Predicting Neoadjuvant Treatment Response in Triple-Negative Breast Cancer Using Machine Learning
Shristi Bhattarai1, Geetanjali Saini1, Hongxiao Li2
1Department of Clinical and Diagnostic Sciences, School of Health Professions, University of Alabama at Birmingham, Birmingham, AL 35294, USA.
Diagnostics (Basel, Switzerland)
|January 11, 2024
Summary
Machine learning models can predict neoadjuvant chemotherapy (NAC) response in triple-negative breast cancer (TNBC) by combining biomarkers. This approach improves patient stratification for better therapeutic decisions.
Area of Science:
- Oncology
- Biomarker Discovery
- Computational Pathology
Background:
- Neoadjuvant chemotherapy (NAC) is standard for early-stage triple-negative breast cancer (TNBC), aiming for pathological complete response (pCR).
- Current pCR rates with NAC in TNBC are suboptimal (30-40%), necessitating improved predictive biomarkers.
- Tumor-infiltrating lymphocytes (TILs), Ki67, and phosphohistone H3 (pH3) are known predictive markers, but their combined value is under-evaluated.
Purpose of the Study:
- To comprehensively evaluate the predictive value of combined biomarkers from H&E and IHC stained biopsy tissue for NAC response in TNBC.
- To develop and assess a supervised machine learning (ML)-based approach for predicting NAC response using histological and molecular features.
- To enable precise patient stratification for guiding therapeutic decisions in TNBC.
Main Methods:
- Serial biopsy sections (n=76) were stained with H&E and IHC for Ki67 and pH3, generating whole-slide image (WSI) triplets.
- Mask region-based CNN (MRCNN) models identified tumor cells, stromal/intratumoral TILs (sTILs/tTILs), Ki67+, and pH3+ cells within WSIs.
- Hotspot regions were identified, and ML models were trained and evaluated using accuracy, AUC, and confusion matrix analyses to predict NAC response.
Main Results:
- Highest prediction accuracy was achieved using hotspot regions defined by tTIL counts, incorporating tTILs, sTILs, tumor cells, Ki67+, and pH3+ features.
- Combining multiple histological features (tTILs, sTILs) with molecular biomarkers (Ki67, pH3) yielded top performance at the patient level, irrespective of hotspot selection.
- The complementary use of various biomarkers significantly enhanced prediction accuracy for NAC response.
Conclusions:
- Prediction models for NAC response in TNBC should integrate multiple biomarkers rather than relying on single markers.
- Machine learning-based models demonstrate strong potential for accurately predicting NAC response in TNBC patients.
- This study supports the clinical utility of ML-driven biomarker combinations for personalized treatment strategies in TNBC.
