Prostate cancer risk prediction remains challenged by the inability to reliably distinguish indolent from aggressive disease, leading to overtreatment or undertreatment. Conventional approaches based on prostate-specific antigen (PSA) and digital rectal examination (DRE) rely on static thresholds or baseline models, failing to capture longitudinal disease dynamics, treatment context, and calibrated risk estimation. a treatment-aware, survival-informed deep learning framework for phenotyping from irregular longitudinal clinical data. The proposed method employs a mask-aware Transformer encoder to model PSA/DRE sequences with irregular sampling, fused with static clinical features to learn a shared latent representation. This representation is optimized through a multi-task learning framework that jointly incorporates survival prediction, discrete-time hazard modeling for calibrated risk estimation, treatment supervision to encode therapeutic context, and outcome-guided deep embedded clustering using Gaussian mixtures (DEC–GMM). Applied to 4,579 men from the PLCO trial, STRATA-PC achieved strong prognostic discrimination (C-index 0.778 for overall survival and 0.853 for prostate cancer–specific mortality) with consistent calibration at 5- and 10-year horizons. Outcome-guided clustering improved latent-space separation (silhouette > 0.15) and stability across refits, while maintaining well-calibrated survival predictions. The learned latent representations consistently separated into three reproducible phenotypes that were robust to missingness patterns and aligned with treatment exposure, indicating that the framework captures meaningful variation in longitudinal disease trajectories and associated outcomes. By integrating longitudinal modeling, treatment-aware supervision, calibrated survival prediction, and mixture-based clustering within a unified framework, STRATA-PC provides a robust approach for outcome-guided phenotyping from sparse clinical time series.