This article describes solutions for an electroencephalographic (EEG) signal-based seizure detection problem, where the class imbalance in the dataset is a severe problem. The 'IEEE ICASSP 2023 Signal Processing Grand Challenge (SPGC): Seizure Detection' posed the diagnostic problem using wearable behind-the-ear (bhe) EEG electrodes. The challenge required using machine learning models to predict seizure events from bhe-EEG signals. We present a data-minimal training strategy to improve the class imbalance in the data, and proposes a projection layer that can map a low dimensional input to a high dimensional space. A fully connected recurrent neural network (RNN) based on Long Short-Term Memory cells (LSTMs) is proposed as the projection layer that finds this optimal projection. We additionally discuss interpretations of the latent representations learned by the classifier RNN network to evaluate if the domain adaptation has facilitated the generalizability of the network. These results suggest that projection-RNN can facilitate robust mappings of the target domain input signal on to the source domain. To illustrate the effects of data engineering and class imbalance, an existing deep–learning model (Chrononet) was optimized using the ’data-minimal’ strategy. Data engineering improved Chrononet's performance modestly (100% sensitivity and 3 false alarms per hour) over baseline models. On the other hand, the domain adaptation scheme resulted in a sensitivity of 89.37% with 94 false alarms per hour in the source domain. Additionally, the network achieved a sensitivity of 42% and a false alarm rate of 19 per hour in the target domain. These solutions were among the best-performing methods presented in the competition.