<p>Definitive diagnosis of intracranial tumors by less-invasive approaches provides diagnostic information to facilitate personalized treatment decisions and avoid unnecessary invasive biopsy operations. However, it remains challenging, largely due to a lack of reliable molecular biomarkers. Here, we demonstrated that molecular profiles of small extracellular vesicles (sEVs) in cerebrospinal fluid (CSF) decoded by surface-enhanced Raman spectroscopy (SERS) revealed highly specific signatures to detect and accurately discriminate common primary intracranial tumors in children that can be challenging to distinguish using standard-of-care imaging. Specifically, the fabrication of silver nanocube-based three-dimensional SERS substrates enabled the acquisition of highly resolved Raman spectra of sEVs in clinical CSF samples. The development of stacking machine learning frameworks to analyze the Raman data sets unveiled differential vibrational modes and generated high accuracy for diagnosing pediatric medulloblastoma (MB), the most common malignant brain tumor in children, with an area under the receiver operating characteristic curve (AUC(ROC)) of 0.963. Furthermore, we observed high discriminative capacity of our Raman spectral classifier to distinguish MB from other brain tumors (AUC(ROC) = 0.906). Finally, we showed that dynamic analysis of the CSF sEV Raman profiles allowed monitoring of the therapeutic response of MB at the molecular level. Our study holds promise for facilitating precision medicine in brain tumors.</p> Graphical Abstract <p></p>