Harnessing AI to optimize patient treatments

Bradley Brimhall, MD, with Yi Zhou, MD, clinical informatics fellow, and Ashley Windham, MD, MSDA, associate professor, around a microscope in Brimhall's lab.
Bradley Brimhall, MD, with Yi Zhou, MD, clinical informatics fellow, and Ashley Windham, MD, MSDA, associate professor.

One machine, multiple targets

By Michael Seringer

Imagine the cancer cell under the microscope as an aerial photograph of a city. The human eye can recognize distinctive features of the city and tell if a building is on fire. An algorithm, backed by ever-increasing processing power, can count every building, stop sign and pothole, take measurements between each, discover new features altogether and even throw in some Euclidean geometry around specific features to help predict problems on city streets.

Bradley B. Brimhall, MD

Professor of Pathology and Laboratory Medicine; Medical Director of Clinical Informatics, Health Care Analytics and Bioinformatics; Staff Pathologist, Hematopathology and Transfusion Medicine
Diagnostic: Predictive artificial intelligence disease detection tool
Indication: Cancer and urinary tract infection
Stage: Pre-seed, proof of concept
Timing: Approximately three years
Funding: $200,000 — awards and grants
The use of AI in medical decision-making is the future, and physicians can lead that change or be swept before it.

Bradley Brimhall and collaborative teams of researchers are harnessing the power of artificial intelligence and predictive modeling to develop leading-edge diagnostics.

“We are using machine learning and artificial intelligence to segment the image, recognize features, give them each a name, count them, calculate some geometric parameters and hopefully make the images more usable and quantifiable,” he said.

The advantage of machine learning-based diagnostics is that the core technology can be used not only to scan slide images, but also to consume large datasets of anonymous patient information to develop accurate prediction models. One of the first disease targets of such algorithms is predicting urinary tract infections using aggregated patient data. The idea is that physicians could use this AI tool to predict — with a high degree of accuracy — a positive urinary tract infection test well before the test results are in. Predictive diagnostics would allow physicians to optimize treatments, improving patient outcomes.

It takes a great team

Brimhall stresses the importance of a cohesive, multidisciplinary team in developing novel diagnostic tools and believes his partnerships with Southwest Research Institute and The University of Texas at San Antonio have been key to the early success of their algorithm. Relying on his past experience leading both for-profit and nonprofit institutions, Brimhall engaged with a multidisciplinary team of skilled scientists from the start.

“It takes a team to develop an AI diagnostic tool,” Brimhall said. “Good data science and analytics requires people. Data are the raw materials and software is the tool, but it’s the people who make things happen.”

A diversity of experience and skill sets improves the overall product, Brimhall said. Each team member approaches the problems from a unique viewpoint, providing more opportunities for creative solutions. For a medical technology startup to succeed, it requires knowledge of both the business and academic worlds. The business, clinical and science team members all need to be aligned to improve the odds of product success.

“Good data science and analytics requires people. Data are the raw materials and software is the tool, but it’s the people who make things happen.”

Even the best AI diagnostic tool is essentially worthless if it does not become part of a medical system’s workflow. The tool won’t make it out of the research institution if the market is too small or the team lacks a broad perspective.

“Research teams can get caught up in an academic myopia where the long view is obscured by small details,” Brimhall said. A balance between the two worlds ensures a more efficient go-to market strategy.

A new regulatory frontier

To gain marketing approval from the Food and Drug Administration for a diagnostic tool, investigators must show the rate of false positives and negatives in a clinical trial. However, AI-based diagnostic platforms have an easier time getting marketing approval because as tools that assist pathologists in reviewing slides, they are considered low risk.

These diagnostic tools are reviewed through the FDA’s “De Novo” classification pathway for novel devices in which two types of classifications, I and II, are assigned based on the risk associated with the tool.

“The FDA pathway for devices is much easier and cost effective compared to medications,” Brimhall said.  Getting FDA device approval is much less than the cost to develop a new drug.”

Another important trait of machine learning is the evolution of the algorithm over time. As the AI-based prediction tool consumes more and more data, it improves. The machine is always learning and sharpening its accuracy.

The FDA has recognized this characteristic of machine learning by providing guides and best practices for AI-enabled diagnostics and devices. The FDA also maintains a searchable database of approved, AI-based algorithms that gives researchers a free resource and technical road map to help them determine market size, understand the competitive landscape and develop a regulatory strategy.

Proving the concept

The lifecycle of a business starts with a proof of concept for the product. This feasibility stage for AI-based diagnostics requires sufficiently large data sets. Data use agreements and patient privacy must be worked out along with specific data requirements.

Once the algorithm is tested using large amounts of data, proof of concept can be determined quickly. Large data volumes yield significant findings to reduce risk and engender confidence for potential investors.

Information gained by running AI-enabled tools also provides confidence for potential partners. Going into talks with electronic medical record companies or leading testing labs is made easier when the diagnostic tool has been proven using large amounts of data.

Information is power, and machine learning and AI generate such power.

“The most dangerous thing is showing up to a negotiation and not having information,” Brimhall said. “More than 90% of a negotiation occurs before you even show up to the table. The key is the preparation that goes into the negotiation.” Another important aspect of medical AI development is to make sure that the unique aspects of the program are well protected by patents.

Medical AI is like the wild west, with many companies with sharp elbows competing for a small amount of intellectual space. Without such intellectual property and the willingness to fight for it, smaller AI companies will get rolled over.

Finally, there is an inherent tension in medical AI development because the end user — physicians — are reluctant to test it. This is true for two reasons. First, many physicians mistrust a computer’s judgment when it contradicts their own, because their own judgment was hard won over many decades of work.

Second, subconsciously physicians may wonder if they will be replaced by AI, and they need to justify their existence. In truth, medicine needs to lead AI or AI will lead medicine, Brimhall said.

Physicians can remain relevant by managing AI if they will take the time to understand it. The use of AI in medical decision-making is the future, and physicians can lead that change or be swept before it.

Share this post!

In the 2022 issue of Future

Future is the official magazine of the Joe R. & Teresa Lozano Long School of Medicine at The University of Texas Health Science Center at San Antonio. Read and share inspiring stories highlighting our medical alumni, faculty and students who are revolutionizing education, research, patient care and critical services in the communities they serve.

View the 2022 issue

Categories for this article :