Context & Study Overview

TNBC is a heterogeneous and aggressive subtype of breast cancer associated with poor outcomes. As a standard of care, patients with TNBC receive NAC therapy, allowing them to undergo breast-conserving surgery and achieve pCR, which is a strong predictor of favorable outcomes.

However, predicting pCR before treatment initiation remains a significant unmet need, showcasing the interest in having a method that can detect treatment response earlier in the patient journey. As an example, FDG-PET/CT has shown potential to detect residual disease early and predict poor outcomes in TNBC.

This proof-of-concept study, designed and developed by SOPHiA GENETICS in collaboration with AP-HP Hôpital Saint Louis and INSERM, aims to develop a multimodal machine learning (ML) model predictive of pCR, integrating multiple patient baseline parameters, such as clinical, histopathological, genomic, and PET imaging.

Dataset and Methods

In this retrospective study, we have analyzed the data from 57 patients with stage II-III TNBC treated at Saint Louis Hospital, in Paris, between 2008 and 2015, with baseline PET and NAC followed by breast surgery.

A total of 241 baseline features were extracted from each patient, out of which 17 were selected for final analysis, encompassing 3 clinical, 2 histopathological, 3 molecular, 3 PET non-radiomic, and 6 radiomic parameters.

Using support vector (SVMs) with nested leave-pair-out cross-validation, the researchers developed an ML model optimized to predict pCR. Model performance was assessed using area under the curve (AUC), and Kaplan-Meier analysis was used to explore associations with event-free survival (EFS).

Key Findings

The data collected resulted in 241 predictors per patient, out of which 17 pretherapeutic features were selected for the final analysis.

The best-performing model was a support vector machine (SVM) algorithm with an AUC of 0.82 (95% CI = 0.74-0.90), using the aggregation of all multimodal features (clinical data, histopathological, molecular features, and PET data, including radiomic features). This estimated AUC was significantly greater when considering the entire set of multimodal data, resulting in a 10% decrease in the AUC when a specific data modality was excluded.

The features with the highest weight in the algorithm included a PET-derived SUVmax, two radiomics features, the genomic grade index (GGIr), and clinical T-stage.

Patients predicted to achieve pCR showed an improved event-free survival (EFS) compared to patients with predicted non-pCR. However, due to the small sample size of the study, it requires further research to validate this result.

Conclusion and Clinical Implications

This study suggests that applying an ML-based method to baseline multimodal data can help predict pCR status after NAC for TNBC patients and may identify correlations with improved long-term outcomes. The integration of clinical, histopathological, molecular, imaging, and radiomic features resulted in a robust AI-powered predictive model.

Identifying patients at greater risk of not achieving pCR earlier could enable a rapid treatment modulation by introducing concomitant treatment with immunotherapy or dose intensification, increasing their chances for better outcomes.
Although the results obtained need an external validation in a large, multicenter population, this study highlighted the importance of adopting a truly multimodal approach to provide better, more efficient, and personalized care to TNBC patients.

Explore this infographic to learn more about this project and the predictive model developed by Groheux et al’s.

Groheux D. et al. Cancers (Basel). 2025. 17(7):1249. doi: 10.3390/cancers17071249.

SOPHiA DDM™ for Radiomics and SOPHiA DDM™ for Multimodal are concepts in development. May not be available for sale.