Falling has become one of the leading causes to both fatal and non-fatal injuries among the elder adults and patients with dysfunctional mobility. The following consequence of a fall will be that the fear of falling can reduce the daily activities, leading to physical deterioration and social isolation. Falls and its related injuries can be predictable and preventable with specific interventions targeting the corresponding risk factors including muscle strength, balance and mobility. The initial step of an effective fall prevention program is to perform fall risk assessment to identify persons at high risk and then target specific interventions to reduce or eliminate falls. Due to several reasons such as unreliable subjective measures, high cost and clinical time constraints, effective fall risk assessment is still not routinely integrated into daily clinical practice. The inexpensive and easy-to-use wearable inertial measurement unit (IMU) provides the promising technique to assess fall risk, however there are still many issues for clinical practice at present. This objective of this dissertation was to investigate the feasibility, reliability and repeatability of using IMU sensors to assess fall risk for clinical use.The accuracy of an IMU sensor was validated on the rigid body and human body using an optical motion capture system. A complimentary filter with gradient descent algorithm was used to verify the accuracy of built-in sensor fusion model. A simple template of magnetic mapping was built to quantify the magnetic disturbances and then a simplified interpolation model was developed to compensate the heading angle error in the complicated lab settings. Then different combinations of sensor configurations were evaluated among three different groups of subjects with six sensors placed on the body. The optimal configurations was explored according to the classification performance from three machine learning techniques. Sixty-five older adults from the senior center with single IMU placed in front of the chest were recruited. A self-report questionnaire was provided as a common clinical assessment to divide sixty-five senior people into two risk levels. Meanwhile, three machine learning models were applied to classify the fall risk of these subjects according to their sensor data during three different physical tests Both of the static and dynamic orientation accuracy from rigid body tests were within 2° achieved by built-in sensor fusion model while the static and dynamic accuracy for sensor orientations on human body were within 2.5°. The accuracy and precision of IMU measurements are sufficient for human motion applications without excluding the soft tissue artifact and unexpected sensor movement on the human body. For the magnetic disturbances, the heading estimation errors from static tests were significantly larger than that of dynamic test. The compensation method did improve the accuracy of heading angles in the static test. When performing the test with IMUs, starting in an undisturbed magnetic field will reduce the heading angle error and the closer IMU is to the floor, the larger the heading angle error will be. The optimal sensor configuration was achieved by the sensors placed on chest, wrist and shank together. A single sensor configuration can also produce very high accuracy like the chest sensor. Timed-Up-and-Go (TUG) test provided the best classification results for fall risk classification and the combination of multiple tests did not improve the classification performance. Support Vector Machine (SVM) model is the best machine learning technique among three models for fall risk classification. Overall, the wearable IMU sensor-based fall risk classification model has potential to improve the diagnosis of elder adults with risk of falling and allow pre-intervention to prevent future falls.