Artificial intelligence (AI)-based machine learning models (ML) for predicting pathological complete response (pCR) in patients with hormone receptor (HoR)-positive/HER2-negative early breast cancer (EBC) undergoing neoadjuvant chemotherapy (NCT): A retrospective cohort study

597 Background: The rates of pCR after NCT in HoR positive/HER2 negative EBC are low. Thus, clinicians need validated, unbiased, and reliable tools to better predict pCR at individual patients’ level. In this regard, AI can generate ML algorithms with these features, using clinical and pathological characteristics. Methods: Medical records were retrospectively retrieved for patients with EBC receiving NCT followed by surgery. Missing data were imputed using chained random forests. pCR was defined as absence of residual invasive cancer on pathologic evaluation of the breast specimen and lymph nodes (ypT0/isN0). Differences between pCR and non-pCR patients were assessed using t test or chi-squared test. 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 to prevent overfitting. Models performance was evaluated using the area under the receiver operating characteristics curve (AUC). Features importance was assessed. Disease-free survival (DFS) was evaluated using the Kaplan-Maier method and hazard ratios (HRs) using Cox regression. Results: 572 patients were included: 332 (58%) were pre-menopausal and median age was 49.0 (IQR 43-57); 330 (58%) were T2 and 273 (48%) were N1; 437 (76%) had a ductal adenocarcinoma and 308 (54%) were grade 3. Median estrogen and progesterone receptor (ER/PgR) expression was 90 (IQR 80-90) and 57.5 (IQR 5-90); median Ki67 was 35 (IQR 25-55) and 230 (40%) patients were HER2 zero. 565 (99%) patients received a combination of anthracyclines and taxanes; 434 (76%) received sequential and 137 (24%) concomitant CT. pCR was achieved in 87 patients (15%, 95% CI 12-18). Ten variables were included in the model: menopausal status, age, histology, grade, clinical T and N stage, ER/PgR status, Ki67 and HER2 status (zero/low). Among the evaluated models, the RF algorithm had the best performance: the AUC was 0.77 (95% CI 0.71-0.83) and sensitivity and specificity were 0.86 (95% CI 0.82-0.88) and 0.56 (95% CI 0.46-0.66). Variables with the highest importance were Ki67, ER/PgR status, age and nodal status. 511 patients were evaluable for survival: patients with pCR had a significant longer DFS compared to those who did not achieve pCR (HR 0.30; 95% CI 0.14-0.65, p = 0.002). Patients whose pCR was predicted by the model had longer DFS compared to those who pCR was not predicted (HR 0.56, 95% CI 0.21-0.87, p = 0.01). Conclusions: The RF-ML algorithm combining clinical and pathological characteristics has the potential to predict pCR in HoR positive/HER2 negative patients undergoing NCT, thus supporting clinicians to individualize treatment for EBC.