B102-02 Deep Learning-based Detection of Homozygous Alpha-1 Antitrypsin Deficiency From HRCT: A Proof-of-concept Study Across COPDGene, SPIROMICS, and GRADS Cohorts

Abstract Rationale Alpha-1 antitrypsin deficiency (AATD) continues to be underdiagnosed. The estimated prevalence of the most common severe deficient genotype, PI*ZZ, is ∼1:3500 or 100,000 Americans but only ∼10,000 such individuals have been recognized. Diagnostic delays persist and these delays are associated with worse disease. High-resolution CT (HRCT) imaging may contain latent phenotypic AATD signatures that are not visibly apparent to clinicians. We hypothesized a deep learning model could identify imaging patterns predictive of AATD from inspiratory HRCT scans. Methods Confirmed genotypes and HRCT scans were pooled from COPDGene, SPIROMICS, and GRADS, yielding 350 MM, 102 MZ, and 69 ZZ participants. Only scans reconstructed with smooth kernels (GE STANDARD, Philips B, Siemens B31f/B35f) and slice thickness =1.25mm were used. The cohort was selected to have similar distributions of quantitative emphysema, sex, and age by matching five MZ and two MM subjects to each ZZ subject (10-year age difference, same sex, and 3% difference in %LAA). Tobacco smoke-exposure was more prevalent among the MM/MZ groups given these were predominantly from the tobacco smoke-enriched COPDGene and SPIROMICS cohorts. A DenseNet121 convolutional neural network (CNN) was trained using the MONAI platform to predict genotype by outputting probabilities for MM, MZ, and ZZ. Input to the CNN was a three-channel LAA-910, LAA-950, and LAA-970 map downsampled to 2.0mm3 resolution to remove bias due to differences in scanner and image reconstruction parameters. Five-fold cross-validation was performed. Area under the curve (AUC) was calculated for classification of ZZ vs other genotypes (MM/MZ) and MZ vs MM, with statistical significance assessed at p 0.05. Results The three groups were similar across age (mean±SD years: MM 57.9±8.8, MZ: 57.5±8.9, ZZ: 55.0±10.7), sex (% female: MM: 48.2%, MZ: 46.7%, ZZ: 47.6), and emphysema (LAA-950%: MM: 13.0±12.2, MZ: 12.4±11.8, ZZ: 13.1±11.9). There was a higher prevalence of ever smokers among MM (100%) and MZ (91.7%), compared to ZZ (47.6%). The model achieved an AUC of 0.806 (p 0.001) for classification of ZZ vs other genotypes, demonstrating robust discriminatory ability for homozygous AATD. Classification of MZ vs other genotypes did not achieve an AUC significantly greater than 0.5. Conclusions This proof-of-concept study demonstrates that deep learning applied to routine inspiratory HRCT can accurately identify homozygous AATD. These findings support the potential role of imaging-based machine learning as a non-invasive screening tool for undiagnosed ZZ individuals. Future work will focus on cohort enrichment, integration with clinical parameters, and explainability of model-derived imaging features. This abstract is funded by: NIH NHLBI K23HL173570