Federated learning for predicting histological response to neoadjuvant chemotherapy in triple-negative breast cancer

Jean Ogier du Terrail ¹, Armand Leopold ², Clément Joly ³

⁰Owkin, Inc., New York, NY, USA. jean.du-terrail@owkin.com.

Nature Medicine +

|

January 19, 2023

View abstract on PubMed

Summary

Machine learning predicts treatment response in triple-negative breast cancer (TNBC) using whole-slide images. Federated learning enhances model performance, offering a privacy-preserving approach for biomarker discovery in rare cancers.

Area Of Science

Oncology
Computational Pathology
Artificial Intelligence

Background

Triple-negative breast cancer (TNBC) presents a significant clinical challenge due to its aggressive nature, high metastatic potential, and limited therapeutic strategies.
Neoadjuvant chemotherapy (NACT) is standard for early TNBC, but patient response heterogeneity necessitates predictive biomarkers.
Current understanding of TNBC treatment response variability is hindered by limited data availability and privacy concerns.

Purpose Of The Study

To develop and validate a machine learning (ML) model for predicting histological response to NACT in early-stage TNBC patients using whole-slide images (WSIs) and clinical data.
To investigate the efficacy of federated learning (FL) in enhancing ML model performance for TNBC response prediction while preserving patient data privacy across multiple institutions.
To demonstrate the interpretability and potential for biomarker discovery of the developed ML model.

Main Methods

A multicentric study involving federated learning was conducted to train ML models on WSIs and clinical data from TNBC patients undergoing NACT.
Local ML models were trained on individual hospital datasets, followed by collaborative training using FL to aggregate learnings without sharing raw patient data.
The performance of FL-trained models was compared against locally trained models and established annotation-based approaches.

Main Results

Local ML models utilizing WSIs demonstrated the ability to predict NACT response in early TNBC.
Collaborative training via federated learning significantly improved the predictive performance of ML models, achieving results comparable to expert-annotated methods.
The developed ML model proved interpretable and sensitive to specific histological patterns associated with treatment response.

Conclusions

Federated learning offers a robust, privacy-preserving framework for developing high-performing ML models for TNBC treatment response prediction using real-world data.
This approach overcomes data limitations in rare cancers and facilitates biomarker discovery through analysis of large, aggregated datasets.
The interpretable ML model holds promise for clinical decision support and advancing personalized medicine in TNBC.