The current study aims to develop and validate machine learning (ML) models for the prediction of cancer status by the non-invasive urinary proteomic in a population-based cohort. In this retrospective study, urinary proteome profiles in 804 cases from the FLEMENGHO cohort were measured by mass spectrometry. After feature selection by LASSO on both clinical variables and urinary proteome profile, benchmark models by clinical variables were built with six different ML algorithms. Proteome-based models and combined models were built and compared with the benchmark models. The models' performance, i.e. area under the curve (AUC) was compared by Delong method. The 95% confidence interval was estimated by the bootstrapping method. The best-performing model was explained by Shapley Additive Explanations (SHAP) method. The predictive role of proteome biomarkers in longitudinal cancer diagnosis was also explored. A clinical model, based on age, blood sugar and blood lipid profile, yielded the best AUC of 0.75 (0.68-0.82), with 0.80 (0.72-0.91) for the proteome model based on 13 selected biomarkers and 0.83 (0.77-0.90) for the combined model (P=0.01 for comparison with clinical model). SHAP on the support vector machine in the combined setting showed that except for age, proteome biomarkers contribute to the final prediction of the model. After adjusting with clinical factors, three proteome biomarkers are independent risk factors for longitudinal cancer development. Urinary proteome profiling, together with fine-tuned machine learning algorithms, demonstrates the predictive potential for cancer diagnosis transparently.