3254 - Investigating the Machine Learning-Derived Nomogram Models and Biomarkers for Side-Specific Prediction of Extraprostatic Extension of Prostate Cancer - Preliminary Results

O. M. Oderinde¹, R. Chimmula², K. Huang³, Y. Yue³, M. Green⁴, M. Tann⁴, M. Koch⁵, K. Collins⁶, C. Njeh³, J. A. Holmes³, and C. Bahler⁵; ¹Advanced Molecular Imaging in Radiotherapy (AdMIRe) Research Laboratory, Purdue University, West Lafayette, IN, ²Advanced Molecular Imagiing in Radiotherapy (AdMIRe) Research Laboratory, School of Health Sciences, Purdue University, West Lafayette, IN, ³Department of Radiation Oncology, Indiana University School of Medicine, Indianapolis, IN, ⁴Department of Radiology and Imaging Sciences, Indiana University School of Medicine, Indianapolis, IN, ⁵Department of Urology, Indiana University School of Medicine, Indianapolis, IN, ⁶Department of Pathology, Indiana University School of Medicine, Indianapolis, IN

Purpose/Objective(s): This study investigated the performance of machine learning-derived nomograms utilizing logistic regression (LR) and random forest (RF) algorithms, along with biomarkers from 68Ga-PSMA-11 PET, MRI, and the Decipher test, in predicting side-specific extraprostatic extension (EPE) in radical prostatectomy (RP) patients. Materials/

Methods: In this retrospective study, data from prostate cancer patients who underwent RP at Indiana University Hospital, alongside their 68Ga-PSMA-11 PET, MRI, and Decipher results were analyzed. This included their clinical information (fraction of positive cores on biopsy, prostate-specific antigen density, and ISUP biopsy grade) alongside results from three biomarkers: 68Ga-PSMA-11 PET, MRI, and the Decipher test. Several machine learning-derived nomogram models were developed using LR and RF algorithms. These models incorporated various combinations of the biomarkers (PET-only, MRI-only, Decipher-only, Decipher-only, PET+Decipher, PET+MRI, MRI+Decipher, and PET+MRI+Decipher) along with the clinical data. The performance of each model was assessed using the area under the receiver operating characteristic curve (AUC-ROC).
Results: In the nomogram development and validation process, we selected 110 patients with a median age of 61.5 years (range 48 –77 years), PSA score of 2.7–7.15 ng/mL, ISUP biopsy grade of 1–5, and a fraction of positive cores of 0.50 ± 0.36. For the PET-only, MRI-only, Decipher-only, PET+Decipher, PET+MRI, MRI+Decipher, and PET+MRI+Decipher variables, the performance of the RF models was 0.82, 0.76, 0.65, 0.83, 0.76, 0.73, and 0.83, respectively. In addition, the performance of the LR models was 0.80, 0.71, 0.62, 0.78, 0.79, 0.78, and 0.63, respectively.
Conclusion: From this study, the RF algorithm generally outperformed LR algorithm in this task. 68Ga-PSMA PET-only variables provided better EPE prediction than MRI or Decipher variables used individually among the evaluated biomarkers. However, the most accurate predictions were achieved by combining Decipher and MRI data with PET data, particularly within the RF models. This suggests that a combined approach using multiple biomarkers holds promise for improving EPE prediction. Hence to guide treatment decisions, such as nerve-sparing RP. Further study with a larger patient population is needed to refine the models accuracy and validate its clinical utility.

Abstract 3254 – Table 1

Variables	RF - AUC Score	LR- AUC Score
Path EPE versus Decipher only	0.65	0.62
Path EPE versus PET only	0.82	0.80
Path EPE versus MRI only	0.76	0.71
Path EPE versus Decipher+MRI	0.73	0.78
Path EPE versus PET+MRI	0.76	0.79
Path EPE versus Decipher+PET	0.83	0.78
Path EPE versus Decipher+PET+MRI	0.83	0.63