Analysis of knee vibration signals using linear prediction.

Clinical methods used at present for the diagnosis of cartilage pathology in the knee are invasive in nature, and carry some risks. There exists a need for the development of a safe, objective, noninvasive method for early detection, localization, and quantification of cartilage pathology in the knee. This paper investigates the possibility of developing such a method based on an analysis of vibrations produced by joint surfaces rubbing against one another during normal movement. In particular, the method of modeling by linear prediction is used for adaptive segmentation and parameterization of knee vibration signals. Dominant poles are extracted from the model system function for each segment based on their energy contributions and bandwidths. These dominant poles represent the dominant features of the signal segments in the spectral domain. Two-dimensional feature vectors are then constructed using the first dominant pole and the ratio of power in the 40-120 Hz band to the total power of the segment. The potential use of this method to distinguish between vibrations produced by normal volunteers and patients known to have cartilage pathology (chondromalacia) is discussed.