Development and Validation of a Clinic Machine Learning Classifier for the Prediction of Risk Stratifications of Prostate Cancer Bone Metastasis Progression to Castration Resistance.

Objective: To explore the predictive factors and predictive model construction for the progression of prostate cancer bone metastasis to castration resistance.

Methods: Clinical data of 286 patients diagnosed with prostate cancer with bone metastasis, initially treated with endocrine therapy, and progressing to metastatic castration resistant prostate cancer (mCRPC) were collected. By comparing the differences in various factors between different groups with fast and slow occurrence of castration-resistant prostate cancer (CRPC). Kaplan-Meier survival analysis and COX multivariate risk proportional regression model were used to compare the differences in the time to progression to CRPC in different groups. The COX multivariate risk proportional regression model was used to evaluate the impact of candidate factors on the time to progression to CRPC and establish a predictive model. The accuracy of the model was then tested using receiver operating characteristic (ROC) curves and decision curve analysis (DCA).

Results: The median time for 286 mCRPC patients to progress to CRPC was 17 (9.5-28.0) months. Multivariate analysis showed that the lowest value of PSA (PSA nadir), the time when PSA dropped to its lowest value (timePSA), and the number of BM, and LDH were independent risk factors for rapid progression to CRPC. Based on the four independent risk factors mentioned above, a prediction model was established, with the optimal prediction model being a random forest with area under curve (AUC) of 0.946[95% CI: 0.901-0.991] and 0.927[95% CI: 0.864-0.990] in the training and validation cohort, respectively.

Conclusion: After endocrine therapy, the PSA nadir, timePSA, the number of BM, and LDH are the main risk factors for rapid progression to mCRPC in patients with prostate cancer bone metastases. Establishing a CRPC prediction model is helpful for early clinical intervention decision-making.