Bold take: a multimodal AI approach combining image, clinical, and molecular data can significantly improve long-term recurrence prediction in early breast cancer—and that could change how we tailor treatments. But here’s where it gets controversial: can algorithms reliably outperform established genomic tests like Oncotype DX in real-world practice? And this is the part most people miss: how do we weigh AI-driven risk against clinician judgment and patient preferences?
A new analysis of the phase 3 TAILORx study shows that the multimodal ICM+ model, which integrates whole slide image data (I), clinical features (C), and enhanced molecular analysis (M+), demonstrated superior prognostic value for long-term recurrence compared with Oncotype DX alone. In both training and validation cohorts, ICM+ improved prediction for overall and late distant recurrence, with pathomic data from tissue slides proving particularly influential for predicting late recurrences. These results suggest potential for more precise risk stratification and personalized treatment decisions in early breast cancer.
Key findings include:
- ICM+ outperformed Oncotype DX for overall distant recurrence (C-index around 0.705 vs 0.617 for Oncotype DX; statistically significant) and for late distant recurrence (C-index about 0.656 vs 0.518 for Oncotype DX).
- The CM+ model (clinical plus molecular data) also improved prognostic performance over traditional methods, but ICM+ generally provided the strongest discrimination, especially for late recurrences.
- In validation, ICM+ continued to show superior prognostic value for overall distant recurrences at 15 years (C-index 0.733 with ICM+ vs 0.631 for Oncotype DX) and for late recurrences after 5 years (C-index 0.705 vs 0.527).
TAILORx data details (high-level):
- The study analyzed 10,273 HR-positive, HER2-negative, axillary node-negative breast cancer patients, stratified by Oncotype DX risk scores. Among them, 69% had intermediate RS (11–25), 17% had RS ≤10, and 14% had RS ≥26.
- In the intermediate-risk group, endocrine therapy alone was noninferior to chemotherapy plus endocrine therapy for invasive DFS overall, with some subgroup benefits from chemotherapy for those over 50 or RS 16–25. Other endpoints, including distant recurrence-free survival and overall survival, were also noninferior.
Design and training highlights:
- AI training used digitized H&E slides (40x) and whole transcriptome data from 4,462 primary tumors in TAILORx. A 5-fold nested cross-validation framework split data into training and validation sets, ensuring robust evaluation.
- Expanded molecular analysis examined multiple gene signatures (Oncotype DX, MammaPrint, Prosigna, EndoPredict, BCI) and 57 high-variance genes, with the final enhanced molecular model (M+) incorporating EndoPredict, BCI, and Oncotype DX gene signatures (42 genes total).
- The aim was to achieve better prognostic information than Oncotype DX alone, especially for late distant recurrences (>5 years).
Performance metrics:
- For overall distant recurrence, ICM+ achieved a C-index of 0.705 (significantly better than Oncotype DX, P < .001).
- For late distant recurrence, ICM+ reached a C-index of 0.656 (P < .001 vs Oncotype DX).
- For early distant recurrence, ICM+ did not significantly outperform Oncotype DX (C-index 0.765 vs 0.738; P = .398).
- In multivariate analysis, ICM+ showed significant risk discrimination for overall, early, and late distant recurrence.
Validation insights:
- In the validation set, ICM+ outperformed Oncotype DX for 15-year overall distant recurrence (C-index 0.733 vs 0.631; P = .00049) and for late recurrences after 5 years (C-index 0.707 vs 0.527; P < .001).
- Pathomic imaging proved particularly valuable for late distant recurrence, with ICM+ outperforming multiple comparative models that lacked pathomic data.
Practical implications:
- AI-based pathomic tools can be deployed on standard tissue slides scanned by common equipment or even smartphones, enabling cost-effective central analysis.
- The added risk stratification from ICM+ could help clinicians tailor adjuvant therapies more precisely, potentially reducing overtreatment and improving long-term outcomes.
Important caveats and questions for the field:
- How will clinicians integrate AI-derived risk scores with patient values, comorbidities, and preferences in real-world decision-making?
- What are the implications for health equity, given potential disparities in access to advanced imaging and computational resources?
- How can we ensure interpretability and clinician trust in AI-driven risk stratification, particularly when recommendations diverge from traditional assays?
Thought-provoking prompt for the audience: Do you think AI-enhanced models like ICM+ should replace, augment, or be reserved for specific scenarios alongside current genomic tests? What safeguards would you want to see before adopting such an approach in routine care?
