People diagnosed with prostate cancer at an intermediate Gleason score can harbour aggressive or non-aggressive disease. These disease subtypes cannot always be accurately distinguished, and this can result in under- or over-treatment of some patients. In this study, we applied machine learning methods to proteomic data obtained from mass spectrometry to assess prostate cancer prognosis. We performed quantitative proteomic analyses of 1,566 prostate tissue samples from 290 patients obtained from the Prostate Cancer Outcomes Cohort (ProCOC). We processed small tissue samples via pressure-cycling technology (PCT) and acquired proteomic data by Sequential Window Acquisition of all Theoretical Mass Spectra (SWATH) mass spectrometry. Tumour tissues were classified by two pathologists, alongside matched benign tissue samples for each patient. We divided the cohort into 31 batches, each containing two control samples, allowing us to evaluate technical reproducibility. Each sample was analysed in technical duplicate at ProCan (Children’s Medical Research Institute) to ensure high reproducibility. Altogether, we quantified 2,800 SwissProt proteins (FDR<1%) in all samples, with an average missing value of 30%. We quantified 73% of proteins in over 50% of samples. We corrected for batch effects and imputed missing values with technical and biological replicates. Using this dataset, we applied machine learning methods to find protein signatures to identify tumour and normal samples, and to assess Gleason scores. Using Random Forest, we could separate tumour and normal samples with an area under the curve of 0.92. We applied a random survival forest model to tumour samples and identified the top 100 proteins that predict survival, using time to recurrence and censoring information. To our knowledge, this is the largest SWATH-based proteomic dataset generated to-date in cancer (>1,500 proteomes). Our study demonstrates the feasibility of SWATH mass spectrometry for the proteomic analysis of prostate cancer prognosis.