Deep learning-based prediction model for pathological complete response to neoadjuvant chemotherapy in different breast cancer subtypes, incorporating peripheral blood inflammatory indicators and clinicopathological features: a multicenter retrospective study

Background We aim to develop and validate a deep learning pathomics clinical model (DLPC) for predicting pathological complete response (pCR) following neoadjuvant chemotherapy (NAC) in breast cancer patients with different subtypes using peripheral blood inflammatory indicators, clinical indices, and hematoxylin and eosin (H&E)-stained biopsy slides. Method In this multicenter retrospective study, eligible participants who had undergone NAC were enrolled from two independent institutions. Handcrafted and deep learning pathomics features reflecting the pathomics tumour nucleus and pathomics microenvironment were respectively extracted from whole slide images (WSIs) of H&E-stained biopsy slides.The DLPC model was constructed using three feature sets associated with pCR, including clinical signature (CS), handcrafted and deep learning pathomics signature (HPS, DLPS) from a training cohort. The performances of DLPC were evaluated using area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). We also evaluated the diagnostic efficacy of the DLCP model for four different molecular subtypes. Result Between January 2020 and August 2023, a total of 531 breast cancer patients were divided into a training cohort (291 patients), an internal validation cohort (101 patients), and an external validation cohorts (139 patients).The DLPC outperformed single-scale prediction models with AUC values of 0.872 (95% CI 0.831-0.913) in the training cohort, 0.840 (95% CI, 0.758-0.922) and 0.867 (95% CI, 0.808-0.925) in the internal and external validation cohort. The DLPC achieved an accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of 82.13%, 77.05%, 85.80%, 79.66% and 83.82% in the training cohort, and 80.20%, 83.05%, 76.19%, 83.05% and 76.19% in the internal validation cohort, and 81.29%, 88.0%, 77.53%, 68.75% and 92.0% in the external validation cohort. The performance of DLPC were evaluated in four different breast cancer subtypes with the best prediction for Luminal-B breast cancer subtype (AUC: 0.856). The decision and calibration curves of DLPC also confirmed the excellent performance of discrimination, calibration and clinical utility. Conclusions DLPC showed excellent performance in predicting pCR after NAC for breast cancer, which highlighted the potential to provide individual decision-making for breast cancer patients,allowing for timely adjustment of chemotherapy regimens.