Prostate cancer is the most common non-skin cancer in men worldwide, with approximately 1.6 million new cases and 366,000 deaths annually. When the cancer spreads beyond the prostate, outcomes worsen dramatically.
Bone metastases (BM) represent a major complication, occurring in about 16.7% of metastatic prostate cancer cases. Patients with bone metastases experience significantly reduced five-year survival, chronic pain, pathological fractures, and severely impaired quality of life.
Current methods for detecting bone metastases -- primarily bone scan and PSMA-targeted nuclear imaging -- carry risks of radiation exposure, are expensive, and are not recommended for all patients. Bone scan is typically reserved for patients who already show signs of skeletal problems, potentially creating diagnostic delays.
There is a pressing need for a low-cost, risk-based screening tool that can identify which prostate cancer patients are most likely to develop bone metastases, so that those at high risk can be prioritized for early imaging and intervention.
The study used the SEER database (Surveillance, Epidemiology and End Results) -- a large, federally funded US cancer registry -- to identify prostate cancer patients diagnosed between 2010 and 2017. After applying inclusion and exclusion criteria, 207,137 patients were included, of whom 6,725 (3.25%) developed bone metastases.
Eight clinical variables were selected as potential predictors: age, race, tumor grade, PSA value, Gleason score, T stage, N stage, and marital status. The data was split 70/30 into training and internal test sets, with an external validation set of 644 patients from a Chinese hospital.
Six different machine learning algorithms were trained and compared: Decision Tree (DT), Random Forest (RF), Multilayer Perceptron (MLP), Logistic Regression (LR), Naive Bayes Classifiers (NBC), and eXtreme Gradient Boosting (XGB). This head-to-head comparison allowed identification of the most accurate approach.
Model performance was evaluated using AUC (area under the receiver operating characteristic curve), accuracy, sensitivity (recall rate), and specificity. The best-performing model was then deployed as a publicly accessible web predictor for clinical use.
All eight variables were significantly associated with bone metastasis in univariate analysis. However, multivariate logistic regression identified four independent risk factors: T stage, N stage, Gleason score, and PSA value. These four factors also ranked as the most important features in most machine learning models.
PSA value showed the strongest individual signal. Patients who developed bone metastases had a median PSA of 77.4 ng/mL compared to 6.4 ng/mL in those without bone metastases -- a more than 10-fold difference. Current guidelines recommend bone scan for patients with PSA above 20 ng/mL.
Gleason score had a dramatic dose-response relationship with bone metastasis risk. Patients with Gleason score 9 or above had a 32-fold higher odds of bone metastasis compared to those with Gleason score 6 or below. At Gleason 8, the odds ratio was still nearly 16 -- confirming that tumor aggressiveness strongly drives metastatic potential.
N stage (lymph node involvement) and T stage (tumor size/local extension) were also significant. Patients with lymph node metastases (N1) had a 2.9-fold higher odds of bone metastasis, and T4 tumors -- the most locally advanced -- carried a 2-fold higher risk than T1 tumors.
Among the six machine learning models tested, XGBoost (XGB) delivered the best predictive performance across all evaluation metrics in both internal and external validation. On the external test set, XGB achieved an AUC of 0.962, accuracy of 88.4%, sensitivity of 90.6%, and specificity of 87.9%.
The XGBoost model also excelled in internal 10-fold cross-validation with a mean AUC of 0.951 -- a technique that tests model stability by training and validating across multiple data partitions rather than just a single split.
Random Forest was the second-best performer (AUC = 0.949 on external validation), followed by Multilayer Perceptron (0.950) and Naive Bayes (0.934). Logistic Regression performed least well overall (AUC = 0.905), highlighting the advantage of non-linear machine learning methods over traditional statistical approaches.
XGBoost's superior performance likely reflects its design: it uses gradient boosting with regularization, trains multiple sequential decision trees where each tree corrects the errors of the previous one, and supports parallel computation for speed. These properties collectively reduce overfitting and improve generalization.
Based on the XGBoost model, the researchers developed a publicly accessible web-based predictor that allows clinicians to enter a patient's clinical variables and instantly receive a bone metastasis risk estimate. The tool requires only routinely available information: age, race, PSA value, Gleason score, tumor stage, lymph node status, grade, and marital status.
The web predictor is designed to function as a decision support tool -- not a replacement for clinical judgment -- to help physicians determine which patients warrant a bone scan or other advanced imaging for bone metastasis screening.
This approach addresses a real clinical gap: because bone scans carry radiation exposure and costs, they cannot be applied to all prostate cancer patients. A risk stratification tool that flags high-risk cases enables targeted, personalized screening rather than blanket testing.
Given the model's high sensitivity (90.6%), it correctly identifies approximately 9 out of 10 patients who will actually develop bone metastases -- minimizing the number of missed cases. High specificity (87.9%) means it also correctly clears the majority of patients who will not develop metastases, reducing unnecessary imaging.
Machine learning models can detect patterns in clinical data that traditional statistical methods miss. Unlike logistic regression, which assumes linear relationships between variables, XGBoost can capture complex, non-linear interactions between PSA, Gleason score, tumor stage, and other factors that collectively predict bone metastasis risk.
The observation that marital status contributed to predictions -- with unmarried men at higher risk -- was unexpected but consistent with prior research. Social factors like marital status may reflect differences in healthcare access, mental health, and adherence to follow-up care, all of which could affect disease detection and progression.
The study used over 200,000 patients from a population-level registry, providing statistical power far beyond most single-center studies. Importantly, external validation on patients from a Chinese hospital confirmed that the model generalizes across different healthcare settings and populations.
The model's limitations include reliance on diagnostic-era SEER data without treatment history, which means it cannot account for how therapies received after diagnosis might influence bone metastasis development. Future iterations could incorporate treatment response data for more dynamic risk assessment.