Purpose <p>Early detection of prostate cancer is crucial for successful treatment; however, uncertainties in prostate-specific antigen (PSA) screening often lead to invasive procedures like prostate biopsy, which can be risky and uncomfortable for patients. This study explores a non-invasive method for prostate cancer detection by analyzing Fourier-transform infrared spectroscopy (FTIR) spectral data from urinary extracellular vesicles (EVs) and patient clinical data to develop a predictive model. The aim of this research is to identify the optimal machine learning model for accurate and reliable prostate cancer prediction.</p> Methods <p>FTIR spectral data of uEVs and clinical data obtained from 22 prostate cancer patients and 31 healthy controls were divided into a 60:20:20 train-validate-test split data analysis. Due to the imbalanced and limited sample size, the Synthetic Minority Oversampling Technique (SMOTE) was applied to generate synthetic data for the minority class. Multiple machine learning algorithms—K-Nearest Neighbors (k-NN), Support Vector Machine (SVM), Naïve Bayes, and Decision Tree—were employed to develop models. Feature selection techniques were applied to retain informative features and reduce noise, enhancing classification performance.</p> Results <p>The integration of FTIR spectral and clinical data, which further enhanced by FTIR feature selection and SMOTE, significantly improved predictive performance. Among the tested models, the linear SVM model validated its promising performance with an AUC of 0.9167 in prostate cancer prediction.</p> Conclusions <p>This multimodal approach highlights the novelty of modality integration of EV-FTIR molecular and clinical data, feature selection, and SMOTE in disease modeling, offering a potent, non-invasive, and reliable tool for prostate cancer detection.</p>