Deep neural networks (DNNs) have the powerful ability to automatically extract efficient features, which makes them prominent in electroencephalogram (EEG) based seizure prediction tasks. However, current research in this field cannot take the model uncertainty into account, causing the prediction less credible. To this end, we introduce a novel end-to-end patient-specific seizure prediction framework via model uncertainty learning. Specifically, we propose a reparameterized EEG-based lightweight CNN architecture and a modified Monte Carlo dropout (RepNet-MMCD) strategy to improve the reliability of the DNNs-based model. In RepNet, we obtain multi-scale feature representations by applying depthwise separable convolutions of different kernels. After training, depthwise convolutions with different scales are equivalently converted into a single convolution layer, which can greatly reduce computational budgets without losing model performance. In addition, we propose a modified Monte Carlo (MMCD) strategy, leveraging the samples-based temporal information in EEG signals to simulate the Monte Carlo dropout sampling. Sensitivity, false-positive rate (FPR), and area under curve (AUC) of the proposed RepNet-MMCD achieve 93.1%, 0.033/h, 0.950 and 81.6%, 0.056/h, 0.903 on two public datasets, respectively. We further extend the MMCD strategy to the other baseline methods, which can improve the performance of seizure prediction by a clear margin.