Gottlieb says artificial intelligence may take over the role of doctors later

Dr. Scott Gottlieb is a contributor to CNBC and has worked with Pfizer, genetic testing startup Tempus, health technology company Aetion Inc. and serves on the boards of biotechnology company Illumina. He is also a partner in the venture capital firm New Enterprise Associates.

Researchers at Harvard have presented a study that demonstrates an achievement that would challenge any medical student. ChatGPT, a large language model, has passed the US Medical Licensing Examination, beating the roughly 10 percent of medical students who fail the test each year.

The inevitable question is not so much when these AI devices can replace doctors. For some positions, this medical future is sooner than we think.

To understand the potential of these tools to revolutionize the practice of medicine, one must begin with a taxonomy of the various technologies and how they are used in healthcare.

AI tools applied to healthcare can generally be divided into two main categories. The first is machine learning, which uses algorithms that allow computers to learn patterns from data and make predictions. These algorithms can be trained on different types of data, including images.

The second category covers natural language processing, which is designed to understand and generate human language. These tools allow a computer to convert human language and unstructured text into machine-readable, organized information. They learn from people’s multiple trial and error decisions and mimic a person’s responses.

The main difference between the two approaches is in their functionality. While machine learning models can be trained to perform specific tasks, large language models can understand and generate text, making them particularly useful for replicating interactions with providers.

The use of these technologies in medicine generally follows one of four different paths. The first involves large language models applied to administrative functions such as medical claims processing or the creation and analysis of medical records. Amazon’s HealthScribe is a programmable interface that transcribes conversations between doctors and patients and extracts medical data, allowing providers to create structured records of encounters.

The second bucket involves the use of supervised machine learning to enhance the interpretation of clinical data. Specialties such as radiology, pathology and cardiology are already using AI to analyze images, read MRIs, evaluate pathology slides or interpret electrocardiograms. In fact, up to 30% of radiology practices have already adopted AI tools. That is, there are other specialties. Google Brain AI has developed software that analyzes images from the back of the eye to diagnose diabetic macular edema and diabetic retinopathy, two common causes of blindness.

Because these tools offer diagnoses and can directly affect patient care, the FDA often classifies them as medical devices and subjects them to regulations to verify their accuracy. However, training these tools on closed datasets where the findings in the data or imaging have been rigorously validated gives the FDA increased confidence when evaluating the integrity of these devices.

A third broad category consists of artificial intelligence tools based on large language models that extract clinical information from patient-specific data and interpret it to deliver it to providers for diagnosis or treatment. Commonly known as clinical decision support software, it evokes the image of an intelligent assistant designed to assist, not replace, a physician’s judgment. IBM’s Watson for Oncology uses artificial intelligence to help oncologists make more informed decisions about cancer treatment, while Google Health is developing DeepMind Health to create similar tools.

As long as a physician is involved and makes an independent decision, the FDA does not always regulate such devices. The FDA emphasizes that it is intended to make a definitive clinical decision, rather than providing information to assist physicians in their evaluation.

The fourth and final grouping represents the holy grail for artificial intelligence: large language models, operating fully automated, analyze a patient’s entire medical record to diagnose conditions and prescribe treatments directly to the patient without a physician.

Currently, there are only a few clinical language models, and even the largest ones have relatively few parameters. However, the power of the models and the data sets available for their training may not be the most significant obstacle for these fully autonomous systems. The biggest hurdle may be creating a suitable regulatory pathway. Regulators are hesitant, worried that the models are prone to errors and that the clinical data sets they are trained on contain incorrect decisions, causing the AI ​​models to repeat these medical errors.

Addressing the barriers to bringing these fully autonomous systems into patient care holds significant promise not only for improving outcomes, but also for addressing financial challenges.

Health care is often cited as an area burdened by Baumol’s cost disease theory, an economic theory developed by economist William J. Baumol that explains that costs in labor-intensive industries tend to rise faster than in other sectors. In fields such as medicine, technological inputs are unlikely to provide large reimbursements for labor costs, as each patient encounter still requires intervention by a provider. In sectors such as medicine, labor itself is a product.

To offset these challenges, medicine has consolidated more non-physician providers to reduce costs. However, this strategy reduces but does not eliminate the central economic dilemma. When technology becomes the doctor, it could be the cure for Baumol’s cost disease.

As the quality and scope of clinical data available for training these large language models continues to increase, so will their capabilities. Even if the current stage of development is not quite ready to completely remove physicians from the decision-making cycle, these tools will increasingly increase the productivity of providers and in many cases begin to replace them.

Leave a Reply

%d bloggers like this: