Retrospective validation of a novel multimodal AI prognostic tool integrating digital pathology and clinical data against real world data and Oncotype DX in a Swiss breast cancer cohort.

Abstract

e13595 Background: Accurate risk stratification in hormone receptor positive, HER2 negative (HR+/HER2-) breast cancer (BC) is critical for guiding personalised treatment. Oncotype DX (ODX) demonstrates limitations in intermediate-risk groups and high-risk subpopulations. The Ataraxis Breast test addresses these gaps by integrating digital pathology-derived features with clinical data, offering a multimodal approach to recurrence risk prediction. Methods: This retrospective study compared Ataraxis Breast, an AI-based model, to ODX in 269 HR+/HER2- BC patients treated at Basel University Hospital (2010–2024). Pathological features from whole-slide images were integrated with clinical data, such as patient age, tumor size, nodal status, ER, PR and HER2 status. The primary endpoint was disease free interval (DFI)defined as time from diagnosis to local or distant relapse. Risk stratification and prognostic accuracy were evaluated using C-indices, Kaplan-Meier analysis, and hazard ratios (HRs) for DFI compared to ODX. Results: From a total of 318 patients, 269 were included in the final analysis. Demographics and tumor characteristics are shown in Table 1. At a median follow-up of 61 months (IQR 46), 9% (n = 25) of patients experienced progression or relapse. The Ataraxis Breast demonstrated superior prognostic accuracy, with Kaplan-Meier analysis (Fig. 1) showing significantly reduced DFI in AI-identified high-risk patients (HR: 2.37, p = 0.032). In the ODX intermediate-risk cohort, survival curves were less distinct (HR: 1.61, p = 0.393) but still highlighted the AI model’s strength in broader populations. The model reclassified risk categories, assigning 33% of low-risk ODX patients to high risk, 77% of intermediate-risk to low risk, and 44% of high-risk to low risk, resolving intermediate-risk ambiguities (Fig. 2). Across subgroups by age (< 50 vs. ≥50 years), histology (IDC vs. ILC), and tumor size (< 2 cm vs. ≥2 cm), Ataraxis Breast consistently outperformed ODX, achieving a C-index of 0.803 in chemotherapy-treated patients and 0.750 in node-positive groups (Fig. 3). Calibration analysis showed strong agreement with observed recurrence probabilities (R² = 0.847). Conclusions: The Ataraxis Breast model provides a valid multimodal approach to BC prognostication compared to ODX. Its superior performance in high-risk subgroups and intermediate-risk reclassification underscores its potential for guiding BC treatment. Prospective validation is needed to confirm its clinical utility. Demographics and tumor characteristics. Demographics Median age 59 (29-85), (IQR 17) Tumor Characteristics Median tumor size 20 mm (3.5 – 130 mm), (IQR 16) pN1 n(%) 107 (40%) pN2 n(%) 8 (3%) Ductal n(%) 210 (78%) Lobular n(%) 47 (17.47%) Grade 2 n(%) 146 (54%) Grade 3 n(%) 92 (34%) Ki-67% low n(%) 181 (67%) Ki-67% high n(%) 88 (33%)