Data mining techniques can be effectively used for major disease prediction and diagnosis. But we argue that diagnosis of some major diseases those are correlated with several small diseases and highly interrelated with each other such as TB and HIV and multi disease diagnosis is tough and challenging task. In Data mining, Outlier detection is useful technique for medical diagnosis. When availability of class labels is there, supervised outlier detection successfully used to detect rare and abnormal objects for disease diagnosis purpose. Classification techniques such as Naïve Bayesian classifier, SVM, Association rule mining, neural networks are used for outlier detection. SVM is one of the best classification techniques for outlier detection as there is no requirement of explicit statistical model for SVM and it avoids dimensionality problem and provides optimum solution for classification. Outlier detection using SVM can be possible with One class SVM and Two-class SVM. But still One class SVM suffers with non availability of accurate class labels and Two-class SVM is suffer when one of the class is under sampled and long training time. This paper provides survey of disease diagnosis using SVM and Outlier detection using SVM on different domains and we have identified some key challenges in this field and introduced a solution of two-phase Outlier detection approach combining both Two-class SVM and One class SVM for prediction and diagnosis of multiple diseases those are highly interrelated with each other using available class labels of both data sets and proposed cross training approach for both SVM classifiers for reduce training time of SVM and accuracy purpose.