Presentation Authors: Steve Zhou*, Alan Priester, Rajiv Jayadevan, Jason Yang, David Johnson, Alex Raman, Karthik Sarma, Corey Arnold, Jorge Ballon, Shyam Natarajan, Leonard Marks, Los Angeles, CA
Introduction: Focal therapies such as hemi-gland ablation (HGA) are evolving treatment options for prostate cancer. However, limitations in current diagnostics--biopsy Gleason score, biopsy-determined laterality, MRI, and PSA--may result in improper candidate selection and lead to suboptimal oncologic control. In machine learning, linear classifiers are simple but powerful tools that use logistic regression models to calculate probability, sorting data into one of multiple possible classes. We trained a linear classifier to better predict patient HGA eligibility using spatial information from fusion biopsy.
Methods: We identified patients who received fusion biopsy and radical prostatectomy (RP) at a single institution between May 2011 and March 2018. In addition to standard clinical features (Fig 1A: III-V), we extracted fusion biopsy features related to core and target locations (Fig 1A: I, II). These features were used to train a linear classifier to classify patients by HGA eligibility (unilateral intermediate-risk cancer on RP) in a randomized set of patients. We assessed the resulting algorithm's predictive performance in an independent set of test data, verified with 5-fold cross-validation. The evaluation metric was the area under the curve (AUC) statistic. Accuracy was compared to current eligibility criteria (biopsy-proven unilateral intermediate-risk cancer).
Results: Cross-validation was performed with 144 patients partitioned into 5 data blocks, each block acting once as a test set for an algorithm instance trained by the other 4. The 5 folds yielded reproducible receiver operator characteristic curves with a mean AUC of 0.83Â±0.07 on test sets (Fig 1B). The 5 most predictive features are summarized in Fig 1A. Model accuracy was 76% compared to 67% with current criteria (p=0.03); sensitivity was 68% compared to 44% (p=0.006). Matching sensitivity to that of current criteria, the model's positive predictive value was 88% compared to 73% (Fig 1B, p=0.08). In other words, in this dataset, 27% of patients selected by current criteria harbored undetected advanced disease for which HGA is insufficient: this incidence is halved when candidates are vetted by our model.
Conclusions: A machine learning classifier, trained with fusion biopsy data, outperforms current eligibility criteria in accurately predicting HGA eligibility.
Source of Funding: Supported in part by Award R01CA218547 from the National Cancer Institute. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Cancer Institute or the National Institutes of Health.