THINKING.

Is this a race-to-the-bottom, or a competition to see who can speed us up so much that we are no longer healers, just fast documenters, since we are so burned out?

Who will we be?

Radio vs TV vs Internet

My optimistic thought is this. Instead of GPT coming to take our jobs, I’m hopeful GPT becomes a useful assistant, sorting through the chaff, sorting and highlighting the useful information in a data-rich, information-poor chart.

Just like the radio industry feared that TV would put them out of business (they didn’t), and TV feared that the Internet would put them out of business (they didn’t), the same, I think, goes for physicians, established healthcare teams, and GPT-automation tools.

Lines will be drawn (with luck, WE will draw them), and our jobs will change substantially. Just like emergent (unpredictable) properties like “GPT hallucinations” have arisen, we must re-invent our work as unexpected curves arise while deploying our new assistants.

Another Bing-Dall-E image of physicians at a computer. In the future, a doctor will apparently have more legs than before.

A possible step-ladder

I think physician thinking really occurs at the assembly of the Assessment and Plan. And that the early days of GPT assistance will begin in the Subjective and Objective sections of the note. GPT could for example:

SIMPLE
  • Subjective: Assemble a patient’s full chart on-demand for a new physician/APP meeting a patient in clinic, or on admission to hospital, focusing on previous events in can find in the local EHR or across a Health information exchange network, into an easily digestible timeline. Include a progression of symptoms, past history, past medications.
  • Objective: Filter a patient’s chart data to assemble a disease-specific timeline and summary: “show me all medications, test results, symptoms related to chest infection in the past year”
  • Then leave the assessment and planning to physician/APP assembly and un-assisted writing. This would leave clinician thinking largely untouched.
MODERATE
  • Subjective and Objective: GPT could take the entire chart and propose major diseases and syndromes it detects by pattern matching and assemble a brief page summary with supporting evidence and timeline, with citations.
  • Assessment and Plan: Suggest a prioritized list of Problems, severity, current state of treatment, suggested next treatments, based on a patient’s previous treatments and experience, as well as national best practices and guidelines. Leave the details, treatment adjustments and counseling to physicians/APPs interacting with the patient. Like Google Bard, GPT may suggest ‘top 3 suggestions with citations from literature or citations from EHR aggregate data’ and have the physician choose.
DREAMY/SCARY
  • Subjective and Objective: GPT could take the Moderate tools, add detection and surveillance for emerging diseases not yet described (the next Covid? the next Ebola? new-drug-associated-myocarditis? tryptophan eosinophilia-myalgia syndrome, not seen since 1989?) for public health monitoring. Step into the scanner for full body photography, CT, MRI, PET, with a comprehensive assessment in 1 simple step.
  • Assessment and Plan: GPT diagnoses common and also rare diseases via memorizing 1000’s clinical pathways and best-practice algorithms. GPT initiates treatment plans, needing just physician/APP cosignature.
  • A/P: Empowered by Eliza – like tools for empathy, takes on counseling the patient, discovering what conversational techniques engender the most patient behavior change. Recent studies already indicate that GPT can be considered more empathetic than doctors responding to online medical queries.

CMIO’s take? First things first. While we can wring our hands about “training our replacements”, there is lots yet to do and discover about our newest assistants. Shall we go on, eyes open?

Virtual Reality: reliving the past for seniors? (nytimes)

Interesting that one of our innovation partners, Rendever, has developed a way for family members to record and annotate video to be viewed by seniors, so that they can see their hometown, where they grew up, where they worked, to reawaken pleasant memories of times past. An interesting, unanticipated way of using virtual reality.

The Xenobot Future Is Coming—Start Planning Now (wired.com)

“…the ability to recode cells, de-extinct species, and create new life forms will come with ethical, philosophical, and political challenges”

https://www.wired.com/story/synthetic-biology-plan/

With CRISPR, the molecular scissors technology ,we are gaining not only read, but WRITE access to our genetic data. Writing code will no longer be limited to computers (and electronic health records), but into living organisms. Are we ready? The technology is racing ahead of our ability to think about and deploy it for the good of all.

Patients View their Own Radiology Images Online: first published experience (UCHealth)

What uses did they find for these images? Does UCHealth recommend this practice? Did CT Lin get fired as a result of these actions? #hcldr #whyinformatics #hitsm #hotoffthepress

From freepixel via JMIR

https://formative.jmir.org/2022/4/e29496

We surveyed patients who had access to view NOT ONLY their radiology reports BUT ALSO their radiology images (including plain film, CT, MRI, PET, etc) online via the EHR patient portal.

What did they think? Were they worried? Did they post the images online? Who did they share with? (hint, 4% shared on social media)

These questions, and more, are answered in the article. Click the link above, dear Reader, and press on.

T Minus 5, 4, 3, 2, 1, and We Have Pharmacogenetic Results in the EHR

For 4000 patients, we now have data and reminder tools to notify clinicians of important drug-gene interactions at the time of prescribing.

by GUEST BLOGGERS: Christina Aquilante PharmD and David Kao MD

The Go Live

The morning of Wednesday, December 1, 2021, members from the Colorado Center for Personalized Medicine (CCPM), UCHealth IT, and BC Platforms teams surrounded their home computers, fixated on a Microsoft Teams channel. It had all the feels of a space shuttle launch. The teams had been working for five months to upgrade the CCPM Biobank pharmacogenetic (PGx) return of results pipeline. Today was the big day – CYP2C19 and SLCO1B1 PGx results were about to be returned to the UCHealth Epic electronic health record (EHR) for Biobank participants.

8:22 am. “Good morning! Happy go-live! Kristy Crooks, Biobank Laboratory Director, will be signing off the first plate at 8:30 am.” typed UCHealth Project Leader, Emily Hearst.

8:30 am. “Please post in the Teams chat when you sign off on the first plate. We know there will be a delay as the plate is being processed,” typed Emily Hearst.

8:32am. “Plate signed off. Not seeing a result in Epic yet,” typed Kristy Crooks.

8:36 am. “PGX molecular was resulted!” typed Kristy Crooks. A flurry of emojis followed.

8:37 am. “Yesssss!!! Strong work all!” typed CCPM Medical Director, Dave Kao.

The teams worked for the next few hours troubleshooting minor technical glitches and testing more plates.

12:21 pm. “We have success!” typed UCHealth Systems Architect, Katie Hess.

The Biobank that returns Clinical Results

The success of December 1st’s go-live was a culmination of years of hard work from many different teams. In 2015, CCPM partnered with UCHealth to establish the Biobank Research Study. As part of the study, UCHealth patients are asked to provide a blood or saliva sample for genetic research. There is also the potential to have clinically actionable results (e.g., PGx) returned to them and their EHR. Prior to 2021, PGx results had been returned for some Biobank participants but the return process was put on hold to upgrade some of the IT infrastructure. After an incredible team effort, the revised IT pipeline launched on December 1, 2021 and

almost 4000 Biobank participants have now had CYP2C19 and SLCO1B1 results returned to their UCHealth EHR and patient portal.

Christina Aquilante, PharmD

CYP2C19 is an enzyme that metabolizes medications such as citalopram, escitalopram, clopidogrel, proton pump inhibitors, and voriconazole. Due to genetics, approximately 60% of patients are not CYP2C19 normal metabolizers, which can influence medication efficacy and safety. SLCO1B1 is a protein that transports statins into the liver. Approximately, 28% of patients have decreased or poor SLCO1B1 transporter function. This can lead to an increased risk for statin-associated musculoskeletal symptoms.

Given that > 30 million Americans take statins annually, this seemingly small risk [genetic variant] can ultimately affect a lot of people.

Christina Aquilante, PharmD

The “Last Mile” Problem

The questions that get asked most often by clinicians are – How will I know if my patient is a Biobank participant? How will I know if they have CYP2C19 or SLCO1B1 results? What do I do with this information clinically?  How often are these alerts going to interrupt what I’m doing?

The good news is that the CCPM and UCHealth teams have built clinical decision support tools to notify clinicians of important drug-gene interactions for Biobank participants at the time of prescribing. In other words – clinicians don’t need to look for it – the tools will tell them when it is important. Currently, PGx CDS tools are live across the UCHealth system for 17 medications affected by either CYP2C19 or SLCO1B1. These tools contain guidance for how to modify drug therapy based on the patient’s PGx results.

In the cable TV industry, this used to be called the “Last Mile” problem, where a cable company could build a terrific network of cable channels, underground cables and signal transmitters, and yet that “last mile” to the customer’s home, determines if the customer gets any benefit.

Importantly, the teams took great care when designing the CDS tools, and most of the tools are highly visible and yet non-interruptive in nature, i.e., they will not stop a clinician’s workflow. As of February 14, 2022,

301 drug-gene interaction alerts have fired in clinical practice for 268 Biobank participants.

David Kao MD

The most common alerts are for proton-pump inhibitors (PPIs), followed by es/citalopram, and then statins. The work to date is just the tip of the iceberg for the CCPM Biobank PGx return of results initiative at UCHealth. The team is in the process of preparing for another gene launch in early summer – this one for DPYD, which affects the chemotherapeutic agents 5-fluorouracil and capecitabine. Simultaneously, the teams are planning for the deployment of a Genomics Module in Epic and testing out new genotyping platforms with more extensive PGx variant coverage. When these pieces are in place, the sky’s the limit for PGx at UCHealth.

Christina Aquilante, PharmD, Professor
Director of Pharmacogenomics, Colorado Center for Personalized Medicine

David Kao, MD, Associate Professor
Medical Director, Colorado Center for Personalized Medicine

Sepsis, Machine Learning and the Centaur (my SMILE conference talk)

Find out: What is a centaur and what does it have to do with healthcare? What are the criteria for a good machine learning project? What is the role of a virtual health center with predictive models? And most importantly: What ukulele song goes with machine learning?

Here are the slides for my talk given at SMILE (Symposium for Machine learning, ImpLementation and Evaluation). The slides are mostly self-explanatory. You can also watch my talk at YouTube. Here is a PDF of the entire deck.

Making clinicians worthy of medical AI, Lessons from Tesla… (statnews)

Novel idea: ensure docs KNOW how to operate AI (!) (image: ETHAN MILLER/GETTY IMAGES, via Statnews)

Here is a different take on AI in healthcare: train and only allow clinicians who understand the limitations of AI, to use AI. Make savvy clinicians better. Don’t give it to all clinicians.

This is a throwback to our experience with Dragon Speech recognition over the past decade: DON’T give Dragon speech to a clinician struggling with computer use; instead, give Dragon to a clinician who is computer-savvy and understands the limitations of Dragon.

But, (in the early years) give the non-computer savvy clinician an “opt out” to dictate their notes by dictaphone or telephone, and gradually bring them along.

Having given several non-computer savvy docs access to Dragon in those early years, our hair stood on end when we ended up reading their notes later: they were clearly NOT proof-reading their work and assuming the Dragon engine was perfect at transcription.

Back to the future.

CMIO’s take? Be careful out there, everyone, both on the road with Tesla, and in healthcare with AI.

Predictive Algorithms Save Lives Sepsis @uchealth: A 5-slide talk

This data dilettante (see previous posts: dilettante #1, dilettante #2) has enjoyed armchair theorizing with all of you, my best (online) friends. Today we explore how our super-smart team scrambled our way to improving sepsis care with a predictive algorithm we built.

The old saying goes: the success of any major project in a large organization follows the 80:20 rule. 20% of the work is getting the technology right, and 80% is the socio-political skill of the people doing the work.

We all underappreciate this fact.

It turns out that we spent months building a sepsis alert predictive tool, based on various deterioration metrics, and a deep analysis of years of our EHR data across multiple hospitals. We designed it to alert providers and nurses up to 12 hours BEFORE clinicians would spot deterioration.

We patted ourselves on the back, deployed the predictive score in a flowsheet row, and in the patient lists and monitoring boards, with color coding and filters, and stepped back to revel in our glory.

Right?

Nope.

Turns out that our doctors and nurses were ALREADY FULLY BUSY (even before the pandemic) taking are of critically ill patients. Adding YET ANOTHER alert, even with fancy colors, did NOT result in a major behavior shift to ordering IV fluids, blood cultures, or life-saving antibiotics any quicker.

Hmph.

See the fancy patient-wearable tech on the left (Visi from Sotera, in this case), and one of our hardworking nurses, with ALL of our current technology hanging off her jacket and stethoscope. She should be the visual encyclopedia entry for “alert fatigue.” 🙁

(right: one of our overburdened hardworking nurses, image used with authorization)

Back to the drawing board

As result of our failure, we huddled to think about transforming the way we provided care. It was time to disrupt ourselves. We decided to implement a Virtual Health Center, mimicking what we had seen in a couple places around the country: we deployed 2 critical care physicians and about a half-dozen critical care nurses on rotation, off-site at an innovative, award-winning Virtual Health Center.

This second time around, we created a cockpit of EHR data and predictive alerts to the VHC clinicians, who were dedicated to watching for deterioration across ALL our hospitals, and responding quickly. This does several things:

  • Takes the load off busy front line clinicians
  • Creates a calm environment for focused, rapid response
  • Dramatically improves the signal-to-noise ratio coming from predictive alerts

This way, the VHC nurses view all the alerts, investigate the chart, and contact the bedside nurse when the suspicion is high for sepsis, and start the sepsis bundle immediately.

Soon, by tweaking the ways our teams worked together, we were able to reduce the burden on bedside nurses and physicians and simplify handoffs.

See chart above: Before the VHC, bedside nurses were responsible for detecting sepsis (infrequent, subtle signals during a busy shift with lots of loud alarms for other things), with many ‘grey box’ tasks, as well as ‘magenta box’ delays.

After implementing the VHC, the VHC nurses took over the majority of ‘green box’ tasks, reducing the bedside ‘grey box’ work and completely eliminating ‘magenta box’ delays.

As a result, we have dropped our “time to fluids” by over an hour, and “time to antibiotics” by 20 minutes, which we estimate has saved 77 more lives from sepsis each year.

CMIO’s take? Predictive analytics, data science, machine learning, call it what you like. This is a paradigm shift in thinking that requires disrupting “business as usual” and is hard, but rewarding work. I can’t wait to see what we all can achieve with these new tools.

Can the Wisdom of Crowds Help Fix Social Media’s Trust Issue? | WIRED

A new study finds that small groups of laypeople can match or surpass the work of professional fact checkers—and they can do it at scale.
— Read on www.wired.com/story/could-wisdom-of-crowds-help-fix-social-media-trust-problem/

%d bloggers like this: