The functional activation of the brain gets affected in conditions such as brain-tumor, localization-related epilepsy (LRE) and lesions. Typical brain activation is such that the left brain is dominant as compared to the right brain. In order to distinguish between the two groups -typical and atypical - the patients undergo functional Magnetic Resonance Imaging (fMRI) test. Based on the processed fMRI maps, nonlinear decision functions (NDF) are used to determine the laterality. Here an alternate algorithm called the 'Iterative Random Training-Testing Algorithm', a modification of the well known gradient descent algorithm, which is used as a means for enhancing the results of the classification, is presented. The algorithm aims at improving the sensitivity of results obtained in earlier studies reported in literature. Improving the sensitivity is of prime importance since sensitivity suggests the proportion of false negatives in the classification results. False negatives are critical in clinical decision making. The algorithm divides the training data set randomly into a pure-training set and cross-validation training set. The decision function is trained with the elements assigned to the pure training set and then tested with the element of the cross validation training set. The whole process is repeated a number of times with the aim that the random division of the data set would take into consideration various formations of the data yielding better results. The results of the algorithm showed an improvement in the sensitivity of 2 to 5% with no significant changes in the accuracy, specificity or precision.