Backpropagation　(BP)　is　a　widely　embraced　method　for　training　artificial　neural　networks　(ANNs)　in　classification　and　regression　tasks.　However,　its　efficacy　diminishes　when　confronted　with　complex　problems　due　to　susceptibility　to　local　optima,　sensitivity　to　initial　parameters,　and　slow　convergence.　Arithmetic　optimization　algorithm　(AOA)　has　emerged　as　a　promising　solution　for　achieving　global　optimization　in　ANNs.　Nonetheless,　similar　to　other　metaheuristic　search　algorithms　(MSAs),　AOA　grapples　with　inconsistent　solution　quality　resulting　from　random　initialization,　impeding　its　effectiveness　in　intricate　optimization　tasks.　This　paper　introduces　a　modified　initialization　scheme　for　AOA　that　incorporates　chaotic　maps　and　oppositional-based　learning　(OL)　to　yield　a　superior　initial　population,　leading　to　a　new　AOA　variant　known　as　AOA-COIS.　Serving　as　a　training　algorithm　for　ANNs,　AOA-COIS　optimizes　the　combination　of　weights,　biases,　and　activation　functions　to　enable　accurate　classification.　Rigorous　simulations　evaluate　AOA-COIS　using　ten　publicly　available　medical　datasets,　comparing　its　performance　with　seven　state-of-the-art　MSA-based　ANN　training　algorithms.　The　findings　substantiate　the　remarkable　capabilities　of　AOA-COIS　in　training　ANN　models,　achieving　outstanding　classification　accuracy　across　most　selected　medical　datasets.　©　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.　2024.