A comprehensive narrative review evaluating prediction tools for prostate cancer progression indicates that while machine learning algorithms report encouraging discrimination scores in initial publications, pervasive methodological weaknesses, scarce external validation, and missing calibration continue to limit clinical adoption. Writing in Frontiers in Oncology, researchers Jun Yuan, Ziwei Jiang, Wenxian Chen, and Shaocheng Wang from Chongming Hospital Affiliated to Shanghai University of Medicine and Health Sciences evaluated five years of studies predicting castration-resistant prostate cancer following androgen deprivation therapy. Although androgen deprivation therapy is standard for locally advanced and metastatic disease, many patients progress to castration resistance within two years of monotherapy. Early identification of high-risk patients could allow tailored monitoring and timely treatment intensification, yet validated, generalisable risk stratification models remain scarce.
Castration-resistant prostate cancer is defined clinically by disease progression occurring despite castrate serum testosterone levels below 50 nanograms per deciliter, or 1.7 nanomoles per liter. Progression appears biochemically as three consecutive prostate-specific antigen rises at least one week apart, with two exceeding fifty percent above nadir and reaching values above 2 nanograms per milliliter, or radiographically through two or more new bone lesions or novel soft-tissue lesions. Current guidelines, including European Association of Urology and National Comprehensive Cancer Network frameworks, distinguish nonmetastatic from metastatic disease for clinical decision-making. Castration resistance arises primarily through androgen receptor reactivation via gene amplification, mutations, or splice variants such as AR-V7. Alternatively, tumors escape hormonal suppression through neuroendocrine differentiation, where disease progresses rapidly despite low prostate-specific antigen concentrations.
Multiple clinical characteristics, circulating markers, and imaging parameters serve as established risk factors for progression to castration resistance. Peripheral blood inflammatory indices, such as the neutrophil-to-lymphocyte ratio and systemic immune-inflammation index, reflect systemic inflammation and microenvironmental conditions influencing therapy resistance. Molecular indicators including circulating tumor cells, circulating tumor DNA fractions, polygenic risk scores, and gene-expression signatures provide additional predictive signals. Furthermore, radiomic features extracted from multiparametric magnetic resonance imaging and positron emission tomography contribute high-dimensional structural data. Because individual markers possess limited isolated predictive value, investigators increasingly integrate these diverse multimodal features into joint multi-parameter prediction models using machine learning and advanced ensemble learning frameworks.
To map the predictive modeling landscape, the review authors executed searches across PubMed and OpenAlex spanning January 2021 to August 2026, supplemented by citation tracking. The core PubMed search retrieved 29 records, which combined with OpenAlex results to yield 65 unique records after deduplication. Screening identified 53 eligible studies, while 12 were excluded for being outside scope, focusing on diagnosis rather than progression, or lacking predictive algorithms. Revision added 17 references, including updated GLOBOCAN projections forecasting 2.9 million global cases by 2040, clinical guidelines, and 11 primary model studies, yielding 70 cited works. Under PROBAST+AI risk-of-bias evaluation, all evaluated core models were rated at high or unclear risk of bias, primarily because external validation was absent and calibration reporting was incomplete.
Traditional prediction tools have historically relied on Cox proportional hazards or multivariable logistic regression formatted as nomograms for bedside clinical interpretation. In advanced disease settings, nomograms have predicted survival after lutetium-177 PSMA radioligand therapy, monitored radium-223 benefits, identified bone metastasis risk in SEER data, and estimated testosterone recovery following hormone therapy cessation. While nomograms offer operational simplicity, published models suffer from noticeable limitations. Most draw on retrospective single-center or registry cohorts, depend on prespecified linear predictors that fail to capture nonlinear relationships or biological interactions, and lack external validation across outside institutions. Consequently, traditional nomograms often struggle to generalise across diverse clinical populations, motivating researchers to explore artificial intelligence and machine learning architectures capable of uncovering multidimensional data patterns.
Machine learning algorithms have been applied across numerous prostate cancer endpoints, including aggressiveness grading, active surveillance tracking, pelvic lymph node assessment, and post-prostatectomy biochemical recurrence prediction. Within the androgen deprivation therapy context, machine learning classifiers frequently achieve higher reported discrimination than standard regression equations within development cohorts. For example, random survival forests applied to longitudinal blood and urine measurements from 340 patients tracked disease trajectories across time, outperforming Cox regression across progression-free, overall, and cancer-specific survival. Multimodal approaches have integrated convolutional deep neural networks such as ResNet-50 with multiparametric magnetic resonance imaging and pathology images. Additionally, trial-based initiatives combined digital pathology with clinical factors to guide hormonal therapy, while population-level models such as Survival Quilts, random forests, and gradient boosting predicted cancer-specific mortality and bone metastasis.
The review synthesized eight core studies published between 2021 and 2026 directly predicting castration resistance after androgen deprivation therapy. In the KYUCOG-1401-A study of 119 advanced patients, pointwise linear algorithms, elastic-net logistic regression, and XGBoost incorporated clinical predictors and single nucleotide polymorphisms to predict two-year castration resistance, yielding area under the curve values of 0.710 to 0.785 for clinical factors alone that improved with genetic data. Huang and colleagues reported a Cox nomogram for metastatic osseous disease, showing time-dependent area under the curve metrics of 0.81 at one year and 0.76 at two years in external validation. Zhu and colleagues analyzed 410 hormone-sensitive patients, where random forest reached an area under the curve of 0.873 with 81.7 percent accuracy. Multimodal deep learning by Zhou and colleagues across 399 patients from three centers achieved an area under the curve of 0.86 combining radiomics, pathomics, and clinical variables, while other studies utilized ultrasound nomograms and deep learning on biopsy specimens.
Despite encouraging discrimination metrics, the authors emphasized that apparent machine learning gains stem from within-study comparisons subject to substantial optimism and bias. Methodological shortcomings pervade the core evidence: external validation was reported in only two of eight models, one of which tested just sixteen patients. Calibration curves assessing alignment between predicted probabilities and observed event rates appeared in only three models, decision-curve analyses evaluating net clinical utility were presented in just two, and only two studies stated their total count of outcome progression events. Predictor timing also poses clinical concerns. Several models included post-treatment variables such as prostate-specific antigen nadir and time to nadir, which are unavailable at baseline therapy initiation and can introduce immortal-time bias if prognostic follow-up begins at treatment onset.
To address variance and model-selection limitations, the authors highlighted Super Learner ensemble learning as a promising architecture for progression prediction. Super Learner is a prespecified stacking framework that avoids picking a single arbitrary algorithm by combining predictions from a library of diverse candidate learners. Using cross-validation folds, the algorithm fits candidate models on training partitions and evaluates loss on held-out folds, allowing a meta-learner to assign optimal weights that minimize cross-validated risk. This stacking strategy has proven stable across varied clinical prediction problems, including hospital readmission after bariatric surgery, COVID-19 severity, cardiovascular events after percutaneous intervention, Long COVID risk, and polygenic risk scores. Although explanation tools such as Shapley values aid model interpretation, they reflect algorithmic attributions rather than underlying disease mechanisms and can become unstable when predictors correlate.
Drawing on these principles, the authors outlined a multi-parameter Super Learner model designed specifically for castration resistance prediction. The candidate library incorporates random forest, XGBoost, AdaBoost, Gaussian naive Bayes, and support vector machines, pooled by a logistic regression meta-learner using ten-fold cross-validation. To avoid data leakage and optimistic performance estimates, the entire weight-estimation procedure must be wrapped within an independent outer cross-validation loop. Because castration resistance is a time-to-event outcome, the framework can be configured as a classification task at a fixed horizon, such as two years, or as a survival ensemble incorporating Cox models or survival forests. A survival formulation must account for right-censoring, handle competing mortality risks, and evaluate performance using time-dependent area under the curve metrics, Brier scores, and calibration at one, two, and three years.
Substantial statistical and methodological challenges must be addressed before machine learning ensembles can transition into routine clinical decision support. Sample size guidelines indicate that multi-layer ensembles with hyperparameterized learners cannot be reliably estimated from modest cohorts of a few hundred patients with limited outcome events. Oversampling techniques like SMOTE should not be applied when calibrated probabilities are desired, as synthetic resampling distorts predicted probabilities and calibration unless confined strictly within cross-validation folds. Additionally, systematic reviews show that machine learning offers no consistent performance advantage over standard logistic regression on structured clinical data. Real-world deployment is further complicated by missing data and unstandardized clinical variables across healthcare centers. Addressing these barriers requires adherence to TRIPOD reporting standards, rigorous preprocessing protocols, and prespecified sample sizes ensuring adequate events per variable to prevent overfitting.
Temporal transportability presents an additional translational hurdle stemming from shifting clinical standards in prostate cancer management. The core models analyzed in the review were developed in historical cohorts receiving androgen deprivation monotherapy. In contemporary practice, metastatic hormone-sensitive prostate cancer is frequently treated with androgen receptor pathway inhibitors added to hormone therapy or triplet combinations with chemotherapy. These modern regimens prolong progression-free intervals and reduce event rates at fixed horizons, meaning historical prediction models will require recalibration or complete redevelopment before clinical use. Furthermore, existing models rarely provide open-access code, are derived primarily from East Asian or European-ancestry cohorts without proven cross-ancestry portability, and lack regulatory evaluation demonstrating prospective safety. The authors concluded that while Super Learner ensembles offer a promising multi-parameter framework, prospective multicenter validation and transparent calibration remain essential before algorithms can guide clinical care.
Subject of Research: Prediction models for progression to castration-resistant prostate cancer following androgen deprivation therapy
Article Title: Prediction models for progression to castration-resistant prostate cancer after androgen deprivation therapy: a narrative review (2021–2026)
Article References: Yuan, J., Jiang, Z., Chen, W., & Wang, S. (2026). Prediction models for progression to castration-resistant prostate cancer after androgen deprivation therapy: a narrative review (2021–2026). Frontiers in Oncology, 16, Article 1964985. https://doi.org/10.3389/fonc.2026.1964985
Image Credits: AI Generated
DOI: 10.3389/fonc.2026.1964985
Keywords: prostate cancer, castration-resistant prostate cancer, androgen deprivation therapy, machine learning, nomograms, Super Learner, ensemble learning, PROBAST, calibration, external validation, prostate-specific antigen, androgen receptor, AR-V7, radiomics, decision-curve analysis, neutrophil-to-lymphocyte ratio, circulating tumor cells, TRIPOD
Cite Scienmag News
Nathaniel Bowman. (November 17, 2026). Prostate Cancer Progression Models Lack External Validation Despite Machine Learning Gains. Scienmag. https://scienmag.com/prostate-cancer-progression-models-lack-external-validation-despite-machine-learning-gains/
Nathaniel Bowman. "Prostate Cancer Progression Models Lack External Validation Despite Machine Learning Gains." Scienmag, 17 November 2026, https://scienmag.com/prostate-cancer-progression-models-lack-external-validation-despite-machine-learning-gains/. Accessed 17 November 2026.
Nathaniel Bowman. "Prostate Cancer Progression Models Lack External Validation Despite Machine Learning Gains." Scienmag. November 17, 2026. https://scienmag.com/prostate-cancer-progression-models-lack-external-validation-despite-machine-learning-gains/








