Beyond the Abstract - Machine learning for improved pathological staging of prostate cancer: A performance comparison on a range of classifiers, by Olivier Regnier-Coudert, et al

BERKELEY, CA (UroToday.com) - Data mining techniques used in predictive tools are often chosen arbitrarily, often based on previous work reported in the literature. In this paper, a range of classifiers are compared with respect to their ability to predict prostate cancer staging. Based on datasets created from patient records collected by the British Association of Urological Surgeons (BAUS), and local data gathered at the Aberdeen Royal Infirmary, the study highlights that Bayesian Networks (BN) may improve prostate cancer staging. Partin tables, the most clinically used tool for this task, showed lower predictive quality when applied on UK patients in comparison with US patients. New tables based on the BAUS data were created that improved the prediction. These tables were built following the same methodology, based mainly on the use of Logistic Regression (LR) and using three predictive variables, namely clinical stage, PSA and Gleason sum score. The comparative performance of these UK-specific tables has recently been published.1 However, when additional variables such as patient age and anaesthetic fitness were added, BN-based models showed a significant improvement over other techniques. This allows the addition of hitherto unknown or newly discovered variables into BNs, thereby increasing its flexibility, and this may enhance the efficiency of predictive models created from BNs.

Finally, it is interesting to note that BNs were learnt from the data using a genetic algorithm, a type of heuristic search. This approach is based on identifying statistical relationships that may exist between variables, and thus eliminates the bias that can be introduced when BN models are created based on expert knowledge alone. Hence, this article represents an example of how medical applications can benefit from the use of computational intelligence in order to improve decision- making by patients and clinicians.

Reference:

  1. T. Lam, O. Regnier-Coudert, J. McCall, S. McClinton. Development and validation of a UK-specific prostate cancer staging predictive model: UK prostate cancer tables. British Journal of Medical and Surgical Urology, 2012. In press. doi:10.1016/j.bjmsu.2011.12.005

 


Written by:
Olivier Regnier-Coudert,a John McCall,a Robert Lothian,a Thomas Lam,b Sam McClinton,b and James N’Dowb as part of Beyond the Abstract on UroToday.com. This initiative offers a method of publishing for the professional urology community. Authors are given an opportunity to expand on the circumstances, limitations etc... of their research by referencing the published abstract.

  1. IDEAS Research Institute, Robert Gordon University, St. Andrew Street, Aberdeen AB25 1HG, United Kingdom
  2. Academic Urology Unit, University of Aberdeen, Health Sciences Building, Foresterhill, Aberdeen AB25 2ZD, United Kingdom

 

 

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

More Information about Beyond the Abstract