Prostate cancer is characterized by an immunosuppressive tumour environment. This work combines Raman spectroscopy with group-and-bases-restricted non-negative matrix factorization (GBR-NMF) and machine learning to assemble models of immune cell densities within the needle-core biopsies of patients undergoing high-dose-rate brachytherapy (HDR-BT). Raman spectral acquisition, as well as immunohistochemistry staining of CD68 \(^+\) , CD3 \(^+\) , and \(\hbox {CD8}^+\) cells, was completed for biopsies collected before and 2 weeks following the first fraction of HDR-BT. Regression techniques, constructed using GBR-NMF scores, that produced the most accurate predictions of immune cell density by metrics of root mean-squared error (RMSE) and R \(^{2}\) were the gradient-boosted trees model of \(\hbox {CD68}^{+}\hbox {+ CD3}^{+}\) density (RMSE: 163 counts \(\hbox {/mm}^2\) , \(\hbox {R}^{2}\) : 0.65) and the elastic net model of \(\hbox {CD68}^{+}\) / \(\hbox {CD8}^{+}\) (RMSE: 0.25, \(\hbox {R}^{2}\) : 0.82). The accuracy of these models, herein defined as the fraction of patient predictions within \(\pm 1\) standard deviation of their measured values was 11/16 and 12/16, for CD68 \(^{+}+\) CD3 \(^{+}\) and CD68 \(^{+}\) / CD8 \(^{+}\) models, respectively. To further delineate which metabolites were most important in the CD68 \(^{+}\) / CD8 \(^+\) model, this ratio was further predicted in stromal and epithelial tissues within the biopsies, and resulting models utilized the GBR-NMF scores of glutathione, collagen, palmitic acid, and the pre- or post-HDR-BT label to produce an optimal performance level according to RMSE and R \(^2\) . In summary, this study illustrates a novel methodology in which supervised machine learning techniques are used to model immune cells, which are prognostic indicators of disease progression.
