With the rapid emergence of K-12 online learning platforms, a new era of education has been opened up. By offering more affordable and personalized courses compared to in-person classrooms, K-12 online tutoring is pushing the boundaries of education to the general public. It is crucial to have a dropout warning framework to preemptively identify K-12 students who are at risk of dropping out of the online courses. Prior researchers have focused on predicting dropout in Massive Open Online Courses (MOOCs), which often deliver higher education, i.e., graduate level courses at top institutions. However, few studies have focused on developing a machine learning approach for students in K-12 online courses. The dropout prediction scenarios are significantly different between MOOC based learning and K-12 online tutoring in many aspects such as environmental modalities, learning goals, online behaviors, etc. In this paper, we develop a machine learning framework to conduct accurate at-risk student identification specialized in K-12 multimodal online environments. Our approach considers both online and offline factors around K-12 students and aims at solving the challenging problems of (1) multiple modalities, i.e., K-12 online environments involve interactions from different modalities such as video, voice, etc; (2) length variability, i.e., students with different lengths of learning history; (3) time sensitivity, i.e., the dropout likelihood is changing with time; and (4) data imbalance, i.e., only less than 20\% of K-12 students will choose to drop out the class. We conduct a wide range of offline and online experiments to demonstrate the effectiveness of our approach. In our offline experiments, we show that our method improves the dropout prediction performance when compared to state-of-the-art baselines on a real-world educational data set.