Machine learning-based optimization of the prostate health index for prostate cancer detection

The Prostate Health Index (PHI) is composed of the prostate-specific antigen (PSA), free PSA (fPSA), and the [-2]pro-PSA isoform (p2PSA). The aim of the study was to modify the PHI using machine learning (ML) and to compare its prognostic performance with traditional PHI for any-, low- and high-grade prostate cancer (PCa) detection on biopsy. To obtain better-balanced data set we used the over-sampling strategy. The extreme gradient boosting was considered as an ML algorithm. Predictive performance was quantified by the area under the curve (AUC), multiclass classification measures, calibration, and decision curve analysis. We used the feature importance and the SHapley Additive exPlanations value for interpretation of the model. There were 200 patients in the initial pool, but after resampling, the dataset had equal number of data points for each cancer grade (114), i.e. 342 data points in total. The AUC of the modified PHI was significantly higher than those of the traditional PHI (0.898 vs 0.808, 0.868 vs 0.581, and 0.983 vs 0.74) for any, low-grade, and high-grade PCa. The modified PHI had higher net benefit compared to the traditional PHI. However, the ML PHI model had sigmoid-shaped reliability plots. The p2PSA had the highest influence, and the fPSA was the most important factor for risk estimate by the ML PHI. The modified ML PHI model for predicting histological grade of PCa showed superior performance in comparison to the traditional PHI. However, additional studies are required before it can be integrated into routine clinical care.