Differentiating multivariate heterogeneous cell populations by Raman spectroscopy based on feature selection

The data obtained from Raman spectroscopy is characterized by high dimension and complexity. In this study, dimensional reduction techniques, including LASSO, random forest, and principal component analysis, were employed to refine the datasets. An optimal model was then constructed using Artificial Neural Networks, which achieved an average accuracy of 0.943 when subjected to 10 times 5-fold cross-validation. Based on the feature importance ranking map of four algorithms, the species differences of five subgroups of leukocytes induced by nucleic acid and protein vibrations were found. The visual representation between feature selection methods and variations in Raman shift characteristics demonstrates the intricate interplay and serves as a robust foundation for our forthcoming "Raman intra-group correlation analysis".