This paper proposes a method, namely MDKS (Kennard-Stone algorithm based on Mahalanobis distance), to divide the data into training and testing subsets for developing artificial neural network (ANN) models. This method is a modified version of the Kennard-Stone (KS) algorithm. With this method, better data splitting, in terms of data representation and enhanced performance of developed ANN models, can be achieved. Compared with standard KS algorithm and another improved KS algorithm (data division based on joint x - y distances (SPXY) method), the proposed method has also shown a better performance. Therefore, the proposed technique can be used as an advantageous alternative to other existing methods of data splitting for developing ANN models. Care should be taken when dealing with large amount of dataset since they may increase the computational load for MDKS due to its variance-covariance matrix calculations.