<p>Facial microexpressions (MEs) reveal genuine feelings that people attempt to mask or conceal. MEs can be used in a diverse range of applications, including public safety, psychotherapy, lie detection, security systems, marketing, and business. Existing reviews on microexpression (ME) recognition highlight several enduring research gaps, which include the lack of high-quality annotated datasets, data imbalance, and limited diversity. Moreover, previous reviews emphasize the need for robust algorithms capable of handling subtle or overlapping expressions, generalizing to real-world scenarios, and incorporating multimodal data. Despite growing interest, these reviews lack systematic methodologies, clear inclusion criteria, and structured comparisons of recent deep learning approaches. In response to these gaps, the present review makes several novel contributions. Firstly, beyond simply reviewing methods, this work focuses on identifying the most prominent factors influencing ME recognition (e.g., duration, context, neurobiological factors, and emotional intelligence), recognition techniques, and addressing key contemporary concerns, such as the critical need for identification of standardized datasets and improved recognition of subtle emotions. Further, this review systematically analyses ME recognition models, with a focus on seven key deep learning paradigms: spatial-only, temporal-only, spatio-temporal models, graph-based models, action unit (AU)-based models, attention-based models, and hybrid models. The study uses the PRISMA framework and expands its scope by examining challenges related to interdisciplinary integration, the computational impact of existing methods. By categorizing models based on learning paradigms and discussing their strengths and limitations, offer a more comprehensive and forward-looking perspective on ME recognition research.</p>