Evaluation of machine learning and deep learning models for the classification of a single extracellular vesicles spectral library

Single extracellular vesicles (EVs) carry molecular signatures from their cell of origin, making them a pivotal non-invasive biomarker for cancer diagnosis and monitoring. However, analyzing the complex data associated with single-EVs, such as fingerprints generated via Surface-enhanced Raman Spectroscopy (SERS), remains challenging. To address this, a thorough comparison of machine learning models' implementations and their accuracy classification optimization is presented. A comprehensive single-EV spectral library collected with a SERS-assisted nanostructured platform including cell lines, healthy controls, and cancer patient samples is used. The performance of different learning models (random forests, support vector machines, convolutional neural networks, and linear regression as reference) was assessed for cancer detection tasks: i) multi-cell line classification and ii) cancerous versus non-cancerous binary classification. To improve their accuracy, we optimized spectra preprocessing, artificially increased the dataset, and implemented feature-driven classification. In sum, these methods enabled more interpretable models to perform on par with the complex one, increasing accuracy up to 12% percent-age points, even with datasets reduced to 66% of the original size. Achieving accuracies of 83% and 91% for Task-i and Task-ii, respectively.