In recent years, non-invasive detection based on near-infrared spectroscopy and artificial intelligence algorithms has received much attention in medicine and biology due to its safety, non-invasiveness, and high efficiency. One key issue is selecting effective input features for intelligent regression models from wide-band near-infrared spectroscopy. This paper establishes a non-invasive near-infrared blood glucose concentration intelligent prediction model by combining near-infrared spectroscopy, genetic algorithm, and support vector regression (GA-SVR) using blood glucose concentration detection as an example. Firstly, according to the OGTT experimental rules, non-invasive dynamic blood near-infrared spectroscopy and corresponding blood glucose concentrations of volunteers were collected. The optimal near-infrared feature wavelength combination was further determined based on a genetic algorithm. Finally, the support vector machine regression model was established to achieve blood glucose concentration prediction. In this paper, comparative experiments were designed to compare the proposed method with the genetic algorithm and multi-layer perceptron regression (GA-MLPR), partial least squares regression (GA-PLSR), and random forest regression (GA-RFR). The experimental results show that the proposed GA-SVR model has the best prediction performance, and the correlation coefficient of the test set is increased by 44% compared with GA-PLSR, the correlation coefficient reaches 99.97%, and the mean square error is 0.000 97. The study shows that the proposed GA-SVR can achieve effective feature selection of near-infrared spectroscopy data, verifying the feasibility of intelligent algorithms for feature selection. The excellent performance of this feature selection model provides a new approach to non-invasive detection.