Multimodal AI model augments Oncotype DX in predicting distant recurrence in HR+/HER2 node negative BC

Share :
Published: 22 Dec 2025
Views: 124
Rating:
Save
Dr Joseph Sparano - Icahn School of Medicine at Mount Sinai, New York, USA

Dr Joseph Sparano speaks to ecancer about multimodal artificial intelligence (AI) models integrating image, clinical, and molecular data for predicting early and late breast cancer recurrence in TAILORx

Using data from thousands of TAILORx participants, researchers developed AI models integrating clinical data, molecular signatures, and digital pathology to improve prediction of distant recurrence risk in HR+/HER2- early breast cancer.

While Oncotype DX remained useful for early recurrence, it showed limited value beyond 5 years.

The multimodal ICM+ AI model significantly improved accuracy for overall and late recurrence prediction compared to traditional tools, helping better identify which patients may need extended endocrine therapy and enabling more personalised long-term risk assessment.

This study evaluated biospecimens from patients enroled on the TAILORx trial and evaluated whole transcriptome and exome sequencing plus pathomic evaluation to determine whether we could develop a diagnostic test that provided improved prognostic risk stratification for distant recurrence compared with the 21-gene Oncotype DX recurrence score. Ultimately we studied primary tumour specimens from approximately 4,600 patients who had whole exome and transcriptome sequencing performed and also H&E slides were digitised for pathomic evaluation.

An artificial-intelligence-based neural network platform was utilised for evaluating the whole slide imaging for a pathomic score associated with recurrence and also evaluated the genes, a total of 187 genes from five commercially available gene expression datasets. The tool that was developed, the diagnostic tool that was found to provide the best risk stratification was one which integrated the imaging pathomic information, the clinical information, namely age, tumour size and grade, as well as the molecular information which included a 42-gene expression set that included genes in the 21-gene recurrence score, the Breast Cancer Index and the EndoPredict assay.

We evaluated the prognostic performance of this newly developed AI-derived score and compared it to the Oncotype DX recurrence score using C-index log-ranked tests and multivariate Cox proportional models.  The major conclusions were that the integrated ICM model, that included pathomic imaging, clinical pathologic features and this expanded molecular set, provided more robust prognostic information than the 21-gene recurrence score used alone or in combination with clinical pathologic features for both overall recurrence at 15 years and late distant recurrence beyond five years.

The imaging pathomic component of the analysis was particularly informative for late recurrence beyond five years, as was the expanded molecular signature that was evaluated in this study.

Our final conclusions were that we developed a multimodal test that integrated pathomic imaging features, clinical features and molecular features that can provide more prognostic information and better risk stratification for distant recurrence at 15 years and also after five years than the Oncotype DX 21-gene recurrence score. We hope that this new diagnostic test will become available in the next year or so for clinical use.