<![CDATA[Dynamic Latent State Models (DLSMs) have become increasingly central to behavioral and physiological prediction due to their ability to represent hidden psychological states and temporal dynamics that static machine-learning models cannot capture. This research conducts a systematic mapping study to analyze the evolution, methodological trends, application domains, and dataset usage of DLSMs published over the last decade. Using a structured search strategy across major scientific databases, studies were screened following PRISMA guidelines, and relevant information was extracted to construct a comprehensive taxonomy of model types, signal modalities, and prediction tasks. The results reveal a significant rise in the adoption of DLSMs, particularly after 2018, driven by advances in deep generative models such as deep Kalman filters and variational state-space models. EEG, HRV, and EDA emerge as the most dominant physiological signals, while stress, emotion, and fatigue prediction constitute the primary application areas. Benchmark datasets including DEAP, WESAD, and DREAMER are frequently used but remain limited in ecological diversity, indicating a continuing need for more realistic, multimodal datasets. Comparison with earlier research shows a shift from interpretable probabilistic models toward more expressive but less transparent deep latent models. This study contributes a consolidated overview of theoretical foundations, research patterns, and methodological gaps in the field. The findings highlight key challenges related to interpretability, dataset diversity, and evaluation consistency, while identifying opportunities for hybrid modeling approaches and more comprehensive data resources. Overall, this mapping study provides a structured foundation to guide future work in advancing dynamic latent-state modeling for behavioral and physiological prediction.]]>
