Suitability of an inter-burst detection method for grading hypoxic-ischemic encephalopathy in newborn EEG

TL;DR

This study evaluates an inter-burst detection method originally designed for preterm infants to classify injury severity in term infant EEGs, demonstrating that simple inter-burst measures can effectively grade hypoxic-ischemic encephalopathy.

Contribution

It adapts and validates an existing inter-burst detection technique for term infants, showing its effectiveness in injury classification using machine learning.

Findings

Best feature, percentage of inter-bursts, achieved 59.3% accuracy.

Combining features with an MLP improved accuracy to 77.8%.

Simple inter-burst interval measures can classify injury grades effectively.

Abstract

Electroencephalography (EEG) is an important clinical tool for grading injury caused by lack of oxygen or blood to the brain during birth. Characteristics of low-voltage waveforms, known as inter-bursts, are related to different grades of injury. This study assesses the suitability of an existing inter-burst detection method, developed from preterm infants born <30 weeks of gestational age, to detect inter-bursts in term infants. Different features from the temporal organisation of the inter-bursts are combined using a multi-layer perceptron (MLP) machine learning algorithm to classify four grades of injury in the EEG. We find that the best performing feature, percentage of inter-bursts, has an accuracy of 59.3%. Combining this with the maximum duration of inter-bursts in the MLP produces a testing accuracy of 77.8%, with similar performance to existing multi-feature methods. These results validate the use of the preterm detection method in term EEG and show how simple measures of the inter-burst interval can be used to classify different grades of injury.

Tables

Table 1. TABLE I : EEG grades for hypoxic–ischemic encephalopathy defined by Murray et al . [ 4 ] .

Grade	Description of EEG
1	Normal/Mild abnormalities: Continuous background pattern
	with mild voltage depression and asymmetric patterns
2	Moderate abnormalities: Discontinuous activity with clear
	asymmetry, IBI $\le$ 10 s
3	Major abnormalities: Discontinuous activity with IBI range
	between 10-60 s, severe depression without wake/sleep cycles
4	Inactive/Severe abnormalities: Background activity pattern
	$\le$ 10$\mu$V or severe discontinuity of IBI $\ge$ 60 s

Table 2. TABLE II : Performance of the multi-layer perceptron using 1 and 2 features.

Features	Classification accuracy
IBI percentage	59.3%
maximum IBI	53.7%
IBI percentage & maximum IBI	77.8%

Table 3. TABLE III : Confusion matrix of the system’s output and actual grade

Actual	System’s Output				Total	Incorrect
grade	1	2	3	4
1	20	2	0	0	22	2
2	4	8	2	0	14	6
3	1	2	9	0	12	3
4	0	0	1	5	6	1
Precision	25	12	12	5	54	12

Table 4. TABLE IV : Comparison of the proposed method with the other techniques.

HIE Grading	Stevenson	Ahmed	Proposed
Method	et al. [10]	et al. [11]	method
Overall accuracy	77.8%	87.0%	77.8%

Equations

$$IBI\%=\frac{{\sum }_{i}IB{I}_i}{L}\times 100$$

$$IB{I}_m=\max (IB{I}_i)-\min (IB{I}_i).$$

Full text

I Introduction

Electroencephalography (EEG) provides an effective non-invasive tool for detecting abnormalities of the brain. Different EEG activity can provide clinical information with respect to changes in sleep [1], maturation [2], neuro-developments [3], and hypoxic-ischemic encephalopathy (HIE) [4], a type of injury caused by lack of blood and oxygen to the brain. In the clinical care of asphyxiated infants, it is essential to correctly identify the degree of HIE severity from EEG recording [4]. But continuous EEG monitoring and review in most neonatal intensive care units is not practical. Automated EEG analysis could help with this review by presenting the physician with useful and timely clinical information about brain function.

A distinct marker of EEG abnormalities, and therefore HIE is a discontinuous pattern in the EEG of term infants. This activity consists of short-duration periods of high-voltage activity, known as bursts, interspersed with periods of low voltage activity, known as inter-bursts [5]. From literature, not many studies have been carried on detecting brain abnormalities in term infants through EEG analysis using inter-burst detection or features based on inter-burst activity [6]. Amongst such is a study by O’Toole et al. which has developed an inter-burst detection approach for preterm infants using multi-feature and machine learning approach [7]. Besides this, there are few studies which shows burst detection among term infants. Matic et al. [6] has used adaptive inter-bursts interval detection by extracting six significant features from the burst interval activities followed by pattern classification. Parisa et al. [8] have recently used non-linear feature extraction techniques with machine learning classification for burst detection. These studies show significant results for detecting burst or inter-burst activity, on preterm or term infants EEG.

An integral part of grading the EEG for HIE injury, is to assess the level of discontinuity in the EEG by visually quantifying inter-burst activity [4, 9]. Automated grading methods developed by Stevenson et al. [10] and Ahmed et al. [11] which use multiple complex time-domain, frequency-domain, and information theory features. These studies do not consider detailed classification based on inter-burst interval analysis, which is an important component of visual grading of the EEG.

Based on the above shortcomings, this paper provides the following contributions:

• •

Validate an inter-burst detection method developed for preterm EEG [7], to investigate its ability to detect inter-burst intervals (IBIs) on the EEG of term infants.

• •

Differentiate the four grades of HIE using an automated EEG system with simple features of the inter-bursts measure followed by a data-driven MLP approach.

II Methods

We implement a burst detection method developed for preterm infant EEG [7] and investigate its ability to detect IBIs on the EEG of term infants. The detector classifies the EEG as bursts or inter-bursts, but for our application we are only interested in characteristics of the inter-bursts. We then extract features from the temporal organisation of the IBIs and use a data-driven approach to differentiate four grades of HIE. The overall structure of the proposed system is shown in Fig. 1.

II-A EEG Data

The EEG was recorded on 54 term infants using a NicoletOne EEG system within the neonatal intensive care unit (NICU) of Cork University Maternity Hospital, Cork, Ireland. This study was approved from the Clinical Ethics Committee of the Cork Teaching Hospital with written and informed parental consent obtained before EEG recording. The EEG recordings was initiated within 6 hours of birth and continued for up to 72 hours to monitor the evolution of the developing encephalopathy with seizure surveillance. The EEG recording used 9 active electrodes T4, T3, O1, O2, F4, F3, C4, C3, and Cz as standard protocol in the NICU. Our analysis used an 8-channel bipolar montage derived from these electrodes as F4-C4, C4-O2, F3-C3, C3-O1, T4-C4, C4-Cz, Cz-C3 and C3-T3.

One-hour epochs of EEG were pruned from the continuous EEG to avoid major artefacts when possible. This epoch was reviewed independently by 2 EEG experts and graded according to the system defined by Murray et al. [4], described in Table I. If the experts disagreed, they reviewed the EEG together and a consensus was reached to determine the final grade. This same dataset has been used by Stevenson et al. [10] and Ahmed et al. [11].

II-B Inter-Burst Detection Approach

In our proposed system architecture, we incorporate an inter-burst detection method which was initially designed for preterm EEG. The method extracts multiple features of amplitude, spectral shape and power, and a frequency-weighted energy features [7]. These features are estimated within four frequency bands 0.5–3, 3–8, 8–15, and 15–30 Hz. A feature selection method is used to maximise relevancy and minimise redundancy, retaining only necessary features. These selected features were then combined using a linear-kernel support vector machine (SVM). This channel-independent system was trained on EEG data from 36 premature infants with a gestational age of $<$ 30 weeks. We test this method on term EEG, processing each channel separately to detect inter-burst activity. Visual assessment of the method demonstrated reasonable performance at detecting inter-bursts, as the example in Fig. 2 shows for different grades.

II-C Preprocessing and Feature Extraction

Each bipolar EEG channel was analysed by the inter-burst detector. As HIE is consider a global brain injury [12], the inter-burst information is compressed over the 8 channels by majority voting of the individual binary masks to form one summary mask. We then extract two features to summarise the binary mask.

II-C1 IBI percentage

IBI percentage ($IBI\%$) is used to detect the overall occurrence of inter-burst activity in an epoch. It is calculated as the summation of duration of the IBIs to the overall epoch length,

[TABLE]

where, $IBI_{i}$ represents the duration of $i$-th IBI and $L$ denotes the overall length of epoch.

II-C2 Maximum IBI

The HIE grading describes various maximum IBIs for different grades. We calculate this measure by subtracting the difference between the maximum and the minimum inter-burst length within an epoch as,

[TABLE]

II-D Classification of HIE Grades

To combine features and to apply a data-driven approach to estimating thresholds for the 4 class separation, we use a multi-layer perceptron (MLP) as the machine learning algorithm.

The MLP was configured with one hidden layer with fourteen neurons and four neurons in output layer, one for each HIE grade to be recognised. The selection criterion was based on the convergence of the learning error from different combinations as described in [13]. Weights and biases were initialised before training from a normal distribution with a mean and variance of 0 and 1, respectively. The training process was stopped when the absolute rate of change in the average squared error ($\theta$) per iteration was sufficiently small ($\theta\leq 0.1$) [13]. The MLP model was trained based on the evaluated feature dataset and while testing, the maximum value within the output layer neuron in MLP was selected as the recognised grade for an evaluated feature set from epoch.

For testing we use a leave-one-subject-out (LOSO) cross-validation method. Based on 54 infants considered in study, system performance was observed by training the MLP classifier with 53 EEGs and testing on the one left out. This was progressed until all 54 infant’s EEGs were tested to determine the overall classification accuracy of system.

III Results

Fig. 3 illustrates distribution of the IBI percentage and maximum IBI features for each of the 4 EEG grades. We find varying degrees of separation between the 4 classes for the different features.

The evaluated IBI percentage shows the lowest percentage of IBI activity for Grade 1 followed by Grade 2 and nearly 30–80% for Grade 3 after Grade 4 showing close to 100% of IBI activity which satisfy the clinical EEG grading condition for HIE defined by Murray et al. Further, the estimated maximum IBI duration for Grade 1 and Grade 2 falls under 10 s, Grade 3 range approximately between 10 – 60 s and the maximum IBI duration for Grade 4 mostly exceeds 60 s which strongly correlates to the clinical grading system defined by Murray et al. in Table I.

Testing results using each IBI feature separately is $<$ 60% accuracy, as shown in Table II. Combining both features, however, greatly improves performance to 77.8%.

Confusion matrix for the LOSO cross-validation using the two features is presented in Table III. It shows that 12 out of 54 infant data were misclassified resulting in the accuracy of 77.8%. Most misclassification occurred between grades 1 and 2, which from Fig. 2 appears to be the harder grades to separate using these IBI features alone.

The proposed system performance is compared with other approaches which uses the same EEG dataset, as presented in Table IV. Stevenson et al. [10] shows similar 77.8% classification accuracy using a complex feature set of 8 values. Ahmed at al. [11] achieve a higher level of accuracy (87%) using Gaussian mixture model supervector combined with an SVM with a larger feature set of 55 values from time, frequency, and information theory features, in addition to probabilistic post-processing analysis. In contrast, our proposed system using only 2 features of the IBI measure, manages comparable accuracy to the Stevenson et al. method.

IV Discussion and Conclusions

We present a novel approach for grading HIE abnormalities in the EEG. The results validate the use of a preterm inter-burst detection method to detect inter-bursts in the EEG of term newborns. Simple measures of the IBI combined using a data-driven approach can achieve comparable accuracy with state-of-the-art HIE grading systems.

A previous study by Stevenson et al. [10] reported an accuracy of 77.8% while testing on same dataset of 54 term infants. This performance improved to 87% by Ahmed et al. using a large feature set, GMM supervectors combined with SVM, and additional post-processing [11]. Both these studies use complex multi feature analysis on each epoch without considering the classification of inter-burst interval based on the clinical criterion. Future work will include developing a more complete feature set to capture other characteristics of the EEG to further improve classification performance.

In conclusion, we validate the use of an EEG inter-burst detection method developed for preterm infants applied to the EEG of term infants. Only two simple features of the temporal organisation of the inter-bursts are used to generate a high-performance system comparable in accuracy with more complex multi-feature methods [10]. Future systems will incorporate these IBI features with a more diverse feature set and optimised machine learning methods to further improve performance.