Recent strategies to predict neonatal seizure risk using machine learning (ML) in combination with quantitative electroencephalography (QEEG) have primarily focused on subject-level predictions over several days during the postnatal period. Time-dependent neonatal preictal state classification with high temporal resolution remains unexplored. In this study, we utilized QEEG feature engineering for ML classification of preictal states and compared it to an end-to-end ML approach. We used two publicly available EEG seizure datasets with a total of 132 neonates containing a total of 281 h of EEG data and segmented data into preictal and interictal epochs of 20 s duration. We employed the Boruta algorithm with Shapley values for QEEG feature selection into ML models. The performance of ML models was assessed with cross-validation with area under the receiver operator characteristic curve (AUROC), area under the precision-recall curve (AUPRC), Matthews Correlation Coefficient (MCC), and F1 score. Feature selection demonstrated statistical moments, spectral power, and recurrence quantification analysis features as robust predictors of preictal states. QEEG feature selection combined with convolutional LSTM outperformed other ML models at preictal versus interictal classification, with AUROC 0.678, AUPRC 0.218, MCC 0.255, and F1 0.334. Our results demonstrate the feasibility of applying ML to facilitate prediction and understanding of neonatal preictal states.