Amplitude-integrated EEG (aEEG) is a popular method for monitoring cerebral function. Although various commercial aEEG recorders have been produced, a detailed aEEG algorithm currently is not available. The upper and lower margins in the aEEG tracing are the discriminating features for data inspection and tracing classification. However, most aEEG devices require that these margins be measured semi-subjectively. This paper proposes a step-by-step signal-processing method to calculate a compact aEEG tracing and the upper/lower margin using raw EEG data. The high accuracy of the algorithm was verified by comparison with a recognized commercial aEEG device based on a representative testing dataset composed of 72 aEEG data. The introduced digital algorithm achieved compact aEEG tracing with a small data size. Moreover, the algorithm precisely represented the upper and lower margins in the tracing for objective data interpretation. The described method should facilitate aEEG signal processing and further establish the clinical and experimental application of aEEG methods.
Electroencephalography (EEG) constitutes a valuable tool for the continuous evaluation of cerebral function. However, conventional EEG has certain difficulties in clinical brain monitoring, mainly because it is inconvenient to discern long-term EEG trends by inspecting the full-size recording (usually 8 - 10 s per page, i.e., paper speed is about 30 mm/s).
To improve the clinical utility of EEG, researchers in the 1960s developed a simplified EEG monitoring system known as the cerebral function monitor (CFM) [
The upper and lower margins of the aEEG shape the top and bottom envelopes of the tracing, reflecting the maximum/minimum peak-to-peak amplitudes of the EEG signals. There are 2 main classification methods for aEEG interpretation. The first method is the simple voltage criterion proposed by al Naqeeb et al. [
In both classification methods, the primary focus is on the upper and lower margins of the aEEG [
To evaluate the aEEG data quantitatively and to establish the aEEG method in a wider application area, this paper introduced a detailed signal-processing method to obtain a compact aEEG with strictly defined upper and lower margins. The algorithm was employed to calculate aEEG tracings from the raw EEG data of 72 normal/abnormal infants. The acquired aEEG results were compared with the corresponding aEEG tracings exported from a commercial CFM to assess the accuracy of the method.
According to Maynard [
It is technically difficult to register a full-band EEG; DC-coupled amplifiers and a DC-stable electrode-skin interface are necessary [
The purpose of the aEEG method is to monitor the brain function by displaying the amplitude trend of brain activity. It is the boundary of the EEG waveform (i.e., the envelope) and not the EEG itself (i.e., the carrier) that characterizes the tendency of amplitude changes. Mathematically, the envelope is defined as the complex modulus of the analytic signal of the carrier, of which the latter can be calculated using the Hilbert transform. Classical envelope detection in CFM is performed using a diode and a resistor-ca-
pacitor pair with a 0.5-s time constant [
In the digital algorithm, a 5-order Butterworth filter was used to acquire the EEG envelope. The process of envelope extraction produces a smooth line approximately drawn through the peaks of the rectified EEG (
To obtain a bird’s-eye view of the cerebral function over a long duration, the EEG envelope was compressed in the scales of time (x-axis) and amplitude (y-axis). Time compression was achieved by laterally compressing the tracing (
In an aEEG tracing acquired by envelope compression, many pixels overlap and cover each other; the pixels on the upper and lower edges of an aEEG tracing contain most of the information on cerebral activity (
In practice, the terminal point of the vertical line does not necessarily equal the maximum or minimum of the epoch. To render the algorithm robust against a noisy environment, the high-amplitude noise (e.g., large, sharp waves produced by some clinical interventions) and low-amplitude noise (e.g., electromagnetic interference produced by other medical devices or mains supply) were depressed. This was done by picking up the terminal point near (but not at) the upper/lower edge in the algorithm. In particular, data in each 15-s epoch (i.e., 1500 points, given a sampling frequency of 100 Hz) were sorted and arranged in an ascending order of amplitude. The position of the selected terminal point in the sorted data was defined as the terminal position with the percentage unit. According to the definition, the upper terminal position is a percentage near and <100% while the lower terminal position is a percentage near and >0%; the specific values depend mainly on the signal-to-noise ratio of the EEG, which will be optimized in Section 3.2.
The upper and lower margins of the aEEG reflect the maximum and minimum fluctuations in cerebral activity. In an aEEG tracing composed of vertical lines, the upper and lower margins were depicted easily using the connecting lines of the terminal points (
To verify the proposed algorithm, 72 infant aEEG tracings and simultaneous EEG data (duration = 137 ± 32 min, sampling frequency = 100 Hz) were analyzed retrospectively. The data were registered using a commercially available CFM (Olympic CFM 6000, Natus, Seattle, WA). Two detecting electrodes were fixed in P3 and P4 in the international 10 - 20 system with a ground electrode placed in Fz. The aEEG tracings covered all 3 aEEG background patterns defined
by al Naqeeb et al. [
All infants were patients in the neonatal ward of Peking University First Hospital, Beijing, China during April to December 2009. The diagnoses of these infants included hypoxic-ischemic encephalopathy (HIE, N = 21), seizures (N = 18), white matter damage (N = 17), intraventricular hemorrhage (N = 11), and normal neurodevelopment (N = 5, control cases). Informed consent was obtained from the parent or legal guardian of each infant. The average postmenstrual age, calculated by adding the weeks of gestational age to the postnatal age, was 42.1 ± 6.5 weeks (range: 31 - 69 weeks). The infant aEEG characteristics are found in [18,20]. Some representative aEEG tracings in the testing dataset are displayed in the left column of
Digital aEEG tracings (represented by the x and y coordinate values of each pixel) were exported from the Olympic CFM 6000 to obtain the upper and lower terminal points of the aEEG vertical line in each 15-s epoch. The results were compared with the corresponding terminal points, which were calculated based on the proposed algorithm using the simultaneously recorded EEG data on the same CFM.
We defined the error rate (ER) of the aEEG algorithm as the sum of the amplitude differences of each terminal point pair (i.e., terminal point in CFM tracing vs. the corresponding terminal point in algorithmic tracing) divided by the sum of the amplitudes of each terminal point in the CFM tracing. Accordingly, the ERs of the upper and lower terminal points, denoted as ERU and ERL, respectively, were defined as
where AU and AL are the amplitudes of the upper and lower terminal points, respectively. The superscripts “CFM” and “a” indicate whether the amplitude was measured in the CFM tracing or in the algorithmic tracing, respectively.
Considering ERU and ERL as objective functions, the upper and lower terminal positions in the algorithm (as described in Section 2.5) can be optimized using an exhaustive method. Keeping other parameters fixed (refer to
The present paper introduces a digital algorithm to acquire a compact aEEG tracing composed of vertical lines. Compared with the full-size tracing obtained by compressing the envelope of the rectified EEG, the compact version of the aEEG is faster to calculate and easier to save. In addition, the upper and lower margins of the aEEG are defined properly in the compact-
tracing based on the 2 terminal points of each vertical line. The smoothed margins depicted on the aEEG tracing provide an objective measurement of the trend of the intensity of the cerebral activity, which can largely improve the quality of the aEEG evaluation.
Nonparametric statistics were used to describe the amplitude of aEEG in this paper. The upper/lower margin was smoothed using the median (not the mean) of every 20 terminal points. We optimized the position of the terminal point in the amplitude-sorted data (as shown in
As one of the most popular commercial CFMs, the Olympic CFM 6000 CFM was selected in the present study as the standard aEEG recorder to collect reliable aEEG and simultaneous EEG signals. The 72 aEEG data in the testing dataset were carefully selected from hundreds of infant aEEGs: 1) to cover all 3 of the aEEG patterns defined in the voltage classification, and 2) to contain data with different levels of signalto-noise ratios. The error rates of the upper and lower terminal points were 2.6% and 4.9%, respectively, which are acceptable values for clinical application and general research use.
The proposed algorithm has been employed to analyze aEEG data from a large population of infants, and a related article was published in a top journal of pediatrics [
The upper and lower margins of the aEEG contain most of the useful information on cerebral function. There was an initial CFM type called CFAM [22,23] that displayed lines representing the maximum, mean, and minimum amplitudes of the aEEG tracing. This CFM type was helpful for objectively assessing the aEEG margins [24,25]. Unfortunately, this methodology currently is not employed in most CFM products; only aEEG tracings are presented graphically, and so visual inspection is needed to determine the tracing margins in these commercial devices. In this paper, we managed to obtain the accurate amplitude of the aEEG margin, which largely enhanced the objectivity of data interpretation. Quantitative aEEG margins can be obtained along with graphical tracing in some new digital aEEG devices (e.g., the NicoletOne monitor [
In summary, a digital signal-processing method was proposed to obtain a compact aEEG tracing with upper and lower margins using raw EEG data. The tracings were compared with those of commercial devices, revealing high consistency. Digitalized aEEG tracing makes it feasible to postprocess the signal and to integrate aEEG features into an automatic classification or diagnosis system. Using the algorithm, researchers can obtain accurate amplitudes of the upper and lower margins for further analysis, thus facilitating the statistical research of large-scale clinical trials. We believe that the step-by-step processing method described in this article provides valuable information on aEEG signal processing. Moreover, this method should help researchers to investigate aEEG tracings without a CFM, potentially promoting the clinical establishment of the aEEG method.
The authors are grateful to Lili Liu, MD (Peking University First Hospital) for her invaluable help in data collecting and sorting.