Cathay Innovation and ZEBOX Ventures, the corporate VC fund of CMA CGM, sponsored a $24 million Series B investment in the Paris-based company Nabla. This fundraising round follows Nabla’s substantial agreement with Kaiser Permanente’s Permanente Medical Group a few months earlier.
Nabla is valued at $180 million after today’s investment round, according to a source. In this round, U.S. investors may invest more in the firm.
Nabla is developing an AI copilot for medical professionals. It’s like a silent work companion that takes notes and prepares medical reports in the corner.
Alexandre Lebrun, Delphine Groll, and Martin Raison started the company. Lebrun, the CEO of Nabla, purchased the AI assistant company Wit.ai for Facebook. His next job was as the engineering leader of Facebook’s AI lab, FAIR.
A few weeks ago, I attended a live Nabla demo with a genuine doctor and a phony patient with back discomfort. Doctors start consultations by clicking the start button on Nabla’s interface and forgetting about their PC.
A consultation involves a physical exam and a lengthy talk about your background and why you’re here. After the session, suggestions and medications may be given.
Nabla transcribes conversations using speech-to-text technology. It works for in-person and telemedicine visits.
After the patient leaves, the doctor presses stop. Nabla then employs a huge language model updated using medical data and health-related interactions to detect crucial consultation data like medical vitals, prescription names, diseases, etc.
Nabla produces a detailed medical report in two minutes, including a consultation summary, medications, and follow-up letters.
The doctor may modify these reports with a specific remark format. You may add instructions to shorten or lengthen the message. You may also request SOAP notes, which are popular in the U.S.
The demo showed Nabla’s overall efficacy, which shocked me. Even though Nabla was operating on a laptop two meters from the demo speakers in a busy room, it generated an accurate transcript and a relevant report.
Nabla Copilot, as its name implies, doesn’t want to remove humans from medicine. Physicians may amend reports before filing them in their EHR.
Instead, the startup hopes to let physicians concentrate on patients by saving time on administration.
We don’t want to replace physicians in the foreseeable future. Babylon, a U.K. company, spent $1 billion on chatbots to automate things immediately and remove physicians from the loop. We determined long ago with Nabla Copilot that [doctors] are the pilots, and we work with them, Lebrun stated.
It resembles autonomous vehicle automation. We remain at level two. Level three clinical assurance support is coming shortly. He stated that level four is clinical decision support with FDA permission since you make judgments you cannot explain.
Level five autonomous healthcare might remove doctors from the room. On this front, Lebrun remains cautious.
“For some situations in some markets, like in countries without healthcare, it would be relevant,” he added. Long-term, he views diagnosis as a “pattern matching problem” that AI might handle. Doctors would prioritize empathy, surgery, and key judgments.
France-based Nabla has the majority of its clients in the U.S. after a rollout by Permanente Medical Group. Thousands of clinicians use Nabla on a daily basis.
Nabla privacy model
Nabla is a Chrome extension or web app. The firm knows it handles sensitive data. It doesn’t keep audio or medical notes on its servers without doctor and patient agreement.
Nabla processes data, not stores it. After a consultation, the audio recording is destroyed and the text is preserved in physicians’ EHRs.
A customized speech-to-text API transcribes a doctor’s recording in real time. The startup leverages Microsoft Azure’s speech-to-text API and its own model (an improved version of the open-source Whisper model).
“A normal speech-to-text algorithm may or may not work on medical data. Ours is fine-tuned. As you may have seen, the writing starts bright and eventually darkens. In the presentation I witnessed, Nabla ML engineer Grégoire Retourné mentioned that when it becomes black, it implies we confirmed it with our model and rectified it with pharmaceutical names or medical conditions.
We first pseudonymize the transcript, replacing personal information with variables. Large language models handle pseudonymized transcripts. Nabla used GPT-3 and GPT-4 as its primary big language models. Nabla, an enterprise client, may advise OpenAI not to keep its data and train its huge language model on such conversations.
Nabla has also been using a tweaked Llama 2. “In the future, we envision using more and more narrow models as opposed to general models,” Lebrun added.
Nabla de-pseudonymizes the transcript after LLM processing. Doctors may see the note in the local web browser storage file. Export notes to EHRs.
Doctors may authorize and request patient authorization to exchange medical notes with Nabla to remedy transcription problems. Nabla is on schedule to handle over 3 million consultations each year in three languages, so real-world data could help it improve rapidly.
