Artificial intelligence for preoperative diagnosis of sentinel lymph node biopsy in early breast cancer.

Purpose: Sentinel lymph node biopsy (SNB) is used for axillary staging in early breast cancer. Benign SNB lacks therapeutic benefit whereas a malignant SNB results in added cost and risks of a second procedure if axillary clearance is undertaken. SNB preoperative prediction could avoid unnecessary surgery or facilitate a single-stage procedure. We aimed to create and evaluate AI models for clinically node-negative patients to predict SN status (benign/malignant). Methods: Data was collected prospectively from the Breast Cancer Foundation National Register NZ. 303 patients that underwent SNB were included. 4 classification algorithms: k Nearest Neighbour (kNN), Support Vector Machine (SVM), Decision Tree (DT), Naïve Bayes (NB) were selected for training. SN histology was used as a target; micrometastasis and larger deposits were defined as malignant. Preoperative data used were: age, tumour type, side, size, grade, location, multifocality and ER/PR/HER2 status. K fold cross validation, partitioning the cohort into k fold subsamples (k=10) was utilised. k-minus-1 fold (90%) was used for training and the remaining k fold (10%) for validation. The partitioning occurred 10 times over. Results: Mean age was 56 and mean tumour size was 21mm. 27% (81/303) had malignant SLN. 22% (66/303) had multifocality. Classifiers that performed well in predicting final SNB histology were: kNN, in one test fold had accuracy, sensitivity, specificity, AUC of 73%, 75%, 73% and 76% respectively and DT's results were 73%, 75%, 73%, 76%. NB and SVM's best accuracy was 69% and 73% respectively. Conclusion: Classifiers showed potential as decision-supporting tools for predicting SNB status. Further work is underway to improve the model for clinical utility.