Ambient scribes in the real world: Tips and insights from an early adopter

Ambient scribes rank among the most promising AI tools I’ve incorporated into my clinical practice. While they are still in their infancy, in my view, they symbolize a gateway for more advanced AI integration into clinical workflows. In this blog, I aim to offer a non-technical introduction to ambient scribes for fellow clinicians, share insights and tips from my eight-month experience with various tools, and discuss the future potential of this technology. Often referred to as ‘digital’ or ‘virtual’ scribes, ambient scribes harness speech-to-text technology and AI to transform clinician-patient conversations into meaningful content. This content can range from simply auto-generating an encounter note to more advanced capabilities, like identifying subsequent administrative actions (e.g., generating a referral). Given the heavy administrative load overshadowing clinical care, it’s not a surprise that ambient scribes are being seen as one of the most immediate and promising AI application in clinical settings. Read: Why explainability is essential for AI to be successful in primary care For years, clinicians have leaned on speech-to-text (aka dictation) in clinical practice. It all began with recording notes into devices or over the phone, later transcribed by someone manually. Jump to the 2010s, and thanks to advancements in automatic speech recognition (ASR), this entire process was digitalized. Most of us have used tools such as Dragon Dictate and M-Modal either during training or continue to use them in practice. They’re time-savers relative to manually drafting notes, but we still find ourselves chained to the task of dictating every note. Shortly thereafter, natural language processing (NLP)—AI that turns our conversations into text while grasping the nuances—entered the mix. This led to the development of early ambient scribe tools, typically born out of academic hubs where data scientists trained AI models from available data. However it wasn’t until late 2022 when ChatGPT and large language models (LLMs) entered the limelight that we saw real potential. A surge in companies building ambient scribe tools ensued, especially in North America, so much so that some companies have already started to fold due to competition and an oversaturated market. While not every company’s jumping on the LLM bandwagon, an increasing number are, looking to supercharge their tools. Read: Top 5 takeaways for med students to be AI-ready Do they work? But the pressing question is, how effective are these tools? The answer isn’t straightforward. Performance varies not only based on the specific tool in question but also hinges on factors like hardware (think microphone type and placement), environmental elements (such as background noise or patient and clinician accents), and clarity of speech and linguistic differences. Moreover, since every clinician has their unique style and preference for documentation, the tool’s output might resonate differently with each user. As for peer-reviewed studies evaluating real-world use, performance, and outcomes of these tools, especially those post-LLM introductions, the literature remains thin. Given the current landscape, anecdotal evidence might be our most reliable source. In my part-time family medicine practice, I’ve engaged with ambient scribes for over eight months. Although my primary experience lies with three Canadian solutions, I’ve also dabbled with US-based options. However, due to privacy and compliance considerations, I settled on continuing with a Canadian product in my own practice. Importantly, the ambient scribe tools I’ve employed are browser-based plugins and lack integration with my EMR, which means I’ve reserved them solely for documentation. My observations Observation #1 — Overall, ambient scribes have been one of the most “sticky” AI tools I’ve introduced into my practice. Given that I’ve continued to use them for eight months, that speaks to their utility. While I haven’t set out to rigorously measure the exact time savings (work on this is underway), I’ve undoubtedly felt a marked reduction in the time I commit to typing. Even when I factor in a few extra steps that need to take place (details later in this post), the overall process feels streamlined. Observation #2 — heightened efficiency aside, I feel more relaxed and present during the visit, creating what I feel is deeper patient connection, and the conversation feels less constrained by the rigidity of documentation templates. Observation #3 — these tools are evolving fast. In just half a year, the accuracy and quality of the tool outputs have seen considerable improvement. Continuous feedback from real-world usage and iterative refinements to these tools have enhanced their reliability. The proportion of encounters where I can use the raw output, without significant edits, has jumped from “a lot of the time” to “almost always”. Observation #4 — I almost always have to make some tweaks to the final output. Most often, this is to emphasize specific details of an encounter I wish to include or to pare down unnecessary information. Less often (perhaps 1 in 5–10 notes), it is to correct an error. Observation #5 — constant “improvements” in the tools can lead to some reliability issues. While frequent updates and improvements are laudable, it occasionally introduces an element of unpredictability in the generated outputs. Sometimes the smallest tweaks can have unexpected consequences on the final transcription. Observation #6 — ambient scribes are way more useful for undifferentiated presentations and visit types. Ambient scribes seem particularly advantageous in settings marked by a diverse array of clinical presentations and visit types. For more structured, repetitive visits, like those in certain subspecialty clinics, traditional templated approaches might still hold an edge. In such contexts, introducing ambient scribes might not offer a substantial time advantage (in speaking with some specialist colleagues, they usually lead to more documentation time ironically). While I don’t possess solid data, my inclination is to believe that primary care settings might exhibit a higher adoption rate of ambient scribes compared to specialized clinics. Observation #7 — Lastly, some workflow changes are needed to make ambient scribes work. These are mostly subtle yet significant behavioral adjustments associated with the deployment of ambient scribes. One pivotal aspect is obtaining patient consent. Over the past eight months, only on a handful of occasions have patients opted out of the ambient scribe feature. I make it a point to obtain and document their verbal consent. Additionally, the reduced necessity to glance at the screen during visits alters the dynamics of the patient interaction. Traditionally, viewing my notes often played a pivotal role in my clinical decision-making process. There’s also a newfound emphasis on verbalizing the details during physical exams. This not only ensures comprehensive documentation by the ambient scribe but potentially augments patient understanding and engagement during visits. Where this is going We’re currently only seeing the tip of the iceberg. Automating documentation is just the foot in the door for this very powerful technology to enter the clinical workflow. In the near term (the next 6-12 months), we will see refining and personalization of clinical documentation. I anticipate more companies will offer the ability to train your tool to your specific notes and allow them to synchronize with your existing templates. In the mid term, as the market consolidates, I believe ambient scribe will be a core EMR functionality (either as an integrated third party asset or more likely one that EMR vendors have built-in natively). This is where greater value will be unlocked—namely, the ability for the natural physician-patient conversation to drive automated administrative and clinical-decision support activities. So it’s not only your clinical note that’s automated, but also all of the subsequent steps that need to take place (patient get’s a copy of the visit, link to book next appointment, referral letter drafting, medical and social history updated, appropriate billing and claim codes applied etc). As the technology matures, the clinical decision support and predictive capabilities are where we will begin to see room for more seismic changes to how care is delivered (beyond simply ameliorating admin steps), including deepening our understanding a person’s health. Imagine Imagine for a moment that you’re having a typical conversation with a patient during a clinical visit where they are presenting with a tick bite. As you take a history for Lyme disease, the EMR pulls up key information such as questions you might want to ask (that you haven’t ask yet), images of Lyme disease target lesions, and public health information in the region (including where tick bites have been positive). As you collect a further history, a differential diagnosis (including % likelihood) is updated in real time on the screen, ultimately suggesting a management plan. They also ask to renew their statin prescription, and as they do, a graph of their cholesterol profile appears, including indication that they are overdue for repeat testing. You discuss a dose adjustment, and the screen models a probability of how their cholesterol profile might improve in six months based on this, along with impact on their cardiovascular risk score. Any subsequent actions for next steps simply require confirmation by the clinician, and they take place. The patient automatically gets a transcription of the visit in patient-friendly language. This fundamentally changes the nature of the clinical interaction and allows for a myriad of ways in which additional AI tools can be integrated into the clinical encounter. In order for this future to happen safely (mitigating bias and error), the clinical community needs to be engaged, including early exposure of clinical trainees to foundational concepts in AI, and a willingness to adapt workflows. The future is at our doorsteps, we simply need to help it through. (Disclaimer: Neither Romina or I have any affiliations, financial or otherwise, with any ambient scribe companies at the time of writing this blog. Both of us work in the digital health industry) Dr. Puneet Seth is a practicing family physician, medical AI researcher and health technologist, having led an EMR company for 10 years. This article originally appeared on Medium.

Ambient scribes in the real world: Tips and insights from an early adopter

Register for free access or login.

Related Topics