To address the challenges posed by the aging process, we designed and validated an LSTM-based automatic remote health risk assessment system for the elderly. This system consists of a wireless physiological parameter sensing unit, a vital sign prediction unit and a pre-defined risk scoring criteria unit. The vital sign prediction module is composed of five 5-input-1-output neural networks based on the LSTM architecture, which are responsible for predicting the vital signs collected by wireless sensors, including: systolic blood pressure (SBP), pulse rate (PR), respiratory rate (RR), temperature (TEMP), and oxygen saturation (SPO2). The pre-defined health risk scoring criteria is a simplified version of the National Early Warning Score (NEWS), which is responsible for calculating the risk level based on the predicted values. This allows the care team to respond to the medical needs of the elderly in a timely manner. Through experiments, our system can achieve a risk identification accuracy of 74% and MAEs of the predicted values for each parameter are in an acceptable range. Our results suggest that an automated remote health risk assessment system for the elderly using deep learning could be a viable new strategy for home-based monitoring systems.