The Digital Scribe: Preclinical but Promising
A system that automatically creates summaries of clinical conversations sounds futuristic, but the 'digital scribe' might be closer than you think. What is already possible and which hurdles still need to be taken to get these digital scribes implemented in clinical practice?
In April 2018, a commentary called 'Reimagining Clinical Documentation Using Artificial Intelligence' was published in Mayo Clinic Proceedings1. This commentary, written by three physicians, asked technicians, data scientists, and physicians to join forces and 'unshackle' physicians from the electronic health record (EHR) using artificial intelligence (AI). This commentary was followed by two perspectives published in npj Digital Medicine. The first one described how clinical documentation would transfer from a human task to a computerized task in the form of a digital scribe2. The second one explained the challenges still facing the development and implementation of such a digital scribe3. This last perspective was published in 2019, so you might be wondering what has happened in the meantime and if we've come any closer to an actual digital scribe. Our recently published scoping review sheds light on this question4.
First things first: what is a digital scribe? A digital scribe is a system that supports physicians by automating a part of the clinical documentation task following a clinical conversation. The digital scribe records the conversation using a microphone and automatically transcribes it using automatic speech recognition (ASR). Then, the digital scribe extracts relevant information from the transcript and forms this into a summary. This last step uses natural language processing (NLP), with techniques such as entity extraction, classification, or summarization. In entity extraction, a model is trained to find specific entities in a text, such as names or dates. For a clinical conversation, you could train this model to find symptoms, conditions, and medications. Classification models can classify parts of the transcript as relevant or not or assign them to a specific part of the clinical note, such as the social history. Summarization can be done using a model that extracts sentences from the text (extractive summarization) or generates new sentences (abstractive summarization). The combination of these components is called a digital scribe (see Figure 1).
How could a digital scribe improve clinical practice? There are several ways in which automating clinical documentation could have a positive impact on healthcare. First of all, it could decrease the administrative burden that is currently weighing down on physicians. Physicians spend an average of 40% of their time on administrative tasks, of which clinical documentation is the most time-intensive task. Decreasing the time spent on clinical documentation would allow physicians to spend more time on their patients or treat more patients. Secondly, clinical documentation doesn't start after the conversation; it is often done while talking to the patient. Since the introduction of the EHR, physicians spend between 25% and 55% of their time looking at the screen during a consultation5. This negatively affects the communication between physician and patient and even negatively affects health outcomes. Lastly, automating clinical documentation makes it possible to create a structured clinical note without an extra time investment from the physicians' side. This could be the first step towards a more ambient intelligence, where medication is automatically ordered, and suggestions for diagnoses are shown to the physician during the conversation. It all starts with collecting structured data on medical conversations. But we're getting ahead of ourselves!
Let's back up a bit: what is the current status of the digital scribe? This is what we wanted to find out with our scoping review. We studied (1) which NLP methods were being used, (2) how accurate these methods were, and (3) if any of these methods had been evaluated in clinical practice. We decided on two approaches: a literature review and an online search for companies already offering digital scribes in the hope of getting some performance data. After screening 2348 articles and contacting six companies, we ended up with 20 articles and performance data from one company. This data led us to the following insights:
- Although the literature is scarce, at least five research groups and ten companies are developing or already offering a digital scribe.
- This field of research has really taken flight since 2019 (see Figure 2), which corresponds to the introduction of contextual word embeddings such as ELMo6 and BERT7, indeed used in most articles.
- The results of the published models are promising! However, none have actually been implemented. On the other hand, several companies offer implemented digital scribes, claiming to save physicians up to hours a day (without scientific evidence).
- The previous point also shows the knowledge gap between development and implementation: all published articles focus on the technical validation of their model, while the companies offering a digital scribe don't publish articles at all. The effect of digital scribes on clinical practice thus remains unknown.
What can we conclude from these insights? Based on our scoping review, there is reason for optimism: it seems possible to create a digital scribe that saves physicians time, and various dedicated research groups and companies are working on it. However, model performance alone doesn't guarantee a lower administrative burden, better communication between physician and patient, or higher quality of clinical notes. We need to clinically evaluate these digital scribes to shed light on their actual effect on clinical practice. This is especially important for the digital scribes that are already being marketed. We cannot be sure that these digital scribes do what they claim to do, are free of bias, and positively affect clinical practice when there is a total lack of scientific reporting. More research should be aimed at studying the clinical validity, usability, and clinical effectiveness of these models.
Based on 'The Digital Scribe in Clinical Practice: a Scoping Review and Research Agenda' by Marieke van Buchem, Hileen Boosman, Martijn Bauer, Ilse Kant, Simone Cammel, and Ewout Steyerberg.
- Lin SY, Shanafelt TD, Asch SM. Reimagining Clinical Documentation With Artificial Intelligence. in Mayo Clinic Proceedings 93, 563-565 (2018).
- Coiera E, Kocaballi B, Halamka J, Laranjo L. The digital scribe. Npj Digital Medicine 1, 1-5 (2018).
- Quiroz, J. C. et al. Challenges of developing a digital scribe to reduce clinical documentation burden. Npj Digital Medicine 2, 1-6 (2019).
- van Buchem MM, Boosman H, Bauer MP, Kant IMJ, Cammel SA, Steyerberg EW. The digital scribe in clinical practice: a scoping review and research agenda. Npj Digital Medicine 4, 1-8 (2021).
- Shachak A, Hadas-Dayagi M, Ziv A, Reis S. Primary care physicians’ use of an electronic medical record system: a cognitive task analysis. J Gen Intern Med 24, 341-348 (2009).
- Peters, M. E. et al. Deep contextualized word representations. in Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies 1, 2227-2237 (Association for Computational Linguistics, 2018).
- Devlin, J., Chang, M.-W., Lee, K. & Toutanova, K. BERT: pre-training of deep bidirectional transformers for language understanding. In Proc. of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL), 1, 4171-4186 (Association for Computational Linguistics, 2019).