Abstract 364P
Background
The objective of this study is to develop prediction models and provide diagnostic results for prostate biopsy among patients.
Methods
This paper presents an improved tree model, called the Light gradient boosting machine model, for building a prostate cancer risk prediction model. The research data used in this study (Waves 2008-2022) are obtained from 1987 patients who underwent transperineal biopsy at West China Hospital. All patients had biopsy indications (high levels of prostate specific antigen (PSA), suspect lesions in digital rectal examination (DRE) or magnetic resonance imaging (MRI) suspect lesions). A total of 13 baseline variables were considered as candidate features, and eight machine learning models were compared in the experiment. SHapley Additive exPlanations (SHAP) was used for explaining the prediction model. The analyses were conducted in 2023.
Results
The Light gradient boosting machine model demonstrated the best performance, achieving an area under the receiver operating characteristic (ROC) curve of 0.76 (CatBoost=0.71, KNN=0.65, LDA=0.61, CART=0.67, SVM=0.64, NB=0.55, XGBoost=0.71). According to the interpretable analysis, the apparent diffusion coefficient (ADC) of MRI was identified as the most important feature in the prediction model. Other important features included prostate volume, red blood cell (RBC) count, neutrophil count, and platelet count.
Conclusions
This study suggests that optimal models show promise in screening out patients at high risk for prostate cancer in prostate biopsies. The decisions regarding prostate biopsies should specifically focus on the ADC of MRI, prostate volume, and blood routine indices such as RBC count, platelet count, and neutrophil count.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.