Inexpensive AI Is The Future Of Medicine

Monetizing Medical AI: Past and Future

Before the public release of ChatGPT in November 2022, most AI applications in medicine were task-specific machine-learning models, known as “narrow AI.”

These models are trained on extensive datasets of “labeled data” where the correct answer is predefined. For instance, in mammography, algorithms are trained using thousands of images labeled as cancerous or benign, enabling the model to identify key differences and estimate the likelihood of cancer in new mammograms.

In an extensive international study published in Nature, an AI system surpassed the performance of the average radiologist, reducing false positives by 5.7% and false negatives by 9.4% in the U.S. dataset.

Despite its effectiveness, narrow AI has significant limitations. A model developed for mammograms cannot read chest X-rays or detect rib fractures, even though the lung and rib cage are adjacent. Although the FDA has authorized over 1,500 AI-enabled medical devices, their integration into everyday medical practice is inconsistent, hindered by narrow applicability, high costs, reimbursement issues, and liability concerns.

Generative AI tools, in contrast, are trained on vast amounts of text and data, including medical information from textbooks and studies. This allows them to address a broad spectrum of medical queries, explaining imaging reports, assessing symptoms, supporting chronic disease management, detailing medication side effects, and helping patients comprehend their chronic conditions. Notably, these tools are accessible to both clinicians and patients.

Research from OpenAI indicates that 55% of U.S. adults already use ChatGPT to explore health-related symptoms.

Despite the growing popularity of generative AI among patients, its commercial application in medical settings remains primarily administrative. Tools like Microsoft’s Dragon Copilot are designed to convert doctor-patient interactions into clinical notes. AI solutions for coding and revenue management assist doctors with billing codes, reducing claim denials, and speeding up payments.

However, there are almost no generative AI applications developed for direct patient use.

Historically, it didn’t make economic sense to build new tools for patients when medical AI was mainly narrow AI models. Acquiring datasets, validating performance, and obtaining FDA approval are time-intensive and costly, leading to prices that most patients cannot afford. Additionally, selling narrow AI tools for specific medical problems directly to patients poses legal risks if harm occurs. By selling to hospitals and health systems, companies can place physicians as the ultimate decision-makers.

This context highlights the significance of the NYU study. Although the specialized clinical tools had medical branding and professional positioning, the more affordable, widely accessible LLMs performed equally well or better.

General-purpose LLMs, unlike medical devices crafted for specific clinical tasks, are not marketed to diagnose or treat specific conditions. Their responses depend on the input, prompt, and follow-up questions.

Consequently, general-purpose LLMs have not required FDA review and have faced limited legal exposure so far, making them an ideal playground for entrepreneurial innovation.

Patients Are the Larger Market

In the healthcare strategy course I teach at the Stanford Graduate School of Business, I emphasize to aspiring entrepreneurs that many start-ups fail not due to poor products or leadership but because they struggle to monetize their ideas effectively.

Upon graduation, some students may choose the traditional technology route. However, I foresee many exploring ways to create affordable educational or assistive tools that empower people to leverage cost-effective LLMs to enhance their health, prevent diseases, and make optimal medical decisions.

Two promising areas emerge:

1. Addressing chronic diseases. Conditions like diabetes, hypertension, obesity, heart disease, kidney disease, depression, and asthma affect 75% of Americans and are often poorly managed. In the U.S., less than half of hypertension patients have controlled blood pressure, and an even smaller fraction of diabetes patients have effectively managed blood sugar levels. Generative AI is well-equipped to empower patients with more control. It can clarify diagnoses and treatments in understandable terms, remind patients to take medications, and analyze daily clinical data to alert individuals when their blood pressure or glucose levels deviate from expectations. Currently, patients may visit their physicians months later, only to find their condition unchanged. With AI support, patients could inform their physicians sooner, enabling medication adjustments or interventions before the next scheduled visit. These tools could also reassure patients when things are progressing well and help avoid unnecessary appointments.
2. Enhancing patient access. Patients often need assistance at night, during weekends, or while waiting for primary care or specialty appointments. When doctor’s offices are closed, the emergency room might be the only option, even for issues that could be managed in a doctor’s office if it were open. Patients are already using large language models in these situations. The key opportunity is to help them use these models more effectively, with safeguards that improve information quality and guidance on when a symptom requires urgent attention. This support would be invaluable for parents concerned about a child’s fever, adults deciphering new symptoms, or families caring for an aging parent with worsening pain.

The Next Healthcare AI Unicorn

Given that affordable LLMs perform on par with costly medical AI, entrepreneurs have a chance to create the next healthcare unicorn by enabling more Americans to utilize existing GenAI models for medical care, rather than developing high-cost specialized applications.

This opportunity, however, involves several challenging decisions. The next wave of entrepreneurs must determine if their product will be an application, a set of agents, or an educational tool. They need to decide whether to sell directly to consumers, charge a subscription fee, partner with employers and insurers, or be compensated based on measurable clinical improvements. They must devise a method to test their product’s safety and reliability, given that accuracy relies on how patients input information and ask follow-up questions. Additionally, they should assess what services or facilities to invest in to provide 24/7 clinical assistance when the LLM identifies a medical issue.

One thing is evident: affordable generative AI solutions designed for patients present an untapped opportunity. The financial and medical rationale for pursuing this is straightforward. Treating patients in a doctor’s office offers one-time help. Teaching patients to effectively use generative AI enhances their ability to manage their health throughout their lives.