A CT-based deep learning screening model identified heterozygous Alpha-1 Antitrypsin Deficiency carriers with an ROC AUC of 0.591, potentially reducing universal genetic testing needs by 32%.
Observational (n=5,327)
Does a CT-based deep learning screening tool accurately identify Alpha-1 Antitrypsin heterozygous carriers in patients with COPD?
A CT-based deep learning model shows promise as an opportunistic pre-screening tool to identify Alpha-1 Antitrypsin heterozygous carriers, potentially reducing the need for universal genetic testing by 32% at 80% recall.
Effect estimate: ROC AUC 0.591
Abstract Background Alpha-1 Antitrypsin Deficiency (AATD) remains one of the most underdiagnosed genetic conditions, with fewer than 10 % of affected individuals aware of their status. Despite guideline recommendations to test all patients with COPD or unexplained emphysema, real-world genetic testing rates remain very low due to cost, access limitations, and lack of awareness. Moreover, heterozygous carriers (MS and MZ) may show subtle imaging or functional abnormalities that go unnoticed in routine clinical care. These challenges highlight the need for non-invasive pre-screening tools. We evaluated the feasibility of a CT-based deep learning approach to identify heterozygous AATD carriers using a foundation model trained on chest CT data. Methods Our model was trained and evaluated on 5,327 COPDGene participants with CT scans and reliable lobar segmentation. A pretrained foundation model generated five lobe-level embeddings per subject, which were then input to a lobe-based Transformer encoder to classify between heterozygous carriers (MS and MZ) and MM subjects. Model validation used 5-fold cross-validation. To address class imbalance given the 9% prevalence of heterozygous subjects, SMOTE was applied during training. Performance metrics—precision, recall, F1 score, ROC AUC, and confusion matrices—were computed per fold. To align with clinical screening priorities, Screen-to-Enroll ratios (SER) were computed at fixed recall thresholds of 70%, 80% and 90%. All other metrics were reported at the 80% recall operating point Results 491 participants were heterozygous carriers. Across folds, ROC AUC ranged between 0.578 and 0.603 (0.591 ± 0.011). Mean precision was stable as recall increased: 0.110 ± 0.007 (@70 %), 0.106 ± 0.003 (@80 %), 0.100 ± 0.002 (@90 %). At a 9.2 % prevalence, these correspond to SERs of 9.1 ± 0.6 : 1, 9.4 ± 0.3 : 1, and 10.0 ± 0.2 : 1, respectively, indicating that approximately 9-10 individuals would require confirmatory genetic testing to identify one true carrier. Weighted F1 scores varied from 0.44 to 0.49, with an aggregated weighted F1 score of 0.46 across the 5 folds (Table 1). Conclusion The pre-trained foundational model and our deep learning screening method demonstrates promise for identifying between heterozygous carriers (MS and MZ) with fixed high sensitivity, crucial for early detection. At 80% recall, imaging pre-screening could reduce the need for universal genetic testing by 32 %, while preserving high sensitivity for carrier detection enabling opportunistic AATD detection from existing imaging data. Further clinical validation and refinement are justified to improve specificity while preserving sensitivity. This abstract is funded by: This work was supported by the National Institutes of Health (NIH 1R01HL149877) and Alpha-1 Foundation (1037165). The COPDGene study (NCT00608764) is supported by the National Heart, Lung, and Blood Institute (NHLBI U01 HL089897 and NHLBI U01 HL089856).
Curiale et al. (Fri,) conducted a observational in Alpha-1 Antitrypsin Deficiency (AATD) (n=5,327). CT-based deep learning screening vs. MM subjects was evaluated on Classification between heterozygous carriers (MS and MZ) and MM subjects (ROC AUC 0.591). A CT-based deep learning screening model identified heterozygous Alpha-1 Antitrypsin Deficiency carriers with an ROC AUC of 0.591, potentially reducing universal genetic testing needs by 32%.