Identification of peak V in brainstem auditory evoked potentials with neural networks

Automated wave labelling of the auditory brainstem response using machine learning

OBJECTIVE

To compare the performance of a selection of machine learning algorithms, trained to label peaks I, III, and V of the auditory brainstem response (ABR) waveform. An additional algorithm was trained to provide a confidence measure related to the ABR wave latency estimates.

DESIGN

Secondary data analysis of a previously published ABR dataset. Five types of machine learning algorithm were compared within a nested k-fold cross-validation procedure.

STUDY SAMPLE

A set of 482 suprathreshold ABR waveforms were used. These were recorded from 81 participants with audiometric thresholds within normal limits.

RESULTS

A convolutional recurrent neural network (CRNN) outperformed the other algorithms evaluated. The algorithm labelled 95.9% of ABR waves within ±0.1 ms of the target. The mean absolute error was 0.025 ms, averaged across the outer validation folds of the nested cross-validation procedure. High confidence levels were generally associated with greater wave-labelling accuracy.

CONCLUSIONS

Machine learning algorithms have the potential to assist clinicians with ABR interpretation. The present work identifies a promising machine learning approach, but any algorithm to be used in clinical practice would need to be trained on a large, accurately labelled, heterogeneous dataset and evaluated in clinical settings in follow-on work.

Machine learning approaches used to analyze auditory evoked responses from the human auditory brainstem: A systematic review

BACKGROUND

The application of machine learning algorithms for assessing the auditory brainstem response has gained interest over recent years with a considerable number of publications in the literature. In this systematic review, we explore how machine learning has been used to develop algorithms to assess auditory brainstem responses. A clear and comprehensive overview is provided to allow clinicians and researchers to explore the domain and the potential translation to clinical care.

METHODS

The systematic review was performed based on PRISMA guidelines. A search was conducted of PubMed, IEEE-Xplore, and Scopus databases focusing on human studies that have used machine learning to assess auditory brainstem responses. The duration of the search was from January 1, 1990, to April 3, 2021. The Covidence systematic review platform (www.covidence.org) was used throughout the process.

RESULTS

A total of 5812 studies were found through the database search and 451 duplicates were removed. The title and abstract screening process further reduced the article count to 89 and in the proceeding full-text screening, 34 articles met our full inclusion criteria.

CONCLUSION

Three categories of applications were found, namely neurologic diagnosis, hearing threshold estimation, and other (does not relate to neurologic or hearing threshold estimation). Neural networks and support vector machines were the most commonly used machine learning algorithms in all three categories. Only one study had conducted a clinical trial to evaluate the algorithm after development. Challenges remain in the amount of data required to train machine learning models. Suggestions for future research avenues are mentioned with recommended reporting methods for researchers.

Automatic removal of sonomotor waves from auditory brainstem responses

We have developed a computerized technique for automatic detection and removal of sonomotor waves (SMWs) from auditory brainstem responses (ABRs). Our approach is based on adaptive decomposition using a redundant set of Gaussian and 1-cycle-limited Gabor functions. In order to find optimal parameters and evaluate the efficiency of the methods, simulated data were first used before applying it to clinical data. Results were good and confirmed by an expert with years of clinical experience in ABR evaluation.

A comparative study between a simplified Kalman filter and Sliding Window Averaging for single trial dynamical estimation of event-related potentials

The classical approach for extracting event-related potentials (ERPs) from the brain is ensemble averaging. For long latency ERPs this is not optimal, partly due to the time-delay in obtaining a response and partly because the latency and amplitude for the ERP components, like the P300, are variable and depend on cognitive function. This study compares the performance of a simplified Kalman filter with Sliding Window Averaging in tracking dynamical changes in single trial P300. The comparison is performed on simulated P300 data with added background noise consisting of both simulated and real background EEG in various input signal to noise ratios. While both methods can be applied to track dynamical changes, the simplified Kalman filter has an advantage over the Sliding Window Averaging, most notable in a better noise suppression when both are optimized for faster changing latency and amplitude in the P300 component and in a considerably higher robustness towards suboptimal settings. The latter is of great importance in a clinical setting where the optimal setting cannot be determined.

Auditory brainstem response classification: A hybrid model using time and frequency features

OBJECTIVE

The auditory brainstem response (ABR) is an evoked response obtained from brain electrical activity when an auditory stimulus is applied to the ear. An audiologist can determine the threshold level of hearing by applying stimuli at reducing levels of intensity, and can also diagnose various otological, audiological, and neurological abnormalities by examining the morphology of the waveform and the latencies of the individual waves. This is a subjective process requiring considerable expertise. The aim of this research was to develop software classification models to assist the audiologist with an automated detection of the ABR waveform and also to provide objectivity and consistency in this detection.

MATERIALS AND METHODS

The dataset used in this study consisted of 550 waveforms derived from tests using a range of stimulus levels applied to 85 subjects ranging in hearing ability. Each waveform had been classified by a human expert as 'response=Yes' or 'response=No'. Individual software classification models were generated using time, frequency and cross-correlation measures. Classification employed both artificial neural networks (NNs) and the C5.0 decision tree algorithm. Accuracies were validated using six-fold cross-validation, and by randomising training, validation and test datasets.

RESULTS

The result was a two stage classification process whereby strong responses were classified to an accuracy of 95.6% in the first stage. This used a ratio of post-stimulus to pre-stimulus power in the time domain, with power measures at 200, 500 and 900Hz in the frequency domain. In the second stage, outputs from time, frequency and cross-correlation classifiers were combined using the Dempster-Shafer method to produce a hybrid model with an accuracy of 85% (126 repeat waveforms).

CONCLUSION

By combining the different approaches a hybrid system has been created that emulates the approach used by an audiologist in analysing an ABR waveform. Interpretation did not rely on one particular feature but brought together power and frequency analysis as well as consistency of subaverages. This provided a system that enhanced robustness to artefacts while maintaining classification accuracy.

Combining Wavelet Analysis and Bayesian Networks for the Classification of Auditory Brainstem Response

The auditory brainstem response (ABR) has become a routine clinical tool for hearing and neurological assessment. In order to pick out the ABR from the background EEG activity that obscures it, stimulus-synchronized averaging of many repeated trials is necessary, typically requiring up to 2000 repetitions. This number of repetitions can be very difficult, time consuming and uncomfortable for some subjects. In this study, a method combining wavelet analysis and Bayesian networks is introduced to reduce the required number of repetitions, which could offer a great advantage in the clinical situation. 314 ABRs with 64 repetitions and 155 ABRs with 128 repetitions recorded from eight subjects are used here. A wavelet transform is applied to each of the ABRs, and the important features of the ABRs are extracted by thresholding and matching the wavelet coefficients. The significant wavelet coefficients that represent the extracted features of the ABRs are then used as the variables to build the Bayesian network for classification of the ABRs. In order to estimate the performance of this approach, stratified ten-fold cross-validation is used.

Anesthesia control using midlatency auditory evoked potentials

This paper shows the development of a system to control inhalation anesthetic concentration delivered to a patient based upon that patient's midlatency auditory evoked potentials (MLAEP's). It was developed and tested in dogs by determining response to the supramaximal stimulus of tail clamping. Prior to tail clamp, the MLAEP was recorded along with inhalational anesthetic concentration and classified as responders or nonresponders as determined by tail clamping. This was performed at a number of different anesthetic levels to obtain a data training set. The MLAEP's were compacted by means of discrete time wavelet transform (DTWT), and together with anesthetic concentration value, a stepwise discriminant analysis (SDA) was performed to determine those features which could separate responders from nonresponders. It was determined that only three features were necessary for this recognition. These features were then used to train a four-layer artificial neural network (ANN) to separate the responders from nonresponders. The network was tested using a separate set of data, resulting in a 93% recognition rate in the anesthetic transition zone between responders and nonresponders, and 100% recognition rate outside this zone. The anesthetic controller used this ANN combined with fuzzy logic and rule-based control. A set of ten animal experiments were performed to test the robustness of this controller. Acceptable clinical performance was obtained, showing the feasibility of this approach.

Latency estimation of auditory brainstem response by neural networks

In the clinical application of auditory brainstem responses (ABRs), the latencies of five to seven main peaks are extremely important parameters for diagnosis. In practice, the latencies have mainly been done by manual measurement so far. In recent years, some new techniques have been developed involving automatic computer recognition. Computer recognition is difficult, however, since some peaks are complicated and vary a lot individually. In this paper, we introduce an artificial neural network method for ABR research. The detection of ABR is performed by using artificial neural networks. A proper bandpass filter is designed for peak extraction. Moreover, a new approach to estimate the latencies of the peaks by artificial neural networks is presented. The neural networks are studied in relation to the selection of model, number of layers and number of neurons in each hidden layer. Experimental results are described showing that artificial neural networks are a promising method in the study of ABR.

Segmentation of auditory brainstem response signals

Auditory brainstem responses are used to detect hearing defects in audiology and otoneurology. The use of computer programs for the analysis of such recordings is increasing. To identify their detailed properties a pattern recognition algorithm implemented in an analysis program must be highly reliable. For the recognition process, some preprocessing phases after recording the necessary, such as filtering and often also segmentation. In the following, we will explore segmentation, which can be used in preprocessing of biomedical signals after filtering. We studied linear segmentation, where slopes of short signal segments are computed and divided into different classes according to their values. A segment length of 8 samples for a sampling frequency of 50 kHz employed was best according to our tests and error criteria. Using clustering, we found that less than 10 segment classes is suitable for pattern recognition.

Modelling techniques and their application for monitoring in high dependency environments--learning models

This paper reviews the use of learning models including Bayesian classifiers and artificial neural networks in monitoring and interpreting biosignals. Generally learning models applied for analysis of biosignals are "black-box' types trained on the basis of measured signals. It is illustrated that the training and application of learning models more or less follow the same sequences. The main focus is the interpretation of electrical signals from the brain (electroencephalogram (EEG) and evoked potentials (EP)). Current analysis of these signals often reveals sudden changes in the EEG or evoked potentials to be the earliest discernible signs of inadequate perfusion of the brain. They may reflect problems such as systemic arterial oxygen desaturation or hypotension arising from other body system failures during critical illness. It is suggested that these brain signals should be recorded in the critical care unit, and that they should form part of the annotated database of biosignals established during the IMPROVE project. This would allow for the development of new methods for on-line warning of impending damage to the central nervous system, such that corrective actions could be taken before permanent damage occurred.