Agitation-related movements in intensive care units (ICUs), such as unintended tube removal and bed exit attempts, pose significant risks to patient safety. The wearable inertial measurement units (IMUs) offer a potential means of capturing such movements. However, the technical feasibility of discriminating ICU-relevant agitation motion patterns from multi-site IMU data remains insufficiently established. To evaluate the technical feasibility of using a convolutional neural network (CNN) applied to multi-site IMU signals to discriminate predefined ICU-relevant agitation-related motion patterns under controlled experimental conditions. Fifteen healthy volunteers performed six scripted movements designed to emulate ICU-relevant agitation-related behaviors while wearing seven IMU sensors on the limbs and waist. A CNN comprising three convolutional layers with kernel sizes of 3, 5, and 7 was trained using 1-s windows extracted from 8-s trials and evaluated using leave-one-subject-out cross-validation. The performance was summarized using macro-averaged accuracy, sensitivity, specificity, precision, and F1 score. Across 135 independent training runs, the CNN achieved a median macro-averaged accuracy of 77.0%, sensitivity of 77.0%, specificity of 95.4%, and F1 score of 77.4%. These results indicate stable window-level discrimination of the predefined motion classes under standardized conditions. This proof-of-concept study demonstrates that multi-site IMU signals combined with CNN-based modeling can technically discriminate ICU-relevant agitation-related motion patterns in a controlled laboratory setting. Although these findings do not establish clinical validity in ICU patients, they provide a methodological foundation for future studies aimed at patient-level validation and real-world critical care deployment.