The Paris-based startup, Nabla, recently disclosed a successful $24 million Series B funding round led by Cathay Innovation, joined by ZEBOX Ventures — CMA CGM’s corporate VC fund. This funding surge arrives shortly after Nabla sealed a significant partnership with Permanente Medical Group, a branch of Kaiser Permanente, a prominent healthcare entity in the U.S.
Based on insider information, Nabla’s valuation has now climbed to $180 million post this funding infusion. The company might even raise supplementary funds from U.S. investors as part of this ongoing round.
Nabla concentrates on developing an AI co-pilot designed for doctors and medical personnel. This technology essentially operates as an unobtrusive work assistant, situated in the room’s background, diligently recording notes and compiling medical reports.
Initially established by Alexandre Lebrun, Delphine Groll, and Martin Raison, Nabla is steered by CEO Alexandre Lebrun, previously the CEO of Wit.ai, an AI assistant startup acquired by Facebook. He later assumed the role of head of engineering at Facebook’s AI research lab FAIR.
In a recent demonstration, I witnessed Nabla in action with a genuine physician and a simulated patient portraying symptoms of back pain. When the consultation commenced, the physician activated Nabla’s interface, enabling the AI to seamlessly manage the note-taking, freeing the doctor from computer-related tasks.
Alongside the physical examination, a consultation involves an extensive discussion, probing into the reasons for your visit and your medical background. Towards the end, there may be recommendations and prescriptions provided.
Nabla employs speech-to-text technology to transform these conversations into written transcripts, adapting to both face-to-face appointments and telehealth sessions.
Once the patient leaves, the doctor concludes the process by stopping the transcription. Nabla, armed with a vast language model enriched with medical insights and health-related discussions, extracts crucial details from the consultation – including vital medical information, drug names, and conditions.
In just a minute or two, Nabla swiftly generates a comprehensive medical report summarizing the consultation, prescriptions, and follow-up appointment directives.
These reports are tailored to suit the preferences of each doctor, allowing customization of formats. Physicians can request concise or detailed notes and even opt for formats like the widely used Subjective, Objective, Assessment, and Plan (SOAP) notes, prevalent in the U.S.
In a recent demo I attended, Nabla impressed with its effectiveness despite being operated from a laptop a few meters away in a crowded room. It accurately transcribed the conversation and provided a useful report.
Nabla’s Copilot, as the name implies, doesn’t aim to eliminate human involvement in medical processes. Physicians retain final control, allowing them to edit reports before filing them into their electronic health record (EHR) systems.
The company’s goal is to streamline administrative tasks, granting doctors more time to dedicate to patient care.
“What we’re foreseeing in the near future is that we don’t aim to replace doctors. We’ve seen companies like Babylon in the U.K. spending a billion dollars trying to introduce chatbots and immediate automation to remove doctors from the equation. With Nabla Copilot, we’ve made a deliberate decision that doctors are the pilots, and we’re here to work alongside them,” explained Alexandre Lebrun, co-founder and CEO of Nabla.
“An analogy could be drawn to the automation journey in autonomous vehicles. Currently, we’re at level two. Soon, we’ll move to level three, offering clinical assurance support. Level four will be clinical decision support, but with FDA approval, as some decisions cannot be easily explained,” he elaborated.
“Looking ahead, there could potentially be a level five in autonomous healthcare, removing physicians from direct involvement. However, I remain cautious about this,” Lebrun cautioned. “While in some situations or markets, such as regions lacking healthcare access, this might be relevant.” In the long run, he envisions the diagnostic process becoming a ‘pattern matching problem’ solvable with AI. Physicians would then focus on empathy, surgical procedures, and critical decisions.
Although Nabla is headquartered in France, the majority of its customer base is in the U.S., primarily following its integration with Permanente Medical Group. Nabla isn’t a mere work in progress; it’s actively employed by thousands of doctors on a daily basis.
Nabla’s Approach to Privacy
Nabla currently operates as a web app and a Google Chrome extension, mindful of the sensitive data it handles. The company refrains from storing audio or medical notes on its servers unless both the doctor and the patient explicitly consent.
Emphasizing data processing over storage, Nabla follows a stringent protocol. Post-consultation, the audio file is promptly discarded, while the transcript is securely stored in the Electronic Health Records (EHR) system that physicians use for their patient documentation.
In more technical detail, Nabla employs real-time transcription using a meticulously tuned speech-to-text API. This involves a blend of a Microsoft Azure off-the-shelf speech-to-text API and their customized speech-to-text model, refined from the open-source Whisper model.
“When you use a standard speech-to-text algorithm, its efficacy with medical data might vary. However, we’ve fine-tuned ours. As you might notice, the text starts light and gradually turns darker. The darker shade indicates that we’ve verified and corrected it using medication names or medical conditions with our proprietary model,” shared Nabla ML engineer Grégoire Retourné during the demonstration I witnessed.
The transcript undergoes an initial pseudonymization process, replacing personally identifiable information with variables. These pseudonymized transcripts are then fed into a large language model (LLM). Historically, Nabla primarily utilized GPT-3 and subsequently GPT-4 as its primary large language models. Being an enterprise customer, Nabla has specified to OpenAI that it cannot retain its data or use it to train the large language model based on these consultations.
However, Nabla has also been experimenting with a specialized version of Llama 2. “We foresee a future where we increasingly utilize more specialized models rather than general ones,” stated Lebrun.
After the LLM processes the transcript, Nabla reverses the pseudonymization to obtain the output. The resulting notes are accessible to doctors and stored in the local web browser storage file on their computer. These notes can then be exported to Electronic Health Records (EHRs).
Nonetheless, doctors can grant their approval and seek patient consent to share medical notes with Nabla. This facilitates the correction of transcription errors. Given Nabla’s projection to handle over 3 million consultations annually across three languages, the platform is poised to improve rapidly, leveraging real-world data.
