Machine learning for improved pathological staging of prostate cancer: A performance comparison on a range of classifiers

Abstract

Objectives

Prediction of prostate cancer pathological stage is an essential step in a patient's pathway. It determines the treatment that will be applied further. In current practice, urologists use the pathological stage predictions provided in Partin tables to support their decisions. However, Partin tables are based on logistic regression (LR) and built from US data. Our objective is to investigate a range of both predictive methods and of predictive variables for pathological stage prediction and assess them with respect to their predictive quality based on UK data.

Methods and material

The latest version of Partin tables was applied to a large scale British dataset in order to measure their performances by mean of concordance index (c-index). The data was collected by the British Association of Urological Surgeons (BAUS) and gathered records from over 1700 patients treated with prostatectomy in 57 centers across UK. The original methodology was replicated using the BAUS dataset and evaluated using concordance index. In addition, a selection of classifiers, including, among others, LR, artificial neural networks and Bayesian networks (BNs) was applied to the same data and compared with each other using the area under the ROC curve (AUC). Subsets of the data were created in order to observe how classifiers perform with the inclusion of extra variables. Finally a local dataset prepared by the Aberdeen Royal Infirmary was used to study the effect on predictive performance of using different variables.

Results

Partin tables have low predictive quality (c-index = 0.602) when applied on UK data for comparison on patients with organ confined and extra prostatic extension conditions, patients at the two most frequently observed pathological stages. The use of replicate lookup tables built from British data shows an improvement in the classification, but the overall predictive quality remains low (c-index = 0.610).

Comparing a range of classifiers shows that BNs generally outperform other methods. Using the four variables from Partin tables, naive Bayes is the best classifier for the prediction of each class label (AUC = 0.662 for OC). When two additional variables are added, the results of LR (0.675), artificial neural networks (0.656) and BN methods (0.679) are overall improved. BNs show higher AUCs than the other methods when the number of variables raises

Conclusion

The predictive quality of Partin tables can be described as low to moderate on UK data. This means that following the predictions generated by Partin tables, many patients would received an inappropriate treatment, generally associated with a deterioration of their quality of life. In addition to demographic differences between UK and the original US population, the methodology and in particular LR present limitations. BN represents a promising alternative to LR from which prostate cancer staging can benefit. Heuristic search for structure learning and the inclusion of more variables are elements that further improve BN models quality.

Introduction

Cancer is a widely spread disease responsible for many deaths all over the world. In 2008, the World Health Organization estimated the number of new cancer cases in the world to be over 7.5 million [1]. Among all types of cancer, prostate cancer is the most frequent in men. In 2008, around 900 000 new cases of prostate cancer were diagnosed, and approximately 260 000 men died from it over the same period [1]. In Britain, the same source shows that 37 000 men were affected with new occurrence of prostate cancer, accounting for nearly a quarter of all male cancer diagnosed annually. It is also the second commonest cause of cancer death in men in the UK after lung cancer [1].

This paper considers the use of different machine learning techniques in order to improve the prediction of pathological stage in prostate cancer. A UK-wide dataset collected by the British Association of Urological Surgeons (BAUS) is used in order to build and assess predictive models. Results are compared with each other, but also against those of tools and methods currently in use clinically. Machine learning gathers a wide range of methods, particularly for classification purposes. Performance studies of such methods on different applications is an essential step towards the optimization of predictive tools. In [2], classifiers are compared with each other with respect to their performance on predicting different outcomes related to pancreatic cancer, including cancer staging. Similarly, in [3], classifiers were applied on breast cancer patient data to improve survivability prediction. In [2], models built using Bayesian techniques and logistic regression presented the best prediction for the different outcomes while decision trees performed best in [3]. This highlights the importance of comparing methods on different domain as the most adapted technique can vary across them.

Partin tables [4] are the most commonly used tool for prostate cancer staging. They were originally created using logistic regression (LR) [5] on a database gathering records of patients that were treated with radical prostatectomy in a single US institution [4]. Since then, the tables have been updated using different up-to-date datasets [6], [7], [8]. The revision takes into account changes in population demographics, advances in health technology and improved health care systems, but the tables are still based on the same fundamental LR-based methodology.

Partin tables are a well-established and most widely used pathological staging tool in the urological community worldwide. However, concerns have been raised regarding their validity on non-US populations [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19]. In some instances, Partin tables were considered to be unsuitable for the target population because of limitations with respect to their predictive quality [9], [11], [16], [17], [18], [19]. The appropriateness of the methodology behind Partin tables, especially in regard to the choice of predictive variables and classifier, was not addressed in those studies. In addition, it is widely recognized that prostate cancer staging is associated with a high level of uncertainty. All these considerations are compelling clinicians to explore alternative means of generating predictive tools, especially those which apply machine-learning techniques which have the potential of improving the quality and accuracy of predictive performance [20].

In this paper, we assess the Partin tables on a British population. Replicate lookup tables, based on British data are built and analyzed following the original approach given in [8]. Additionally, we propose and compare alternative classification techniques, including Bayesian networks (BNs), which by simplifying the probability distribution, are recognized as a reference method to reason under uncertainty [21]. In addition, we consider more variables to include in the model. The paper is divided as follows. Section 2 presents important details related to the understanding of prostate cancer and its treatments. Section 3 describes the methodology of the study. It presents the different objectives and the data, provides technical background knowledge on classification and describes the experiments. Results are presented in Section 4 and discussed in Section 5. Finally, we describe our conclusions and introduce ideas for further work in Section 6.