Artificial intelligence (AI) can potentially reduce health disparities if current biases are recognized and addressed, according to the experts who spoke during Sunday’s Fran Comi Keynote Lecture at the ATS 2024 International Conference.

Fatima Rodriguez, MD, MPH, associate professor of medicine at Stanford University, described her work using AI in opportunistic screening for cardiovascular disease. Matthew Decamp, MD, PhD, associate professor at the University of Colorado, discussed two key ethical issues in AI: bias and explainability.

Michael Howell, MD, MPH, chief clinical officer at Google, moderated the session.

Using AI to reduce disparities in cardiovascular disease prevention

Specialized gated computed tomography (CT) scans can detect coronary artery disease and prompt timely preventive actions. However, these scans are often inaccessible to historically marginalized groups.

“Like all new technologies, AI can inadvertently exacerbate health disparities,” said Dr. Rodriguez. “It is a promising tool, but not a standalone solution in cardiovascular disease prevention. Our implementation of AI should prioritize health equity and mitigate structural bias.”

She sees an opportunity to use AI to evaluate the nearly 20 million non-gated chest CT scans performed every year for reasons other than cardiovascular risk, including cancer screening and staging.

The AI algorithm deep learning coronary artery calcium (DL-CAC) predicts cardiovascular risk scores across diverse groups of patients using non-gated CT scans. In a quality improvement study, Dr. Rodriguez used DL-CAC to identify high-risk patients and reach out to the patient and primary care provider, recommending that the patient go on a statin. She showed that among patients with incidental coronary artery calcium who were not notified, only seven percent received statins; among those identified by DL-CAC and alerted to their risk, more than 50 percent received statins.

Dr. Rodriguez described patients’ reactions to their images being analyzed by AI as mixed.

“Patients expressed optimism and hope around AI but were concerned about privacy and the use of their data,” she explained. “They felt strongly about having actionable insights and wanting their physician to be included in the conversation and not just have the AI speak for itself.”

Addressing bias and explainability in AI

Dr. Decamp described how bias is introduced into AI through biased datasets, processing, and output. Most of the focus has been on addressing bias within datasets by increasing the diversity of individuals represented in large biobanks, cohorts, and clinical trials. However, Dr. Decamp emphasized that researchers and clinicians must keep the bigger picture in mind.

“Biased data and processing may be just the tip of the iceberg,” he warned. “We wonder about latent biases, biases in the system waiting to happen.”

Addressing bias in AI requires engaging underrepresented communities, ensuring diversity in teams involved with data processing, and acknowledging underlying social and structural inequities. Dr. Decamp warned that proposed solutions, such as simulating diversity or establishing “fair AI” labels, were insufficient.

“These are short-term fixes to circumvent long-term social justice and community engagements,” he said. “We must identify fairness issues and tradeoffs and make sure that they are measured and reported, not hidden.”

He noted that explainability, understanding how an AI algorithm comes up with its interpretation, won’t fix bias either.

“Bias in AI is more than a data problem; it’s a social problem,” he said. “And while explainability is important, it won’t fix bias.”