AI-ECG Model Detects COPD Months Before Diagnosis, With Monica Kraft, MD

A deep learning model applied to routine electrocardiograms (ECGs) identified chronic obstructive pulmonary disease (COPD) months before clinical diagnosis, according to a new study.1 The approach, validated across multiple cohorts, suggests a potential role for artificial intelligence (AI)-enhanced ECG analysis as a scalable screening tool in routine clinical care.

In an interview with HCPLive, investigator Monica Kraft, MD, from the Icahn School of Medicine at Mount Sinai, emphasized the clinical importance of earlier recognition. She noted that identifying patients with early or subclinical COPD could create opportunities for timely intervention, particularly smoking cessation and initiation of therapies aimed at slowing disease progression.

“I think what's happened in the past is when we've done lung function screening [and] testing, oftentimes the lung function…in the clinic looks very normal,” Kraft said. “You can imagine [a] scenario where patients are smoking, they have their lung function done, it looks normal, and they think,’ Hey, this is great. I can keep smoking.’ Whereas, if we have a mechanism that demonstrates that there may actually be some lung damage going on because of what the EKG shows, then there may be an opportunity to really educate the patient that it's really important that they make a lifestyle change now in order to avoid future problems down the road.”

COPD remains a leading cause of morbidity and mortality worldwide, with delayed diagnosis representing a persistent challenge. Symptoms are often nonspecific in early disease, and confirmatory testing, primarily spirometry, can be underutilized due to logistical and resource constraints.2 Current guidelines do not recommend routine screening in asymptomatic individuals, further contributing to underdiagnosis.3

In this study, Kraft and colleagues developed a convolutional neural network (CNN) model trained on > 760,000 ECGs from > 67,000 patients across the Mount Sinai Health System, with additional validation in an independent cohort from the UK Biobank.1 The model demonstrated consistent performance, with AUROC values of 0.80 in internal testing, 0.82 in external validation, and 0.75 in the UK Biobank cohort.

The model’s predictive signal appeared to precede clinical diagnosis. Analyses showed that ECG-based predictions could identify COPD-related changes approximately 6 to 15 months before formal diagnosis, suggesting the model is capturing early physiologic alterations rather than late-stage disease alone.¹ Kraft highlighted that this window may be clinically actionable for high-risk individuals, such as active smokers, where earlier counseling and intervention could alter disease trajectory.

The model leverages subtle ECG features associated with pulmonary vascular and cardiac changes seen in COPD. Explainability analyses indicated a focus on P-wave morphology, consistent with early right atrial and pulmonary vascular alterations.1 While such changes are typically recognized in advanced disease, the application of AI enables detection of patterns not readily discernible to clinicians interpreting individual ECGs.

The clinical implications of this approach lie in its potential integration into existing workflows. ECGs are widely available, low-cost, and routinely performed across inpatient and outpatient settings. As Kraft noted, this creates an opportunity for patients undergoing ECGs for unrelated indications to be flagged for further pulmonary evaluation, which is valuable in health systems with limited access to spirometry or populations less likely to undergo formal testing.

However, the model is not intended to replace spirometry, which remains the diagnostic standard for COPD. Rather, it may function as a triage or decision-support tool, identifying patients who warrant confirmatory testing.

Investigators highlighted several limitations, including reliance on ICD-coded diagnoses, declining performance over time, and modest correlations with spirometry, suggesting the model may reflect cumulative cardiopulmonary changes more than early airflow limitation alone.

The team noted that prospective validation is needed to determine whether earlier detection via AI-ECG translates into improved clinical outcomes. Kraft underscored this, noting that the next phase of research will involve identifying at-risk patients in real time, conducting comprehensive clinical evaluations, and assessing whether early interventions, such as smoking cessation or pharmacologic therapy, alter disease progression.

“We still have some work to do. The algorithm, just creating it and applying it, is step 1,” Kraft said. “Step 2 is now to see how it applies to patients moving forward so [not] quite in the hands of clinicians yet, but we hope it will be very soon.”

References

1. Vaid A, Sharma J, Jiang J, et al. Automated diagnosis of chronic obstructive pulmonary disease using deep learning applied to electrocardiograms. eBioMedicine. 2025;106066. https://doi.org/10.1016/j.ebiom.2025.106066

2. World Health Organization. Chronic obstructive pulmonary disease (COPD). https://www.who.int/news-room/fact-sheets/detail/chronic-obstructive-pulmonary-disease-(copd)

3. US Preventive Services Task Force. Screening for chronic obstructive pulmonary disease: recommendation statement. JAMA. 2022;327(18):1806-1811. https://doi.org/10.1001/jama.2022.5692