A Neural Network Model Combining [-2]proPSA, freePSA, Total PSA, Cathepsin D, and Thrombospondin-1 Showed Increased Accuracy in the Identification of Clinically Significant Prostate Cancer

SIMPLE SUMMARY: The widespread use of PSA for prostate cancer diagnosis significantly contributed to the high rate of overdiagnosis and overtreatment. For several years, the Prostate Health Index (PHI) has been proposed as a tool able to improve PSA specificity. More recently, the Proclarix test has been developed. The combination of these two tests promises to ameliorate risk stratification of PCa patients at initial diagnosis. In this study, we evaluated the performance of an artificial-neural-network-based model combining kallikrein markers included in PHI and the cancer-related markers of Proclarix for the prediction of positive biopsy and high-grade cancers. Our findings suggested that the combined model had an increased accuracy in the identification of pathological aggressive PCa at initial diagnosis. ABSTRACT: Background: The Prostate Health Index (PHI) and Proclarix (PCLX) have been proposed as blood-based tests for prostate cancer (PCa). In this study, we evaluated the feasibility of an artificial neural network (ANN)-based approach to develop a combinatorial model including PHI and PCLX biomarkers to recognize clinically significant PCa (csPCa) at initial diagnosis. Methods: To this aim, we prospectively enrolled 344 men from two different centres. All patients underwent radical prostatectomy (RP). All men had a prostate-specific antigen (PSA) between 2 and 10 ng/mL. We used an artificial neural network to develop models that can identify csPCa efficiently. As inputs, the model uses [-2]proPSA, freePSA, total PSA, cathepsin D, thrombospondin, and age. Results: The output of the model is an estimate of the presence of a low or high Gleason score PCa defined at RP. After training on a dataset of up to 220 samples and optimization of the variables, the model achieved values as high as 78% for sensitivity and 62% for specificity for all-cancer detection compared with those of PHI and PCLX alone. For csPCa detection, the model showed 66% (95% CI 66–68%) for sensitivity and 68% (95% CI 66–68%) for specificity. These values were significantly different compared with those of PHI (p < 0.0001 and 0.0001, respectively) and PCLX (p = 0.0003 and 0.0006, respectively) alone. Conclusions: Our preliminary study suggests that combining PHI and PCLX biomarkers may help to estimate, with higher accuracy, the presence of csPCa at initial diagnosis, allowing a personalized treatment approach. Further studies training the model on larger datasets are strongly encouraged to support the efficiency of this approach.