Automatic recognition of the source of temporomandibular joint sounds.

Sounds evoked in the temporomandibular joint (TMJ) during jaw movements may indicate pathology. They are in dental clinics usually recorded by auscultation and noted in protocols by verbal, subjective descriptions. Time-frequency analysis of electronically recorded TMJ sounds makes possible a more objective and sophisticated analysis. Such sounds were recorded from four subjects and grouped into two sets. One was used for training a classifier, while the other was used for testing its ability to relate a given sound to the subject from which it was recorded. Both scale and time-shift invariant representations, as well as only time-shift invariant representations of the Reduced Interference Distributions of the TMJ sounds were used for pattern recognition. The nearest neighbor, zero-subspace and nearest constrained linear combination classification methods were employed. It was observed that TMJ sound patterns could be very typical for a given person. This indicates that our classification approach can be developed into a useful diagnostic tool by obtaining training sets from patients where a definitive diagnosis of TMJ pathology has been obtained.