The rapid evolution of deepfake generation techniques based on Generative Adversarial Networks (GANs), autoencoders, and diffusion models has produced highly realistic facial forgeries that increasingly evade visual artifact detectors. Recent works have therefore turned toward physiological signals in particular remote Photoplethysmography (rPPG)—as a robust, hard-to-forge biometric cue. This paper presents a systematic review of rPPG-based deepfake detection methods, synthesizing key works that exploit heartbeat-related signals, spectral fingerprints, multi-region coherence, and multi-modal fusion. The literature is organized along three axes: signal domain (time, frequency, and hybrid), model architecture (traditional Machine Learning (ML), Convolutional Neural Networks (CNNs), Temporal Convolutional Networks (TCNs), transformers, and hybrid pipelines), and target task (binary detection and source model). Beyond summarizing performance, a micro-level analysis is performed of how different methods construct rPPG features Photoplethysmography (PPG) cells, Multi-Scale Spatio-Temporal Maps (MST-maps), Matrix Visualization Heatmaps (MVHM), Beats Per Minute (BPM) statistics, and similarity metrics), motion artifacts, compression distortions, and illumination variations along with cross-dataset generalization, which is analyzed. Some advanced deepfakes can contain realistic heartbeat rhythms, shifting the field from detecting the mere presence of a pulse to analyzing spatial distribution, temporal stability, and cross-region synchronization of physiological signals. Dataset bias (e.g., skin-tone variability), robustness to compression and adversarial rPPG manipulation are examined. Further research directions including multitask learning, multimodal fusion, compression-resilient pipelines, and fairness-aware dataset design, are discussed.