In order to improve the prediction accuracy of medical diagnosis, a feature learning and classification stage combination method based on enhanced sparse self encoder and softmax regression is designed. In the feature learning of the sparse self encoder (SAE) network, the sparsity is realized by punishing the weight of the network, and the change is transmitted backward and the cost function is optimized iteratively. In the stage of softmax regression classification, the cross entropy of softmax classifier is optimized by small batch gradient descent method, and the model error is calculated with small batch data, and the model parameters are updated to achieve convergence. The prediction accuracy of the proposed method is 91%, 97% and 98% respectively for heart disease, cervical cancer and chronic kidney disease (CKD) data sets, and shows high feature learning and robust classification performance.