<p>Fatal disorders associated with cancer are prevalent in many industrialized and developing nations across the world. It is especially true for women, where the daily incidence of breast cancer rises partly as a result of early diagnosis errors and general ignorance. Early detection and accurate classification of the cancer must occur in order to provide an appropriate initial course of treatment for breast cancer. Artificial neural networks (ANNs) are a powerful process of classifying for medical data and are essential for diagnosing illnesses. The learning algorithm selection has a significant impact on how well ANN training works and a number of approaches have shown potential. To train ANNs for classifying breast cancer data, present work carries out an empirical investigation that focuses on performance of several metaheuristic (MH) algorithms as learning techniques. The experimental in this article makes use of the well-known Wisconsin breast cancer dataset for binary classification (benign or malignant). Eight newly developed parameter-less swarm-based MHs, namely GAO, COA, OOA, WOA, ZOA, POA, NGO, and TDO, have been used as learning algorithms, and their performance is evaluated against the traditional backpropagation (BP) neural network. Performance is evaluated based on several factors, including accuracy, sensitivity, specificity, precision, geometric mean (GM), F-measure, and false-positive rate (FPR), using box plots. Also, binary entropy loss versus iteration has been explored using convergence graph. Finally, the ideal ANN model for accurate classification is found using the TOPSIS method, a multi-attribute decision-making (MADM) technique. According to studies, the northern goshawk optimization (NGO) algorithm performs better than any other algorithm. These results highlight how MHs, particularly the NGO, can improve ANN classification performance for medical data, especially when it comes to breast cancer diagnosis.</p>