Prostate cancer is one of the most common causes of tumor-related deaths among men worldwide. Advanced and highly aggressive mCRPC develops during ongoing ADT, showing how ADT eventually fails to suppress disease progression; hence, other treatment modalities must be sought. Second-generation androgen receptor inhibitors such as Abiraterone and Enzalutamide have shown promise in improving survival. However, identifying the best treatment for specific subgroups of patients remains a major challenge in ongoing studies.

The following objectives will be pursued in this study by analyzing the clinical and demographic data of patients with mCRPC: Identification of critical determinants of death, such as frailty, comorbidities, and clinical signs; comparison of the effectiveness of Abiraterone and Enzalutamide in real-world settings, develop predictive models to improve risk stratification and guide therapeutic decisions. Metastatic castration-resistant prostate cancer (mCRPC) management has evolved significantly with the introduction of second-generation androgen receptor (AR) inhibitors, such as abiraterone acetate and enzalutamide. These therapies have transformed the treatment landscape by targeting androgen receptor-axis signaling pathways, which are essential factors in the development of prostate cancer [1]. Despite their efficacy, therapeutic challenges such as treatment resistance and the need for personalized therapeutic approaches exist. Both enzalutamide and abiraterone were shown to be effective in landmark clinical trials. In post-chemotherapy settings, abiraterone plus prednisone demonstrated improved overall survival compared to placebo (15.8 vs 11.2 months; HR 0.74) in the COU-AA-301 trial [2], and comparable populations exhibited similar benefits from enzalutamide (18.4 vs 13.6, and 0.63, respectively) [3]. Among chemotherapy-naïve patients, the COU-AA-302 validated abiraterone’s survival benefit (34.7 vs 30.3 months, HR 0.81) [4]. Enzalutamide also significantly extended radiographic progression-free survival (rPFS) and reduced the risk of death by 29% (HR 0.71) [5].

Comparative analyses of these therapies have yielded variable outcomes. La et al. found that enzalutamide improved prostate-specific antigen (PSA) response rates and provided modest survival benefits (0.90 additional months > four years), especially in patients who never received chemotherapy [6]. Despite enzalutamide’s benefit in delaying the disease progression and PSA elevation. However, Tan et al. found no significant difference in overall survival despite the advantage of enzalutamide in delaying disease progression and PSA elevation. Tan et al. did not find any significant differences in overall survival [7]. The average survival benefit of enzalutamide over abiraterone was shown by Gopukumar et al. using a causal survival forest analysis in a veteran cohort. The treatment effect heterogeneity was influenced by age, BMI, PSA levels, and comorbidities [8].

Clinical factors such as disease presentation and patient characteristics were considered when choosing treatment. Although statistical significance favored non-visceral disease subgroups, Goodman et al. observed that abiraterone was efficacious across visceral disease subgroups [9]. The similar efficacy and safety profiles of both drugs in elderly populations were validated by age-related analyses [10][11].

The interplay between frailty, comorbidities, and survival has been well established in mCRPC research. Frailty and baseline comorbidities had a substantial impact on therapy outcomes, with cardiovascular events being more common in patients treated with abiraterone, according to Lin et al.’s investigation of treatment-emergent comorbidities and survival outcomes in mCRPC patients receiving abiraterone or enzalutamide [12]. These findings emphasize the importance of individualized treatment plans that account for patient health profiles.

Fiala et al. conducted a long-term evaluation of mCRPC patients treated with enzalutamide or abiraterone in a real-world clinical context, emphasizing the heterogeneity in medication efficacy based on patient characteristics, such as baseline performance status and previous treatment history [13]. Despite the efficacy of both medications and patient-specific factors affecting survival rates, suggesting the need for structured treatment approaches.

With sequencing studies showing decreased efficacy in subsequent treatment lines, cross-resistance among these therapies poses a serious challenge. Loriot et al. and Noonan et al. reported modest progression-free survival with abiraterone after enzalutamide failure (2.7 months and 15.4 weeks, respectively) [14][15], while Bianchini et al. and Brasso et al. observed limited PSA responses (≤13.3% achieving ≥50% PSA decline) when enzalutamide followed abiraterone [16][17]. Similar findings by Thomsen et al. and Badrising et al. in heavily pretreated patients underscore the need for optimized sequencing strategies [18][19]. Recent technological advances have reshaped the management of prostate cancer. Although issues with data quality and standardization still exist, artificial intelligence has demonstrated potential for enhancing diagnosis, risk assessment, and treatment selection [20]. Machine learning algorithms for predicting survival outcomes in bone-metastatic disease have achieved promising predictive performance, with AUC values up to 0.78 for overall survival [21].

Phan et al. proposed incorporating androgen-to-PSA ratios into predictive models to account for androgen dependency in PSA expression, potentially improving assessment accuracy during disease progression and treatment resistance development [22]. For nonmetastatic castration-resistant prostate cancer (nmCRPC), Ni et al. developed a machine learning-based prognostic model using data from phase III clinical trials and identified eight key prognostic factors, including novel hormone therapy application, Gleason score, previous treatments, race, PSA doubling time, hemoglobin levels, and PSA values. Their model demonstrated robust predictive performance with a C-index of 0.724 in internal validation and good calibration in external testing [23].

Nakata et al. developed a deep learning model for metastatic hormone-sensitive prostate cancer (mHSPC) using biopsy images to estimate time-to-resistance to hormone therapy before progression to castration-resistant disease. Their research demonstrated a strong correlation between this predicted time-to-resistance and overall survival, achieving nearly 80% accuracy in identifying patients likely to benefit longer from standard hormone therapy [24].

Integrating predictive modeling with clinical datasets offers new opportunities for personalized mCRPC management. Lin et al. and Jarimba et al. emphasized the potential of data-driven approaches to refine treatment decisions, improve risk stratification, and enhance survival [12][25]. The use of predictive analytics aligns with these findings by identifying key predictors, such as frailty, PSA levels, and systemic health markers, paving the way for precision oncology. These technological developments collectively suggest an emerging paradigm shift toward personalized treatment selection through advanced computational approaches across the spectrum of prostate cancer from hormone-sensitive to castration-resistant stages.

The project’s data was collected by Dr. Martin Schoen’s lab. The dataset contains patient-level data on 32 features and is grouped as follows.

Demographics:Age(0%missing),sex(0%missing),andrace(6patients;<0.1%missing).ClinicalData:PSAlevels(1481patients;5.9%missing),albumin(1571patients;6.3%missing),hemoglobin(1253patients;5.0%missing),andGleasonscore(0%missing).

TreatmentDetails:AdministrationofAbiraterone(0%missing),Enzalutamide(0%missing),anddocetaxeltherapy(0%missing).

Indices:CharlsonComorbidityIndex(4599patients;18.4%missing),frailtyscores(0%missing),andElixhausergroups(0%missing).

DataPreprocessing:

MissingData: Addressed using predictive modeling techniques.Scaling: The Standard Scaler ensured that no feature dominated the others during model training.

Preprocessing procedure: Missing continuous variables were imputed using median imputation, while categorical and binary variables were imputed using the most frequent (mode) value. To prevent data leakage, imputation and z-score standardization were performed independently within each training fold of the five-fold cross-validation procedure, and the learned parameters were subsequently applied only to the corresponding validation fold. This preprocessing pipeline was identical for all machine learning models (Table 1).

CohortConstruction: This retrospective cohort study included 25,070 Veterans with metastatic castration-resistant prostate cancer (mCRPC) who received care within the Veterans Health Administration (VHA) between 2011 and 2022. Eligible patients had confirmed mCRPC and available baseline demographic, clinical, and treatment information. Baseline variables included demographics (age, sex, and race), clinical characteristics (prostate-specific antigen [PSA], albumin, hemoglobin, and Gleason score), treatment history (abiraterone, enzalutamide, and docetaxel), and comorbidity measures, including the Charlson Comorbidity Index, frailty scores, and Elixhauser comorbidity groups. Patients with insufficient baseline information or missing outcome data were excluded from the analysis. Missing predictor values were addressed using predictive imputation techniques, and continuous variables were standardized using the Standard Scaler before model development to ensure balanced feature contributions. Patients were followed from the index date until death or the end of follow-up. The primary outcome was all-cause mortality, which was evaluated over the entire follow-up period as well as at predefined 1-, 2-, 3-, 4-, and 5-year follow-up intervals.

Table 1. Missing data summary for variables included in the analysis.

The dataset integrates demographic, clinical, treatment, and comorbidity variables to enable development of predictive models for mortality risk assessment in patients with mCRPC by integrating demographic, clinical, treatment, and comorbidity data, the machine learning models provided improved risk stratification and individualized mortality prediction. These findings support the growing role of computational approaches in precision oncology and highlight a paradigm shift toward personalized treatment selection and clinical decision-making across the continuum of prostate cancer, from hormone-sensitive disease to metastatic castration-resistant prostate cancer.

ModelDevelopmentandValidation

The predictive modeling workflow followed a standardized machine learning pipeline for all evaluated algorithms to ensure fair model comparison and minimize bias. A total of 25,070 eligible patients were included in the analysis after applying the predefined inclusion and exclusion criteria (Figure 1). The study cohort was randomly partitioned into a training dataset (80%, n = 20,056) and an independent testing dataset (20%, n = 5014) using stratified sampling based on mortality outcome to preserve the class distribution in both datasets. Model development, data preprocessing, feature engineering, and hyperparameter optimization were performed exclusively on the training dataset, while the independent testing dataset was reserved for final model evaluation. For the final performance assessment, only 1165 patients from the testing dataset had complete data available for all predictors and outcomes required for model evaluation after preprocessing; therefore, the confusion matrix and associated performance metrics were calculated using this evaluable subset.

Figure 1.Patient selection flowchart.

Within the training dataset, five-fold cross-validation was employed for model training and hyperparameter optimization. During each fold, missing predictor values were imputed using predictive imputation methods, and continuous variables were standardized using the Standard Scaler. All preprocessing parameters were estimated solely from the training portion of each fold and subsequently applied to the corresponding validation fold, thereby preventing information leakage. The final optimized model was retrained using the complete training dataset and evaluated on the independent testing dataset, with the confusion matrix reported for the subset of 1165 evaluable patients.

Multiple supervised machine learning algorithms were evaluated, including logistic regression, random forest, gradient boosting, extreme gradient boosting (XGBoost), support vector machine, k-nearest neighbors, decision tree, and multilayer perceptron neural network. Hyperparameters for each algorithm were optimized using a grid-search strategy within the cross-validation framework. Representative hyperparameters included the number and depth of trees for ensemble methods, the regularization parameter (C) and kernel parameters for support vector machines, the number of neighbors for k-nearest neighbors, and learning rate and network architecture for neural networks.

Because mortality outcomes were imbalanced, model performance was assessed using multiple complementary metrics, including accuracy, sensitivity, specificity, precision, and F1-score. These measures were used to evaluate both overall predictive performance and the ability of models to correctly identify mortality and survival outcomes.

No explicit feature-selection algorithm was applied before model training. Instead, all 32 clinically relevant variables identified from demographic, clinical, treatment, and comorbidity domains were included in model development to preserve their potential prognostic value. Variable importance was subsequently evaluated using model-specific feature importance measures and SHapley Additive exPlanations (SHAP) analyses to identify the most influential predictors of mortality.

OutcomeDefinition: The primary outcome was all-cause mortality. Patients were classified into two groups: deceased patients (death = 1) and surviving patients (death = 0). The binary classification models were trained to predict mortality risk based on baseline demographic, clinical, treatment, and comorbidity features.

Table 3 shows the train and test accuracy of regression and classifiers. Based on this table, Random Forest and SVM had the best performance.

Table 3. Performance comparison of machine learning models for mortality prediction in mCRPC.

This study developed and evaluated machine learning models for mortality prediction in patients with metastatic castration-resistant prostate cancer using demographic, clinical, treatment, and comorbidity features. Among the evaluated algorithms, Random Forest demonstrated the strongest predictive performance and provided interpretable feature importance measures. PSA levels, frailty indices, and comorbidity burden emerged as consistent predictors of mortality, suggesting that systemic patient health plays an important role in survival outcomes.

The findings support the integration of machine learning approaches into mCRPC risk stratification by complementing traditional clinical assessment with data-driven prognostic insights. Future studies should focus on external validation, incorporation of longitudinal clinical information, and integration of patient-reported outcomes to further improve personalized oncology care.

This retrospective study was conducted using de-identified data from the Veterans Health Administration (VHA). The study was reviewed and approved by the appropriate Institutional Review Board (IRB). Because the study involved retrospective analysis of existing clinical data with minimal risk to participants, the requirement for informed consent was waived in accordance with applicable regulations and institutional policies.

The data used in this study were obtained from the Veterans Health Administration (VHA) and are not publicly available because of institutional policies and patient privacy regulations. Access to the data may be granted to qualified investigators upon reasonable request and with approval from the Veterans Health Administration and the appropriate regulatory authorities.

During manuscript preparation, the authors used Claude solely for limited language editing of selected sentences to revise to improve clarity. No AI tool was used to generate research content, analyze data, interpret results, or prepare references; the ideas, arguments, and intellectual contribution of the paper are entirely original. All AI-assisted suggestions were reviewed, revised where necessary, and approved by the authors, who take full responsibility for the final manuscript. We used AI to format references and find the doi numbers.

Bahareh Rahmani: Supervision, Administration, Methodology, Review of the Paper;

Pavan Sai Gajula: Data Analysis, Methodology, Writing—Original Paper;

Mustaid Ahmed: Methodology, Writing—Original Paper;

Naga Pushpa Latha Ginni: Data Analysis, Visualization, Writing—Original Paper;

Vanitha Ajmeer: Visualization, Data Analysis;

Jason Doherty: Data Collection, Preprocessing, Review of the Paper;

Sailaja Vatsalya Madhurapantula: Literature Review, Visualization;

Payam Norouzzadeh: Validation, Methodology Advisor, Review of the Paper;

Leslie Hinyard: Validation, Review of the Paper;

Mostafa Yazdi: Funding, Review of the Paper;

Martin Schoen: Supervision, Data Collection, Review of the Paper.