Abstract GS1-08: Multimodal artificial intelligence (AI) models integrating image, clinical, and molecular data for predicting early and late breast cancer recurrence in TAILORx

Abstract Background: TheTAILORx trial established that adjuvant endocrine therapy (ET) is non-inferiorto chemotherapy plus ET (CET) in T1-2N0 HR+/HER2-negative early breast cancer(EBC) with an Oncotype DX (ODX) 21-gene recurrence score (RS) of 11-25, RS 0-10is associated with low distant recurrence (DR) rates with ET alone, and RS26-100 is associated with higher DR rates despite CET. Although ODX isprognostic for 10-year DR, it has limited value for late DR >5 years. Inorder to develop improved DR risk prognostication, we developed multimodal AImodels integrating clinical, molecular, and histopathology features to provideprognostic information for early (<5years), late (>5 years), and overall DR at 15 years using primary tumorsamples and clinical data from volunteers who participated in TAILORx. Methods: DigitizedH&E slides were generated and nucleic acids extracted and sequenced from∼5000 primary tumor samples using Caris MI Tumor Seek-Hybrid, of which 2,808(56%) were used for model training and 5-fold nested cross validation (60/20/20split). Models included single-modality (clinical [C], image [I], molecular [M:ODX genes]), dual-modality (CM, IC), and multimodal (ICM). An expandedmolecular model (M+) model was also derived from expanded molecular inputsderived from whole transcriptome sequencing (WTS) data, including 5 commercialgene signatures (ODX, MammaPrint, Prosigna, EndoPredict, BCI) and 57 high-variancegenes; only EndoPredict, BCI, and ODX gene signatures were retained for model M+.Model-CM+ combined clinical plus expanded molecular features, while ICM/ICM+integrated image, clinical, and (expanded) molecular inputs. Risk scores were dichotomized into high vs. low risk groups usingthresholds chosen to yield risk group proportions closely matching thosedefined by the ODX RS cutoff of 25 for distinguishing high vs. low genomicrisk. Theprimary endpoint was distant recurrence (DR). Prognostic performance wasassessed using truncated concordance index (C-index, 90th percentile of eventtimes), hazard ratios (HRs), and log-rank tests. DNA-level whole exomesequencing (WES) features were also tested but did not provide additionalprognostic information and thus excluded from final models. Results: Theactual ODX continuous RS achieved a C-index of 0.625 for overall DR and 0.743early DR, but no prognostic value for late DR (0.52). Similar results for theC-index were observed for ODX+C for overall DR (0.619), early DR (0.724), andlate DR (0.514). Of all modelsevaluated, model-ICM+ gave the best overall prognostic accuracy for overall DR (C-index0.713; hazard ratio [HR] 3.56 for comparison of high vs. low risk, p<0.001)and late DR (C-index 0.655; HR 2.36, p<0.001), and ranked second for earlyDR (C-index 0.772; HR 6.59, p<0.001) behind the CM+ model (C-index 0.782). Amongsingle-modality models, Model-M+ improved early DR prognostication comparedwith a molecular model including only ODX genes(C-index 0.767 vs 0.744 for M),whereas Model-I was strongest for late DR prognostication (C-index 0.635). Conclusions: Molecularfeatures primarily drove prognostic accuracy for early DR within 5 years,whereas histopathology features strengthened the prognostic accuracy for lateDR after 5 years. Multimodal models captured these complementary signals, withICM+ giving the best overall accuracy for individualized prognostic assessment.Additional validation of the optimized ICM+ model and other models is beingperformed in an intendent validation sample of 2000 TAILORx volunteers, withresults to be presented at the meeting.Supported by the Breast CancerResearch Foundation and the U.S. NIH/NCI U10CA180820 and 5U24CA196172, and theU.S. Postal Service Breast Cancer Stamp Fund. Citation Format: J. A. Sparano, V. Wang, R. J. Gray, D. F. Makower, K. S. Albain, D. F. Hayes, C. E. Geyer, E. C. Dees, M. P. Goetz, J. A. Olson, S. S. Badve, T. J. Saphner, T. J. Whelan, M. Radovich, V. G. Kaklamani, G. W. Sledge. Multimodal artificial intelligence (AI) models integrating image, clinical, and molecular data for predicting early and late breast cancer recurrence in TAILORx [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr GS1-08.