Using AI to identify optimal clinical, genomic, and radiographic prognostic features and novel risk classifiers compared to routinely available risk classifiers.

Abstract

396 Background: This study investigated the use of artificial intelligence (AI) and traditional statistical techniques to identify and validate clinical (Clin), MRI-report-derived (MRIr), and genomic classifier (GC) features important for prognostic biomarker development and staging systems for prostate cancer metastasis. AI models were compared with conventional risk classifiers for prognostic accuracy. Methods: Data from 576 patients in a prospectively collected research registry, including 35 Clin, 1 GC, and 7 MRIr features, were used. Random forest (RF) classification and artificial neural networks (ANNs) using Leave-One-Out Cross-Validation evaluated the key features predictive of metastasis. Models with 35, 9, 6, or 0 Clin features were compared to those ± MRIr or GC features. ANN prognosticators from each model were fed into competing-risk regressions and compared to other established risk classifiers (CAPRA, STARCAP, 22-gene genomic assay (Decipher), Cell-Cycle Progression (CCP, Prolaris) score, and Combined Clinical Cell-Cycle Risk (CCR, Prolaris)) by the 5-year AUC (AUC5) for prognostic accuracy. Results: AI-derived models combining Clin ±GC +MRIr features (AUC5 range 0.72-0.76) outperform models excluding MRIr features (AUC5 0.66-0.71). AI models perform similarly to contemporary risk classifiers like Prolaris, CAPRA, STARCAP (AUC5 of 0.79. 0.77, 0.76, respectively (Table)). Models with 6 Clin, 4 MRIr, and a GC (Clin: Age, BMI, PSA, cT, alcohol use, and primary Gleason score; MRIr: prostate volume, max tumor diameter, extracapsular spread, and lymphadenopathy suspicion; GC: CCP score) performed as well as conventional classifiers. The classifier with the best ability to prognosticate early metastasis was Prolaris (AUC5 0.79). Conclusions: The findings indicate that AI biomarker discovery pipelines, integrating Clin, MRIr, and GC features, can successfully identify new and previously identified prognostic factors. AI-derived models have similar accuracy in predicting metastasis compared to contemporary risk classifiers. Adding MRIr and GC features improves risk prediction of early metastasis in AI-trained models. Lifestyle factors may have an important role in metastasis. Total # Features Genomic Features Radiographic Features Clinical Features Time-Dependent Competing Risk Regression 3-Year AUC 5-Year AUC 35 Clinical Feature AI Models (n=576) 43 1 7 35 0.70 0.72 42 7 35 0.69 0.72 36 1 35 0.64 0.66 35 35 0.62 0.65 11 Clinical Feature AI Models (n=576) 16 1 6 9 0.72 0.76 15 6 9 0.71 0.75 10 1 9 0.67 0.71 9 9 0.67 0.70 6 Clinical Feature AI Models (n=576) 11 1 4 6 0.73 0.75 10 4 6 0.73 0.75 7 1 6 0.61 0.64 6 6 0.62 0.65 0 clinical feature AI Models 7 1 6 0 0.66 0.68 6 6 0 0.65 0.67 Prolaris (n=570) 6 1 5 0.75 0.79 STARCAP (n=566) 5 5 0.72 0.77 CAPRA (n=570) 6 6 0.72 0.76 CCP score (n=576) 1 1 0.73 0.74 Decipher (n=25) 1 1 0.40 0.29