Online learning environments cannot often evaluate students’ attentiveness and engagement in minute detail. Our project addresses this gap by developing a real-time predictive engine to better understand students’ learning states, enabling more active learning and adaptive teaching in online settings. This engine leverages a combination of digital tools to predict cognitive load and engagement in real-time. In this work, we outline the procedures and some preliminary findings from developing and testing this predictive model using data from wearables and webcams. These devices captured physiological signals, facial expressions, eye gazes, and head poses to measure engagement levels in content-heavy classrooms. To build the model, we manually selected and extracted features from a pre-existing database containing sensory data. The model categorized students into two engagement levels: low and high. Following model development, we conducted an experimental study involving two 30-minute machine learning (ML) lectures with 12 college students. Alongside physiological data, self-reported survey data on engagement were collected as the ground truth. The engagement levels predicted by the engine were found to be highly associated with their self-reports on engagement (approximately 50% of variance shared). Results demonstrated that electrodermal activity (EDA), recorded via E4 wearable devices, was higher during challenging lecture segments than in easier ones. Additionally, variability in EDA levels during lectures correlated with self-reported stress, suggesting a potential link between stress and engagement. Through this work, we present insights into developing multimodal learning analytics based on real-time psychophysiological signals and learning behavioral data. These insights highlight the potential for integrating wearable and webcam data to enhance online learning.