[Background] Infantile epileptic spasms syndrome (IESS) is a developmental epileptic encephalopathy in infants, which is often difficult to be predicted long-term seizure outcomes at the time of onsets. The aim of this study is to predict its long-term outcome by analyzing EEG data at the onset of IESS of unknown etiology. [Methods] The study included eighteen patients with IESS of unknown etiology. Thirteen patients in whom seizures disappeared after initial treatments were categorized into a good outcome group, and five patients with continuation or relapse of seizures were into a poor outcome group. We trained a machine learning (ML) model from clinical EEG data of patients in the good outcome group only utilizing an anomaly detection framework. The delta and the beta bands, constructing basis of hypsarrhythmia were extracted from scalp EEG data during sleep by bandpass filtering, and the phase of each band was used as features of the ML model. Long-short term memory autoencoder (LSTM-AE), which copes with anomaly detection with time series data, was adopted as the ML model. We tested its performance of long-term outcome prediction. This trial was repeated ten times randomly exchanging training, validation, and test datasets for precise performance evaluation. [Results] The trained LSTM-AE model achieved a sensitivity, specificity, and accuracy of 0.82 ± 0.08, 0.80 ± 0.10, and 0.81 ± 0.10, respectively, when patients the poor outcome group were detected, which may contribute to clinical decision. [Conclusion] The developed ML model enabled highly accurate prediction of seizure outcomes of IESS of unknown etiology from the EEG data at its onset.