To achieve sensing non-invasive blood glucose concentration, this paper proposes a non-invasive blood glucose spectrum algorithm based on support vector machine (SVM) model. The algorithm firstly uses a PPG device to sample red and infrared signal of volunteers to obtain the PPG signal, and then extracts the blood glucose concentration by a minimally-invasive glucometer. Compared with traditional near-infrared spectrum and regression method, the proposed algorithm could effectively improve the prediction accuracy by adopting the SVM model to find regression relationship between signal and glycemic. And the SVM model uses the Gauss-function as the kernel function. The algorithm is independently of the individual and environmental factors. The experimental results demonstrate that compared to partial least squares regression, the proposed algorithm could improve the predictive accuracy by 10%–15%. And the algorithm's forecast accuracy could reach to 98%.