Falling is an important health issue that occurs in elderly people, which becomes a major problem that needs to be addressed urgently. To address this problem, a wearable fall detection system based on a micro-electromechanical system (MEMS) inertial sensor is proposed. The identification of four falling behaviours (forward, backward, left, and right falls and six normal behaviours (walking, running, hopping, up-and-down, stooping, and sitting) was successfully performed using this detection system. Quaternion complementary filtering attitude analysis and multi-level threshold algorithm are applied to determine the thresholds of the combined acceleration, angular velocity, and attitude angle of the fall behaviours, which are 4.8 g, 180 deg/s, and 100 deg, respectively. The proposed system can effectively distinguish falling behaviour from normal behaviour and give early warning before falling. The sensitivity, specificity, and accuracy for predicting the falling are calculated to be 91.0%, 93.3%, and 92.3%, respectively.