To solve the problem that the number of hidden nodes in regularized extreme learning machine(RELM) affects classification accuracy, sensitive regularized extreme learning machine(SRELM) algorithm is proposed. Firstly, based on the output of hidden layer activation function and its corresponding output layer weighting factor, the formula of computing the sensitivity for hidden node is deduced by residual between actual value and hidden nodes output. Then different hidden nodes are sorted according to sensitivity. And minor hidden nodes are deleted based on classification accuracy of optimization samples. As a result, SRELM classification accuracy is increased effectively. A case study of MNIST handwritten digit database shows that, compared with common SVM and RELM, time consuming of SRELM is almost the same as RELM, and is obviously lower than SVM. Meanwhile SRELM recognition accuracy for handwritten digit is the highest.