Comparative analysis of machine learning-derived nomogram and biomarkers in predicting side-specific extraprostatic
Ruchika Reddy Chimmula1, Mark Green2, Mark Tann2
1Advanced Molecular Imaging in Radiotherapy (AdMIRe) Research Laboratory, School of Health Sciences, Purdue University, West Lafayette, IN, United States of America.
Aim:
This study aimed to assess and compare the performance of nomograms and machine learning (ML) techniques using preoperative biomarkers for predicting side-specific extraprostatic extension (EPE) in prostate cancer, which is linked to poor outcomes and early recurrence. Accurate preoperative prediction can guide clinical decisions and improve treatment.
Materials And Methods:
A retrospective analysis was conducted using data from 108 prostate cancer patients undergoing radical prostatectomy. Clinical, imaging, and genomic data were collected, including PSA density, ISUP biopsy grade, fraction of positive biopsy cores, 68Ga-PSMA-11 PET, MRI, and Decipher Genomic Classifier (DGC) scores. Predictive models were built using logistic regression (LR) and extreme gradient boosting (XGBoost) algorithms, incorporating different combinations of these inputs. Model performance was evaluated using area under the ROC curve (AUC).
Results:
The median patient age was 61.5 years. XGBoost outperformed LR across most biomarker combinations. PET+DGC models had the highest AUC (0.85 for XGBoost), followed by PET+MRI + DGC (0.83). XGBoost consistently achieved higher AUCs than LR, particularly for DGC and combined input models. PET-only predictions were stronger than those based solely on MRI or genomics, but multi-modal combinations significantly enhanced prediction accuracy.
Conclusion:
This is the first study to integrate PSMA-PET, MRI, and genomics in ML-based nomogram models for side-specific EPE prediction. XGBoost models demonstrated superior predictive power, especially when combining PET and DGC. These findings highlight the potential of a multi-biomarker, machine learning approach to improve preoperative risk stratification and support personalized treatment planning. Further studies will validate this model in larger cohorts.
