Simulation optimization of PSA-threshold based prostate cancer screening policies

Abstract

We describe a simulation optimization method to design PSA screening policies based on expected quality adjusted life years (QALYs). Our method integrates a simulation model in a genetic algorithm which uses a probabilistic method for selection of the best policy. We present computational results about the efficiency of our algorithm. The best policy generated by our algorithm is compared to previously recommended screening policies. Using the policies determined by our model, we present evidence that patients should be screened more aggressively but for a shorter length of time than previously published guidelines recommended.

Fig. 1

Receiver Operating Characteristic (ROC) curve illustrating the imperfect nature of PSA testing based on longitudinal data for a regional population in Rochester, MN

Fig. 2

Health-state transitions diagram where NC is no cancer, C is cancer not detected, M is metastasis, and D is death. The treatment state, T, is an aggregate state

Fig. 3

Histogram of PSA observations from Olmsted County, MN data set

Fig. 4

Flowchart describing the major steps of the simulation optimization method

Fig. 5

Performance of the GA using different probabilities of correct selection, P(CS), in the Rinott method

Fig. 6

Improved policy generated by the GA in raw form (with noise)

Fig. 7

Improved policy generated by the GA with noise filtered using the heuristic

Fig. 8

Comparing the Ross policies to the best policy generated by the GA

Fig. 9

Comparing the effect of the Ross policies and the best GA policy on subpopulation of males who develop cancer between the age of 40 and 50