When prostate cancer spreads beyond the gland itself to regional lymph nodes, it is classified as node-positive disease. Lymph nodes are small immune system structures throughout the body that act as relay stations -- cancer cells that reach lymph nodes have taken a significant step toward systemic spread. According to NCCN and EAU guidelines, lymph node involvement (LNI) categorizes patients as a single high-risk group.
Approximately 5 to 10% of newly diagnosed prostate cancer patients have clinically suspected lymph node invasion without distant metastasis. Among patients who undergo radical prostatectomy, pathological confirmation of LNI is found in 0 to 37% depending on risk category. Once LNI is detected, the risk of cancer recurrence and death increases substantially.
Despite sharing a high-risk classification, patients with LNI have highly variable outcomes. Some survive for many years with aggressive treatment, while others experience rapid progression. Existing prognostic tools for this group are limited -- most were built on small, single-institution cohorts and apply only to patients who underwent surgery, excluding those treated non-surgically. A more accurate, broadly applicable survival prediction tool was needed.
The standard statistical tool for survival prediction in oncology is Cox regression, a model that estimates how individual patient characteristics affect the risk of an event over time. Cox regression has been the backbone of survival analysis for decades, but it assumes that the relationships between predictors and survival are linear and proportional -- assumptions that biological systems often violate.
Machine learning survival algorithms offer important advantages. They can model complex nonlinear relationships between variables, automatically discover interactions between predictors, and are not constrained by the proportional hazards assumption. This makes them particularly well-suited for medical prognosis, where multiple variables interact in ways that simple linear models cannot capture.
This study from Xi'an Jiaotong University used a large national database to develop and compare three machine learning survival models against Cox regression for predicting overall survival in lymph node-positive prostate cancer patients. The best-performing model was then translated into a publicly accessible web tool that any clinician can use to generate individualized survival predictions for their patients.
Data came from the SEER (Surveillance, Epidemiology, and End Results) database, a large publicly available collection of cancer registry data from 18 population-based registries in the United States. SEER captures diagnosis, treatment, and outcome information for hundreds of thousands of cancer patients, making it one of the most powerful resources for population-level cancer research.
Patients diagnosed with prostate cancer with lymph node involvement between 2000 and 2019 were identified. After applying inclusion and exclusion criteria, 3,280 non-metastatic patients with LNI were included. The dataset was split 80/20 into a training cohort (2,624 patients) for model development and a validation cohort (656 patients) for performance evaluation. The primary outcome was overall survival (OS) from the date of diagnosis to death from any cause.
Nine predictor variables were used: age at diagnosis, race, marital status, clinical T stage, PSA level at diagnosis, Gleason Score, number of positive lymph nodes, whether radical prostatectomy was performed, and whether radiotherapy was administered. Variable selection used forward and backward stepwise selection, though PSA was retained even when not statistically significant on stepwise analysis because of its known clinical relevance and because ML models can capture its contribution through nonlinear pathways.
Gradient Boosting Survival Analysis (GBSA) builds predictions by sequentially adding simple models, each one correcting the errors of the previous one. This iterative process creates a powerful composite model that captures complex nonlinear patterns. Gradient boosting has consistently demonstrated top performance across a wide range of prediction tasks in medicine and elsewhere.
Random Survival Forest (RSF) trains a large ensemble of decision trees, each built on a random subset of the training data and predictor variables. The final prediction is the average across all trees. The randomness injected during training reduces overfitting and makes the ensemble robust to variability in the training data.
Extra Survival Trees (EST) is similar to RSF but introduces additional randomness by selecting split thresholds randomly rather than optimally at each tree node. This further reduces variance at the cost of slightly increased bias, often yielding excellent overall performance. All three algorithms were implemented using the scikit-survival Python library, which handles the statistical complexity of censored survival data -- a crucial advantage over simpler machine learning classification approaches that ignore patients lost to follow-up.
All three machine learning models showed better discrimination than Cox regression. The mean time-dependent AUC for GBSA was 0.782, RSF was 0.779, and EST was 0.781, all higher than Cox regression at 0.770. The concordance index (c-index) showed a similar pattern: GBSA 0.745, RSF 0.743, EST 0.745 versus Cox regression at 0.734. While the numerical differences are modest, they are consistent and meaningful in a clinical context.
Model calibration -- how accurately the predicted probabilities match observed survival rates -- was similar across all four models. The integrated Brier score (IBS), a measure of calibration over time (lower is better), was 0.114 for all ML models versus 0.115 for Cox regression. These low scores indicate that all models made well-calibrated probability estimates, not just discriminating between high and low risk patients.
The GBSA model performed best overall and was selected as the basis for the web prediction tool. However, the EST model showed competitive performance at certain time points, illustrating that time-dependent AUC can reveal differences hidden by summary c-index values. No single algorithm dominated at all time points, suggesting that clinical application should consider the specific prediction horizon of interest.
SHAP analysis of the best-performing GBSA model identified Gleason Score as the most important predictor of overall survival -- the higher the Gleason Score, the higher the cancer grade, and the greater the risk of death. Patients with GS of 9 or higher had more than three times the mortality risk of those with GS of 3+4 or lower in Cox regression analysis, confirming the established prognostic importance of tumor differentiation.
The number of positive lymph nodes was the second most important variable. Patients with three or more involved lymph nodes had a 63% higher risk of death compared to those with only one positive node (HR 1.631). This dose-response relationship between lymph node burden and survival is clinically intuitive -- more extensive nodal spread indicates more aggressive systemic disease.
Notably, marital status emerged as the third most important variable in SHAP analysis, with unmarried patients having a 64% higher mortality risk (HR 1.640). This finding echoes a well-established literature showing that social support -- which marriage typically provides -- influences cancer screening, treatment adherence, and recovery. The inclusion of such social determinants of health in AI models highlights factors that clinicians and support systems can address beyond pure tumor biology.
An important finding from the Cox regression analysis is that both radical prostatectomy and radiotherapy were independently associated with improved survival in node-positive patients. Patients who did not undergo prostatectomy had 52% higher mortality risk, and those without radiotherapy had 54% higher risk. This supports emerging evidence that local treatment -- surgery or radiation -- provides benefit even in patients with lymph node metastases, challenging earlier assumptions that nodal involvement made local treatment futile.
The debate over surgery versus radiation for node-positive prostate cancer remains unresolved. Some studies show no significant difference in overall mortality between the two approaches, while others suggest a cancer-specific survival advantage for surgery. A large prospective randomized trial (SPCG-15) is currently underway to provide definitive evidence. The GBSA model developed here could help identify which patients are most likely to benefit from aggressive local treatment.
The freely accessible web-based prediction tool built on the GBSA model allows clinicians to enter nine patient characteristics and generate individualized survival probability curves. Such tools can improve shared decision-making by showing patients and physicians not just a generic prognosis for their risk category but a personalized estimate that reflects their specific combination of tumor and demographic factors.
The SEER database, while large and population-representative, has important gaps. Androgen deprivation therapy (ADT) information -- a critical treatment for node-positive prostate cancer -- was not available, meaning its prognostic impact could not be modeled. Since ADT is the cornerstone of systemic therapy for this patient group, future models that incorporate ADT data could be substantially more informative.
The database also lacks detailed clinical information such as comorbidities, performance status, and pathological details beyond what is captured in registry fields. The combination of broad population scope and limited clinical depth is an inherent tradeoff of registry-based research. External validation on independent patient cohorts from other institutions and countries is needed before the tool can be confidently used in clinical practice globally.
Finally, the study's primary endpoint was overall survival -- a measure that captures all deaths including those from causes unrelated to prostate cancer. Including cancer-specific survival and progression-free survival as additional endpoints in future work would provide a more nuanced picture of disease outcomes and enable more precise identification of patients who need more intensive treatment versus those who can be safely managed with less aggressive approaches.