Background: Given the low specificity of the current standard of care diagnostic tests for prostate cancer (PCa), there is an unmet clinical need for higher specificity tests to counter overdiagnosis of grade group (GG) 1 PCa. The aim of the study was to create optimized neural network risk models using PSA, free PSA, and other useful clinical features, and to validate the accuracy of the risk models to predict GG ≥2 PCa using real-world samples and data. Methods: Men aged 40-75 years with PSA >=3ng/mL and a biopsy referral were recruited into cohorts from sites in Canada (Kipnes Urology Centre (KUC), Edmonton, AB, and Prostate Cancer Centre (PCC), Calgary, AB) and the United States (Johns Hopkins University (JHU), Baltimore, ML and UCLA, Los Angeles, CA). Risk models to predict all GGs or GG ≥2 PCa were fit using data from KUC and JHU (train cohort n = 1037) while fixed models were validated on PCC (n = 401) and UCLA (n = 945). Prediction models were created using neural networks to generate the patient’s risk score. To compare the risk model test with PSA, the high-grade cancer detection sensitivity was fixed, and the number of biopsies needed to achieve that sensitivity was evaluated. Threshold values for the training cohort were determined using at least 95% sensitivity and maximum specificity. Threshold values for other clinical features and risk calculator outputs were set to match the test sensitivity when possible. Results: The optimized neural network risk models test had the highest area under the curve (AUC 0.81) for predicting GG ≥2 PCa on the validation cohorts compared to four other risk calculators; Prostate Cancer Prevention Trial Risk Calculator 2.0 (PCPTRC) with free PSA (0.78, p-value<0.001), Prostate Biopsy Collaborative Group risk calculator (PBCG, 0.73, p-value<0.0001), European Randomized study of Screening for Prostate Cancer risk calculator 3 (ERSPC-3, 0.72, p-value<0.0001), and PCPTRC with no free PSA (0.71, p-value<0.0001) and PSA (0.66, p-value<0.0001). At a threshold of 18.6%, this test provided 94% sensitivity, 37% specificity, 49% positive predictive value, and 90% NPV for predicting GG ≥2 PCa. If a biopsy was performed only when the tests’ risk score prediction was ≥17.8% for GG ≥2 PCa, 36-44% of unnecessary prostate biopsies could be avoided while missing 4-12% of patients with GG ≥2 prostate cancer. Conclusions: This neural network risk model is a tool that can be used to inform the physician of a patient’s risk of having GG ≥2 PCa on prostate biopsy. Using this novel test in the clinic is expected to significantly reduce the number and burden of unnecessary prostate biopsies.