Leveraging AI to Overcome Underrecognition of AATD, with James Stoller, MD

Alpha-1 antitrypsin deficiency (AATD), a genetic condition that can lead to serious lung and liver disease, remains vastly underdiagnosed in the United States—despite clinical guidelines and an increasing need to identify patients eligible for emerging therapies.

Often, symptoms of individuals who have chronic obstructive pulmonary disease (COPD) and/or liver disease due to AATD are indistinguishable from those without deficiency, necessitating additional tools to establish diagnosis. Research presented at the American Thoracic Society (ATS) International Conference 2025 by James Stoller, MD, MS, a pulmonary/critical care physician at the Cleveland Clinic and Chairman of the Cleveland Clinic Education Institute, suggests a novel AI model may offer a potential tool for identifying symptomatic patients who likely have AATD but are unrecognized.1,2

“[Recognition] is particularly important now, because, somewhat uniquely in the history of alpha-1, there's a pipeline of drugs being developed with very novel approaches. In order to do registrational trials with those drugs, companies need adequate numbers of patients to recruit,” Stoller explained to HCPLive. “With only 15,000 patients recognized in the United States, most of whom are receiving augmentation therapy, many of whom, like my patients, are unlikely to go off augmentation therapy to participate in a randomized, placebo-controlled trial, there won't be enough patients to be able to conduct registrational trials.”

Leveraging claims data from the United States Komodo Health database, Stoller and a team of investigators previously reported an AI model to identify symptomatic patients who likely had AATD but were unrecognized. In the research presented at ATS, they used retrospective deidentified patient data generated from the Cleveland Clinic Electronic Medical Record (EMR) between January 1998 and August 2024 to define and characterize patient cohorts utilized to calibrate and refine the model.1

Of 794 patients with AATD, 50.4% of patients were male and the median age was 63 years. A total of 290 (36.5%) patients were identified by genotype (genetic), and the remaining patients were characterized by serum AAT level in the severe (n=117, 14.7%) and mild (n=387, 48.7%) categories.1

Investigators noted the prevalence of common comorbidities was lower in the mild subgroup versus the total population and other subgroups. Additionally, they pointed out 12.5% of patients had any AATD augmentation therapy drug in their records.1

Investigators further analyzed the AI model’s performance using 21,166 records from the Cleveland Clinic EMR. The calibration and the validation datasets consisted of 582 positive and 5864 negative patients.2

Results showed the calibrated model achieved high levels of performance (ROC-AUC=0.89, PR-AUC=0.62) on the hold-out validation subset in distinguishing AATD positive from negative cases. At a discrimination threshold of 0.90, the model achieved high sensitivity=0.70, high specificity=0.92, and balanced accuracy=0.81.2

“Like any study, you know, it invites further inquiry,” Stoller said, describing plans to further test the model in clinical practice to determine its utility in these settings as well as to confirm its generalizability beyond this Cleveland Clinic dataset.

Editors’ note: Stoller has relevant disclosures with CSL Behring, Grifols, Metronic, Sanofi, Takeda, and Vertez.

References

1. Stoller JK, Khandelwal N, Esmaeili V, et al. Application of an Artificial Intelligence Model to Detect Alpha-1 Antitrypsin Deficiency: Characterizing the Study Population [abstract]. Am J Respir Crit Care Med 2025;211:A4200. https://doi.org/10.1164/ajrccm.2025.211.Abstracts.A4200

2. Stoller JK, Khandelwal N, Liakoni V, et al. Application of an Artificial Intelligence Model to Detect Alpha-1 Antitrypsin Deficiency: Model Performance [abstract]. Am J Respir Crit Care Med 2025;211:A4201. https://doi.org/10.1164/ajrccm.2025.211.Abstracts.A4201