Abstract Background: Lung cancer is the leading cause of cancer death, and better early detection is critical to lowering its burden. Sybil, a deep learning model, predicts 6-year lung cancer risk from a single low-dose CT (LDCT). Yet screening patients typically receive annual LDCTs that capture temporal changes in nodules and disease. We hypothesized that fusing longitudinal imaging features across scans would improve long-term risk prediction and generalizability over single-timepoint models. Methods: We re-implemented Sybil’s backbone and feature embedding layers to enable the extraction of intermediate feature representations while preserving the original inference behavior by loading the publicly released pretrained checkpoints unchanged. We developed six longitudinal deep learning architectures trained on the National Lung Screening Trial (NLST; n = 3,977participants, n = 991 cancers over 6 years): (1) Bi-Fusion (MLP and Transformer), integrating baseline + year-1 LDCTs; (2) Tri-Fusion (MLP and Transformer), integrating baseline, year-1, and year-2 LDCTs; (3) Hybrid (Temporal Hybrid and Masked Hybrid), combining CNN backbone, vision transformer, and temporal transformer modules and supporting up to three timepoints with missing scan handling. All models estimated time-to-lung-cancer using a hazard layer and were evaluated using time-dependent AUCs at yearly prediction horizons and Harrell’s C-index. External evaluation used Moffitt Lung Screening data (n = 1876 participants, n = 469 cancers). For Hybrid models, only the hazard head was fine-tuned on a Moffitt adaptation subset, followed by evaluation on a held-out external test set. Results: Hybrid longitudinal models demonstrated the highest performance and stability across follow-up horizons. During internal testing of the NLST, the Masked and Temporal Hybrid models achieved near-perfect discrimination at Year 1 (time-dependent AUC ≈ 0.99), gradually decreasing to 0.91 and 0.90 by Year 6, with a C-index of 0.75 (95% CI: 0.73-0.78). After external fine-tuning and evaluation, Hybrid models generalized robustly to Moffitt, with an AUC of 0.89 at Year 1, declining to 0.81 at Year 6, and a C-index of 0.81 (0.74-0.87). Tri-Fusion architectures peaked at an AUC of ∼0.92 at Year 3 and declined to 0.86 by Year 6 in NLST, but demonstrated limited generalizability (0.63 to 0.57, C-index 0.60 0.53-0.67). Bi-Fusion performance was intermediate (AUC 0.83 to 0.77 in NLST; ∼0.70-0.62 on Moffitt). Conclusion: Longitudinal fusion of LDCT scans improves risk prediction vs single-timepoint models. Hybrid models using up to three scans show superior accuracy, robust external performance, and may enable more precise risk-based care while reducing unnecessary imaging and invasive procedures; clinical utility will be tested via decision-curve and implementation analyses. Citation Format: Hanieh Ajami, Asim Waqas Waqas, John Michael Templeton, Matthew B. Schabath, Ghulam Rasool. Enhancing lung cancer risk prediction with longitudinal fusion of LDCT data abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 6727.
Ajami et al. (Fri,) studied this question.