Automated Vibration and Acoustic Crepitus Sensing in Humans

Crepitus vibrational and acoustic signal analysis of the human facet joints of the lumbar spine has historically been a difficult problem due to the inhomogeneous and varied signal characteristics. Here we improve upon our previous automated computational method, now enhancing it for analysis of human crepitus. Compared with this group's previous studies using a mechanical model; human crepitus is extremely complex. Moreover, there is no existing availability of large numbers of human crepitus data to enable effective machine learning approaches. Therefore, we proposed an automated method (AM) of analysis that, analogous to machine learning, used a test set (n = 16) and an experimental set of data (n = 48). The advantage of beginning with this approach was that we identified characteristics of the signal that are unavailable or otherwise not easily obtained in more advanced methods, such as “black box” machine learning methods. However, we did not have the high fidelity that a machine learning approach would provide. This was shown by only a fair level of inter-rater agreement (Kw = 0.367; SE = 0.054, 95% CI = 0.260-0.474) between the AM and human observers before adjustments were made in the AM. Following adjustments to the AM, inter-rater agreement improved to a substantial level of agreement (Kw = 0.788; SE = 0.056, 95% CI = 0.0.682-0.895). In the future, we recommend a machine learning study with a high number of subjects, that can better capture the nuances of varying types of human crepitus.