St. Joseph Health Medical Director's Review of AI Scribe Technologies
DeepScribe is the only truly automated natural speech-to-note software that does not require some degree of physician dictation or speaking directly to the system to get information into the note. It is the most fully automated, most scalable, and most cost-efficient solution I could find on the market.
WORTH THE HYPE? Medical AI Scribe Software Demoed and Reviewed
Physician burnout is a major problem plaguing our modern healthcare system. There has been a robust amount of work published on investigating the extent of the problem, its costs to the system, and its causes. Unfortunately, there has been very little work (other than opinion pieces and editorials) on designing real-world solutions that can alleviate physician burnout and thereby increase access, affordability and quality of care within our healthcare system.
This is a short information piece about new real-world solutions that I as a practicing clinician am very excited about. I have been practicing general Internal Medicine for over 20 years, both in inpatient and outpatient settings, and in urban, suburban, and rural communities. During that time I have seen my professional energy getting more and more hijacked by meaningless, clinically useless and often frankly insulting clerical work. I understand intellectually the scope of the problems related to physician burnout, and I have experienced it personally.
There is general consensus that the widespread adoption of EMR systems has been a significant contributor to physician burnout and a primary cause for the documented increase in numbers of physicians who report symptoms of burnout. Certainly, my experience in clinical medicine has been that the use of EMR systems has created more meaningless clerical work and reduced the amount of face-to-face time available for clinical care, all while failing to provide any significant improvements in safety or clinical decision-making support. Does my EMR remind me when my patient is due for his pneumococcal vaccine or her mammogram? No. I still have to sort through paper printouts from other EMR systems when I see a new patient coming from a different doctor, find when they had their pneumococcal vaccine or mammogram, and then manually input that information into my EMR. If I fail to do this, then I get penalized on quality metrics.
Is this what I went to medical school for? Is this why I spent my precious second and third decades of life staying up all night long caring for patients with sepsis in the ICU?
Hell no. It most certainly is not.
Until recently, the best commercially available solution to the problem of EMR intrusion into patient care has been the medical scribe. This service involves hiring a person who stands at the physician’s side during every patient encounter. The scribe enters the relevant information into the EMR for the physician, presents available information from the EMR such as medication lists and test results, and enters orders for the physician. This has been found to be helpful for some physicians, but only for the physicians who are already high performers. It is often offered as a solution to help those physicians who are chronically behind in completing their clinical documentation and entering the charges for their encounters, but in general it has been found that the chart completion rates of these poorly performing physicians does not significantly improve with the help of a medical scribe. This model has other shortcomings as well. Many patients are not comfortable having another person in the room during a medical examination. Medical scribe services have high turnover rates among the people hired as scribes, which often interrupts service. And it is not a solution that can scale efficiently: it requires hiring a scribe for each physician, and the cost is prohibitive at about $50,000 per year for each physician in a practice.
Enter the burgeoning age of AI, or artificial intelligence. AI, along with its software cousin Natural Language Processing, is now being used to develop software systems that can record the physician-patient interaction and enter that information into the EMR. This represents a significant breakthrough in how technology can be used to improve the process of care. Recently, I had the opportunity to demo and review the major AI software options available in an effort to find the best one for my practice.
There are several companies developing products in this realm, and they all claim to have features that will only make the physician’s workflow more efficient, more accurate, and more patient-focused. In my review of the five major products that are currently available, I found that while all of them are improvements over the current state of EMR documentation, they have significant differences in user interface and workflow design.
I will start with my personal favorite, DeepScribe. DeepScribe is a cloud-based software product that uses either an iOS app or a small listening device (think Amazon’s Alexa or Google’s Home) placed unobtrusively in the examination room. DeepScribe listens to the dialogue between physician and patient, and the software uses AI, deep learning and Natural Language Processing algorithms to translate that interaction into a clinical note written in the style that physicians normally use for their progress notes. It records the physical examination simply by having the physician say out loud to the patient things like, “your heart examination is normal,” or “you have a small, easily reducible inguinal hernia on the left.” Most impressive about this software is its ability to sift out all the small talk and other non-clinically relevant chatter that is a part of most interactions, and boil the conversation down to the most clinically relevant details. That way you can still keep the small talk with your patients, without spending the extra cognitive energy required to remember the clinically salient facts for documentation later on. At the end of the visit, usually within about 20 minutes, the program has created a HIPPA-compliant encounter note in the EMR which accurately reflects the interaction, the examination, as well as the assessment and plan, and includes all required data points to comply with coding and billing regulation. Once the note is complete, the physician simply has to review it, make any adjustments desired, and sign off. And because it uses the most advanced machine learning algorithms, it can learn with each individual physician and get more and more precise to each physician’s individual documentation style over time. DeepScribe allows physicians to integrate any template or macros in their models to emulate the way that they practice as well. They also recently released an iPhone, iPad, and Apple Watch version of their scribe.
This product results in a completed, accurate, and compliant clinical note without the physician having to manually enter any information, or dictate any information directly into the system. It translates normal colloquial speech into the medical terminology that is required for clinical documentation. It learns the unique style and thought process of each physician that uses it. In addition, DeepScribe is nearly completely automated in several specialties, with little need for human oversight of the note production. This makes it easily scalable to large group practices, and allows for the most cost-efficient pricing among everyone I’ve reviewed here.
Saykara is a close competitor of DeepScribe, but has some significant differences in how it interfaces with the user. Saykara operates off of an iOS app, so it requires that the physician carry an Apple device (iPhone or iPad) into the examination room. It has two modes of listening: “ambient” mode which is listening during the physician-patient interaction, and “summary” mode which the physician uses after the visit to complete dictation of the clinical note. During the ambient mode the system is listening only to the physician, and transcribes based on what the physician says, so the physician has to make sure to repeat important elements of the history to get them into the HPI section of the note. In this way, it is a mix of automated natural speech-to-text and dictated speech-to-text. Saykara has not tried to solve the problem of filtering out all the unrelated colloquial speech the way that DeepScribe has. In this way, Saykara is using AI technology to support the clinical documentation, but the physician still has to “summarize” most of the information directly into the system rather than having the system create the note based on the natural interaction between the physician and the patient.
While almost every product on this list uses in-person reviewers, Saykara’s human involvement seems to have the most downstream effects - primarily the increase in turn-around time for production of a complete note for the physician to review and sign. They have multiple pricing tiers based on the turn-around time of the notes, with options ranging from 2-3 hours to the end of the day. It also makes it more challenging and costly to scale to large numbers of users, as the company has to hire more note reviewers to accommodate larger numbers of users. However, Saykara reports that this is not a significant problem because there is a large available labor pool of medical transcriptionists who have lost their more traditional employment roles. This increases the per-physician cost of the service compared to DeepScribe.
Saykara has two levels of service. The basic level is just the clinical documentation. The higher level service has some unique and potentially very useful features. It provides the ability for the system to retrieve information from the EMR, which allows the system to find clinically relevant information and incorporate it into the note automatically. In addition, there is an added feature whereby a physician can dictate orders into the system for ordering tests and prescribing medications. This could be a huge time-saver, and would reduce the mental fatigue we all feel when clicking endless boxes to order tests and medications. For those practices with a higher budget, the additional features provided by Saykara make it a very attractive option.
Notable is another interesting competitor, although it is not just a clinical note documentation service. Notable describes themselves as “an automation company.” Their software is designed to help with online scheduling, pre-visit paperwork completion over the internet, and also includes a dictation feature. It is accessed via a smartwatch that the physician wears while seeing patients. The software can pre-populate the clinical encounter note with information from pre-visit forms or previous chart notes, and it listens in on the conversation but does not have a Natural Language Processing feature that generates the note. In this way, the physician still has to dictate the note after the visit. The physician can also order tests or write prescriptions using the voice command feature. It is a mostly automated process of note generation, so can be scaled quickly for larger groups or systems. However, EMR access is variable, especially for some of the smaller or older EMR systems in use. Overall it can help improve the efficiency of the entire visit using the pre-visit coordination features with patients, but the note completion still takes extra time, requiring the physician to dictate into the system after the visit.
Sopris is another software product that is used as an app on a mobile device, but unlike Saykara it is not iOS-specific. Similar to Saykara and Notable, Sopris is an advanced dictation system, and does not fully automate the process of generating the clinical note. The physician dictates into the app, and the system uses automated processes to structure and complete the note. Note structure can be customized to each practice or group using the software based on predefined parameters that are set up at the initiation of service. This allows physician groups to structure the notes based on the unique needs of their practice, but it does require achieving agreement from all the physicians within the group and limits ability to change note structure for individual physicians within the group. The software does have the ability to ask structured questions to the physician to help complete the note documentation, which has interesting possibilities for point-of-care reminders. The note generation is fully automated, and therefore easily scalable. It has some limited order entry and data-mining capabilities which no doubt will be expanded as the product develops.
Suki is the final product that I researched. Suki runs on an app that can be kept on a mobile device, or it can be accessed on a laptop or desktop. It did not interface with the EMR our practice uses, so it would have had to create a WordPerfect document that would be scanned into the chart. It is essentially an advanced dictation system. The physician dictates the note into the system, and can use voice commands to manipulate the note and add or subtract pre-defined documentation features. It has no data mining capability and limited deep learning features. It is automated and easily scalable, and is an affordable option for physicians or practices looking for a limited solution.
These are the major innovative solutions currently available for the problem of clinical documentation. Despite these available solutions, burdensome and time-consuming documentation requirements are baked into our healthcare system, and encoded in the largest and most widely used EMR systems. This documentation burden is a direct cause of increasing physician burnout, which is estimated to cost the healthcare system around $4.6 Billion per year.
But let’s remember, this is not just a documentation problem. In addition to the financial costs, it results in human costs as well. Patients have less access to care, and the care they can manage to access is of unacceptably poor quality. Many patients relate stories to me of seeing other doctors who “spent the entire 10 minutes of my visit just staring at his laptop and clicking boxes.” They feel rushed and unheard, unable to ask their most important questions about their own health issues. And at the end of that rushed visit, there is still an unacceptably high chance that they will get prescribed the wrong medication, or have the wrong test ordered.
And perhaps most alarming is the fact that physicians as a group have some of the highest suicide rates of all the current American professions. This is not a trivial problem.
What would an ideal solution to this problem look like? And are we approaching a critical transition in our available technology that might allow us to design an effective solution?
The ideal solution to our clinical documentation problem would use technology in a way that enhances the personal human interaction between the physician and the patient. We use important scientific knowledge in the practice of medicine, but at the end of the day true healing requires a meaningful and trusted human relationship between patient and practitioner.
Technology, if present in the room, should not be the primary focus of attention. It should be a more subtle feature which provides the ability to document the physician-patient interaction in a way that is both clinically meaningful and also efficient, without requiring the physician to directly enter that information into the record.
My ideal solution would also have the capacity to provide important clinical information that may be available in the record, and present that information to the physician in a timely and unobtrusive manner such that it can help support appropriate and effective clinical decision making. It would also seamlessly take care of any administrative tasks required by those decisions, such as ordering a test or prescribing a medication. It would provide meaningful and actionable information to the patient about their episode of care and any plans or decisions that were made. During the visit, this technology would allow the physician to focus on the particular needs of the patient without feeling rushed or distracted. At the end of the visit, the patient would leave with a feeling that their voice was heard, their needs were attended to, and a clear understanding of their plan of care. The physician would leave with a relaxed state of mind knowing that the entire interaction has already been documented completely. This would allow an easy transition of his or her mental focus to the next patient.
All of the current AI scribe software companies are focused on creating solutions that move towards the ideal as I outlined it, but there are significant differences in their approaches. For me, the most important consideration is which product allows for complete and accurate documentation without requiring the physician to dictate information after the visit, click on boxes, or spend a lot of time editing notes after-hours. Second, as a member of a large healthcare system, scalability and cost were very important factors. And looking ahead, I am also interested in which system has the deep learning capabilities to expand its use of the information collected during the visit to help the physicians access better and more reliable clinical information.
It seems clear to me that DeepScribe is the current solution which comes closest to this ideal. DeepScribe is the only truly automated natural speech-to-note software that does not require some degree of physician dictation or speaking directly to the system to get information into the note. It is the most fully automated, most scalable, and most cost-efficient solution I could find on the market. These are critical components that work best if built into the product from the beginning, and are more difficult to add on later. DeepScribe also has the deep learning capacity to expand its services over time in ways that will better serve the physician-patient interaction and lead to better and more cost-effective care. For my primary care practice, it was a clear choice.
Paul Laband MD, HMD-C
Internal Medicine, Primary Care, Hospice and Palliative Care
May 19, 2020
Find the original article here.