For this edition of the AAAIH Newsletter we are focusing on one of the most active areas of clinical AI innovation. Digital scribes have only recently appeared and are now an intensely contested part of the digital health marketplace. Touted for their potential to reduce documentation burden and increase note quality, the speed of their appearance has left professional and regulatory communities scrambling to understand their benefits and risks, and the best way to safely use the technology in a dynamic regulatory space.
What are digital scribes?
Digital scribes are a class of software application, designed to assist health professionals create documentation like medical records, discharge summaries, imaging reports and professional letters.
The functions found in scribe products vary. Software that transcribes dictation verbatim has been with us for a while. More advanced natural language processing driven versions, made possible with large language models like ChatGPT, aim to summarise dictation and conversations. Scribes have been researched for several years but have only become commercially feasible with the arrival of generative AI in the form of enabled large language models (LLMs).
Placing scribes at the centre of a clinical workflow changes the way health professionals use electronic record-keeping systems. It also permits changes in doctor-patient interactions, allowing for a more natural engagement without a keyboard in the middle of the consultation.
Further reading:
The Digital Scribe
The digital scribe in clinical practice: a scoping review and research agenda
Evidence synthesis, digital scribes, and translational challenges for artificial intelligence in healthcare
The view from regulators and professional bodies
Peak health bodies are closely examining the rise of AI scribe technologies, as this is a new class of technology with new benefits and risks. The market is currently deploying technology ahead of solid research evidence for safety and benefit, which increases the level of vigilance required of scribe regulators and users.
Several key advisories have appeared in the last 12 months:
In November 2023, the AAAiH National Policy Roadmap (the Roadmap) emphasised the need to “Urgently communicate the need for caution in the clinical use of generative AI when it is currently untested or unregulated for clinical settings, including the preparation of clinical documentation.”
In May 2024, the Therapeutic Goods Administration (TGA) released updated guidance indicating that LLMs – e.g. GPT 4 and Bard – supplied for a medical purpose (like diagnosis, prevention, monitoring, prognosis, and treatment) are subject to medical device regulations and must obtain approval. Basic AI scribes currently fall outside of TGA medical device regulatory processes and are considered to be simple record-keeping tools rather than medical devices. However, the addition of software functions that support clinicians beyond record keeping can quickly make a scribe subject to TGA regulation. This includes software for “clinical decision support that uses generative AI to create symptom and diagnosis summaries”.
The RACGP released a fact sheet on AI scribes July 2024, detailing important considerations for GPs and health professionals when using an AI scribe. They highlight the need for GPs to carefully check AI generated output given the risk of medical errors (such as filtering out pertinent information, ‘mishearing’ medical terms, and incorrectly categorising data).
The Australian Health Professions Regulatory Authority (Ahpra) has released an August 2024 update to its shared code of practice, explaining how health professionals must use AI in a way that meets their professional obligations.
More broadly, the RACGP and AMA have released position statements on AI in healthcare, outlining key considerations and policy parameters that should be addressed before integrating AI tools into healthcare delivery.
Potential benefits and risks
The purported benefits from using AI scribes are reduced time spent on manual documentation, more time for clinicians to devote to direct patient care during consultations, increased patient satisfaction, and reduced clinical burnout.
There is however, currently limited evidence around the performance of AI scribes in real-world clinical settings and several risks have been identified. One of the first large-scale studies on AI scribes in real-world clinical practice was published in NEJM Catalyst in February 2024. The Permanente Medical Group in California implemented a pilot ambient AI scribe for 10,000 of its clinicians. The study revealed that clinicians reduced their time spent using the EMR outside of 7:00am and 7:00pm (“Pyjama time”) and there was an enhanced quality of interaction between patients and doctors. 81% of patients reporting that clinicians spent less time looking at their computer than in past visits. But the study is also among the first to document real-world instances of AI hallucination including incorrectly diagnosed conditions, omitted crucial details, and summarisation inconsistencies. Other studies report varying levels of concern around errors in notes with current generation scribes and large language models. A study in JAMA that looked at electronic messages between physicians and patients found generative AI was not associated with reduced time for writing a reply but was associated with longer read time, longer replies, and perceived value in making a more compassionate reply.
Scribes and the law
In an recent advisory piece, Avant’s Head of Medico-legal Advisory Kate Gillman highlights that not all AI scribes are the same. She urges users to ensure that their AI scribe is trained to identify clinically significant information, and to check with colleagues and colleges for reputable tools.
Legally in Australia, patient consent must be obtained before using digital scribes. Patients also need to be given meaningful information to support informed consent, such as the limitations of AI, adverse events, and alternative (non-AI) solutions where possible. MDA National, a leading Australian provider of medical indemnity insurance and medico-legal services, recommends that patient consent should be collected on a written form. They state that providing signs in the practice about the AI technology, whilst helpful for information and expectation setting, does not substitute for consent.
Protecting data when using AI scribes is a major concern, with the risk that sensitive patient information is exposed and harm caused to individuals and companies. Users should consider: Where are the recordings being stored (e.g. in Australia in a secure environment, or overseas beyond local jurisdiction), how are data being protected, and what unconsented uses might the data be put to?
Medico-legal liability is clear. Healthcare professionals are ultimately responsible for reviewing and amending any AI-generated health records, ensuring that patient health records are accurate and free from error.
What is happening in the AI scribe industry?
In the last five years, digital scribes have developed from basic speech-to-text technologies into advanced tools that assist with generating patient health records, referral letters, and treatment summaries. The AI scribe market in Australia is evolving quickly, with companies offering varied products and capabilities that differ in the LLMs used, and the software’s accuracy, speed, EHR integration, localisation, and cost.
Scribe products are available for clinicians working in general practice, specialist care, and hospital settings, and for allied health professionals. As part of rapid consolidation in the marketplace locally, partnerships are forming between AI scribe companies and larger health or tech enterprises, including EHR vendors. A partnership between Lyrebird Health and Best Practice Software was announced in February 2024. This was notable because Best Practice is used by an estimated 65% of GPs. This partnership integrates Lyrebird’s generative AI scribe capabilities directly into Best Practices Software. Another big story is Telstra Health’s new launch of Smart Connected Care, an ecosystem of digital health solutions that will soon include a Smart Scribe tool. Telstra Health acquired MedicalDirector, the second leading GP practice management software company, in 2021.
AI scribes have also been released by many other companies including i-Scribe by Akuru, Astra Health AI, Heidi Health, mAIscribe, Patient Notes, MedNotes.ai, Telecare’s AutoTranscribe, and Silknote.
In the US the competitive landscape for scribes is intense. “There are 42 companies in this space right now that are offering generative AI medical documentation solutions,” Augmedix CEO Krakaris told Forbes in May 2024, “So there’s a lot of noise in the market and a lot of trials that are ongoing.” US AI scribe companies include DeepScribe, Abridge, Scribeberry (launched in April 2023), Amazon Web Services’ HealthScribe (released in 2023), and Nuance Communication’s Dragon Medical platform (DAX) (acquired by Microsoft in 2022 for $18.8 billion, and used by over 550,000 physicians according to Nuance).
It should go without saying that mentioning a product or company in this Newsletter does not constitute endorsement by the AAAiH.
Next issue of the AAAIH Newsletter
If you like this special themed issue of our newsletter, please let us know, and suggest topics for future issues. We are also looking for stories for the next issue, so please send these in.
Contact – Pippa Dorricott, Executive Officer: info@aihealthalliance.org