Recent improvements in data learning techniques have catalyzed the development of various clinical learning systems. However, for clinical applications, training from noisy data can cause significant misleading results, directly leading to potentially dangerous clinical decisions. Given its importance, this work targets to present a preliminary effort to identify corrupted vital sign data by analyzing the patient motions on hospital beds. Specifically, we design an embedded sensor-based motion detection platform to capture and categorize different noise-causing motion on intensive care unit beds through a pre-deployment study at the Ajou University Hospital. We design light-weight and low-resource demanding software for motion sensor data processing and evaluate its performance from real-patient traces collected at the ICU. Evaluation results using a ~200 minute data set show that our system detects and classifies patient motion states with 76% accuracy and well-identifies vital sign time-series regions affected by motion noise.