276P - Development and external validation of an artificial intelligence (AI)-based machine learning model (ML) for predicting pathological complete response (pCR) in hormone-receptor (HoR)-positive/HER2-negative early breast cancer (EBC) undergoing neoadjuvant chemotherapy (NCT)

Background

In HoR-positive/HER2-negative EBC pCR rates are low and ML algorithms could help to predict pCR using clinical and pathological characteristics.

Methods

Patients with EBC receiving NCT followed by surgery were included in the training/validation cohorts (TC/VC). pCR was defined as absence of residual invasive cancer on pathologic breast specimen and lymph nodes (ypT0/is ypN0). Eight ML models (c5.0, k-nearest neighbour, random forest [RF], neural network, support vector machine [linear/radial], boosted trees and boosted logistic regression) were trained and tuned. A stratified ten-fold cross-validation was used and models’ performance was evaluated using the area under the receiver operation characteristics curve (AUC), sensitivity and specificity.

Results

572 patients were included in the TC and 87 achieved pCR (15%, 95% CI 12-18%). Ten variables (menopausal status, age, histology, grade, clinical T/N, ER/PgR, Ki67 and HER2 [zero/low]) were included according to multivariate analysis and clinical relevance. The RF algorithm had the best performance: AUC 0.77, 95% CI 0.71-0.83; sensitivity 0.86 (95% CI 0.82-0.88) and specificity 0.56 (95% CI 0.46-0.66). Variables with the highest importance were Ki67, ER/PgR status, age and nodal status. 293 patients were included in the VC and 24 achieved pCR (8%, 95% CI 6-12%). The RF algorithm had an AUC of 0.71 (95% CI 0.61-0.81), sensitivity of 0.29 (95% CI 0.13-0.51) and specificity of 0.90 (95% CI 0.86-0.94). With a median follow-up of 53.3 and 57.7 months for DFS and OS, patients whose pCR was predicted by the model had longer DFS (HR 0.51, 95% CI 0.31-0.83, p=0.003) and OS (HR 0.48, 95% CI 0.29-0.79, p=0.001). A one percentage increase in the probability of predicted pCR was significantly associated with longer DFS (HR 0.992; 95% CI 0.986-0.999, p=0.02) and OS (HR 0.992, 95% CI 0.985-0.998, p=0.01).

Conclusions

The ML algorithm has the potential to predict pCR in HoR-positive/HER2-negative patients undergoing NCT, thus supporting clinicians to individualize treatment for EBC.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.