In this work we explore the problem of multiclass classification where the classifier may abstain from classifying on some observation. We derivate a new surrogate loss function and a multiclass decision rule by using a reject threshold on posterior probabilities in the Bayes decision rule, known as Chow's rule. The goal of the decision rule is to minimize the value of given misprediction and rejection cost functions specified by the user. We suggest a general training algorithm by plug-in the surrogate loss in to Support Vector Machine (SVM) structure. We then test the algorithm on various real-life problem in the photonic medical sensing field where accuracy is critical. We present an example of a non-invasive way of detecting glucose level in blood to help patients with Diabetes mellitus diseases while the sensing is performed with speckle-based approach to analyze remote sensing of biomedical parameters. The results will show that the value of the reject threshold has importance in determining how many samples to reject and in the overall accuracy of prediction. As the threshold grow so does the number of samples rejected and overall accuracy, meaning that only samples with strong confidence are outputted in the classification process. A very important point in working with reject option is that there is a tradeoff between the number of samples being rejected and the accuracy of the labeled samples. High precision comes with high rejection rate, while low rate of rejection derogates from the general correctness of the output.