Removing technology barriers between clinicians and patients

In this opinion piece, Cat Miller (Chief Technology Officer, Flatiron Health) explores the future of healthcare technology and the ways in which we can seamlessly integrate AI into healthcare. For AI applications to be successful within healthcare, Cat stresses the importance of AI being an extension of the clinician, rather than a replacement.

Removing technology barriers between clinicians and patients

Sometimes the best technology is the one that is invisible. While gathering user feedback for OncoEMR, Flatiron’s cloud-based electronic health record (EHR) product, we gave a room full of experienced clinicians a blank piece of paper each and asked them to draw the future of medicine - a challenging prompt to answer with 10 minutes and a pencil. All but one of them drew the same thing in the end: themselves and the patient - nothing else.

Clinicians want to keep computers from intruding, yet as technologists we constantly find ourselves building more novel tools and applications for them. Physicians use EHRs to keep track of patient data, an email client or chat tool to communicate, a program like Medscape to quickly access information about an unusual case or complication, a calculator for dosage and another to quickly estimate the risk of complications like heart disease. Some of these tools have proven their value for practitioners, driving adoption; others are required by regulators or organizations. Because of gaps in integration and data access, we have created a reality where a physician is armed with an array of tools, each requiring its own investment of time and resources.

Here is where the problem lies: for a clinician that already spends only a quarter of their day face to face with patients , any additional time in front of a screen comes at a cost to both the physician and the patient. Even before the onset of the COVID-19 pandemic, physician burnout was severely on the rise, approaching 50% in the United States . The time physicians spend on administrative activities has increased, while the time they spend with patients stays flat. Unfortunately, technology has added burden to clinicians even as it seeks to do the opposite.

In order to truly support clinicians and to achieve the vision of patient and physician alone in a room, we need to address specific areas within the health technology ecosystem. It is not the task of any single organization or group to solve these challenges in isolation, but a direction of intentional investment we should take on together.

Enabling seamless information flow

One of the key activities that takes physicians’ undivided attention away from patients is entering information in the EHR. Currently the strongest attempts to reduce the burden of data entry have been with the use of medical scribes: humans who sit in the exam room and act as the record keepers for the clinician. Depending on setting, scribes increase physician productivity by 10% and face-to-face time with the patient by over 50% . AI-aided transcription has the potential to mirror those benefits without the cost and potential discomfort of having another person in the room. That revolution is already underway: in March 2023, Microsoft announced that an early commercial use of GPT-4 will be its AI scribe platform Dragon Ambient Intelligence.

Currently, the problems with voice-assistance in the medical realm are similar to its issues in general use, but the stakes are higher. When used to record notes, errors in transcription can be difficult to identify and fix smoothly and errors that persist create risk to patients. Moreover, medical records do not consist purely of unstructured dictation-style notes: data that is entered via structured forms is essential in powering the capabilities of an EHR system. To take over scribe duties, voice transcription needs to navigate complex interfaces, not just write notes.

Easing physicians’ burdens is not just about information flow into systems, it is also about ensuring that information flows freely between systems. The need for interoperability for medical records has been clear for over thirty years: the HL7 clinical data standard, which establishes a common way for healthcare systems to communicate with each other, has been in common use since the early 1990s . More recently, the 21st Century Cures Act and subsequent Interoperability and Patient Access final rule are in the process of making Fast Healthcare Interoperability Resources (FHIR), a protocol for exchanging electronic healthcare data between systems, a standard implemented by effectively all EHRs.

This push towards making data available and integrated is critical to shifting physicians’ time back toward patients. Currently, tools ranging from phone apps to clinical trial electronic data capture (EDC) systems expect that a person will manually copy over data that has been entered in another place - a poor use of an expert’s time that also increases the likelihood of error. A dose calculator should already know the height and weight of a patient in the system; a clinical decision support tool should already know any relevant lab and prior treatment information. FHIR expansion is making this increasingly possible, but we need to view that expansion as absolutely essential. A given data point should be entered by a physician, at most, once.

Prioritizing acceleration

A well-suited but possibly underinvested area in healthcare is helping clinicians do tasks more efficiently. Much of the recent AI buzz is around helping people perform existing tasks better: in the medical space, that might include reducing errors in diagnosis and navigating complex treatment options. Beyond entering and retrieving data from an EHR, physicians are often performing repetitive tasks like ordering labs or scans; saving a handful of clicks on a repeated workflow have meaningful implications for how much time and attention they ultimately have for patients.

Machine learning opens up far more potential to reduce toil than classic rule-based systems. Instead of a standard menu of options, an algorithm could populate a default lab order on the basis of clinical characteristics, treatment and what has been ordered for that patient in the past. There is tremendous opportunity for time-saving by taking what the physician already knows they want to do and shifting the labor of it onto technology.

Extending the expertise of the experts

The rate of increase in medical knowledge is outpacing Moore’s law - on track to double faster than every 2 years  - and keeping up with the literature is rapidly becoming an impossible task. The volume of regulatory documents and guidances to be considered is also increasing apace. This is a clear place for expert systems to take up the burden, and indeed clinical decision support (CDS) is one of the hottest areas in health technology today. While it is self-evident that such systems need to be accurate, that is not the only requirement for success.

In the future, a clinician will need to be able to tap into that extended knowledge base in a way that does not come between them and the patient. This is a substantial UX challenge, without a single, clear solution.

The voice systems discussed earlier may be part of the answer, but when it comes to spoken interaction, most of our experience revolves around Siri-like systems that respond to isolated commands and queries. The vast majority of these interactions - like asking about the weather or the capital of Peru - involve a single response . For a voice CDS system, the UX will need to support far more interaction. In order to truly support a diagnosis or treatment, the system will need to get information from the clinician and generate potential pathways.

Better visualizations may also be part of the answer. A page full of tiny words is hard to see and generally unapproachable, but perhaps an elegant Sankey diagram or novel information display would allow patient and physician to explore and discuss options together. The display would become the physician’s whiteboard, helping them better explain rather than being an attention void that pulls them away from the patient.

Building trust

To reap the benefits of fluent and efficient AI in medicine, the users must trust it. Not just physicians, but patients too, must have enough trust to let it help, while maintaining sufficient credulity to question when it errs. This trust may evolve naturally over time: In the 1980s spell check was an optional tool that a user could choose to run on a document; now autocorrect is an expectation of text entry. It may require transparency on the part of our models, which could fundamentally limit the types of solutions we can bring to bear. Or it might require an evolution in the relationship between humans and AI.

The scenario of a doctor sitting alone with a patient while an AI listens and records the information, providing support but never intruding does not have to be as far in the future as it seems. Consumer companies such as Amazon have funneled billions of dollars into the development of core capabilities like voice recognition, which could be the foundation of leaps forward in the health technology ecosystem. Harnessing those capabilities requires us to collectively focus on the current barriers technology erects around physicians. A future of healthcare centered around the physician–patient relationship is possible.

‍Disclaimers:

The opinions expressed in this feature are those of the interviewee and do not necessarily reflect the views of Future Medicine AI Hub or Future Science Group.