The early detection of cancers can be crucial to achieving a positive therapeutic outcome and in recent times, machine learning methods adopting image-based algorithms have been used, with varying degrees of accuracy. In this paper, the focus is on the early detection of breast cancer using biological samples known as exosomes, which are subsequently processed using Surface-Enhanced Raman Spectroscopy to transform the samples into datasets suited to a wide range of machine learning functions. As this represents a novel approach to cancer detection, it is not clear if the raw featureset delivers best predictive performance or if feature engineering can improve the performance of detection algorithm. Furthermore, the process required to generate the datasets is time consuming with the result that only small datasets, which may not be suitable to some machine learning functions, are available. In this paper, we present a study which attempts to find both an optimal featureset and best performing algorithm in early stage breast cancer detection.