This study developed a visualized machine learning XGBoost model to assess the PCA risk in moderate or severe BPH patients. This model was based on non-invasive parameters including STK1P, age, FPSA, and FTPSA. The XGBoost model performed better than other machine learning and logistic models. The XGBoost model helps to accurately select patients at high risk for PCA on subsequent biopsies.
The values of STK1P, age, FPSA, and FTPSA in PCA risk assessment were consistent with previous studies. Several studies have shown that STK1P levels are greater in PCA patients than in patients with benign prostate disease or healthy individuals11,13,14,15. Previous data on the global cancer burden show that the risk of prostate cancer in China increases with age16. Similarly, the present study showed that age is positively associated with the risk of prostate cancer. FPSAs have been extensively studied in relation to PCA detection17. Studies have also shown that combinations of various parameters, including FTPSA, can significantly increase PCA detection rates (Wang et al. 2006). Similar associations have also been reported in the current study. Our study also highlights the importance of establishing standardized protocols for the measurement and interpretation of biomarkers, ensuring consistency and reliability across different populations.
Other biomarkers were found to be associated with PCA in previous studies, but not in this study. For example, elevated levels of CA153 have been found in a variety of cancer types, including PCA.18and may be used as a potential marker. NSE is an enzyme found in neuroendocrine cells. Its role in PCA is not well defined, but has been reported as a marker in certain circumstances, particularly in progressive and metastatic castration-resistant PCA19. CEA is a tumor marker that can rise in many types of cancer, including progressive PCA.20. The reasons why we found no such associations in this study could include small sample sizes, different measurements, and different populations. Therefore, a larger study sample size is required to further explore these observations.
A single predictor, such as TPSA, may not be sufficient to accurately assess PCA risk3. Therefore, this study developed a new multivariable model that could assist in accurate identification of high-risk individuals to optimize PCA screening strategies. Currently, the Chinese PCA Screening Guidelines recommend using TPSA as a predictor of early screening, with a sensitivity of 91%.3,4. However, the specificity of TPSA in early PCA screening in the Chinese population is as low as 41%, which corresponds to a high percentage of false positives and unnecessary biopsies.4. In this study, we used a new machine learning algorithm XGBoost, which is highly efficient, flexible and descriptive in our previous model development research.twenty one. The XGBoost model performed better than traditional logistic models in PCA risk assessment in the current study. The XGBoost model has both high sensitivity (95%) and specificity (98%) and may serve as a robust, accurate, interpretable tool that can improve clinical outcomes by promoting accurate detection of PCA risk. Additionally, the Xgboost decision tree was visually presented to illustrate model decisions. This is important for healthcare professionals to understand the model in clinical practice.
Error analysis of the XGBoost model identified two major types of errors. We diagnosed false positives (2%) when the model was falsely diagnosed as BPH and PCA, and false negatives (5%) when the PCA was misclassified as BPH. When examining misclassified cases, we observed that false positives tend to occur in patients with elevated levels of TPSA but without the presence of other discriminant factors such as STK1P. False negatives were observed when PCA presented atypical biomarker profiles. The importance of features indicates that STK1P, FTPSA, and FPSA are the most influential among the forecasts and collaborates with current medical understanding. To address these errors, we plan to refine the model by adjusting decision-making thresholds, investigating alternative functional engineering techniques, and increasing the diversity of the training datasets to better capture variation in PCA and BPH presentations. Future work will focus on external validation of the model using additional datasets to further evaluate and improve prediction accuracy.
This study has several limitations. The small number of samples and centers included in this study may have led to a lack of representativeness, which further expands the survey sample size. Furthermore, small sample sizes did not allow performance of subgroup analyses based on Gleason scores and other variables. To improve PCA risk assessment, subsequent studies will further analyze the performance of different subgroups of models. There is no possibility that this study cannot include these factors due to lack of data on PCA risk factors, such as ethnicity, family history, and lifestyle factors in daily health care settings. We employ questionnaires to improve the collection of these data and investigate whether these indicators can further improve the performance of the model. Our study did not include data on digital rectal examination (DRE), prostate-specific antigen (PSA) density, and magnetic resonance imaging (MRI).twenty two. However, these measures may have influenced the inclusiveness of the results as they were not routinely used for PCA detection at research centres. A further limitation of our study is that all participants were recruited from two centers in China. The distribution of biomarkers such as STK1P and PSA may differ in Western or other Asian populations, which may limit the generalizability of the findings. We encourage future research to conduct external validation of findings in multiethnic international cohorts to assess the generalizability of results across different populations. Furthermore, due to limited sample sizes, we were unable to perform analyses separating Gleason ≥7 vs Gleason 6 or clinically significant PCA (CSPCA) and non-CSPCA. Future studies with larger sample sizes are needed to investigate these clinically meaningful distinctions.
