Online asynchronous courses have become a common way of learning, especially after the COVID-19 epidemic. Despite their convenience and flexibility, these courses often suffer from low engagement and/or poor course quality, ultimately reflected by a low rating. A low rating is often noted at the end of a course, by when it is often too late to intervene or improve the course. The goal of this study is twofold. First, we aim to identify key factors influencing the ultimate ratings for online asynchronous courses. Second, we then build an AI model to predict the risk of ultimate low ratings before a course is fully completed. Inspired by Moore’s model of interaction, which includes two main factors: learner-content and learner-instructor interactions, we derived two features from the dataset: course complexity and course engagement. We used a dataset of 191,849 Udemy courses collected from May 2024 to August 2024. To predict the ultimate low ratings in advance so mid-course interventions are possible, we applied machine learning models, including Logistic Regression, Random Forest, and XGBoost. In our experiments, the XGBoost model achieved the highest performance with 71% accuracy, with course complexity and the number of reviews left by learners being the two most influential factors in predicting a low course rating. Constantly monitoring course complexity and the number of course reviews can help with early detection. The findings can provide valuable, actionable insights and guidance for online education platforms, especially for instructors to adjust their courses before the courses are complete.