I appreciate your point that AI scribes are more like a band-aid than an actual solution to the documentation problem. If we have to keep using tech (AI scribe) to address issues with tech (EMR), seems like we're just building on a bad foundational.
We've been in discussions at our practice about "Where to find the story?" in our EMR (to your point about A/P as a "now" screenshot vs. "story summary" for covering providers). Too many documents and difficulty adding to the A/P quickly for clinical updates makes for a sporadic and fractured story at times. But, as you mentioned, creating these summaries is time consuming and challenging for providers who already have enough going on. If only, if only...
Maybe someday we’ll trust the AI to extract the deltas from the current visit and then add a bullet point to the problem list summary, detailing the update into the longitudinal view. Maybe someday…
I love your thought-provoking article. I am reminded of thoughts I had in medical school, now 10 years ago, about how EMR documentation should be more of a sophisticated wiki-based system rather than just an electronic filing cabinet. You edit a health summary collaboratively, and the "change note" indicates the reason why certain changes were made. You can then easily generate diffs, track history, and understand rationales. I hope that work to redesign the underlying documentation systems goes on in parallel and does not stop short because AI scribes will save us time now - and that they will continue to support interim improvements in documentation. My organization will soon have them and I have not yet decided how to integrate them with my notes that include the brief longitudinal assessments that I currently copy forward and update.
As a retired consultant pod I recall the structural limitations with my 'early adoption' of an emr in my office practice. I felt comfortable with dictated notes (transcribed by my secretary) which might capture the essence of the initial problem and focus on current situation and needs. This was my progress info but also the info that went to the referring clinician.
From the outset I felt uncomfortable with the way the emr seemed to want me to go and your lucid depiction of the limitations of SOAP and those of AI to 'aid' are important and help me to understand why. Thanks.
Thanks Derek. The SOAP challenge was there on paper as well; the advent of EMRs should have surpassed that, but instead were designed to imitate it - and inherit all the challenges. And they are so hard to change...
Yes! This!..... "The long-term answer is to redesign EMRs with longitudinal documentation and planning as a core architectural feature. The SOAP note should be something that the EMR generates as needed, not the primary data construct for encounters." Thanks James for another thought-provoking post!
As a lead designer working for an EHR organization in ambulatory care, I’ve encountered the challenges of problem-based charting. Implementing any modern EMR/EHR app within a landscape of legacy technology is no small feat.
Currently, while medications, charges, and orders can be tied back to a diagnosis code (problem), narration remains a challenge. Narration doesn't seamlessly link to specific diagnosis codes. While AI could help here, I’m hesitant to use technology for the sake of it. I believe it’s possible to do this without AI, though it might increase the time required for documenting a patient visit. However, the added time could be minor and potentially worth it.
We should expect diagnoses to become more specific over time, but what happens to the connection between data and the original diagnosis? Should we show the evolution of a diagnosis, or should the original connection be overridden by the more specific one? I believe it's essential to show how the problem evolves over time.
For a problem-based chart to work, everything—narration, orders, medications, charges, etc.—must be tied to the diagnosis, which in turn needs to evolve as the diagnosis becomes more specific. This isn’t an easy task. AI could handle it, but I don’t think we should wait for that. Instead, we can work with doctors and staff to research and analyze how best to present this information in a way that’s intuitive and easily consumed.
As an industry, we must do a better job of learning, shadowing, observing, and interacting with our users (doctors, administrators, staff) on the front lines to truly understand and meet their needs.
As a physician who worked closely with Larry Weed, the inventor of the SOAP note, i can share that his invention has been corrupted and not understood by students, residents and clinicians today. In the 1960's and 70's luminaries in Internal Medicine (such as Willis Hurst MD cardiologist to President Johnson) were excited by the idea of bringing a common format to the medical note and embraced the POMR and SOAP format. In 1968 Larry wrote in the NEJM "Medical Records That Guide and Teach" The idea of the POMR and SOAP format was to bring rigor and precision to clinical reasoning. Unfortunately the note has turned into a billing instrument and there is little discussion about the cognitive mistakes that are tied to the EHR today, the same cognitive biases Larry was trying to reduce through the SOAP format. No conversation about what to do with the EHR should occur without considering biases such as premature closure, anchoring, representative bias and dozens more. Our goal should be to reduce these biases and become more accurate in our clinical care delivery. The conversation has been about burnout and making documentation more efficient which is tremendously important but we must not be delusional and think if we could just document faster medicine is fixed! It is almost two separate worlds, the folks concerned about diagnostic and therapeutic accuracy and those in informatics/industry trying to save docs time......the tremendous quality and safety problems we have must inform changes to the EHR not just the burnout problem.
Dr. Weed also pointed out that the bedrock for improving medical decision-making is to have feedback loops. Without them we will make the same mistakes over and over. There is incredible promise in AI, ML and LLM's to perhaps build in feeedback loops but if we don't get it right, we are just "computerizing the chaos" as Larry would frequently say. The other important point which might seem quaint to many is that the problem list was not intended to be a past medical history list. A chief complaint was a chief complaint in the problem list until you had a basis for the diagnosis. It is rare to see actual symptoms/problem such as "unexplained vomiting" in the EHR. Prematurely listing a diagnosis forces anchoring. Hope some of this is helpful. On this page i posted his 1971 Emory IM Grand Rounds, its the black and white video. https://www.visualdx.com/about-visualdx/larry-weed/. It's basically the first TED talk before there were Ted talks. If you have never watched it. Watch it.
Dr. Papier, thank you for your thoughtful reply! Your connection to Dr. Weed is amazing, I feel lucky to see your comments - and validated by the congruence of our opinions on so many things.
I would add to what you said: the SOAP note has been coerced for billing purposes, it's true. But it equally has been coerced for primary care. The same logical thought process applies across the board, but the data we need to get there is quite different in chronic outpatient care than in hospital care (as probably you know). In my experience, hospital care, and specialist care, is a bit like taking a seat in a theater, reading the program, and then viewing the play. Primary care is more like jumping in at episode 6, season 3 of a TV series. The play will have within it a summarized backstory, however complete or incomplete; but to pick up watching the series, we need to see a running account of what is going on. The most important things are the current characters, the ones acting right now, how their individual stories relate to the plot. The analogy is far from perfect, but captures the feeling. This is especially, perhaps exponentially, true in the modern shared-care environment of a community health center. The car team model, the shortage of clinicians, and the health literacy level of the typical patient population all drastically change the model of the old days, when a single family doctor followed patients through the medical story of their lives.
Your penchant for considering biases mirrors my own. Representative and anchoring bias are big ones. Availability bias is one of the most important. Computer intelligence could help there, by helping us balance likelihood (via prevalence, incidence, sensitivity/specificity/predictive values of tests etc.) with our own diagnostic acumen. This would help us tell horses from zebras, and gather stray diagnoses we might miss. But a system that does that, once it does it well enough for clinicians to trust it, introduces what some call technological bias: the increased likelihood of accepting an answer because it came from a technologic source. This is on the rise since Google got smart, and could become rampant with the expansion of medical AI. As more and more barely-trained, mid-level clinicians (NPs, PAs) increasingly cover primary care - often forced to do so unsupervised - those sources will be come even more of a mainstay.
PCPs often can't use the EMR "Medical History", that is usually an outdated, unmaintained, mixed cache of procedures, ancient history, real medical diagnoses and bizarre billing diagnoses. At best it is a corrupted and incomplete backstory. And the "recent A&P" approach leaves out important characters and events left and right. This problem is ultimately solvable, but the redesign of how EMRs deal with continuity of care will have to be radical and deep. In my own articles and videos I take the tach of partial solutions that layer reasonably on top of existing paradigms, as these are palatable enough to interest today's EMR designers.
I am downloading Dr. Weed's talk now, to listen on a plane tomorrow. His enthusiasm is infectious from the first moment of the video. Thank you for that.
I appreciate your point that AI scribes are more like a band-aid than an actual solution to the documentation problem. If we have to keep using tech (AI scribe) to address issues with tech (EMR), seems like we're just building on a bad foundational.
We've been in discussions at our practice about "Where to find the story?" in our EMR (to your point about A/P as a "now" screenshot vs. "story summary" for covering providers). Too many documents and difficulty adding to the A/P quickly for clinical updates makes for a sporadic and fractured story at times. But, as you mentioned, creating these summaries is time consuming and challenging for providers who already have enough going on. If only, if only...
Maybe someday we’ll trust the AI to extract the deltas from the current visit and then add a bullet point to the problem list summary, detailing the update into the longitudinal view. Maybe someday…
I love your thought-provoking article. I am reminded of thoughts I had in medical school, now 10 years ago, about how EMR documentation should be more of a sophisticated wiki-based system rather than just an electronic filing cabinet. You edit a health summary collaboratively, and the "change note" indicates the reason why certain changes were made. You can then easily generate diffs, track history, and understand rationales. I hope that work to redesign the underlying documentation systems goes on in parallel and does not stop short because AI scribes will save us time now - and that they will continue to support interim improvements in documentation. My organization will soon have them and I have not yet decided how to integrate them with my notes that include the brief longitudinal assessments that I currently copy forward and update.
Exactly right, agree.
I agree with you, I don’t want an AI scribe to write my notes for me. I would love an AI thing that would sign off VNA forms and Same forms though!
Supposed to say DME forms.
As a retired consultant pod I recall the structural limitations with my 'early adoption' of an emr in my office practice. I felt comfortable with dictated notes (transcribed by my secretary) which might capture the essence of the initial problem and focus on current situation and needs. This was my progress info but also the info that went to the referring clinician.
From the outset I felt uncomfortable with the way the emr seemed to want me to go and your lucid depiction of the limitations of SOAP and those of AI to 'aid' are important and help me to understand why. Thanks.
Thanks Derek. The SOAP challenge was there on paper as well; the advent of EMRs should have surpassed that, but instead were designed to imitate it - and inherit all the challenges. And they are so hard to change...
Yes! This!..... "The long-term answer is to redesign EMRs with longitudinal documentation and planning as a core architectural feature. The SOAP note should be something that the EMR generates as needed, not the primary data construct for encounters." Thanks James for another thought-provoking post!
I love what you're doing, James! Keep it going.
As a lead designer working for an EHR organization in ambulatory care, I’ve encountered the challenges of problem-based charting. Implementing any modern EMR/EHR app within a landscape of legacy technology is no small feat.
Currently, while medications, charges, and orders can be tied back to a diagnosis code (problem), narration remains a challenge. Narration doesn't seamlessly link to specific diagnosis codes. While AI could help here, I’m hesitant to use technology for the sake of it. I believe it’s possible to do this without AI, though it might increase the time required for documenting a patient visit. However, the added time could be minor and potentially worth it.
We should expect diagnoses to become more specific over time, but what happens to the connection between data and the original diagnosis? Should we show the evolution of a diagnosis, or should the original connection be overridden by the more specific one? I believe it's essential to show how the problem evolves over time.
For a problem-based chart to work, everything—narration, orders, medications, charges, etc.—must be tied to the diagnosis, which in turn needs to evolve as the diagnosis becomes more specific. This isn’t an easy task. AI could handle it, but I don’t think we should wait for that. Instead, we can work with doctors and staff to research and analyze how best to present this information in a way that’s intuitive and easily consumed.
As an industry, we must do a better job of learning, shadowing, observing, and interacting with our users (doctors, administrators, staff) on the front lines to truly understand and meet their needs.
As a physician who worked closely with Larry Weed, the inventor of the SOAP note, i can share that his invention has been corrupted and not understood by students, residents and clinicians today. In the 1960's and 70's luminaries in Internal Medicine (such as Willis Hurst MD cardiologist to President Johnson) were excited by the idea of bringing a common format to the medical note and embraced the POMR and SOAP format. In 1968 Larry wrote in the NEJM "Medical Records That Guide and Teach" The idea of the POMR and SOAP format was to bring rigor and precision to clinical reasoning. Unfortunately the note has turned into a billing instrument and there is little discussion about the cognitive mistakes that are tied to the EHR today, the same cognitive biases Larry was trying to reduce through the SOAP format. No conversation about what to do with the EHR should occur without considering biases such as premature closure, anchoring, representative bias and dozens more. Our goal should be to reduce these biases and become more accurate in our clinical care delivery. The conversation has been about burnout and making documentation more efficient which is tremendously important but we must not be delusional and think if we could just document faster medicine is fixed! It is almost two separate worlds, the folks concerned about diagnostic and therapeutic accuracy and those in informatics/industry trying to save docs time......the tremendous quality and safety problems we have must inform changes to the EHR not just the burnout problem.
Dr. Weed also pointed out that the bedrock for improving medical decision-making is to have feedback loops. Without them we will make the same mistakes over and over. There is incredible promise in AI, ML and LLM's to perhaps build in feeedback loops but if we don't get it right, we are just "computerizing the chaos" as Larry would frequently say. The other important point which might seem quaint to many is that the problem list was not intended to be a past medical history list. A chief complaint was a chief complaint in the problem list until you had a basis for the diagnosis. It is rare to see actual symptoms/problem such as "unexplained vomiting" in the EHR. Prematurely listing a diagnosis forces anchoring. Hope some of this is helpful. On this page i posted his 1971 Emory IM Grand Rounds, its the black and white video. https://www.visualdx.com/about-visualdx/larry-weed/. It's basically the first TED talk before there were Ted talks. If you have never watched it. Watch it.
Dr. Papier, thank you for your thoughtful reply! Your connection to Dr. Weed is amazing, I feel lucky to see your comments - and validated by the congruence of our opinions on so many things.
I would add to what you said: the SOAP note has been coerced for billing purposes, it's true. But it equally has been coerced for primary care. The same logical thought process applies across the board, but the data we need to get there is quite different in chronic outpatient care than in hospital care (as probably you know). In my experience, hospital care, and specialist care, is a bit like taking a seat in a theater, reading the program, and then viewing the play. Primary care is more like jumping in at episode 6, season 3 of a TV series. The play will have within it a summarized backstory, however complete or incomplete; but to pick up watching the series, we need to see a running account of what is going on. The most important things are the current characters, the ones acting right now, how their individual stories relate to the plot. The analogy is far from perfect, but captures the feeling. This is especially, perhaps exponentially, true in the modern shared-care environment of a community health center. The car team model, the shortage of clinicians, and the health literacy level of the typical patient population all drastically change the model of the old days, when a single family doctor followed patients through the medical story of their lives.
Your penchant for considering biases mirrors my own. Representative and anchoring bias are big ones. Availability bias is one of the most important. Computer intelligence could help there, by helping us balance likelihood (via prevalence, incidence, sensitivity/specificity/predictive values of tests etc.) with our own diagnostic acumen. This would help us tell horses from zebras, and gather stray diagnoses we might miss. But a system that does that, once it does it well enough for clinicians to trust it, introduces what some call technological bias: the increased likelihood of accepting an answer because it came from a technologic source. This is on the rise since Google got smart, and could become rampant with the expansion of medical AI. As more and more barely-trained, mid-level clinicians (NPs, PAs) increasingly cover primary care - often forced to do so unsupervised - those sources will be come even more of a mainstay.
PCPs often can't use the EMR "Medical History", that is usually an outdated, unmaintained, mixed cache of procedures, ancient history, real medical diagnoses and bizarre billing diagnoses. At best it is a corrupted and incomplete backstory. And the "recent A&P" approach leaves out important characters and events left and right. This problem is ultimately solvable, but the redesign of how EMRs deal with continuity of care will have to be radical and deep. In my own articles and videos I take the tach of partial solutions that layer reasonably on top of existing paradigms, as these are palatable enough to interest today's EMR designers.
I am downloading Dr. Weed's talk now, to listen on a plane tomorrow. His enthusiasm is infectious from the first moment of the video. Thank you for that.