Healthcare is emerging as a prime candidate for more AI applications. This is to support clinical operations and alleviate some of the more time-consuming administrative burdens associated with clinical care. Now his Nabla, a Parisian digital health startup co-founded by AI entrepreneur Alexandre Lebrun, is the first to use GPT-3 to build tools to help doctors do their jobs, more specifically paperwork. claims to be a company of
Nabla’s new service, called Copilot, is launching today as a digital assistant for doctors. Initially accessed as a Chrome extension, it helps you transcribe and repurpose information from video conversations. We plan to launch a face-to-face consultation tool in the coming weeks.
As doctors see patients, Copilot automatically transforms those conversations into a variety of document-based endpoints such as prescriptions, follow-up appointment letters, and visit summaries. It is based on GPT-3, the language model used to generate human text built by OpenAI, powering hundreds of applications including his ChatGPT on OpenAI itself.
Nabla was one of the first companies to experiment with GPT-3 when it was released in 2020. Nabla currently uses his GPT-3 as the basis for Copilot (as a paying customer), but Lebrun tells me his long-term goals are on the way. It builds its own large-scale language models customized for specific languages and medical and healthcare needs, powering Copilot, powering everything Nabla builds in the future, and potentially other Apply it to your application as well.
Initial versions have already gained some popularity, the startup says. Used by medical practitioners in the US and France, and about 20 digital and in-person clinics “with critical care teams.”
What large-scale, long-term use will we see for generative AI technologies, and whether they, and the large-scale language models that drive them, will bring net gain or net loss to our world. have not yet come to a conclusion. and Whether they make money in the process.
In the meantime, healthcare has been one of the big industries that people have followed with interest to see how they respond to these developments, following roughly two development corridors. if it can be used for For example, using ChatGPT to diagnose patients, as described in this article co-authored by physicians and academics from Harvard Medical School. The second is automating more repetitive functions. This is shown in this article on the future of her Lancet discharge summary.
Much of that work is still in its infancy, especially since healthcare is particularly sensitive.
“Every large language model has risks,” Lebrun said in an interview. “It’s incredibly powerful, but 5% of the time it’s completely wrong and there’s no way to control it. [literally] A 5% error rate is unacceptable. “
But in many ways, healthcare seems like a prime area for AI adoption. Globally, we face a chronic shortage of doctors. One of the reasons for this is that so many people are leaving the profession and the demands placed on doctors. We have many very specific and formal documents in place not only to see patients but also to record appointment data and to plan upcoming appointments that are required not only by rules and regulations but also by the patients themselves. , you have to take your time as an administrator. In addition to all this, there are unfortunately occasional instances of human error.
But on the other hand, many steps in healthcare have already gone digital, paving the way for patients and clinicians to use more digital tools to help them do the rest.
That’s one of the reasons Alexandre LeBrun first started Nabla and focused Copilot on helping doctors manage administrative tasks (rather than patient examinations, counseling, or other clinical tasks).
LeBrun has experience building language-based applications. In 2013, he sold his VirtuOz startup, then called “his Siri for enterprises,” to his Nuance to lead the development of digital his assistant technology for businesses. He then founded his next startup, his Wit.ai, which he eventually sold to his Facebook. He and his team then worked on the foray of his social network into Messenger chatbots. He then joined He FAIR, Facebook’s AI research center in Paris.
While these early tools for businesses to interact with their customers were primarily marketed as marketing and customer loyalty aids, Lebrun believed they could be applied to less obscure scenarios as well.
“In 2018, we could already see how much time doctors were spending updating patient records, so we introduced AI technology to [advanced] Machine learning, especially for healthcare, is meant to help with that,” said Lebrun.
Interestingly, Lebrun didn’t mention this to me, but he must have made the observation at the same time that RPA (robotic process automation) was gaining momentum in the market.
RPA has made automation in the enterprise pop into people’s minds. But assisting physicians with live consultations is a more complex problem than automating menial tasks. The relatively limited set of linguistic and subject variables used in physician-patient consultations made it an ideal scenario to be assisted by an AI-based assistant.
Lebrun discussed the idea with his boss at the time, Yann LeCun, who is still chief AI research scientist at Facebook. LeCun supported his idea, so Lebrun left and LeCun became one of his first investors in Nabla.
It took Nabla a few more years to disclose that and other funding — raising about $23 million — and the startup held off on announcing it at the same time as its first product. A Q&A “super app” that tracks a variety of health-related questions and combines that information with other data as a way to help understand what people want in remote health interactions It seemed designed primarily. And what can you build from there?
This was followed up last year with the more generalized Health Tech Stack for Patient Engagement. This is interesting because it had a small impact on Lebrun’s previous product’s core metric: engagement.
You might be somewhat skeptical of a startup looking to fix something broken in healthcare that doesn’t have a medical expert among its founders. Mr. Martin Raison.
That was also a difficult point for Lebrun. He told me he considered pausing his early ventures to go to medical school himself.
He chose not to, instead drawing on feedback and input from doctors and other clinicians and hiring them to work with startups to steer the roadmap.
“The Nabla Copilot is designed for clinicians who want to be on the cutting edge of medicine,” said Nabla Chief Medical Officer Jay Parkinson, MD, M.P.H., in a statement. “As a doctor, I know that doctors are always short on time and have better things to do than filling out forms. [electronic health record]With Nabla’s super-powerful clinical notes, doctors can look their patients in the eye during a consultation and send them a summary of their consultation so they can make sure they remember every word they say. became. Parkinson, who recently joined the startup, is an entrepreneur whose telemedicine startup Sherpaa Health was acquired by Crossover.
AI improvements are generally based on taking in more information to train, and that was the hard part of building Copilot. The company is not only HIPAA and GDPR compliant, but also opt-in to data sharing entirely without storing data on its servers. Those who agree to share their training information run the data through a “pseudonymization algorithm” built in-house. And for now, there are no plans to build clinical assistants. No diagnostic suggestions.
Lebrun said it’s easier said than done. During construction, his AI at Nabla continued to try to automatically provide diagnostics to users even when engineers didn’t ask for it and tried not to provide it, Lebrun said.
“We don’t want to step over and make a diagnosis,” he said. no to do it.
He said that in the distant future, it might become something, “another product,” but more development and foolproofing would need to be accomplished first.
“We don’t believe in medical chatbots,” he added. “We want to improve doctors’ lives by saving them time.”
