Improving the reliability of the pattern electroretinogram with adaptive filtering

Abstract

Electrophysiological testing of the eye and visual system is important in the diagnosis of some diseases. Unlike the electrocardiogram and other more common tests, the amplitudes of the signals are very small and subject to noise and interference. Innovative signal processing algorithms have been shown to be helpful for the electroretinogram (ERG). In this paper, we extend that work to the pattern ERG (PERG), which has utility in the assessment, diagnosis and management of central retinal and optic nerve disease. The use of the PERG has been limited by the small amplitude signal and low signal-to-noise ratio. The aim of this study was to use adaptive denoising techniques to improve the reliability and utility of the PERG. PERGs were performed on 70 eyes of 36 patients following ISCEV protocols. After a short break of approximately 2–5 min, a second recording was obtained to assess test-retest reliability of denoising. Denoising was performed on both recordings using five different adaptive filtering algorithms. Denoising algorithms generally improved test-retest reliability measured with intraclass correlation coefficient (ICC) and coefficient of repeatability (CoR). Empirical Wavelet Transform provided the best overall improvement in reliability, improving ICC from 0.79 to 0.84 (p = 0.049) and CoR from 2.12 µV to 1.69 µV (p = 0.02). Adaptive denoising algorithms may improve the reliability of PERG.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-30929-y.

Introduction

The electroretinogram (ERG) is the evoked potential measured at the cornea of the eye relative to the retina¹. It is usually evoked with a luminous flash stimulus and is generated by the outer layers of the retina (the photoreceptors – rods and cones, and the bipolar cells)². The ERG has utility in the diagnosis and monitoring of certain diseases of the eye, especially those in which function is lost before the structure is changed^3,4. Diseases such as paraneoplastic retinopathy and genetic disorders of the outer retina can manifest symptoms and ERG changes before any structural changes can be observed – for example, with Optical Coherence Tomography (OCT)⁵. The ERG is typically collected with a differential pair electrode setup. The most commonly used is a Dawson Trick Litzkow (DTL) fibre electrode with an EEG skin electrode as the indifferent, located at the lateral canthus of the eye. Contact lens electrodes (unipolar and bipolar) were previously very popular but are less used today. The amplitude of the ERG b-wave is typically 100–200 µV and can be observed without signal averaging.

The electroretinogram can be evoked by a patterned stimulus, which can be either pattern-onset-offset (where a pattern appears and disappears from a grey background) or pattern-reversal. This study focused on the pattern-reversal electroretinogram (PERG) which produces a much lower amplitude signal (typically 5–10 µV) compared to the flash ERG. The stimulus, typically an alternating checkerboard is carefully chosen to have a constant luminance to avoid any embedded flash ERG contaminating the signal. The checkerboard should be designed with a check width of 0.8° of visual field when viewed from 1 m, maintaining an aspect ratio between 1:1 and 4:3. As the number of dark and light squares are the same, the luminance should be constant over time. Traditional cathode ray tube displays or plasma display screens are still used for this purpose because constant luminance is easily achieved. Liquid crystal displays can be problematic as the on and off transitions are not of equal speed. In recent years, organic light-emitting diode displays have become popular as pattern stimulators, as the on and off transitions are very fast⁶. A consistent ambient light or dark environment should be maintained for all tests⁷.

The PERG has been shown to be generated by the ganglion cell layer (GCL) of the retina and is dominated by the macula. The test has utility in diagnosing diseases of the inner retina, such as glaucoma and certain genetic disorders^8–11. Evaluated in combination with the ERG, the exact location of the pathophysiology can be ascertained. At low pattern reversal frequencies (for example, 2–4 Hz) a transient waveform is detected. This is known as the transient PERG or simply PERG. At higher frequencies, a regular harmonic waveform is detected, known as the steady state PERG (ssPERG) and is often analysed with transformation into the frequency domain. The PERG consists of an initial negative deflection (N35) followed by a positive peak (P50) and a second negative wave (N95). The N35 is measured relative to the pre-stimulus baseline and the P50 is measured relative to the N35 trough. P50 reflects RGC function but also has more distal origins that are not yet fully understood¹². N95 reflects pure RGC function and this signal may be diminished in primary RGC disease (e.g. glaucoma) and secondary to optic nerve disease (e.g. demyelination, optic nerve compression) where selective loss of N95 may be seen. The P50 is the measure most commonly used in clinical practice. Figure 1 demonstrates a typical transient PERG.

Fig. 1.

A transient PERG demonstrating the typical waveform. The N35, P50 and N95 are shown.

Standards for the PERG have been published by the International Society for the Clinical Electrophysiology of Vision (ISCEV) and were most recently updated in 2024. The standards specify the distance, luminance, pupillary dilation and other stimulus and acquisition parameters for the test¹³.

Collection of the PERG can be challenging – especially when compared to the ERG. The signal is very small compared to the types of noise contaminating the recording. Chief sources of noise are the electromyogram from the eyelids, induced voltages due to movement of the fibre electrode within the tear film, the electrocardiogram, the electrooculogram, as well as amplifier noise, powerline interference and high-frequency noise from switch mode power supplies¹⁴. As impedances and time constants between the differential pair electrodes may not be perfectly matched, this sort of noise is often not completely eliminated by common-mode rejection.

Another troublesome sort of noise is ultra-low frequency drift and baseline wander. Where very large, individual traces can be rejected, but if of smaller amplitude, can induce large errors in the signal. This sort of noise is very challenging to remove by conventional analogue or digital filter techniques. Adaptive filtering has shown potential for removing this sort of noise in the ERG^15,16. An adaptive filter is one in which the parameters change for each trace, based upon the trace itself. Examples of adaptive filters include empirical mode decomposition, variational mode decomposition and empirical wavelet transform¹⁷.

We have previously explored adaptive filters for the removal of ultra-low frequency noise from the ERG when targeting the photopic negative response (PhNR)¹⁶. The PhNR is a slow negative-going wave following the b-wave of the ERG. It is usually elicited with a red flash on a blue background. Because the PhNR has a long and variable implicit time, it is substantially affected by drift and baseline wander. In this work, we investigate whether adaptive filters are also useful for the PERG given its large noise-to-signal ratio.

Our study is novel in applying and comparing multiple adaptive denoising algorithms to the PERG in a clinical setting, with a specific focus on their effect on signal reliability. Importantly, we also consider the feasibility of implementing these methods in real-time or near-real-time applications.

State of the art

Various adaptive denoising methods have been used throughout electrophysiology^18–20. The general aim of these methods is to decompose the recorded signal into a limited number of components that may be summed to reproduce the original signal. Often, the lowest frequency component is attributed to baseline drift and is removed from each trace. The ensemble average of filtered traces thus produces a denoised signal¹⁹. The following filters have been applied to various electrophysiological signals, but have yet to be studied for the PERG. We chose to investigate the following five adaptive denoising methods:

Empirical mode decomposition (EMD), ensemble empirical mode decomposition (EEMD) and complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN)

The EMD was first described by Huang et al. in 1998 and has been applied to various biomedical signals, including electroencephalogram (EEG), electromyogram (EMG), electrocardiogram (ECG) and ERG, as well as other fields such as seismology^20–24.

The EMD is an adaptive, data-driven iterative process that separates the signal into intrinsic mode functions (IMFs) and a residue²⁵. To exist, IMFs must satisfy two criteria:

1. The number of extrema and zero crossings differ by no more than one.

2. The mean of upper and lower envelopes must be zero.

The residual signal (with a maximum of one turning point) is considered to contain ultra-low frequency noise and is removed. See Supplementary Material for mathematical formulae and theory of how this is performed^26,27.

A known limitation of the EMD is its susceptibility to mode mixing. This occurs when the algorithm cannot extract two closely spaced frequency components.

To overcome issues with mode mixing, the EMD was extended to produce the EEMD and CEEMDAN, which are slightly more sophisticated adaptive noise filters. Both EEMD and CEEMDAN add white noise to the signal to help overcome this issue but differ in how they proceed between each decomposition level.

EEMD incorporates finite variance (Gaussian) white noise into the original signal. The ensemble of these is then used to define modes as the average of corresponding IMFs. Each decomposition is performed independently with no connection between the residues at each stage of the process, causing the EEMD to suffer from reconstruction error.

CEEMDAN recombines the signal and noise after every IMF extraction to address the issue underlying the EEMD. This maintains a connection between decomposition levels and allows for perfect signal reconstruction, which is not the case for EEMD.

Key tuneable parameters of the EMD-based functions are summarised in Table 1.

Table 1.

Description of the parameters that can be optimised for each adaptive denoising method. Parameter settings used for signal analysis are also listed.

Denoising function	Parameters	Parameter settings used
EMD/EEMD/CEEMDAN	S-Number: The sifting number controls the stopping criterion in the sifting process of EMD. A higher S-number results in more accurate IMFs but increases computational time and may overfit noise. A default S-Number of 4–5 is typical but may vary from 3–10	S-Number – 4
	Gaussian White Noise Strength: Controls the amplitude of Gaussian noise added to the signal. Higher white noise helps reduce mode mixing, although excessive white noise may distort IMFs and reduce signal fidelity. As a proportion of signal amplitude, the strength may typically be set to 0.1–0.4	Gaussian White Noise Strength – 0.2 relative to standard deviation of input signal
	Ensemble Size: Represents the number of noise-added trials. A larger (100+) ensemble size provides more stable and accurate decomposition at the cost of computational time	Ensemble Size – 250
EWT	Number of modes: More modes yield finer frequency resolution, fewer modes achieve more smoothing.	Number of modes – 5
	Gaussian filter width (σ): Larger σ suppresses minor peaks helping avoid spurious boundaries. Smaller σ captures finer spectral variation. (Default σ = 5)	Gaussian filter width – 5
	Filter length: Longer filters improve smoothing but reduce spectral locality (Default filter length is 10)	Filter length – 10
	Boundary transition width (τn): Controls overlap between adjacent filters. Smaller τ yields sharper transitions but may increase spectral leakage	Boundary transition width (τn) – 1
	Normalisation method: Whether the spectrum is normalised before maxima detection. Common options include L1, L2 norm max normalisation. Affects relative peak strength and thus influences segmentation	Normalisation method: LocMaxMin was used to detect wavelets
VMD	Number of modes (K): Underestimation may cause mode mixing. Overestimation may lead to noise or redundant modes	Number of modes (K) – 10
	Initialisation method (ωk₀): Initial estimate of mode centre frequencies. Influences speed and solution quality	Initialisation method – 1
	Balancing parameter (α): Inversely proportional to bandwidth of decomposed modes	Balancing parameter (α) – 2000
	DC parameter: Includes zero-frequency mode. Helps isolate baseline trends and is useful for zero-mean signals	DC parameter – True

Further theoretical discussion of EMD-based functions in the context of the ERG is provided by Sarossy et al.¹⁰.

Empirical wavelet transform (EWT)

The EWT was described by Gilles and performs a wavelet decomposition of a signal into a fixed number of components based on the wavelet transform¹⁷. The exact wavelet ensemble is empirically derived from the signal itself. Refer to Supplementary Material for further details²⁸.

A wavelet transform decomposes a signal into multiple frequency bands using basis functions (wavelets). The EWT computes the Fourier spectrum of the signal and performs an adaptive discrete Meyer wavelet transform rather than using predefined wavelets like traditional wavelet transforms¹⁷. This allows effective separation of modes into signals with non-stationary frequency counts. Local maxima and minima of the spectra are arranged in order of magnitude. A pre-specified number of frequencies are retained with the aim of eliminating noise from the removed frequencies. The inverse Fourier transform is used for each component to recover a signal in the time domain.

An illustration of how the EWT constructs wavelets is seen in Fig. 2.

Fig. 2.

Empirical Wavelet Transform decomposing a signal into 10 wavelets. The top panel shows the power spectrum derived from the original signal. Local maxima and minima are identified. From these maxima and minima, the ensemble of wavelets is constructed in the frequency domain – shown in the lower panel. The first wavelet filter (wavelet 1 in this example) is a high-pass filter and is what is used in this algorithm to remove the ultra-low frequency noise.

Variational mode decomposition (VMD)

The VMD extends the theory of the EMD with a more robust basis and obtains the decomposition without iteration. It separates the signal into modes which have specific sparsity characteristics. The VMD determines frequency ranges for each oscillatory mode by minimising the variational model, ensuring each mode is narrowband. As a non-recursive method, the VMD avoids errors that can occur with EMD-based methods during the calculation of recursion. It also avoids the strict filter bank boundaries of EWT¹⁸. VMD theory is discussed further in Supplementary material²⁹.

Boundary functions

Boundary functions or conditions refer to how the algorithm handles the edges of the signal. Decisions as to how to handle the data at the boundary of the signal may attenuate, distort or falsely smooth the 20ms post-stimulus region and affect the signal latency or amplitude.

The EMD-based functions’ reliance on local extrema to construct upper and lower envelopes can make the boundary signal less reliable where the extrema are sparsely distributed, such as in the PERG signal³⁰. Methods used to overcome this issue have included mirroring the extrema close to the edges or the addition of a sinusoidal wave for three periods beyond the data span.

The VMD implicitly treats the signal as periodic, connecting the end of the signal to the beginning. This may introduce discontinuities near the start or end, particularly for the transient PERG, as it is not a periodic signal.

The EWT uses mirror extension at the ends to avoid discontinuities in the transformed domain. This can potentially introduce reflected artefacts near the signal boundaries³⁰.

Methods

This research was conducted in accordance with the Declaration of Helsinki and was approved by the Human Research Ethics Committee of the Royal Victorian Eye and Ear Hospital. Participants were recruited from an ophthalmology practice and all gave written informed consent.

Electroretinography

Pattern electroretinograms were recorded using an Espion E2 system (Diagnosys LLC; Lowell, MA) and followed the extended ISCEV protocols with the exception of the automatic rejection threshold¹³. Dawson-Trick-Litzkow (DTL) electrodes were positioned in contact with inferior bulbar conjunctiva. Participants’ skin was cleaned and conductive paste applied adjacent to the lateral canthus, where reference gold cup electrodes were placed, and a common ground electrode was placed on the central forehead, aiming for impedance < 5 kΩ.

A cathode ray tube (CRT) monitor displayed a checkerboard pattern with a check-width of 0.8° when viewed from 1 m. Consistent ambient lighting was maintained throughout recordings. Pattern reversal rate was four times per second (2 Hz).

An automatic rejection system was employed to remove any sweeps with significant artefact or baseline drift, with a threshold of 200 µV, which is higher than the ISCEV-suggested 100 µV. This threshold was selected to determine if denoising algorithms could provide good reliability despite introducing more noise (baseline drift). A total of 200 non-rejected sweeps were collected for each recording¹³.

After a short break of approximately 2–5 min, a second recording was performed to enable comparison and reliability measurements of denoising algorithms within a recording session.

Signal analysis

All sweeps were exported from the Espion E2 system and signal processing was undertaken in R and Python. Data was stored in comma-separated value (CSV) format as a de-identified database. PERG tracings were manually inspected by an expert electrophysiologist and poor-quality traces were excluded. Denoising consisted of decomposing each sweep and removing the lowest frequency component, which typically corresponds to baseline drift and low-frequency noise. The PERG was then extracted as the ensemble average of the denoised sweeps, enhancing the signal-to-noise ratio.

EMD/EEMD/CEEMDAN

The R package Rlibeemd³¹ was used to decompose the sweeps with EMD, EEMD and CEEMDAN. Ten IMFs were extracted, including the residue. Parameters consisted of an S number of 4 and for EEMD and CEEMDAN the ensemble size was set to 250 with a Gaussian White Noise strength of 0.2 relative to the standard deviation of the input signal.

EWT

The Python package ewtpy³² was used to decompose the signal through EWT. Five levels were specified. The scalespace method was used for boundary detection, and no trend extraction or regularisation took place. The Length Filter was set to 10 and the Sigma Filter was set to 5. LocMaxMin function was used to detect wavelets.

VMD

The R package VMDecomp³³ was used to decompose the signal via VMD. Ten modes were specified. The following parameters were used: α = 2000 (controlling the bandwidth of decomposed modes), τ = 0 (indicating no bias towards smoothness), the initialisation method was set to 1 and the DC parameter was set to true.

Statistics

The P50 peak amplitude was the main measure used to assess reliability. P50 was considered the highest amplitude point occurring between 30 and 85ms on the PERG. Its amplitude was set as the difference in raw amplitude between it and the N35, the lowest amplitude point occurring prior to the P50 peak. Change in amplitude of N35 peak and P50 implicit time were also assessed to ensure waveform morphology was maintained.

The two statistics calculated to assess test-retest repeatability were Intraclass Correlation Coefficient (ICC) and Coefficient of Repeatability (CoR).

ICC is a statistic which uses a pooled mean and standard deviation to assess the extent of agreement between two groups of data, with values closer to 1 indicating higher reliability^34,35. ICC data was calculated selecting a single eye from each participant at random to control for correlation between eyes.

CoR was also calculated to assess reliability. CoR represents the maximum expected difference between two measurements with 95% confidence³⁶. In the PERG, the CoR represents the number of µV absolute difference that contains 95% of test-retest pairs. It is measured in the same units as the primary measure – in this case, µV – and is directly related to the difference between serial measurements in a clinical scenario that can be considered significant. In other words, a difference of more than the CoR has less than 5% chance of occurring by chance. CoR data was calculated including both eyes (except those excluded) from each participant.

Bootstrap replication (n = 1000 resamples) with replacement was used to determine significance of differences in CoR and ICC. For each method, the ICC and CoR compared the raw signal which had no extra denoising beyond what is included in the manufacturer’s software³⁷. There was no dedicated control group in this project because participants’ unfiltered data was used as a within-subject control.

Results

A total of 70 eyes of 38 patients were tested. Age of participants ranged from 19 to 83 (µ = 66.4) years with a variety of healthy and co-morbid eyes. Two eyes were excluded (grounded) due to poor signal or high impedance at the time of recording. Four eyes were excluded after manual inspection due to poor waveform, likely unrelated to a disease process. Characteristics of the study cohort are summarised in Table 2.

Table 2.

Summary of cohort characteristics including total participants, eyes, sex, age, exclusions and eyes with glaucoma.

Participants	38
Female participants (%)	13 (34%)
Male participants (%)	25 (66%)
Total eyes	70
Female eyes	26 (37%)
Male eyes	44 (63%)
Mean age, range (years)	66.4, 19–83
Eyes excluded at time of recording	2
Eyes excluded post-hoc	4
Eyes with glaucoma	6

Raw data comparing the two PERGs P50 peak amplitude had a CoR of 2.12 µV, which was significantly improved with EWT (1.69 µV, p = 0.02) and CEEMDAN (1.85 µV, 0.03), and insignificantly with EEMD (1.99 µV, p = 0.19). CoR was higher for EMD (3.11 µV, p = 0.04) and VMD (2.26 µV, p = 0.04).

ICC significantly improved from raw data (0.79) with EWT (0.84, p = 0.049). However, all other filters failed to significantly improve ICC including EMD (0.75), EEMD (0.79), CEEMDAN (0.78) and VMD (0.73).

Results are summarised in Table 3. The effect of improved reliability from the various denoising algorithms is visually demonstrated in Fig. 3.

Table 3.

Results table comparing P50 amplitude (primary outcome) reliability measures of adaptive denoising techniques with Raw data.

	P50 amplitude CoR (µV)	P50 amplitude ICC
Unfiltered	2.12	0.79
EMD	3.11*	0.75
EEMD	1.99	0.79
CEEMDAN	1.85*	0.78
VMD	2.26*	0.73
EWT	1.69*	0.84 *

* Indicates significance at p < 0.05 tested with bootstrap resampling, n = 1000.

Fig. 3.

Test vs. retest PERG trace from one participant’s eye with EMD, EEMD, CEEMDAN, VMD and EWT denoising algorithms compared to original data.

Secondary outcome measures

Changes in N35 amplitude and P50 implicit time were also calculated and are summarised in Tables 4 and 5. Compared with unfiltered data, EEMD, CEEMDAN and VMD demonstrated statistically significant increases in N35 amplitude (p < 0.05), whereas EMD and EWT showed no significant difference.

Table 4.

Results table comparing the effect of filters on change in amplitude of N35 peak compared to unfiltered data. * Indicates significance at p < 0.05.

	N35 amplitude difference (µV)	P-value
Unfiltered vs. EMD	0.07	0.24
Unfiltered vs. EEMD	0.12	0.03*
Unfiltered vs. CEEMDAN	0.27	0.001*
Unfiltered vs. VMD	0.24	0.0001*
Unfiltered vs. EWT	0.03	0.63

Table 5.

Summary of change in P50 implicit time with each adaptive denoising filter.

	P50 implicit time (ms)	Change P50 implicit time (ms)
Unfiltered	48.0
EMD	53.4	5.3*
EEMD	53.9	5.9*
CEEMDAN	52.8	4.8*
VMD	46.4	-1.6
EWT	53.8	5.8*

* Indicates significance at p < 0.05.

P50 implicit time was affected by all denoising algorithms. All filters except VMD resulted in a significant lengthening of P50 implicit time compared with unfiltered data, ranging from 4.8 to 5.9 ms longer compared to unfiltered data average peak at 48.0 ms (p < 0.05). VMD was associated with a small, non-significant reduction in P50 implicit time (1.6 ms, p = 0.07). These findings indicate that while denoising improves overall signal reliability, it may also introduce systematic shifts in waveform timing, which should be considered when interpreting absolute implicit time values.

Collectively, these results demonstrate that adaptive denoising algorithms, particularly EWT, enhance the reliability of the PERG without substantially distorting the waveform, although P50 implicit time may be affected.

Discussion

Most of our methods of adaptive filtering improved the reliability of the PERG signal, as measured primarily with P50 amplitude. Similar to what we found in earlier work with the photopic negative response of the ERG, intrusive EMG, electrochemical and mechanical movements contaminate the response and make identification of the peaks more difficult and less reliable. Tiny lid movements can have a large impact on the DTL fibre electrodes and adaptive filtering shows promise in removing the contamination. Unlike the VMD and the variants of the EMD, the EWT is essentially a linear high-pass filter but with the corner frequency tuned to each sweep.

Despite this improvement in the reliability of the P50 of the pattern-reversal PERG, the coefficient of repeatability remains quite high relative to the amplitude of the waveform. This means that even with the enhanced reliability from adaptive filtering, a change in PERG amplitude between visits needs to be around 40% to fall outside the 95% confidence limit. This could still represent a substantial amount of disease progression in a disease such as glaucoma before detection confidence can be reached. There is thus still much room for further improvement in the filtering algorithms to improve clinical practice. With such improvements, the PERG may have increased utility in the investigation of retinal dystrophy, optic atrophy and ocular toxicity from agents such as quinine^38,39. The PERG has also been proposed for monitoring of glaucoma, papilledema and ocular inflammatory disorders such as Birdshot chorioretinopathy^40,41.

Implementation of the algorithms is relatively straightforward. Open source libraries written in C exist for all of the algorithms which could be incorporated into commercial software. Alternatively, the algorithms are not difficult to refactor from that code into the developer’s preferred environment. With the increasing availability and utility of machine learning and GPU equipped computers, the next step from these types of adaptive filters may be implementation through deep learning architecture⁴³.

The cohort of eyes used in this study consisted of a mixture of eyes with and without glaucoma and of those with glaucoma, there was a range of severities. The method was designed to emulate the sort of mix of eyes that would be tested in a typical laboratory. The number of eyes in this study was too small to perform meaningful subgroup analysis of the performance of the test in the presence or absence of disease although we do anticipate to do this in the future when more data has been collected.

Further work will also investigate optimisation of more of the tuneable parameters such as the number of wavelets/IMFs and the length of the pre-stimulus data collection. Another interesting extension of this work would be to examine the denoising strategies on the steady-state PERG. Our expectation is that the techniques will have less utility in this application because the ultra-low frequency noise is removed as part of the Fourier transform analysis that is routinely done on this sort of signal⁴⁴. There may yet be some benefit in the removal of the drift and wander from the signal there.

Finally, although most test subjects cope with testing protocol, some are very sensitive to the idea of electrophysiological testing or are unable to tolerate the DTL electrodes without large artefacts. Our denoising, by relaxing the blink artefact threshold, may make testing possible in such patients or may reduce the number of sweeps required to get useful results in that context. Similarly, it might help in more difficult cases such as those with nystagmus, tremors or other neurological issues and paediatric patients⁴⁵.

Limitations

While our study was limited to adult participants, the utility of these denoising techniques in paediatric populations, where obtaining a quality PERG recording is often more challenging, warrants further investigation. The number of eyes was also relatively small. The time between test and retest was fairly short and did not allow for physiological diurnal variation in the response. As such, our test-retest reliability figures may overestimate what can be achieved in a real-world setting. The adaptive filters slightly alter the N35 and P50 amplitude and implicit time. Lab-based normative values may need recalibration with filters applied. Nonetheless, the overall wave morphology is maintained with denoising filters applied.

The analysis was performed with a mix of healthy and comorbid eyes. Further work could investigate differences in the reliability of PERG traces with denoising between particular disease groups (e.g. advanced glaucoma) and healthy eyes. In such cases, the amplitude of the signal could approach that of signal noise, which could potentially limit adaptive denoising methods. It was felt that there was insufficient statistical power to separate participants into disease groups. In this study, we did not assess how these denoising methods impact diagnostic sensitivity or specificity. A logical next step is to conduct a prospective clinical study comparing diagnostic outcomes in filtered vs. unfiltered PERGs across disease groups.

Conclusion

Adaptive denoising filters were applied to transient PERGs to assess their ability to improve signal reliability. Among the five tested methods, EWT was the best-performing filter which significantly improved the reliability of P50 amplitude as measured with ICC and CoR compared to unfiltered data. These results indicate that adaptive denoising can meaningfully reduce the influence of low-frequency noise while preserving the underlying waveform morphology. With further validation, such adaptive denoising methods could be integrated into clinical PERG recording systems for improved signal clarity, reduced testing burden and enable more confident interpretation of results – both in real-time and through post hoc analysis of noisy recordings. Future studies should explore the optimisation of filter parameters, boundary conditions and performance in specific patient populations such as paediatric groups, as well as investigate the feasibility of implementing these algorithms directly into acquisition software for real-time noise suppression.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Zachary Angus Conceptualisation, Methodology, Validation, Investigation, Writing - reviewing & editing. Alexander Sarossy Conceptualisation, Investigation, Formal analysis, Data curation, Writing - reviewing & editing. Puneet Parihar Methodology, Validation, Investigation, Writing - reviewing & editing. Marc Sarossy Conceptualisation, Methodology, Investigation, Validation, Formal analysis, Supervision, Writing - reviewing & editing.

Funding

This research received no funding or grants from any funding agency in the public, commercial or not-for-profit sector.

Data availability

Data is available upon request. Please email Zachary Angus at [zacangus87@gmail.com](mailto: zacangus87@gmail.com) to request data.

Competing interests

The authors declare no competing interests.

Ethics approval

This research was approved by the Royal Victorian Eye and Ear Hospital HREC.