Epilepsy is a chronic neurological disorder affecting over 50 million people worldwide, characterized by recurrent seizures that require timely and accurate detection for effective management. Traditional seizure monitoring systems often rely on high-power computing and cloud-based processing, which introduce latency, privacy concerns, and power consumption challenges—particularly in wearable or ambulatory settings. Tiny Machine Learning (TinyML), which enables the deployment of lightweight, low-power models on microcontrollers and edge devices, has emerged as a promising paradigm for real-time, on-device seizure detection. This paper presents a comprehensive review of recent state-of-the-art applications that employ TinyML for epileptic seizure prediction. We analyze a range of approaches based on EEG and other biosignals, covering model architectures, datasets, optimization techniques (e.g., quantization, pruning), deployment platforms, and performance metrics. Our analysis highlights key trends such as the dominance of neural networks optimized via TensorFlow Lite or Edge Impulse, and the use of resource-constrained platforms like Raspberry Pi, Arduino Nano 33 BLE, and STM32 MCUs. Despite promising results in terms of accuracy and efficiency, we identify critical challenges, including limited dataset diversity, lack of clinical validation, and the absence of standardized evaluation benchmarks. Finally, we outline future research opportunities such as personalized TinyML, federated learning, multimodal sensor fusion, explainability, and real-world validation.