Tang et al. propose a classification scheme based on EEG and facial analysis of video features that detects episodes of mind wandering, which is a shift in attention away from the primary task. The dataset for this model was obtained from participants watching a 90 minute introductory video broken into 18 minute chunks with self-paced breaks in between. Thought probes, which asked participants to self-report whether they had been paying attention to the video and their confidence in that answer, occurred at intervals ranging from 20 to 120 seconds, along with some that were explicitly triggered by the experimenter. The EEG feature extraction used multiscale permutation entropy and frequency band powers. Multiscale permutation entropy(MPE) compares neighboring values in a time series and additionally constructs a coarse-grained time series by averaging non-overlapping windows. The frequency band powers used in this model were delta, theta, alpha, and beta; alpha frequencies are especially important, as they have been shown to increase during episodes of mind wandering. Within the video analysis, features from facial landmark locations (FLL), facial action units (AU), point distribution model (PDM), gaze angle (GA), and eye-region landmarks (ERL) were extracted. A random forest classifier was trained on these features using either within-participant tenfold cross-validation, across-participant LOPO cross-validation, and across-participant augmented LOPO cross-validation (a-LOPO). Across all validation schemes, MPE, PDM, and band power were selected the most frequently. In tenfold cross-validation, most of the features were video, which suggest that video-derived features may better analyze individualized characteristics. The detection results for all models showed that video features alone performed better than EEG features alone, but performance overall was better with both video and EEG features.