B-373 Interpretable Steroid Profile-Related Diagnostic Models for Prostate Cancer Based on Random Forest Model and Shapley Value

Abstract Background The role of steroids in the prostate remains sparse and inconsistent, limiting their application in diagnosing prostate cancer (PCa). Thus, we aimed to establish a steroid-related explainable random forest (RF) model for advancing PCa diagnosis. Methods 108 PCa patients consecutively referred for prostate biopsy and 112 age-matched patients with benign prostate hyperplasia (BPH) were included, with demographic data and parameters of prostate size collected. 80% of cases were randomly selected for modeling and 20% were used for validation. Serum steroid profile and oxidative stress related biomarkers including 8-oxo-7,8-dihydroguanosine (8-oxoG) and 8-oxo-7,8-dihydro-2’-deoxyguanosine (8-oxodG), were determined by liquid chromatography-tandem mass spectrometry. Comparisons between groups and correlation analyses were conducted. Interpretable steroid profile-related diagnostic models were established based on RF method, and the Shapley additive explanation (SHAP) method was applied to interpret the estimation. Results Compared to the BPH group, PCa patients had less common lower urinary tract symptoms, smaller prostate volume, and lower levels of gonadal hormones, but higher levels of total prostate-specific antigen (tPSA) and free PSA (fPSA). Correlation analysis revealed that some androgens were significantly correlated with parameters of prostate size, 8-oxoG and 8-oxodG. The RF model containing ten features yields a high accuracy of 88.37% and an area under the curve of 0.913 in validation. And the ratio of fPSA and tPSA, tPSA and testosterone, testosterone and estrone were the top three most important features. Visualization of the effect of incorporated features in a single instance as well as total patients was achieved by using Shapley values. Conclusion Steroid profile varied between groups and a well-performed RF model integrating steroids and PSA was established and then interpreted by SHAP method to help advancing PCa diagnosis.