This study investigates the feasibility of predicting future human states to enhance ergonomic interventions using experimental data. Two datasets were analyzed: Dataset 1 involved VR-induced motion sickness with discomfort reported every 10 s, while Dataset 2 involved MR-based 3D puzzle assembly with discomfort reported every minute. Physiological indicators such as heart rate and skin conductance were monitored. A convolutional neural network model was employed to predict discomfort. For Dataset 1, accuracy declined from 83.8% at 0 s to 57.8% at 50 s, stabilizing around 56.2%. The F1-scores for predicting “Nothing” and “Discomfort” showed similar declines. For Dataset 2, accuracy began at 89.9% at 0 and 1 min, dropping to around 77.7% from 6 to 9 min, with F1-scores following similar patterns. The study confirms that predictive modeling is feasible, achieving higher accuracy than chance level. However, Dataset 1’s predictions in seconds were less accurate over time due to fewer features, indicating a need for more indicators. In contrast, Dataset 2 suggested that physiological reactions might precede conscious symptoms, emphasizing the importance of incorporating temporal characteristics into predictive models.