AI offers real promise in the fields of pulmonary medicine and critical care, experts shared, but also features hype, uneven maturity, and some considerable barriers to implementation. Despite these challenges, AI technology is advancing rapidly across these fields. Applications in radiology still outpace the rest, accounting for about 76% of the approximately 1400 AI medical devices cleared by the FDA.

Within pulmonology, AI is changing or poised to transform diagnostics in general and for chronic obstructive pulmonary disease (COPD), interstitial lung disease (ILD), and pulmonary hypertension (PH) in particular.

“Pulmonary medicine is particularly ripe for this because we deal with a lot of physiological data,” Shirin Shafazand, MD, MS, professor of clinical medicine in the Division of Pulmonary, Critical Care, and Sleep Medicine at the University of Miami Miller School of Medicine in Miami, said during a panel presentation on AI at the American Thoracic Society (ATS) 2026 International Conference.

“We deal with a lot of images, complex patients, and patients that require long-term follow-up management. 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. Hopefully,” said Shafazand, who also leads the Office of AI in Medical Education at University of Miami.

One current limitation is that most FDA-approved AI technology has a narrow focus. Machine learning algorithms focus on specific indications such as triage and detection of lung nodules, pulmonary emboli, or pneumothorax, for example. Therefore, unlike humans, AI cannot yet read an entire x-ray for different abnormalities, Shafazand said. “Although that is coming, and many radiologists are excited about that possibility.”

AI Flags COPD Presence but Not Severity

For COPD detection, AI is able to accurately predict whether the condition is present or not. Deep learning models yielded a pooled sensitivity of 0.87, specificity of 0.88, and an area under the curve (AUC) of 0.93 for this binary classification in a January 2026 systematic review and meta-analysis. However, the analysis of 56 studies with 886,753 participants also showed that accurate staging remains a challenge for AI in COPD, Shafazand said.

Combining wearable monitors with AI technology could reduce emergency department (ED) visits, Shafazand added. The dual strategy predicted COPD exacerbations about 7 days in advance by combining oximetry, heart rate, genomics, and environmental data, a 2025 review revealed. AI models that integrated these data achieved an AUC of 0.80 or more, lowered ED visits by up to 98%, and reduced readmission rates by 25%-48%.

In another study, AI outperformed both senior and junior pulmonologists when presented with the same pulmonary function test (PFT) and clinical data for 50 patients. Results from the AI model trained on European Respiratory Society guidelines were contrasted with interpretation from 120 pulmonologists. The study included a choice of nine diagnostic categories, including COPD. Pulmonologists made correct diagnoses in 45% of the cases (range, 24%-62%). In contrast, the AI-based software correctly diagnosed 82% of all cases (P < .0001). One caveat is that AI technology has evolved rapidly since the study was published in 2019.

A more recent study, published in 2023, revealed that when pulmonologists used AI assistance to interpret PFT results, they outperformed AI alone in terms of accurate diagnoses.

Work in Progress: ILD and PH

AI-supported tools to detect ILD or PH could accelerate these diagnoses, Shafazand said. “The technology could be handy in centers where you do not have expertise for the ILD diagnosis or, in particular, idiopathic pulmonary fibrosis.” However, “there’s a lot more work to be done.”

AI could shorten the time until a patient is referred to a tertiary care center for appropriate therapy, she added. “These are the kind of things to look forward to as many of these tools improve.”

Regarding PH, right-sided heart catheterization is considered the benchmark for diagnosis, Shafazand said. However, it is invasive. “The question is: can we use a tool like an AI to look at the patterns on echo and be able to confidently diagnose pulmonary hypertension?,” said Shafazand.

According to one study, the answer is yes. Researchers assessed 7853 consecutive patients with both right-sided heart catheterization and transthoracic echocardiography. They found machine learning accurately diagnosed PH with an AUC of 0.83, an accuracy of 82%, sensitivity of 88%, positive predictive value of 89%, and negative predictive value of 54%.

Reality Check

“But, like everything else, there’s hope and hype,” Shafazand said. One limitation is that models are often trained on patients from smaller, single- or two-center studies, leaving generalizability to other centers and different populations an open question. For example, up until 2025, only a minority of these AI-supported tools reported the demographics of their training populations, including race and ethnicity.

Another concern is degradation of performance over time. These AI tools “need to be updated and constantly checked to make sure that they’re still doing what they’re doing in the most accurate way,” she said.

Furthermore, there is the well-known “black box” problem with AI, or the opaqueness of how it reaches its conclusions. Shafazand said this characteristic erodes clinician trust. Explainable AI, also called trustworthy AI, is emerging, but it is not yet the standard, she added.

A Critical Role in the ICU?

AI is “a growing reality in the ICU…in sepsis prediction, physiologic monitoring, triage and risk stratification, and in documentation,” Rania Esteitie, MD, an assistant professor of internal medicine at Central Michigan University in Mount Pleasant, Michigan, said during the AI panel presentation at ATS 2026.

Detection of sepsis is a main focus of AI in the ICU. For example, the Epic Sepsis Model V2 features an AUC of 0.82-0.92, but the number needed to treat ranges from 21 to 35 patients to detect one event, Esteitie said. One potential drawback is the “heavy alert fatigue” from frequent alarms, she added. Another approach, the Sepsis ImmunoScore (Prenosis), is the first FDA-authorized machine learning sepsis device. The technology uses up to 22 predetermined inputs from a patient’s electronic health record to generate a risk score and assign the patient to one of the four risk-stratification categories.

Esteitie also highlighted some AI technologies designed to predict shock. For example, the FDA cleared the Acumen Hypotension Prediction Index (HPI, Edwards Lifesciences). This approach assesses arterial waveform morphology and creates a hypotension score from 0 to 100.

“This reflects the probability of hypotension in the next 5-15 minutes. And the nice thing about this is that there was a randomized controlled trial that showed that it reduces the total time that’s spent in hypotension [during cardiac surgery and subsequent ICU stay]. That’s a real, clinically meaningful reduction, not just a statistical one,” added Esteitie, who is also a pulmonary and critical care physician at Covenant HealthCare in Saginaw, Michigan. 

When the Acumen HPI software is used with an Acumen IQ sensor, Esteitie said, it automatically calculates key parameters every 20 seconds, including cardiac output, stroke volume, stroke volume variation, pulse pressure variation, dynamic arterial elastance, and contractility. “What I find particularly useful about this one is that it doesn’t just flag high-risk patients. It also identifies patients at low risk and gives the clinicians the confidence to de-escalate certain treatments or management.”

Overall Themes Emerge

One of the themes from the AI panel is that AI, once deployed in a healthcare setting, needs ongoing monitoring and evaluation, said Laurah Turner, PhD, associate dean for artificial intelligence and educational informatics at the University of Cincinnati College of Medicine, Cincinnati.

“I think the other one that was very clear throughout all talks was the idea of interpretability and reliability,” added Turner, who was tasked with summarizing the panel presentations. “Both of these are very important issues from a technical standpoint,” she added.

Turner added that solving these issues will require new teams and infrastructure. “There is going to be new and emerging expertise in skill sets, which are needed to deploy, monitor, and evaluate these technologies.”

Shafazand is a member of the advisory board for ResMed and the End Point Review Committee for IQVIA. Esteitie and Turner reported having no relevant financial relationships.

Damian McNamara is a freelance contributor to Medscape Medical News. He worked full-time for Medscape and WebMD from 2018 to 2024. Damian has a BA in chemistry and an MA in science, health and environmental reporting/journalism.