This study investigates the automatic classification of developmental disorders in children based on pose estimation and Long Short-Term Memory (LSTM) model. The dataset consists of 223 videos of children diagnosed with developmental disorders according to DSM-5 criteria and 50 videos of typically developing children. All videos depict eating behavior, from which the hand joint coordinates were extracted as time-series features. To address class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied. Using the Keras library, the LSTM model was trained with various window sizes (10 to 400). The best classification accuracy (79%) was achieved with a window size of 40, suggesting the model effectively captured short-term temporal patterns. Accuracy declined with window sizes over 100, indicating the model’s limitation in handling long-term dependencies. Misclassification mostly occurred in typically developing children, likely due to age-related variability in movement. These findings demonstrate the potential of LSTM-based modeling as a quantitative tool for developmental state assessment based on behavioral patterns.