While studying medicine at UC San Francisco, Dereck Paul noticed a troubling trend: innovation in medical software was not keeping pace with advancements in other sectors, such as finance and aerospace. He felt strongly that patients benefit most when doctors have access to cutting-edge technology, leading him to envision a company that prioritizes the needs of patients and physicians over hospital administrators or insurance providers.
In 2021, Paul partnered with his friend Graham Ramsey, an engineer at Modern Fertility, to launch Glass Health. The company offers a digital notebook for physicians to store, organize, and share their diagnostic and treatment approaches throughout their careers. Ramsey describes it as a "personal knowledge management system" tailored for medical learning and practice.
“During the pandemic, Ramsey and I were struck by the immense pressures on our healthcare system and the increasing incidence of burnout among healthcare providers,” Paul shared. “I faced provider burnout firsthand during my medical rotations and later as an internal medicine resident at Brigham and Women’s Hospital. Our compassion for these frontline workers inspired us to build a company dedicated to maximizing technology in the practice of medicine.”
Glass Health quickly gained traction on social media platforms, especially X (formerly Twitter), among physicians, nurses, and medical trainees. This buzz helped secure the company’s first funding round—a $1.5 million pre-seed investment led by Breyer Capital in 2022. Later, Glass Health was accepted into Y Combinator’s Winter 2023 batch. However, earlier this year, Paul and Ramsey made a strategic pivot towards generative AI, aligning with a growing trend in healthcare technology.
Today, Glass Health provides an AI tool powered by a large language model (LLM), similar to the technology behind OpenAI’s ChatGPT, which assists doctors in generating potential diagnoses and “evidence-based” treatment options. Physicians can input patient summaries, such as “71-year-old male with a history of myocardial infarction presents with subacute progressive dyspnea on exertion,” and the AI returns likely prognoses and clinical plans.
“Clinicians input a patient summary, detailing demographics, medical history, symptoms, and relevant lab and radiology findings—essentially the information they would share with a fellow clinician,” Paul explained. “Glass Health analyzes this summary and suggests five to ten possible diagnoses that the clinician might want to explore further.”
Additionally, Glass Health can draft a case assessment paragraph for clinicians, including relevant diagnostic study recommendations. These drafts can be edited and utilized in clinical documentation or shared with the broader Glass Health community.
On the surface, Glass Health’s tool appears incredibly beneficial. Yet, even the most advanced LLMs have faced criticism for their limitations in providing reliable health advice. Babylon Health, an AI company supported by the U.K.’s National Health Service, has repeatedly come under fire for claims that its disease-diagnosing technology outperforms human doctors.
In a notable misstep, the National Eating Disorders Association (NEDA) partnered with AI company Cass to create a chatbot aimed at supporting individuals with eating disorders. Unfortunately, an upgrade led the chatbot to promote harmful “diet culture” suggestions, prompting NEDA to discontinue the tool.
Moreover, Health News recently asked a medical professional to assess ChatGPT’s health advice across various subjects. The expert concluded that the chatbot missed recent studies, made misleading claims (including that “wine might prevent cancer”), and even plagiarized content from health news outlets.
In a more favorable review published in Stat News, researchers found that ChatGPT provided correct diagnoses in the top three choices for 39 out of 45 clinical vignettes. However, they cautioned that the vignettes resembled those typically used to test medical students and might not accurately capture how patients—particularly those with language barriers—express their symptoms in real-world scenarios.
While bias in LLMs hasn’t been widely studied, it raises concerns about potential inaccuracies in patient diagnoses. Medical LLMs, including Glass Health’s, often learn from health records that reflect only what medical professionals document—generally about patients who can afford care. This can lead to gaps in knowledge, as doctors may inadvertently inject their own biases regarding race, gender, or socioeconomic status into these records, skewing the AI’s outputs.
Paul seems acutely aware of the scrutiny surrounding generative AI in healthcare and confidently positions Glass Health’s AI as a superior alternative to existing solutions.
“Glass connects LLMs with clinical guidelines created and peer-reviewed by our team of academic physicians,” he stated. “Our team comprises members from top medical centers across the country, who work part-time for Glass Health, akin to their roles on medical journals, refining our guidelines and enhancing our AI’s application of them. We emphasize that our clinician users must closely supervise all outputs from our AI, treating it as a supportive assistant rather than a replacement for their clinical judgment.”
Throughout our discussion, Paul emphasized that Glass Health’s AI, while geared toward suggesting diagnostic possibilities, should not be interpreted as definitive solutions. This caution may stem from a desire to avoid increased legal scrutiny or potential FDA regulation.
Paul is not alone in exercising caution. Google is carefully marketing its Med-PaLM 2, a medical-focused language model, ensuring that it doesn’t imply that the technology can replace healthcare professionals’ expertise in clinical environments. Similarly, Hippocratic is developing an LLM tailored for healthcare applications, though not specifically for diagnosis.
Despite the challenges, Paul argues that Glass Health's approach enables “fine control” over its AI outputs, ensuring alignment with current medical knowledge and best practices. This strategy involves collecting user data to enhance Glass’s underlying LLMs, a move that may concern some patients.
Paul assures that users can request to delete all stored data at any time.
“Our LLM application draws on physician-validated clinical guidelines at the moment it generates outputs,” he noted. “Glass is distinct from LLM applications like ChatGPT, which solely depend on pre-training data, making it more susceptible to inaccuracies or outdated medical information. We exercise stringent oversight over the guidelines and information that inform our AI, striving to mitigate bias and support health equity.”
As the landscape evolves, it will be interesting to see how these claims hold up.
Currently, Glass Health is attracting early adopters, with over 59,000 users signed up and a direct-to-clinician monthly subscription model in place. This year, Glass will pilot an enterprise offering that integrates with electronic health records and complies with HIPAA; 15 unnamed health systems and companies are already on the waitlist, according to Paul.
“Hospitals and health systems can provide Glass Health to their doctors, empowering clinicians with AI-driven clinical decision support, including diagnosis recommendations, diagnostic studies, and treatment options,” Paul explained. “We also have the capability to tailor Glass’s AI outputs to align with specific clinical guidelines or care practices within a health system.”
With $6.5 million in total funding, Glass Health aims to invest in the development, review, and updating of its clinical guidelines, refine its AI technology, and engage in research and development. Paul asserts that the company has four years of operational runway ahead.
Ranking
