Human Activity Recognition (HAR) systems play a crucial role in enhancing eHealth applications by providing real-time monitoring and analysis of patient’s physical activities and behaviors. In this paper, a novel lightweight activity recognition model leveraging Tiny Machine Learning (TinyML) tailored for embedded devices is proposed. Our proposed model is designed to execute complex activity recognition tasks enabling the monitoring of patient activity and mobility, which can be critical for fall prevention, rehabilitation assessments, and ensuring adherence to exercise regimens. Specifically, data is collected and trained for five different activity labels, including the “idling”, “walking”, “running”, “stepping upstairs” and “stepping downstairs”. The AI model demonstrates impressive accuracy in distinguishing various human activities, showcasing its potential in applications requiring precise activity recognition. It achieves a flawless 100% accuracy rate in identifying idle states, ensuring that periods of inactivity are consistently recognized without error. Meanwhile, for walking recognition, the model attains an accuracy of 99.5%, indicating a high level of reliability in monitoring the common daily activity. For running, the model maintains a commendable accuracy of 99.5%, effectively capturing the dynamics of more vigorous movement. In more complex activity scenarios, such as stepping up and stepping downstairs, the model performs with accuracies of 99.1% and 93.9%, respectively. Moreover, our lightweight model is deployed and validated on low powered and low-resource microcontroller that allows for scalability and accessibility, as the technology can be integrated into a wide range of embedded devices, making advanced features available to various for healthcare monitoring, fitness tracking, and rehabilitation gaming. Accurately interpreting activity data, our HAR model can be extended to monitor the activity levels of elderly patients, aiding in fall detection and prevention.