The aim of this paper is to construct a data model applicable to youth sports training in the IoT environment and develop efficient pattern recognition algorithms to achieve accurate analysis and assessment of youth sports training status. The features of youth sports training data collected by IoT technology are extracted through a combination of deep learning and feature decomposition. The feature vectors obtained from feature extraction are inputted into the Long Short-Term Memory (LSTM) network to generate the data model of youth sports training in this paper and predict the state of youth sports training. The prediction results are input as features into the Support Vector Machine (SVM) algorithm, and these features are extracted using the Empirical Modal Decomposition (EEMD) method, and at the same time, the hierarchical idea is utilized to realize the recognition of youth sports training patterns. The results of the study showed that the errors of the results of predicting youth sports training states using the LSTM model were mostly within 0 ± 0.5. The prediction accuracies of the model on the test set for the three athletic training state metrics were 96.80%, 99.40%, and 98.80%, respectively. Meanwhile, the performance of the SVM model for youth athletic training state pattern recognition using the SVM model was significantly superior, with 100% accuracy on the test set for four models, including pattern 2.