Prediction of factors associated with upgrade/upstage status after prostate biopsy in prostate cancer based on machine learning models

Prostate cancer (PC) is a significant disease affecting men's health and is the second most common cancer in men. Accurate risk classification is crucial for transitioning from diagnosis to treatment in PC. There are important parameters available for performing risk classification, including the patient's clinical characteristics, age, serum prostate-specific antigen (PSA) levels, prostate-specific antigen density (PSAD), free to total PSA ratio (F/T PSA), magnetic resonance imaging (MRI), and pathology results from prostate biopsy (PB). A review of the literature reveals significant discrepancies between the ISUP grades for PB and the T stages in MRI compared to the final pathology results from radical prostatectomy (RP). These discrepancies indicate that accurate risk classification for patients cannot be performed at the time of diagnosis. In this study, we retrospectively evaluated the data from patients who underwent RP at our clinic. We examined the following information: age of the patients, S/T PSA, PSAD, T stage in MRI, ISUP grade in PB, number of positive cores and clinically significant positive cores, presence of HGPIN, ASAP, perineural invasion, cribriform pattern, and existence of intraductal morphology. The aim was to develop machine learning (ML) based models using these features to predict the factors associated with upgrade and upstage statuses. In our study, fourteen ML-based models were applied for predicting upgrade and upstage. The model with the highest performance for upgrade prediction was the Random Forest Classifier, which achieved AUC: 0.7, accuracy: 0.78, recall: 0.43, precision: 0.62, and an F1 score of 0.48. These results are comparable to those obtained in the existing literature. Among the models applied for upstage prediction, the Logistic Regression (LR) model was found to be the most successful. The LR model yielded results of AUC: 0.79, accuracy: 0.73, recall: 0.7, precision: 0.71, and an F1 score of 0.71, demonstrating high performance. There are very few studies in the literature focused on upstage prediction. The absence of significant features such as adverse pathology for PB, HGPIN, ASAP, and perineural invasion among the parameters in those studies distinguishes our model. The success of our model may be attributed to the inclusion of these parameters. In conclusion, we developed a model for predicting upgrade and upstage statuses. We believe that this will enable more accurate risk stratification for patients and facilitate appropriate treatment planning.