This study bridges subtle facial cues and macro-engagement to improve lecture difficulty and unexpected event prediction using two-class lecture attitude metrics— \(r_{\text {class}}\) and \(\rho _{\text {class}}\) . Previous work has integrated individual facial emotion recognition (FER) outputs to estimate classroom engagement, yet the relationship between subtle cues and macro-level engagement remains unclear. We extend prior methods by incorporating soft output values to capture subtle facial cues more effectively. Evaluation experiments in simulated lectures highlight that \(\rho _{\text {class}}\) , derived from soft output probabilities, outperforms the previous metric, \(r_{\text {class}}\) , in detecting unexpected events (0.317 vs. 0.000). These results demonstrate the advantage of soft outputs in capturing subtle facial cues and their potential for real-time classroom assessment and intervention.