Ultrasound-based machine learning models for predicting response to neoadjuvant chemotherapy in breast cancer: A meta-analysis

Abstract

Background and aims

Breast cancer remains the most common cancer among women globally, with neoadjuvant chemotherapy (NAC) serving as a critical pre-surgical intervention. Ultrasound-based radiomics and machine learning (ML) models offer potential for early prediction of NAC response, aiding personalized treatment strategies. This study systematically reviews the efficacy of ultrasound-based ML models in predicting NAC response in breast cancer patients.

Methods

We conducted a systematic review and meta-analysis following PRISMA-DTA guidelines, searching PubMed, Scopus, Web of Science, and Embase up to August 30, 2023. Studies developing ultrasound-based radiomics or deep learning (DL) models to predict NAC response were include. Models for complete and partial response were analyzed separately.

Results

Twenty-two studies were included. For models predicting complete response, pooled sensitivity, specificity, and AUC were 85.1 % (95 % CI: 79.2–89.6 %), 85.8 % (95 % CI: 76.7–91.8 %), and 86 % (95 % CI: 82 %–94 %), respectively for internal validation and 82.9 % (95 % CI: 76.2 % - 88.1 %), 89.4 % (95 % CI: 84.7 %–92.9 %), and 93 % (95 % CI: 82 %–94 %), respectively for external validation. For partial response, analysis could only be performed on internal validation and the pooled sensitivity was 87.5 % (95 % CI: 85.1–89.6 %) with pooled specificity of 82.3 % (95 % CI: 75.6–87.5 %), and pooled AUC of 88 % (95 % CI: 85 %–92 %).

Conclusion

Ultrasound-based ML models show strong potential for predicting NAC response in breast cancer, with delta radiomics enhancing predictive accuracy. Further research is needed to develop clinically generalizable models.