The effect of missing values using genetic programming on evolvable diagnosis

Abstract

Medical databases usually contain missing values due the policy of reducing stress and harm to the patient. In practice missing values has been a problem mainly due to the necessity to evaluate mathematical equations obtained by genetic programming. The solution to this problem is to use fill in methods to estimate the missing values. This paper analyses three fill in methods: (1) attribute means, (2) conditional means, and (3) random number generation. The methods are evaluated using sensitivity, specificity, and entropy to explain the exchange in knowledge of the results. The results are illustrated based on the breast cancer database. Conditional means produced the best fill in experimental results.