Computer based pattern recognition of carotid artery Doppler signals for disease classification: prospective validation.

A computer based pattern recognition method has been developed to classify the percent diameter reduction in nonoccluded internal carotid arteries. Using a combined B-mode/pulsed Doppler unit, the system utilizes spectral waveforms obtained from the low common and proximal internal carotid artery locations. The ECG-R wave is used as a time reference to synchronize the averaging of Doppler spectra from 20 heart cycles. An averaged waveform is generated and represents the spectral data from which features are extracted for analysis. A stepwise selection algorithm identifies a feature subset for partitioning the entire range of disease into two states, less than and greater than a decision point. Three such partitions are made, leading to the following categories: Normal, 1-20, 21-50 and 51-99% dia. reduction. A classifier was trained, tested prospectively against unknown data and the results compared to angiography. Of the 170 vessels tested, 141 (82%) were classified in the same category by angiography and the computer system. Agreement for each category was 93% (27/29) for the normals, 81.5% (44/54) for the 1-20% lesions, 78% (29/37) for the 21-50% lesions and 82% (41/50) for the 51-99% lesions. The computer method and angiography differed by more than one category in only one of the 170 tests. The level of agreement corrected for chance (Kappa +/- SE(K] was 0.769 +/- 0.039. Future efforts will be directed toward dividing classification of disease further (especially in the 51-99% category), developing a dedicated microprocessor for on-line analysis of the signals and using the system for prospective epidemiological studies of various populations.