Prostate cancer is the second most common cancer in men worldwide and the fifth leading cause of cancer death among men. Treatment options span a wide spectrum -- from active surveillance and watchful waiting for low-risk disease, to surgery, radiation, hormone therapy, chemotherapy, and targeted drugs for more aggressive cases. Each option carries its own risk-benefit trade-off for cancer control, side effects, and quality of life, making treatment selection one of the most consequential decisions in oncology.
A key practical question is whether to pursue surgical treatment -- specifically radical prostatectomy (surgical removal of the prostate) or cryotherapy (freezing to destroy the tumor) -- versus non-surgical options such as radiation, hormone therapy, or observation. This binary decision depends on a complex combination of factors including cancer stage, tumor aggressiveness, patient age, and overall health, which interact in ways that are difficult to weigh systematically using conventional clinical intuition alone.
Prior machine learning studies on prostate cancer treatment recommendations have been limited by small datasets from single institutions, a narrow range of input variables, and reliance on just one or two algorithm types. These constraints reduce the reliability and generalizability of results and limit how much clinicians can trust model-driven recommendations in diverse real-world settings. This study addresses all three limitations directly by using a national population-based database, incorporating 14 validated clinical and demographic variables, and systematically comparing 10 different machine learning algorithms.
The study analyzed 140,294 prostate cancer patients diagnosed between 2010 and 2015 drawn from the SEER (Surveillance, Epidemiology, and End Results) database, maintained by the National Cancer Institute. SEER covers approximately 28% of the US population and is one of the most comprehensive public cancer registries available, making it well-suited for developing generalizable treatment prediction models. All patients were confirmed by pathological diagnosis, were male and aged 40 or older, and had complete records of key clinical variables.
Fourteen clinical and demographic features were retained after feature selection using two complementary methods: LASSO regression, which shrinks less important feature coefficients toward zero to eliminate redundant variables, and recursive feature elimination with cross-validation (RFECV), which iteratively removes the least important features and selects the optimal subset automatically. The retained features included age, cancer stage, tumor stage (T stage), lymph node status, metastasis status, primary and secondary Gleason grade, PSA level, cores positive percentage (CPP), race, income, and population density.
Ten machine learning algorithms were trained and compared: support vector machine (SVM), k-nearest neighbors (KNN), decision tree (DT), logistic regression (LR), random forest (RF), AdaBoost, XGBoost, LightGBM, CatBoost, and multi-layer perceptron (MLP). This range spans linear models, tree-based models, ensemble boosting methods, and a neural network, providing a comprehensive evaluation of different algorithmic approaches on the same dataset. The data were split 70% for training and 30% for testing, with performance measured by AUC, sensitivity, specificity, accuracy, and predictive values.
To make the best-performing model interpretable, the researchers applied SHAP (Shapley Additive Explanations) analysis, which quantifies how much each feature contributed to any individual prediction. Derived from cooperative game theory, SHAP assigns each feature a value representing its share of the model's output for a given patient -- making it possible to explain not just which features matter overall, but exactly how and in which direction they influence treatment recommendations for any specific case.
Three gradient-boosting algorithms -- XGBoost, LightGBM, and CatBoost -- achieved the highest AUC of 0.939 on the test set, outperforming all other approaches. Among these three, CatBoost demonstrated superior generalization: while XGBoost and LightGBM showed declining AUC on the validation set as the number of decision tree estimators increased (a sign of overfitting), CatBoost's performance remained stable and continued to improve. In 10-fold cross-validation, CatBoost consistently outperformed both competitors when more than 100 estimators were used.
CatBoost's final performance metrics were an AUC of 0.939, sensitivity of 0.877, specificity of 0.878, and accuracy of 0.877, with a positive predictive value of 0.801 and negative predictive value of 0.927. This means the model correctly classified patients into surgical versus non-surgical treatment groups in approximately 88% of cases, with balanced detection of both treatment categories -- an important property because imbalanced sensitivity and specificity would indicate systematic bias toward one treatment type.
Comparing the 35% of patients who received surgical treatment versus the 65% who did not revealed clear differences in their clinical profiles. Surgical patients were younger (median around 60 years), had significantly lower PSA levels (mean 8.6 vs. 14.8), and overwhelmingly had localized disease (T2 stage), while non-surgical patients were older and more commonly had T1 or metastatic disease. These differences are consistent with standard clinical practice and validate that the models are learning clinically meaningful patterns.
T stage had by far the largest influence on treatment prediction, with a mean SHAP value of 1.72 -- more than twice the influence of any other feature. Higher T stage (more locally advanced disease extending beyond the prostate into nearby structures) strongly pushed predictions toward non-surgical treatment. This reflects clinical reality: patients with T3 or T4 disease have cancer that has grown beyond the prostate, making surgical cure less likely and typically favoring systemic approaches.
After T stage, the next most influential factors were cancer stage (SHAP 0.62), age (0.47), and cores positive percentage (CPP, 0.46). CPP represents the proportion of biopsy cores that tested positive for cancer -- a higher percentage indicates more extensive tumor involvement throughout the gland. Higher CPP, along with higher age and higher PSA and Gleason score, pushed predictions toward non-surgical treatment. These associations reflect the guideline-recommended practice of reserving surgery for younger patients with more contained disease.
A notable finding was that PSA and Gleason score had a smaller relative influence on treatment assignment than T stage and cancer stage in this model. This may reflect the complexity of real-world treatment decisions in the SEER data, where anatomical stage tends to drive surgical eligibility most directly. This contrasts with some clinical literature that emphasizes PSA and Gleason score as the primary drivers of treatment aggressiveness, and may reflect differences in how treatment decisions are distributed across population-level registry data versus academic medical centers.
The SHAP analysis also validated the model by confirming that its feature influences aligned with current clinical guidelines from the European Association of Urology (EAU) and American Joint Committee on Cancer (AJCC). Guidelines consistently recommend surgery for younger patients with localized disease and favor non-surgical approaches for older patients with extensive or metastatic disease -- exactly the patterns the SHAP analysis revealed the model had learned from the data without being explicitly programmed with these rules.
Kaplan-Meier survival analysis comparing the three treatment groups -- no surgery, cryotherapy, and radical prostatectomy -- showed striking differences in long-term outcomes. Radical prostatectomy had the highest 5-year survival rate (96.2%) and the highest 10-year survival rate (89.2%). No surgery had a 5-year survival rate of 84.6% and a 10-year survival rate of 67.8%. This represents a 20.36% improvement in 10-year survival for surgical patients overall -- a clinically and statistically substantial difference observed consistently across all age groups.
Cryotherapy showed a 5-year survival rate of 86.1% and a 10-year survival rate of 64.4% -- notably lower than radical prostatectomy and comparable to the non-surgical group overall. However, for patients aged 75 and older specifically, cryotherapy showed a survival advantage over other surgical approaches, suggesting it may be a preferred intervention for older patients who are candidates for local treatment but not ideal candidates for the physiological demands of radical surgery.
These survival differences must be interpreted with caution because of potential confounding. Patients who received radical prostatectomy were systematically younger and had lower-stage disease than those who did not -- the very factors that both predict longer survival and make surgery the appropriate treatment choice. Disentangling the independent effect of the surgical procedure itself from the baseline characteristics that led to its selection requires more sophisticated causal analysis than survival curves alone can provide.
The clinical implication is that this ML model could serve as a decision support tool to standardize treatment recommendations across different clinical settings, reducing the variability in care that occurs when physicians apply different interpretive frameworks to the same clinical data. The combination of the CatBoost prediction with SHAP explanations provides the transparency necessary for clinicians to review, challenge, and contextualize any individual recommendation -- a critical requirement for real-world clinical adoption.
The most significant limitation is that SEER does not capture all clinically relevant variables. Patient comorbidities, performance status, quality of life preferences, urological function, and other factors that routinely influence treatment decisions in clinical practice are absent from the registry. A patient's preference for preserving urinary or sexual function, for example, is one of the most important considerations in the surgical decision for localized prostate cancer -- yet such patient-reported factors are entirely missing from the model's inputs.
Because the model was built from historical treatment assignments in a registry, it learns the patterns of how treatments were allocated in the past -- not necessarily which treatment would produce the best outcomes for each patient. This is an observational study, not a randomized trial, and the models reflect clinician behavior patterns rather than optimized evidence-based recommendations. SHAP values describe how each feature influenced the observed treatment choice, but this should not be interpreted as evidence that those features should causally determine the optimal treatment for each patient.
The research team plans to address the current limitations by incorporating imaging data, genomic information, and more granular clinical variables in future iterations of the model. Integrating multiparametric MRI findings, biopsy spatial mapping, genetic markers, and patient-reported outcomes alongside the demographic and pathological variables already in the model could substantially improve predictive accuracy and clinical relevance. Such a model would align more closely with the multimodal, individualized approach that prostate cancer care is moving toward.
Despite these limitations, the study makes a meaningful contribution by demonstrating that a large-scale ML model with interpretability tools can learn clinically sensible treatment patterns from population data, produce predictions that align with established guidelines, and quantify the relative importance of each factor in a transparent way. As registry data, electronic health records, and ML infrastructure continue to mature, models of this type have genuine potential to support more consistent and personalized treatment planning in prostate cancer care.