3116 - Multimodal Prediction of Hormonal Therapy Response in PSMA-Positive Prostate Cancer Using PSMA PET Deep Features and Clinical Variables

12:30pm - 01:45pm ET

Poster Hall - Exhibit Hall A

Screen: 33

POSTER

Presenter(s)

Omar Awad, MD, MS - Baylor College of Medicine, Houston, TX

A. M. Saad¹, M. Abdelhalim², Y. A. Elsaid³, A. K. Morsi¹, A. H. Algendy¹, K. A. Belal⁴, A. S. Hassan⁵, A. S. Eldardiry⁵, Y. Y. Basha⁵, S. A. Almalky⁵, Z. Elmahdy⁶, A. O. Esayed⁷, A. M. Eldrieny⁸, A. S. Eldesoky⁵, F. M. Aboumadawy⁹, A. Nawar¹⁰, O. Awad¹¹, H. M. Hegazy¹², B. El Sabaa¹³, S. H. Gamie¹⁴, H. El Mansy¹⁵, A. A. Ismail¹⁶, Y. Elkerm⁹, and A. A. ElSaid²; ¹Alexandria University, Faculty of Medicine, Alexandria, Alexandria, Egypt, ²Clinical Oncology Department, Faculty of Medicine, Alexandria University, Alexandria, Egypt, ³Diagnostic and Interventional Radiology department, Faculty of Medicine, Alexandria University, Alexandria, Egypt, ⁴Alexandria University, Department of Oncology, Alexandria, Alexandria, Egypt, ⁵Alexandria University, Department of Oncology, Alexandria, Egypt, ⁶Alexandria University, Faculty of Medicine, Alexandria, Egypt, ⁷Pharos University, Faculty of Computer Science and Artificial Intelligence, Alexandria, Egypt, ⁸Faculty of Applied Health Sciences Technology, Pharos University in Alexandria, Alexandria, Egypt, ⁹Department of Cancer Management and Research, Medical Research Institute, Alexandria University, Egypt, Alexandria, Egypt, ¹⁰Department of Radiation Oncology Kansas University Medical Center, Kansas, KS, ¹¹Baylor College of Medicine Houston, TX, United States, Houston, Texas, Egypt, ¹²Clinical Oncology and Nuclear Medicine Department- Alexandria University, Alexandria, Egypt, ¹³Faculty of Medicine Alexandria University, Alexandria, Egypt, ¹⁴UC San Diego Health Department of Radiology, San Diego, CA, ¹⁵Cancer Management and Research Department, Medical Research Institute, Alexandria, Egypt, ¹⁶Department of clinical oncology and nuclear medicine, Faculty of Medicine, Alexandria University, Alexandria, Egypt

Purpose/Objective(s): To develop and evaluate a multimodal machine learning framework to predict complete response (CR) to hormonal therapy in PSMA-positive prostate cancer by integrating deep representations from baseline PSMA PET with structured clinical and treatment variables, and to determine whether feature fusion provides measurable predictive signal in a limited cohort.

Materials/Methods:

A retrospective cohort of PSMA-positive prostate cancer patients treated with hormonal therapy was analyzed, with patient-level splits of 120 for training and 30 for independent testing (CR prevalence 36.7% in both cohorts). Baseline PSMA PET volumes were encoded using a three-dimensional ResNet initialized with MedicalNet weights, generating 8,192-dimensional embeddings per patient. These were concatenated with structured clinical, pathological, treatment, and metastatic variables to create a high-dimensional multimodal feature space (65,566 features). Supervised classifiers were trained using repeated stratified five-fold cross-validation with randomized hyperparameter search. Preprocessing included scaling, variance filtering, supervised feature selection (SelectKBest), and optional dimensionality reduction with principal component analysis (PCA). Class imbalance was addressed using balanced weighting and XGBoost scale_pos_weight. Evaluated models included Random Forest, XGBoost, logistic regression, k-nearest neighbors, and a multilayer perceptron. Performance was assessed using ROC-AUC as the primary endpoint, with PR-AUC, F1-score, and accuracy as secondary metrics.

Results: Across models, cross-validation demonstrated moderate discrimination with best cross-validated ROC-AUC values of 0.680 (logistic regression), 0.679 (XGBoost), 0.660 (random forest), and 0.659 (multilayer perceptron). On independent testing (n=30; 11 CR, 19 non-CR), the highest test ROC-AUC was achieved by Random Forest (ROC-AUC 0.679; PR-AUC 0.635; accuracy 0.667; F1-score 0.375 for CR). The multilayer perceptron demonstrated comparable discrimination (ROC-AUC 0.675; PR-AUC 0.617) with higher CR-class F1-score (0.552) but lower overall accuracy (0.567). XGBoost achieved ROC-AUC 0.632 with PR-AUC 0.593, accuracy 0.667, and CR-class F1-score 0.444. Logistic regression yielded ROC-AUC 0.622 and PR-AUC 0.602 on testing.

Conclusion: Multimodal integration of PSMA PET imaging features with clinical variables demonstrates measurable predictive signal for complete response to hormonal therapy, achieving test ROC-AUC up to 0.679 despite limited sample size. These findings support the potential role of multimodal modeling and transfer learning for individualized response stratification. Larger multi-institutional validation is required to confirm robustness and clinical utility.