Latent class analysis of conventional ultrasound features: a novel approach to predicting non-response to neoadjuvant chemotherapy in breast cancer.

OBJECTIVE

To determine if holistic patterns derived from conventional ultrasound features are associated with pathological response to neoadjuvant chemotherapy (NAC) in breast cancer, providing a potential basis for non-invasive treatment monitoring.

METHODS

This study included 509 female breast cancer patients with unilateral primary lesions who underwent NAC followed by surgery. We employed three clustering methodologies to integrate conventional ultrasound features into holistic patterns: the model-based latent class analysis (LCA) and the distance-based K-modes and hierarchical clustering. The association between the resulting patterns and pathological non-response was assessed using Firth logistic regression, with internal validation performed through 10-fold cross-validation and 10, 000 bootstrap resampling.

RESULTS

When comparing the three approaches, LCA was identified as the superior classification framework, providing the best balance of statistical model fit, class separation, and meaningful association with clinical outcomes. The optimal LCA model identified two distinct sonographic patterns, referred to as Pattern A and Pattern B. Pattern B was significantly associated with a 2.51-fold higher (95% CI: 1.05-5.99) likelihood of pathological non-response compared to Pattern A. In subgroup analyses, among defined high-risk groups, Pattern B was linked to an increased risk of pathological non-response ranging from 12.84% (95% CI: 5.74%-23.59%) to 21.34% (95% CI: 13.63%-28.32%), whereas in non-high-risk groups the corresponding increase ranged from 1.12% (95% CI: 0.79%-1.51%) to 3.45% (95% CI: 2.60%-4.47%).

CONCLUSION

Patterns of conventional ultrasound features derived through LCA may offer a preliminary indication for identifying breast cancer patients less likely to benefit from NAC, potentially informing more individualized treatment planning following external validation.