Machine learning models for predicting prostate cancer and clinically significant prostate cancer at biopsy: An updated analysis of an expanded Japanese cohort
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Our institution reported in 2019 on a multilayer artificial neural network (ANN) that predicted prostate cancer at biopsy in 334 patients. That model was trained with TensorFlow 1.x and scored at three fixed step counts, and hyperparameter selection was not kept separate from test evaluation. Here we revisit an expanded cohort from the same hospital using current machine-learning practice.
Methods
All biopsy episodes available in the institutional database were pooled (n = 526; 524 after one non-binary outcome code and one record with missing digital rectal examination [DRE] data were dropped). The seven predictors of the original report were retained: age, prior biopsy history, PSA, prostate volume, DRE, and MRI diffusion-weighted imaging findings in the peripheral and transition zones. Twenty-seven patients contributed more than one biopsy episode, so folds were built with patient-ID-grouped, stratified k-fold cross-validation (StratifiedGroupKFold; scikit-learn 1.8.0) at 3 and 5 folds, repeated over 10 random partitions, so that no patient had episodes in a training fold and a test fold at the same time. Four classifiers were compared: L2-regularized logistic regression, gradient boosting, random forest, and a shallow (single hidden layer) multilayer perceptron. Two outcomes were modeled: detection of any prostate cancer, and detection of clinically significant prostate cancer (Gleason score ≥ 7).
Results
Any-cancer prevalence was 55.7% (292/524); Gleason score ≥ 7 prevalence was 39.7% (208/524). Under repeated 5-fold cross-validation, gradient boosting discriminated best, both for any prostate cancer (mean AUC 0.826, 95% CI 0.823–0.830) and for Gleason score ≥ 7 (mean AUC 0.855, 95% CI 0.852–0.859). Random forest and logistic regression followed closely (AUC 0.81–0.85). The shallow multilayer perceptron did worse, and less consistently so (any-cancer AUC 0.671; Gleason score ≥ 7 AUC 0.742), falling below the deeper five-hidden-layer ANN of the 2019 report as well. Results with 3-fold cross-validation were essentially unchanged.
Conclusions
In the expanded cohort, regularized logistic regression, gradient boosting and random forest each discriminated prostate cancer at biopsy at least as well as the multilayer ANN reported earlier, with much simpler models and with hyperparameter tuning kept apart from performance estimation. At this sample size the shallow network gained nothing over the simpler alternatives. This is a preprint; the study has not undergone external peer review.
