To develop and validate a machine learning model for individualized prediction of bone metastasis in patients with newly diagnosed prostate cancer.
Approach:
Data Collection: Retrospective data from 327 patients with newly diagnosed prostate cancer were collected from two tertiary hospitals.
Model Development: Patients were randomly assigned to a training set (n = 229) and an internal validation set (n = 98). The Boruta algorithm identified significant predictors, and seven machine learning models were developed and evaluated.
Model Evaluation: Model performance was assessed using ROC curves, calibration, and decision curve analysis. The best-performing model was interpreted using SHAP and deployed as an online prediction tool.
Key Findings:
Six predictors were identified: clinical T stage, Gleason score, total prostate-specific antigen (tPSA), alkaline phosphatase (ALP), regional lymph node metastasis, and fibrinogen.
The random forest model achieved an area under the curve (AUC) of 0.902 in the training set and 0.906 in the internal validation set.
Calibration curves showed good agreement between predicted and observed outcomes.
Interpretation:
An interpretable random forest model for predicting bone metastasis in newly diagnosed prostate cancer was developed and validated.
Limitations:
The study is retrospective and conducted at two centers, which may limit generalizability.
External validation in prospective multicenter studies is needed.
Conclusion:
An interpretable random forest model for predicting bone metastasis in newly diagnosed prostate cancer was developed and validated.
