Objective detection of brainstem auditory evoked potentials with a priori information from higher presentation levels

Abstract

This paper describes a brainstem auditory evoked potentials (BAEPs) detection method based on supervised pattern recognition. A previously used pattern recognition technique relying on cross-correlation with a template was modified in order to include a priori information allowing detection accuracy. Reference is made to the patient’s audiogram and to the latency–intensity (LI) curve with respect to physiological mechanisms. Flexible and adaptive constraints are introduced in the optimization procedure by means of eight rules. Several data samples were used in this study. The determination of parameters was performed through 270 BAEPs from 20 subjects with normal and high audiometric thresholds and through additional BAEPs from 123 normal ears and 14 ears showing prominent wave VI BAEPs. The evaluation of the detection performance was performed in two steps: first, the sensitivity, specificity and accuracy were estimated using 283 BAEPs from 20 subjects showing normal and high audiometric thresholds and secondly, the sensitivity, specificity and accuracy of the detection and the accuracy of the response threshold were estimated using 213 BAEPs from 18 patients in clinic.

Taking into account some a priori information, the accuracy in BAEPs detection was enhanced from 76 to 90%. The patient response thresholds were determined with a mean error of 5 dB and a standard deviation error of 8.3 dB. Results were obtained using experimental data; therefore, they are promising for routine use in clinic.

Introduction

Brainstem auditory evoked potentials (BAEPs) are a non-invasive measure of the subcortical auditory pathways’ functional integrity. They are produced in response to a brief acoustic stimulation and recorded using scalp electrodes. Because of poor signal to noise ratio, it is necessary to average several hundreds of single responses to get a recognizable BAEP. Normal response signal consists of five waves which can be characterized by their latency (time coordinate). The BAEP’s waveform is deeply modified when the stimulation intensity is decreased. The waves’ amplitude is reduced and their latency is increased. When the stimulation intensity is further decreased, the first waves are no longer visible on the signal, which comprises wave V only (see in Fig. 1). The plotted latency–intensity (LI) curves reflect this evolution. The lower stimulation intensity that still allows signal identification represents the subject response threshold. Precise determination of this threshold is necessary for accurate hearing assessment. The most widely used BAEPs are responses to clicks. They allow a fast evaluation of hearing impairment at high frequencies. If a more specific frequency evaluation is required, especially at the low frequencies, then a notched noise can be superimposed on either clicks or brief tones.

The conditions under which BAEPs are recorded are always difficult and often render their interpretation difficult and dependent on subjective analysis by the clinician. It is therefore desirable to develop objectively reproducible BAEP detection methods. Several statistical approaches (reviewed by Dobie [12]) have been proposed. Both analysis in the time domain and in the frequency domain give interesting performances. The result is a pass or fail response based on comparison of the BAEP trace under evaluation with either a signal reference or a noise reference. The main principles of the detection methods in the time domain are correlation of two BAEP traces [21], variance ratio of the BAEP to a noise reference [21], [25] and correlation with a template [25], [26]. In the frequency domain, the detection methods mainly are evaluation of the squared magnitude coherence and analysis of the phase distribution. Stürzebecher and co-workers compared the efficiency of several statistical tests and found that the best one was in the frequency domain [5], [35].

Other approaches are match filtering [41], autoregressive modeling [14], adaptive filtering [6] and classification of BAEPs using neural networks [1], [31]. Sanchez et al. reported high efficiency when combining several quantitative measurements to serve as feature vector for automatic detection [31].

As the BAEP signal comprises five waves, the detection of a response can also be related to a peak identification problem. A great variety of peak identification methods have been proposed based on syntactic pattern recognition [17], [20], band pass filtering [10], [17], [30], expert system design [3], neural networks [15], [29], [38], fuzzy logic [27], or wavelet transform [23], [29]. These methods show good performance when applied to well designed BAEP waveforms which are obtained with normal subjects at high stimulation intensity. The weakness generally encountered is poor generalization to degraded BAEP waveforms which may be obtained either with patients or at low stimulation intensity. Another difficulty to cope with is inter-subjects variability and possible misleading artifacts near threshold response. Fig. 2 shows another series of records for threshold determination. Note that the wave V LI curve has a different shape from that of Fig. 1, and signals with no response are not necessarily flat. Recently, special attention has been addressed to the identification of unclear IV/V complexes [29]. Therefore, interest for a supervised method taking into account higher level information has been pointed out since the eighties but was not further developed [10], [13], [28]. Nevertheless, several methods use heuristic criteria trying to reproduce human analyzing procedure [3], [27], [28], [29]. Control rules appear all the more useful than the wave V identification has to be performed at low stimulation intensity. Fig. 3 shows two similar BAEP waveforms with opposite interpretations by an otologist. This example illustrates the strong limitation of a single BAEP analysis apart from the recording context.

In this paper, we propose a new BAEP detection method in the time domain, based on supervised pattern recognition. We adopt an intermediate position between BAEP pass or fail detection and peak identification, addressing a wave V detection problem. Pattern recognition is therefore almost suitable to handle identification and detection problems in the same time. The interest of a time domain BAEP representation lies in the reference to human analyzing procedure which can help the determination of control production rules. We modified a template matching technique for BAEP detection and latency measurements in order to introduce a priori information to increase the detection accuracy. A priori information is driven from the subject audiogram and from physiological evolution of the latencies with decreasing stimulation intensity. Results are obtained with experimental data.