Collecting data in many clinical settings is faced with several challenges and difficulties that results in large proportions of missing values in the data. To address this issue, we study the effect of mining association rules from the training data then applying these rules (of high confidence) to replace some of the missing values in the training and testing data. The rationale is to add knowledge from the domain, which in practice can be generated by clinical guidelines. In this work, we present experimental results of the effect of adding noise to synthetic data, masking attributes that are known to be related to the classification task, generating missing proportions of values, and replacing some of those missing values using association rules discovered from the training data.