Abstract PS2-08-11: Artificial Intelligence (AI)-based spatial assessment of tumor-infiltrating lymphocytes (TIL) and outcome in early HER2+ breast cancer (BC): ancillary study of the phase III adjuvant ShortHER trial

Abstract Background: We previously demonstrated the prognostic role of TIL (manual score) in patients with early HER2+ BC enrolled in the ShortHER trial at long follow up (Dieci, JAMA Oncol 2025). Here, we analyze the added prognostic value of AI-based spatial TIL assessment. Methods: The ShortHER study randomized 1254 pts with HER2+ early BC to 9 weeks vs 1 year of adjuvant trastuzumab combined with chemotherapy. Manual TILs were scored according to international accepted standard developed by the TILs-WG. A zero-shot, explainable AI pipeline (Case45), was used to analyze H immune hotspots (normalized fraction of immune cell aggregates in relation to cancer and tissue). Immune metrics were considered as continuous and as categorical (cut-offs: 20%/20% for manual TIL; 75%/75% percentile for AI-TIL and immune hotspots). Survival endpoint was DDFS (distant relapse, death). Cox regression models were adjusted for clinicopathological factors (age, stage, hormone receptor, treatment arm, menopause, histologic grade). Likelihood ratio test evaluated the added prognostic value of TIL metrics to prognostic models. Results: 858 were cases evaluable by AI, median FU was 9.0 years. AI-TIL showed a moderate positive correlation with manual TIL (p0.001, Spearman’s rho 0.500). Patients with high AI-TIL (n=211) experienced a better DDFS vs low AI-TIL: 8-yr rate 93.2% vs 86.7%, log-rank p=0.003, adj HR 0.47 (95%CI 0.26-0.82). Added prognostic value of AI-TIL was observed for patients with intermediate TIL scores between 10% and 29% (n=205): 8-yr DDFS rate was 94.8% for high AI-TIL vs 87.8% for low AI-TIL, p=0.062, mostly driven by the subgroup of patients with manual TIL 20-29% (8-yr DDFS 100% for high AI TIL vs 82.9% for low AI-TIL, p=0.017). Patients with high immune hotspots had a significantly better DDFS vs others (8-yr rates 92.9% vs 86.8%, p=0.028, adj HR 0.66, 95% CI 0.39-1.10). The prognostic role of immune hotspots was confined to patients with low TIL, either manual or AI. In low manual TIL, 8-yr DDFS was 92.9% vs 85.2% for high vs low immune hotspots (p=0.029; adj HR 0.67, 95% CI 0.36-1.25). In low AI-based TIL, 8-yr DDFS was 92.6% vs 84.7% for high vs low immune hotspots (p=0.024; adj HR 0.67, 95%CI 0.38-1.17). Table shows the added prognostic information provided by TIL metrics to multivariable models (Model 1: clinico-pathological factors + manual TIL; Model 2: clinico-pathological factors + AI-TIL). Conclusions: TIL AI-based spatial analysis complements manual TIL scoring in refining prognostic stratification early HER2+ BC. This added value is more pronounced in cases with low to intermediate manual TIL scores. Citation Format: M. Dieci, K. AbdulJabbar, G. Bisagni, S. Bartolini, L. Cavanna, A. Musolino, F. Giotta, A. Rimanti, O. Garrone, E. Bertone, K. Cagossi, A. Ferro, F. Piacentini, E. Orvieto, M. Sanders, F. Miglietta, H. Yan, S. Balduzzi, R. D'Amico, P. Conte, R. Salgado, V. Guarneri. Artificial Intelligence (AI)-based spatial assessment of tumor-infiltrating lymphocytes (TIL) and outcome in early HER2+ breast cancer (BC): ancillary study of the phase III adjuvant ShortHER trial 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 PS2-08-11.