Spend a day in the shoes of a doctor and you’d see that, in addition to diagnosing and treating disease, physicians spend an inordinate amount of time on administrative work. One study estimated that 24% of a physician’s working hours are spent on such administrative duties. That may be an underestimate.
As a physician who cared for hospitalized patients, it would not be uncommon for me to spend just 5 minutes with a patient and another 40 digging through the chart, answering questions, entering orders, documenting and billing. I often felt more like a clerk than a doctor.
New research gives us an early glimpse of how effective artificial intelligence (AI) could be in handling these sorts of administrative tasks. In the study, researchers compared the responses of human physicians to 195 separate patient questions to those of ChatGPT. Then, a panel of human experts reviewed the responses on a blinded basis to determine who gave the best medical responses and who had the better “bedside manner”.
The results? The panelists preferred AI responses in 78.6% of the 585 evaluations, finding them often superior to physician responses in both quality and empathy.
Investors and entrepreneurs have understood this promise for quite some time. We need to level the playing field so that smaller companies that are driving AI innovation in healthcare can compete.
In 2022, healthcare AI companies received more than $17.8 billion in funding from venture capitalists. Many of these companies are focused on the administrative aspects of healthcare.
For example, one company, SukiAI, simply listens to a recording of an encounter between a patient and provider and automatically creates written documentation, promising to save doctors tens of hours per week, enabling them to focus on higher value tasks.
Another company, Banjo Health, is tackling the problem of prior authorizations, where a physician needs approval from an insurance company before pursuing a certain test or intervention. Its software automatically pulls data from clinical notes to drive more efficient and accurate decision-making.
Of course, there’s a big difference between what these tools can do in theory and what happens in real world settings. In fact, the experience of many medical groups with AI is decidedly mixed. For example, earlier this year, buzzy startup Olive AI, laid off more than a third of its staff after the company’s promised cost savings failed to materialize.
Olive was focused on automating high-volume, repetitive tasks. Among the challenges facing Olive is that it needed data from other technology applications and, in the absence of more robust mechanisms, relied on “screen-scraping” to do so.
Unfortunately, screen scraping is notoriously fragile as small changes in say the placement of a button can cause the integration to break and lead to errors.
The challenges facing Olive are emblematic of many AI companies who operate in a world where information is owned by a few health tech giants. For these companies to be effective they need large volumes of data, often in real-time. AI companies extract these data and use it to make predictions. Those predictions are then fed back to users in the form of recommendations—much like spell-check.
But what if these AI companies can’t effectively access data or integrate recommendations into the day-to-day workflows of busy clinicians? Well, they won’t work reliably or be effective.
Today both data and workflow largely live within the confines of what is known as the electronic medical record. According to KLAS, just 3 companies control nearly 69 percent of the US hospital market share for electronic medical records. And, despite significant advances over the years, getting access to real-time data and real estate within these systems remains an obstacle.
Even if the electronic medical record formally offers a pathway toward integration, as most are now required to do, the dirty little secret in healthcare IT is that it often doesn’t hold up to the data-intensive needs of today’s applications.
As the Co-Founder and CEO of a health tech company that leveraged electronic medical record data, integration challenges were our most common cause for errors and the number one reason for customer dissatisfaction. In our case, we focused on recommending the best time to schedule a patient’s appointment. But those recommendations required very specific information about a provider’s availability or key details about the patient (such as clinical information or insurance), information that wasn’t reliably available from all the systems we worked with.
This, unfortunately, is often exactly what the electronic medical records makers want. These companies are announcing their own AI solutions, often in partnership with well-known companies. They might offer a “check-the-box” solution that many customers will adopt, regardless of how effective they are. The result will be that doctors won’t have access to technologies that deliver real results, AI will fail to achieve its promise, and medicine will not progress.
The only way to fully realize the potential of AI is to make sure these early-stage companies have a field to play on. We need a thriving marketplace of AI solutions so that we can make sure the best ones are in the hands of users.
This begins by lowering the cost of connecting to electronic medical records and enhancing how data flows. The Cures Act prohibited information blocking by electronic medical records. The Office of the National Coordinator should continue to leverage its authority to make it easy for systems to connect in ways that are real-time, integrated into workflows, and work reliably. If not, healthcare providers won’t get access to the tools they need to address the administrative challenges that are choking medicine.
To be sure, simply integrating into electronic medical records won’t ensure we get the best AI in the hands of doctors. We need to make sure we actually study the tools we implement and share those results, not unlike what we do for clinical interventions. And there are real concerns about data security with so many companies touching a patient’s data. These need to be addressed and can be.
But better access to data within electronic medical records is a first step. Without it, we can’t have the kind of competition that will help us crown the best companies in healthcare AI.