Author: admin

Multimodal artificial intelligence (AI) models using a combination of molecular, imaging, and clinical features improved the individual prognostic assessment of patients with early breast cancer’s risk of distant recurrence, according to an analysis presented at the 2025 San Antonio Breast Cancer Symposium (Abstract GS1-08). 

The team assessed various models and found that the addition of molecular features significantly strengthened the prognostic accuracy for early distant recurrence, while histopathologic data drove improvements in accuracy for late distant recurrence. 

 

“This study shows the potential for how AI can be leveraged to develop better diagnostic tests that may more accurately estimate recurrence risk and individualize treatment decisions,” stated Joseph A. Sparano, MD, Chief of the Division of Hematology and Oncology at the Mount Sinai Tisch Cancer Center. 

Background and Study Methods 

The TAILORx trial established the use of Oncotype Dx 21-gene recurrence scores for guiding treatment with endocrine therapy with or without chemotherapy for patients with T1-2, N0 hormone receptor–positive/HER2-negative early breast cancer. Oncotype Dx is considered a prognostic indicator for 10-year distant recurrence, but has limited value for late distant recurrence for more than 5 years. 

Researchers developed several single and multimodal AI models to improve distant recurrence risk prognostication. The models combined clinical, molecular, and histopathological features to determine prognosis for early (<5 years), late (>5 years), and overall distant recurrence (15 years) with the use of primary tumor samples and clinical data from patients in the TAILORx trial who volunteered for the analysis. 

“Our goal was to develop a new diagnostic test that provides better prognostic estimation of recurrence risk, including late recurrence risk, by studying tumor specimens from the TAILORx trial,” Dr. Sparano said. “We developed an AI model that evaluates both the images of digitized slides used for routine pathologic assessment, plus the molecular and clinical characteristics of breast cancer to provide better prognostic information about cancer recurrence risk out to 15 years, including early recurrence within 5 years after diagnosis, and late recurrence after 5 years.” 

The researchers digitized slides from 4,462 primary tumor samples and nucleic acids were extracted and sequenced using Caris MI Tumor Seek–Hybrid. They used 63% of the samples for model training and 37% were saved for an independent validation set. 

Single-modality models included separate clinical, image, and molecular features; there was also an expanded molecular model, dual modality for combinations of the three modalities, and a multimodal model for all three combined. The expanded molecular model included 42 genes from the Oncotype DX, MammaPrint, Prosigna, EndoPredict, and Breast Cancer Index commercial gene signatures plus 57 high-variance genes. 

Continuous risk scores were separated into high-risk vs low-risk groups than aligned with the Oncotype DX recurrence score cutoff for distinguishing between high and low genomic risk. 

Key Findings 

Continuous Oncotype DX recurrence scores achieved a concordance index (C-index) of 0.617 for overall distant recurrence and 0.738 for early distant recurrence, but did not provide prognostic value for late distant recurrence (C-index = 0.518). When Oncotype DX was combined with clinical features, the C-index scores were similar for overall (0.600), early (0.706), and late (0.512) distant recurrence. 

The multimodal AI model combining clinical, molecular, and histopathological features performed best of all tested models in terms of prognostic performance for overall distant recurrence (C-index = 0.705; hazard ratio [HR] for high vs low risk = 3.6; P < .001) and late distant recurrence (C-index = 0.656; HR = 2.84; P < .001). For early distant recurrence, it performed second best after the dual-modality model of clinical and expanded molecular features (C-index = 0.776; HR = 5.6; P < .001). 

The expanded molecular model performed best of all single-modality models for early distant recurrence prognostication (C-index = 0.757), while imaging features were stronger for late distant recurrence prognostication (C-index = 0.637; HR = 1.9; P = .001). 

In the validation set, the multimodal AI model again outperformed Oncotype DX for overall distant recurrence through 15 years (C-index = 0.733 vs 0.631; P = .00049) and late distant recurrence after 5 years (C-index = 0.705 vs 0.527; P = .000031). 

“AI-based pathomic tools that rely on evaluation of tissue sample slides routinely generated from clinical practice can be captured with scanners or even widely available smartphones, uploaded electronically, and analyzed centrally—with minimal cost,” Dr. Sparano said. 

Disclosure: This research was a public-private partnership between the federally funded ECOG-ACRIN Cancer Research Group and Caris Life Sciences, supported by the Breast Cancer Research Foundation, the National Cancer Institute of the National Institutes of Health, and the U.S. Postal Service Breast Cancer Research Stamp Fund. Dr. Sparano serves as a consultant for AstraZeneca, Delphi Diagnostics, Genentech, Genomic Health/Exact Sciences, Novartis, and Pfizer; is a member of the scientific advisory board for PreciseDX; and receives institutional research support from Olema Oncology. For full disclosures of the other study authors, visit abstractsonline.com.