Introduction: Pathogenic variant in the KCNQ2 gene is a common genetic etiology of neonatal convulsion. However, it remains a question in KCNQ2-related disorders that who will develop into atypical developmental outcomes. Methods: We established a prediction model for the neurodevelopmental outcomes of newborns with seizures caused by KCNQ2 gene defects based on the Gradient Boosting Machine (GBM) model with a training set obtained from the Human Gene Mutation Database (HGMD, public training dataset). The features used in the prediction model were, respectively, based on clinical features only and optimized features. The validation set was obtained from the China Neonatal Genomes Project (CNGP, internal validation dataset). Results: With the HGMD training set, the prediction results showed that the area under the receiver-operating characteristic curve (AUC) for predicting atypical developmental outcomes was 0.723 when using clinical features only and was improved to 0.986 when using optimized features, respectively. In feature importance ranking, both variants pathogenicity and protein functional/structural features played an important role in the prediction model. For the CNGP validation set, the AUC was 0.596 when using clinical features only and was improved to 0.736 when using optimized features. Conclusion: In our study, functional/structural features and variant pathogenicity have higher feature importance compared with clinical information. This prediction model for the neurodevelopmental outcomes of newborns with seizures caused by KCNQ2 gene defects is a promising alternative that could prove to be valuable in clinical practice.