ESMO Congress 2023 | OncologyPRO

Abstract 332P

Background

Neoadjuvant chemotherapy (NAC) has been a standard treatment for locally advanced breast cancer (BC). However, the high heterogeneity of BC results in various tumor response. Therefore, there is a growing need for early assessment of tumor response, particularly for insensitive BC, to guide personalized treatment. We aim to develop a longitudinal radiomics multitask model to accurately predict RCB score in the early stage of NAC for breast cancer.

Methods

A total of 1048 patients with breast cancer receiving NAC across four institutions were retrospectively enrolled. We collected longitudinal MRI sequences at the pre-NAC and mid-NAC timepoints, and extracted 21804 radiomics features per patient. We used a multitask learning strategy to predict RCB score (RCB 0-I, II and III). The Mann-Whitney U-test, Spearman analysis, the least absolute shrinkage and selection operator regression and Boruta method were used to perform feature selection. We developed various base machine learning models, followed by an ensemble stacking method to integrate the base model outputs. The multitask learning model was subsequently verified in three independent external validation cohorts.

Results

Of the total patients, 442 (42.18%) reached RCB 0-I, 462 (44.08%) reached RCB II and 144 (13.74%) reached RCB III. 17 and 19 significant features were selected for two independent tasks. For identifying RCB 0-I and RCB II-III, the multitask model reached an area under the curve (AUC) of 0.932 in primary cohort, and AUCs of 0.890, 0.919 and 0.911 in the external validation cohorts. It also identified RCB II and RCB III with an AUC of 0.916 in primary cohort, and AUCs of 0.870, 0.899 and 0.871 in external validation cohorts.

Conclusions

The longitudinal radiomics multitask learning model is a noninvasive tool to predict RCB score for breast cancer, and help clinical decision-making in the early stage of NAC.