Real-time seizure prediction using EEG holds significant clinical promise for enhancing patient safety, yet deployment remains constrained by the computational demands of deep learning models in resource-limited healthcare and edge IoT environments. This paper presents a hardware-independent compression framework combining structured and unstructured pruning with post-training quantization for efficient EEG-based epilepsy prediction. Applied to a CNN-BiLSTM-Attention architecture (2.19\,M parameters) evaluated on the CHB-MIT dataset, our approach achieves 93.3\% parameter reduction ($14.8\times$ compression) while preserving clinical-grade performance: 96.7\% segment-level accuracy and 94.7\% event sensitivity with minimal false alarm rates using SPH/SOP protocol. Cross-subject generalization improves substantially through lightweight fine-tuning using only a single seizure per patient. System-level benchmarking demonstrates consistent acceleration across deployment scenarios: up to $2.7\times$ speedup on cloud servers with sub-millisecond latency via ONNX Runtime, and $11\times$ speedup on edge hardware (Raspberry Pi~4) achieving sub-5\,ms inference with 450+ segments/s throughput using INT8 quantization. This work establishes that compression-aware pruning effectively reconciles high predictive accuracy with stringent deployment constraints, enabling scalable cloud-edge seizure prediction systems for continuous patient monitoring.