AI in Pulmonary Medicine: What Works Now and What Still Needs Proof
Dr. Shirin Shafazand, the inaugural faculty director for the Miller School of Medicine’s Office of AI, argues artificial intelligence is helping clinicians analyze pulmonary imaging and other complex data, but local validation, human oversight and stronger evidence are essential.
Key Takeaways
- Pulmonary imaging is among the most mature areas for clinical artificial intelligence applications.
- Strong performance in research or vendor studies does not guarantee the same results in local clinical practice.
- Health systems should validate AI tools using their own scanners, workflows and patient populations before deployment.
- Clinician oversight remains essential to assess accuracy, reduce alert burden and monitor real-world performance.
- Future advances may come from multimodal systems that combine imaging, pulmonary function tests, laboratory data and other clinical information.
Artificial intelligence is rapidly moving from research laboratories and industry into clinics, imaging suites and hospital workflows. But while enthusiasm about AI’s potential is widespread, pulmonologists must balance excitement with a clear-eyed assessment of what these technologies can and cannot do today.
That was the central message delivered by Shirin Shafazand, M.D., M.S., FAASM, ATSF, professor of pulmonary, critical care and sleep medicine at the University of Miami Miller School of Medicine, during a recent discussion of the current state of AI in pulmonary medicine at the American Thoracic Society (ATS) 2026 International Conference in Orlando, Florida.
Dr. Shafazand’s presentation explored where AI is already demonstrating real clinical value, where important limitations remain and how health systems should approach adoption responsibly.
Why is Pulmonary Medicine Well Suited for AI?
According to Dr. Shafazand, faculty director for the Miller School’s Office of AI, pulmonary medicine is particularly well positioned to benefit from AI because the specialty generates enormous amounts of imaging and physiologic data. Millions of chest CT scans and chest X-rays are performed annually, creating datasets well suited for machine learning tools capable of identifying complex patterns that may be difficult or time-consuming for humans to detect.
“We deal with a lot of images, complex patients and patients with chronic disease that require long-term follow-up management,” she said. “Wearables are coming into play right now, and there is a shortage of pulmonary specialists, so there is potential for these tools, as they mature, to actually have benefit for many populations.”
Some of the most mature applications are already appearing in lung cancer screening and pulmonary imaging.
Dr. Shafazand highlighted studies showing that AI-assisted systems for lung nodule detection can achieve high levels of performance accuracy and may help radiologists identify potential cancers earlier. She also reviewed evidence suggesting AI can assist with pattern recognition in interstitial lung disease and support more consistent interpretation of pulmonary function tests.
“Radiology-embedded tools, especially CT lung nodule detection and characterization and triage of acute radiologic findings such as AI worklist prioritization for pulmonary embolism, have the most mature evidence base and the largest number of FDA clearances behind them,” she said. “They are already deployed in real workflows and designed as augmenting radiologists’ work, with the radiologist retaining final interpretive authority, which keeps the regulatory and medicolegal barriers comparatively low.”
Strong Technical Performance Doesn’t Guarantee Better Care
Dr. Shafazand repeatedly emphasized that strong performance in research or vendor studies does not automatically translate into real-world clinical impact.
“Performance metrics from development studies may not transfer to practice,” she said. “Local validation is essential before clinical deployment. Before going live, every health system should deploy a silent or shadow mode period where the tool runs on live data without influencing care, so you can compare its output against the actual results before anyone acts on it.”
Dr. Shafazand added that prospective randomized clinical trials demonstrating improvements in patient outcomes remain limited. Assumptions about cost effectiveness often have not been rigorously evaluated.
Among the most significant concerns are issues of bias and generalizability. Many AI models have been trained using datasets from academic medical centers that may not adequately represent the diverse patient populations encountered in everyday practice. Dr. Shafazand highlighted evidence showing that demographic information is often poorly reported in regulatory submissions for AI-enabled medical devices, raising questions about how these tools will perform across different populations and care settings.
She also discussed the challenge of trust. Many advanced deep-learning systems operate as “black boxes,” producing recommendations without clearly explaining how conclusions were reached. While emerging approaches in explainable AI may help, clinicians remain reluctant to rely on recommendations they cannot independently verify.
What Could the Next Generation of Pulmonary AI Look Like?
Looking ahead, Dr. Shafazand believes the next wave of innovation will extend beyond image analysis. She pointed to multimodal foundation models, wearable monitoring technologies, ambient clinical documentation systems and AI-assisted disease management platforms as areas poised for major growth during the coming decade.
“Narrow, single-task tools like what we have right now each solve a specific problem,” she said. “A model that reasons across all of it at once, when it happens, is a transformative change.”
At the same time, she argued that successful implementation will require careful oversight. Her recommendations included establishing institutional governance structures, validating performance on local patient populations, continuously monitoring outcomes after deployment and ensuring transparency with patients about how AI is being used in their care.
AI tools, Dr. Shafazand suggested, have already demonstrated meaningful capabilities in pulmonary medicine. Responsible adoption will require rigorous evaluation, clinician education and continued attention to ethics and equity.
She concluded with a thought-provoking adaptation of a frequently cited observation about the future of medicine.
“AI will not replace physicians,” Dr. Shafazand said. “But physicians who use AI will replace those who don’t.”
More from the Miller School of Medicine
Miller School researchers used AI and laboratory experiments to identify TM184C, a protein that helps cells connect and share resources.
Sylvester Comprehensive Cancer Center is developing an AI-powered platform to provide personalized support for breast cancer survivors.
AI models using clinical and biomarker data identified Parkinson’s patients at risk for rapid cognitive or motor decline years in advance.
AI-tested visuals and cancer-specific messages may improve awareness of alcohol-related cancer risks, according to a Sylvester study.
Tags: AI, artificial intelligence, Division of Pulmonary, Critical Care and Sleep Medicine, Dr. Shirin Shafazand, medical education, technology