PQA 09 - PQA 09 Hematologic Malignancies and Digital Health Innovations Poster Q&A
3420 - A Machine-Learning Model to Predict Axillary Node Positivity from Clinical Characteristics and Large-Scale DNA Organization Features in Breast Cancer Biopsy Specimens
N. Alabi1, M. Guillaud1, S. El Hallani2, F. Inaba1, and A. Nichol3,4; 1BC Cancer Research Centre, Vancouver, BC, Canada, 2Alberta Precision Labs, Edmonton, AB, Canada, 3University of British Columbia, Vancouver, BC, Canada, 4Department of Radiation Oncology, BC Cancer, Vancouver, BC, Canada
Purpose/Objective(s): Breast cancer is the leading cause of cancer-related fatalities among women globally. Although sentinel lymph node biopsy (SLNB) is used routinely for clinically node-negative patients, over 60% lack axillary lymph node metastases, so SLNB upstages only a minority. Consequently, there has been ongoing research into non-invasive methods, such as ultrasound, MRI, and PET, which are increasingly used in staging axillary lymph nodes. Large-scale DNA organization (LDO) analysis involves measuring various characteristics of nuclear DNA organization, including size, shape, and chromatin organization texture. LDO analysis has shown correlations with diverse clinical outcomes in multiple cancers, including survival in breast cancer patients. We hypothesized that a machine learning-based tool to predict axillary lymph node metastases using LDO would perform better than nomograms based on routine clinical features.Materials/
Methods: This study utilized a training cohort of tissue microarray (TMA) core samples from 347 breast cancer patients and an external validation cohort from a different hospital comprising TMA core samples from 157 patients. The TMA slides underwent staining with Feulgen-thionin, followed by automated segmentation of cell nuclei and selection of intact, in-focus nuclei. LDO features of each nucleus were calculated and input into a Random Forest model to distinguish between nuclei from patients with and without lymph node metastases. The predictions for each patient were amalgamated into a score and combined with clinical features in the Random Forest model to classify their lymph node metastasis status. Results: During model training, several LDO features, such as nuclear radius and intensity, were identified as crucial in predicting axillary lymph node metastases. The LDO features emerged as the most important variables in the Random Forest model trained on LDO and clinical features. On the external validation set, the area under the receiver operating characteristic curve (AUC) for the model trained using LDO features alone was 0.711, and the AUC for the model trained on clinical features alone was 0.741. The AUC of our final model that used both clinical and LDO features was 0.775, outperforming the established Memorial Sloan Kettering and MD Anderson nomograms for node positivity on the same data. Conclusion: Our model identified several key nuclear features that can predict lymph node metastasis status from a biopsy. Our study presents the first histology-based, explainable machine-learning model capable of predicting axillary lymph node metastases.
