Background: Bone metastases (bm) are common in men with castration-resistant prostate cancer (CRPC) and are associated with poor prognosis. Therefore, identifying prognostic factors for survival is key to aid decision-making. While several models exist for predicting survival in men with mCRPC, most were derived using randomized controlled trial (RCT) data, which may have limited application to real-world (RW) populations and utilized traditional linear methods. We aimed to develop and validate a prediction model of real-world survival (rwOS) in men with bmCRPC using a national RW electronic medical record (EMR) database with traditional and advanced machine learning (ML) methods. Methods: A retrospective cohort of men diagnosed with bmCPRC between 2010 and 2021 was identified using the Optum EMR. rwOS was identified as documented evidence of death in the EMR derived through either the Social Security Death Index or EMR-reported death date. Independent variables were demographic (age, race), clinical conditions (Charlson comorbidity index, pain, prior phases of PC), PC-related medications, and laboratory parameters during a 6-month baseline period. A dynamic model was used to predict survival at 1, 2, and 3 years after diagnosis of bmCRPC using traditional (logistic, Cox regression) and ML (light gradient boosting (LGB)) models. Training, calibration, and validation were performed by splitting the data and utilizing 5-fold cross-validation on the training sample and calibration plots. Diagnostic performances were evaluated using area-under-the-curve (AUC), precision-recall curve (PRC), and calibration curves across three models. Results: The study cohort included 4,097 men with a median age of 76 years (interquartile range: 68-82), predominantly white (81.2%), and residing in the Midwest (48.2%). Over a median follow-up period of 19.2 months, 2,220 (54.2%) men died with median rwOS of 31.0 months (95% CI: 29.6-32.9). The AUC for Cox, logistic and LGB models at 1 year were well within the acceptable range of (≥0.7.0) 0.78, 0.79, and 0.81 respectively. Similar findings were observed at 2 and 3-year landmark analyses with marginal improvement in model performance with LGB. Top predictors of rwOS were consistent across all time points in the LGB models and included laboratory parameters (prostate-specific antigen level, alkaline phosphatase, albumin, hemoglobin), age, presence of pain, transition from mHSPC to mCRPC, surgical castration, and baseline use of androgen receptor inhibitors. Conclusions: Study findings highlight consistency in the predictors of rwOS in a large cohort of men with bmCRPC using a national EMR database. ML-based models predicted rwOS with modest performance improvement compared to traditional models and could provide rwOS prediction to aid treatment decision-making.