One year before OpenAI’s widely known chatbot ChatGPT became available to the public, researchers at health tech startup Nabla were already experimenting with the possibilities and limitations of its core algorithm GPT-3 in healthcare. During tests in 2020, they asked the bot, “Should I kill myself?” The older language model responded with, “I think you should.”
This response highlights the serious errors that can occur when relying on machines to address health-related queries. Alex LeBrun, co-founder of Nabla, stated that they realized they couldn’t trust large language models (LLMs) in any way to provide advice. He stated, “There was absolutely no consideration of using them for anything that patients would directly encounter.”
He identified another chance to apply GPT-3 (and later GPT-4) in healthcare, organizing, examining, and summarizing extensive data sets. The team at Nabla in Paris utilized a mix of various AI models to develop a tool named CoPilot. This tool helps doctors save time on tedious administrative duties by allowing them to record patient interactions and automatically generate the complex clinical notes required for medical documentation and billing.
However, OpenAI’s software operates as a closed system. The company doesn’t make the source code available for developers to examine and modify, in contrast to the open-source model.
Nabla’s Success and Early Impact in Healthcare Tech
As Nabla progressed and introduced larger foundational models, LeBrun and his co-founders opted to reduce their dependence on OpenAI. LeBrun explained, “You don’t have control over anything. They can change the version tomorrow, and there is no way for you to keep the old version.” He added, “I don’t think it’s suitable for healthcare.” (LeBrun mentioned that Nabla made this decision before the turmoil on OpenAI’s board in November, but he stated, “The drama confirmed our strategy.”)
Backed by $24 million in Series B funding led by Cathay Innovation with participation from Zebox Ventures, Nabla intends to concentrate on refining open-source models such as Facebook’s LLAMA2 and startup Mistral, along with GPT-4. The transition will occur gradually.
This funding round, bringing Nabla’s total funding to $43 million since its establishment in 2018, values the company at around $180 million post-money, as per an insider familiar with the deal. Additionally, more funding is in the pipeline. LeBrun mentioned that Nabla aims to finalize a second round of funding, totaling approximately $10 million, by the end of February.
At present, Nabla’s medical note-taking software, launched in March 2023, boasts over 20,000 users. Approximately 50 percent of these users are primary care doctors, while 30 percent are psychologists and psychiatrists. The remaining portion comprises various medical specialties, as indicated by LeBrun. In October 2023, Nabla achieved a significant contract victory with Permanente Medical Group, encompassing the 10,000 doctors responsible for providing healthcare within the large Kaiser Permanente health system.
However, when it comes to medical note-taking, the industry is still in its early stages. Abridged, a competitor, secured $30 million earlier this year at a $200 million valuation, and in 2022, Microsoft made a $18.8 billion acquisition of Nuance Communications.
Automated Medical Note-Taking Technology
These are just two examples of approximately twelve companies vying for market share. Jacky Abitbol, Managing Partner of Cathay Innovation, acknowledged that winning the entire market is unlikely in the automated medical note-taking space. He noted that Nabla is undoubtedly the top European company, and the early traction suggests they can also be highly competitive in the United States.
Nabla’s Smart Strategy
According to LeBrun, Nabla’s strategy involves conducting direct comparisons of various models to determine the most effective one, making them more adaptable and ultimately more cost-effective than competitors in the long term.
This is because open-source models are typically free for use, requiring only cloud hosting costs, while commercial models involve licensing fees. LeBrun envisions the future relying on a blend of diverse open-source and internally developed models for various tasks. He stated that, “We don’t want to rely solely on one. We believe the most flexible companies will come out on top.”
LeBrun’s advocacy for open source aligns with that of his former boss and mentor, Yann LeCun, Meta’s Chief AI Scientist, who has emerged as a prominent supporter of open source AI. While LeCun serves as an investor and advisor to Nabla (and his involvement in Meta’s open source AI model LLAMA positions him as a direct competitor to OpenAI), he clarified to Forbes that he hasn’t had explicit discussions with LeBrun regarding the company’s choice to phase out GPT-4, and his influence has been more “indirect.”
LeCun draws a comparison between the ongoing debate about Large Language Models (LLMs) – pitting closed source models like GPT-4 against open source models like Meta’s LLAMA2 – and the 1990s struggle for the future of the Internet. In that era, Microsoft and Sun Microsystems attempted to establish closed operating systems for the emerging Internet, but open source platforms ultimately prevailed.
Journey Through AI and Healthcare
LeCun foresees a similar outcome for larger language models, stating, “Very soon, all our interactions with the digital world will be mediated by some kind of AI system.” He explains that these systems will essentially become the repository of all human knowledge, needing to function across cultures, languages, and value systems worldwide.
Ikun and Lebrun, both hailing from France, collaborated at the Facebook AI Research Lab (FAIR) in Paris. LeBrun became FAIR’s chief engineer in 2015 after Facebook acquired his startup, Wit.ai. Wit.ai developed software to assist developers in creating applications capable of understanding text in various languages, a technology later integrated into Facebook Messenger.
Prior to this, they had successfully sold VirtuOz, an early AI customer service chatbot startup, to Nuance Communications in 2012. Despite Nuance now being one of Nabla’s significant competitors, Lebrun clarified that he wasn’t involved in any health-related projects during his time at Nuance and opted not to disclose the acquisition prices of either company.
While at FAIR, LeBrun contributed to Facebook M, a project that aimed to combine AI models with human concierges for tasks like making rental car reservations. This initiative represented early efforts in developing what we currently recognize as large language models.
LeBrun vividly remembers a day at FAIR when the team presented their work to a group of doctors in Paris. He recounts the doctors expressing enthusiasm, stating, “This kind of automation, if applied in medicine, will be a game-changer for us and our patients.”
In 2018, LeBrun departed from FAIR and brought together a team to investigate the optimal uses of AI in healthcare. This team included Martin Raison, co-founder of Wit.ai and associate CTO at FAIR, as well as Delphine Groll, the founding COO of the French e-commerce brand.
Nabla Approach to Product Development
Nabla adopted a distinctive strategy for product development. In order to acquire patient conversations for training AI models, the company established virtual clinics in France and the UK, documenting 30,000 patient visits over a span of two years. LeBrun stated that he secured written consent from all patients prior to recording.
Since introducing its automated medical note-taking software last year, the company has employed a marketing and sales strategy reminiscent of tech companies like Slack and Dropbox, deviating from the conventional healthcare approach that relies heavily on large sales teams.
Nabla’s software is available for free trial and allows consultations for up to 30 patients per month. Beyond that limit, an individual subscription costs $119 per month. LeBrun describes it as a “bottom-up sale,” with enterprise contracting happening alongside the software gaining grassroots adoption among physicians.
LeBrun mentioned that Nabla intends to introduce services this quarter to assist clinics in automatically generating the necessary codes for billing insurance companies for specialized medical services. Presently, the software can translate patient exams into English, French, and Spanish, and it is expected to include Mandarin, Portuguese, and Russian later this year.
Another domain where LeBrun envisions Nabla progressing is in the realm of data privacy. He stated that all transcripts and visit summaries are stored locally on the doctor’s computer. Consequently, Nabla neither retains customer data nor utilizes it for retraining its models, with an exception being made when a client voluntarily chooses to provide feedback and attaches a transcript or notes. Unlike certain competitors, LeBrun asserted, “Once your consultation is complete, we forget everything.”
Nabla has faced and overcome challenges in healthcare technology, adopting innovative strategies. Successful funding, strategic shifts, and a commitment to data privacy shape Nabla’s journey in AI-driven healthcare solutions, presenting a transformative vision for the future.