The evolution of Digital Twin technology led to a new way for innovative applications, including the replication of human beings through Human Digital Twin. Human Digital Twin enables real-time monitoring, simulation, and analysis of individual states and activities, with significant potential in areas such as healthcare, fitness, and manufacturing. However, their development presents unique challenges, particularly in accurately modeling human behavior and ensuring the adaptability of machine learning models. This work addresses these challenges by integrating Human in the Loop methodologies into the Human Digital Twin system. Human in the Loop leverages human to enhance the performance of ML models. Specifically, we explore the application of Active Learning techniques for Human Activity Recognition, a critical component of Human Digital Twin. We propose a Human Digital Twin architecture that incorporates Active Learning to optimize the Human Activity Recognition model. Through this approach, we demonstrate how the integration of Human in the Loop and Active Learning can refine the virtual representation of individuals, advancing the capabilities and real-world applicability of the Human Digital Twin system. Moreover, we show that it is key to engage users in reporting their activities, as well as having appropriate datasets to train models.