Objective: Prediction of comatose patients’ neu-rological outcome after cardiac arrest (CA) is essentialto prevent unnecessary continuation of treatment (withexpensive and scarce resources) and prevent emotionalburdens on families. Electroencephalography (EEG) canaid the prediction of outcome after CA. Visual EEGanalysis is subjective, time-consuming, and might missimportant information. Existing EEG-based machinelearning models are more reliable but lack interpretationof temporal characteristics of EEG. Long-short-term-memory networks (LSTMs) do take advantage of suchtemporal characteristics. Therefore, I hypothesized thatan LSTM would outperform time-insensitive models.Methods: Patients’ neurological outcome at six monthsafter CA was classified as good (no/mild neurologicaldamage) or poor (severe neurological damage/vegetativestate/death). Twelve quantitative EEG features wereextracted from five-minute EEG epochs recorded 12(n=78) or 24 (n=176) hours after CA. A time-insensitivebaseline logistic regression model (LR) and an LSTMwere developed and trained. The performance was eval-uated using the area under the receiver operator curve(AUC) and the sensitivity at 100% specificity (SeSp100)for poor outcome prediction. The LSTM was comparedto the current LR and previously published models.Results and conclusion: The LSTM predicted pooroutcome with AUC=0.90 and SeSp100=0.66, meaningit did not significantly improve performance over LR(AUC=0.89, SeSp100=0.67). However, the LSTM andLR outperformed almost all previously reported mod-els, likely due to the features’ high prognostic power.The SeSp100 was even higher for the LSTM (0.79) andLR (0.78) when using only epochs at 12 hours after CA,suggesting that earlier EEG might further improve prog-nostication.