Improved automated detection of embolic signals using a novel frequency filtering approach.

Abstract

Asymptomatic embolic signal detection with the use of Doppler ultrasound has a number of potential clinical applications. However, its more widespread clinical use is severely limited by the lack of a reliable automated detection system. Design of such a system depends on accurate characterization of the unique features of embolic signals, which allow their differentiation from artifact and background Doppler speckle. We used a processing system with high temporal resolution to describe these features. We then used this information to design a new automated detection system. We used a signal processing approach based on multiple overlapping band-pass filters to characterize 100 consecutive embolic signals from patients with carotid artery disease as well as both episodes of artifact resulting from probe tapping and facial movement and episodes of Doppler speckle. We then designed an automated detection system based both on these embolic signal characteristics and on the fact that embolic signals have maximum intensity over a narrow frequency range. This system was tested in real time on stored 5-second segments of data. The value of peak velocity at maximal intensity discriminated best between embolic signals and artifact and allowed differentiation with 100% sensitivity and specificity. Relative intensity increase, intensity volume, area under volume, average rise rate, and average fall rate appeared to discriminate best between embolic signals and Doppler speckle. For the majority of embolic signals, the intensity increase was spread over a narrow frequency or velocity range. The automated system we developed detected 296 of 325 carotid stenosis embolic signals from a new data set (sensitivity, 91.1%). All 200 episodes of artifact from a new data set were differentiated from embolic signals. Only 2 of 100 episodes of speckle were misidentified as embolic signals. Using a novel system for automated detection, which utilizes the fact that embolic signals have maximum intensity over a narrow frequency range, we have achieved detection with a high sensitivity and high specificity. These results are considerably better than those previously reported. We tested this initial system on short 5-second segments of data played in real time. This approach now needs to be developed for use in a true online system to determine whether it has sufficient sensitivity and specificity for clinical use.