Optimizing retinal prosthesis stimulation: Enhancing the avoidance of unintended axonal activation

1. Introduction

Age-related macular degeneration (AMD) and retinitis pigmentosa (RP) are two retinal degenerative diseases that significantly contribute to global vision loss and blindness (Bourne et al., 2013). Over 8% of individuals aged 45 and older exhibit signs of macular degeneration, while RP affects approximately 1 in every 3,000 individuals (Ferrari et al., 2011; Wong et al., 2014). RP and AMD are the primary etiological causes of irreversible visual impairment, which are guided through different pathogenic mechanisms. RP is mostly a monogenic disease where the initial loss of rod photoreceptors leads to further cone dysfunction that is caused by oxidative stress and microglial response (Jadeja & Martin, 2021; Vingolo et al., 2022). On the other hand, AMD is a multifactorial, age-related disease that is characterized by impairment of the retinal pigment epithelium (RPE), the accumulation of drusen, chronic inflammation, and disruption of the complement cascade (Ochoa Hernández et al., 2025; Wang et al., 2025). Even though the result of both is the destruction of photoreceptors, RP is predominantly a result of inherited mutational defects, and AMD is an outcome of a complex interaction between genetic defects, the aging process, as well as exogenous environmental insults, which damage the viability of RPE and destabilize retinal homeostasis. As a result of these conditions, the retina’s ability to convert light into visual signals is diminished, as photoreceptor cells deteriorate.

Current therapies for vision loss have limited effectiveness in halting disease progression or restoring sight. Recent advancements like gene therapy and cell transplantation show promise (Scholl et al., 2016), but both face significant challenges. Gene therapy struggles due to insufficient knowledge of treatable mutations, while cell transplantation is hindered by issues related to the functionality and connectivity of the transplanted cells (Hartong et al., 2006; Tong, Meffin, et al., 2020). Additionally, there are challenges in gene therapy and cell transplantation for retinal diseases due to surgical procedures. Gene therapy requires accurate subretinal injection, which is associated with the risk of retinal detachment and inflammation, requiring high-technology surgical skills and live images (Davis, 2018; Davis et al., 2019). Cell transplantation becomes hindered by the trauma of needle insertion, which causes inflammation, cell death, and poor integration of the cells in host retinal circuitry (Coco-Martin et al., 2021; Ozaki et al., 2025). Both modalities involve extremely specific surgical procedures to maintain the integrity of the cells and provide working results, and the success of both modalities depends on the reduction of potential risks of the procedure.

Over the past twenty years, retinal prostheses that electrically activate the remaining retinal neurons have given hope for vision restoration in individuals with visual impairments (Goetz & Palanker, 2016; Weiland et al., 2016). There are unique benefits of retinal implants in the case of advanced retinal degeneration: it does not need the underlying genetic defect (as with gene therapy), has immediate functional value and no risk of rejection (as with cell transplantation), and can be surgically reversed (Mills et al., 2017; Palanker et al., 2025). They do not induce regeneration of retinal tissue, but their extensive application and standardized procedures make them useful in cases where biological therapies are ineligible. There are three distinct types of these prosthetic devices, distinguished by the location of their electrode arrays. Epiretinal devices are composed of electrode arrays positioned on the outer surface of the retina, near the layer of retinal ganglion cells (RGCs). Subretinal implants are situated beneath the retina, adjacent to the damaged photoreceptor layer. Suprachoroidal implants, meanwhile, are located in the space between the choroid and the sclera. Patients have been implanted with various devices, including the epiretinal Argus II by Second Sight (Stronks & Dagnelie, 2014), the suprachoroidal devices by Bionic Vision Australia (Ayton et al., 2014), the subretinal Alpha AMS by Retina Implant AG (Stingl et al., 2013), and the epiretinal IRIS II (Tong, Meffin, et al., 2020) and subretinal PRIMA by Pixium Vision (Tong, Meffin, et al., 2020). The basic principle of epiretinal stimulation is that an external electric field is produced, and this depolarizes the membrane of RGCs. When adequate depolarization occurs, this triggers the voltage-gated sodium channels, and thus an action potential is triggered. The site of stimulation is very important; although the axon initial segment (AIS) contains a high concentration of sodium channels, the long structure of the axon also makes it very vulnerable to the exerted electric field (Corna et al., 2024; Paknahad et al., 2020). This type of inherent biophysical property leads to the occurrence of unintended axonal activation, which significantly impairs spatial resolution by generating elongated phosphenes.

Clinical studies indicated that patients have reported good clinical outcomes, including improved light perception, object distinction, and letter identification (Ayton et al., 2014; Humayun et al., 2012; Stingl et al., 2017; Zrenner et al., 2011). Nevertheless, the visual resolution obtained from existing devices is very limited. Problems like face recognition, which are crucial, are still not solved (Ayton et al., 2020; Tong, Meffin, et al., 2020). Ayton et al. (2020) recently provided detailed information about the up-to-date advancements and the clinical trials of retinal prosthetics (Ayton et al., 2020). On the Snellen acuity scale, 20/20 indicates normal vision, and 20/200 is the legal blindness threshold. This scale is the standard method for measuring visual acuity. The best vision that current devices can provide, like the Alpha-AMS (20/546), the Argus II (20/1260), and the Bionic Vision Australia suprachoroidal devices (20/8397), is still considered legally blind based on clinical trials (Ayton et al., 2020).

The accuracy of neural activation is the key determinant of the spatial resolution and functional success of these devices. The ability of each electrode to activate a small group of retinal cells is determined by various factors, such as the design and the placement of the electrodes and the stimulation parameters. Cellular resolution targeting with high-density microelectrode arrays and spatial control, as well as RGC subtype selectivity with smaller electrodes, are achievable (Tsai et al., 2017). Sophisticated electrode designs, such as hexapolar arrangements, enhance spatial confinement, while novel materials reduce the power required to activate neurons (Chung et al., 2024; Duvan et al., 2024; Samba et al., 2015). Finally, discrete neuronal activation in the retina is obtained by optimized electrode design, customized stimulation strategies, and proper attention to the anatomy of the retina.

Current retinal devices have undergone several attempts to improve their efficiency. However, a number of issues continue to impede the spatial resolution of these devices. The unwanted passing axon fiber activation is a serious problem with epiretinal stimulation (Behrend et al., 2011; Beyeler et al., 2019; Nanduri et al., 2012). A single stimulating electrode can cause a patient to perceive an elongated phosphene that lines up with the RGC axon pathway, according to clinical investigations involving patients who had epiretinal implants (Nanduri et al., 2012). Different retinal electrical stimulation techniques have been suggested to solve this issue (Beyeler et al., 2019; Chang et al., 2019; Corna et al., 2024; Esler et al., 2018; Freeman et al., 2010; Jensen et al., 2005; Rattay & Resatz, 2004; Tong, Hejazi, et al., 2020; Vilkhu et al., 2021; Weitz et al., 2015; Werginz et al., 2022). Scientists have discovered that placing long rectangular electrodes parallel to passing axon fibers (Rattay & Resatz, 2004; Tong, Hejazi, et al., 2020) or using low-frequency sinusoidal stimulation (Corna et al., 2024; Freeman et al., 2010) or long pulse durations (25 ms) (Weitz et al., 2015) or using very short pulses (Chang et al., 2019; Tong, Hejazi, et al., 2020) can specifically achieve focal activation. While these studies show promising results, implementing these techniques in retinal prostheses faces significant challenges. The parallel placement of long rectangular electrodes along axonal pathways faces significant design constraints. Firstly, the inherent spatial constraints of the retina necessitate smaller electrodes to achieve sufficient density for replicating physiological visual processing, conflicting with the geometric requirements of this design. Secondly, achieving and maintaining precise parallel alignment with axons in vivo remains technically difficult. Additionally, both low-frequency sinusoidal stimulation and long-pulse stimulation risk exceeding safe charge injection limits. Long-pulse stimulation faces further challenges, including desensitization of RGCs during repetitive pulses, which diminishes effectiveness in standard stimulation protocols (Freeman & Fried, 2011; Jensen & Rizzo, 2007). Conversely, direct RGC stimulation with ultrashort pulses (≤120 μs) preferentially activates soma while largely sparing axons. This approach triples the current threshold required for axonal activation as pulse duration decreases (Beyeler, 2019; Chang et al., 2019; Tong, Hejazi, et al., 2020; Werginz et al., 2022), a phenomenon consistently observed in both wild-type (WT) and retinal degeneration (RD) models (Chang et al., 2019).

In this study, we developed a computational model of OFF RGC to systematically investigate how pulse duration, waveform, and electrode-to-RGC distance influence cellular responses during epiretinal stimulation. Our goal was to determine optimal stimulation parameters for selectively activating RGC while avoiding unwanted passing axon fiber activation. By quantifying activation thresholds across varying pulse durations, waveforms, and electrode-cell distances, we demonstrated that shorter pulse durations and specific waveform configurations critically enhance stimulation selectivity. This selectivity ratio improved significantly with greater electrode-to-RGC distances, highlighting the interplay between spatial targeting and pulse parameter optimization in avoiding unwanted passing axon fiber activation.

2. Methods

2.1 Model preparation

In our computational study, we employed realistic multicompartment models of rabbit OFF RGC (Fig. 1a) obtained from experiments and sourced from (Guo et al., 2016) to investigate their response to epiretinal electrical stimulation. This morphology was imported into the COMSOL Multiphysics finite element simulation software as equivalent-cylinder cable models, where 1D edge segments represent the RGC. The RGC is composed of five segments: cell body (soma), axon hillock (AH), AIS, dendrites, and axon.

Close

Fig. 1.

(a) The geometry of the realistic rabbit OFF RGC obtained from (Guo et al., 2016), (b) AIS activation settings, and (c) axonal activation settings. Red circle is the stimulating electrode.

Export to PPT

The OFF RGC cell was placed in a semi-ellipsoid shape to mimic an extracellular environment. To deliver the extracellular electric stimulation, we used the monopolar configuration, which involved placing a disk electrode with a 200 μm diameter on the upper surface (epiretinal side) and designating the lower boundary of the semi-ellipsoid as the ground.

We examined two distinct electrode placement configurations for RGC electrical stimulation: (1) placement over the AIS, the sodium channel-dense proximal region of the RGC (Fig. 1b); and (2) placement over the distal axon (Fig. 1c). Focal activation (AIS activation preceding axon activation) occurs when the AIS activation threshold is lower than that of the axon. Conversely, passing axon fiber activation (axon activation preceding AIS activation) occurs when the axon activation threshold is lower. Therefore, focal activation is achieved when the current threshold for AIS activation is lower than that for axon activation. We varied the z-distance between the RGC and the electrode across three distances: 120 μm, 240 μm, and 480 μm. We tested four types of rectangular pulse waveforms: monophasic cathodic (MC), monophasic anodic (MA), biphasic symmetric charge-balanced cathodic first (BC), and biphasic symmetric charge-balanced anodic first current pulses (BA) (Fig. 2). These pulse waveforms were evaluated using two durations: 50 μs and 0.5 ms. We determined the minimum current threshold required to activate the RGCs for both AIS and axon stimulation scenarios, considering the three electrode distances and the four pulse waveforms associated with the two durations.

Close

Fig. 2.

The four pulse waveforms used in this study for stimulating the RGC.

Export to PPT

3. Results

3.1 Spatiotemporal activation patterns of RGCs

In our RGC model, action potential initiation depended critically on electrode placement: when positioned above the AIS, initiation occurred in the AIS (1.5 ms post-stimulus) with bidirectional propagation toward soma/dendrites and distal axon (Fig. 3); when positioned above the axon, initiation occurred in the axon with bidirectional propagation toward soma/dendrites and distal optic nerve (Fig. 4). This site-specific dependency was confirmed by compartmental traces showing resting potential until stimulation and electrode-determined initiation sites (Fig. 5).

Close

Fig. 3.

Spatiotemporal activation of the OFF RGC when the stimulating electrode was located above the AIS. A BC waveform, 0.5 ms duration, and 120 µm RGC-Electrode distance was used.

Export to PPT

Close

Fig. 4.

Spatiotemporal activation of the OFF RGC when the stimulating electrode was located above the axon. BC waveform, 0.5 ms duration, and 120 µm RGC-Electrode distance was used.

Export to PPT

Close

Fig. 5.

Action potential traces from various compartments of the OFF RGC obtained when (a) the stimulating electrode was located above the AIS, and (b) the stimulating electrode was located above the axon. Both are taken when a BC waveform, 0.5 ms duration, and 120 µm RGC-Electrode distance was used.

Export to PPT

3.2 The influence of pulse duration on RGC activation

Fig. 6 presents data on the stimulation thresholds of OFF RGC using a biphasic symmetric charge-BC pulse waveform, which is the most commonly used pulse waveform in retinal implants, at two different pulse durations: 0.5 ms and 50 µs. The distance between the electrode and the RGC was 120 µm. For both pulse durations, the axon requires a lower current amplitude compared to the AIS. This means that with the biphasic symmetric charge-balanced cathodic first (BC) waveform, avoiding passing axon fiber activation is not achieved at a distance of 120 μm. This finding indicates that adjustments in the stimulation parameters or different waveform strategies may be necessary to achieve focal activation.

Close

Fig. 6.

Current threshold for RGC activation using a biphasic cathodic (BC) waveform at two pulse durations. The vertical axis (Y-axis) shows the current threshold in microamperes (µA), and the horizontal axis (X-axis) shows the pulse duration (50 µs and 0.5 ms).

Export to PPT

3.3 The influence of pulse waveform on RGC activation

Fig. 7 illustrates the RGC activation thresholds for four different current pulse waveforms at two pulse durations and at an electrode-to-RGC distance of 120 μm. At 0.5 ms, all waveforms yielded lower thresholds for axon activation than for AIS activation, indicating a high risk of unwanted passing axon fiber activation. Notably, the MA and BA waveforms at 0.5 ms produced nearly identical AIS and axon thresholds, indicating less effective focal activation. In contrast, at 50 μs, the BA waveform exhibited a substantially higher threshold for axonal activation relative to AIS activation, making it the most effective waveform for preferentially targeting RGC AIS while avoiding unwanted passing axon fiber activation. Under this short-pulse condition, MC and BC still showed lower axon thresholds than AIS thresholds, and the MA waveform remained non-selective, yielding nearly identical AIS and axon thresholds. Overall, the BA waveform at 50 μs provided the greatest AIS-axon selectivity, with the axon activation threshold approximately 1.6-fold higher than the AIS threshold. This pronounced threshold separation suggests that BA pulses at short durations can effectively activate RGC while largely avoiding unwanted passing axon fiber activation. These findings underscore the importance of optimizing both waveform shape and pulse duration to achieve focal neuronal activation, particularly in retinal stimulation applications where avoiding unintended passing axon fiber activation.

Close

Fig. 7.

Comparison of current thresholds (Y-axis) for RGC activation using the four pulse waveforms with the two pulse durations (X-axis): (a) 0.5 ms and (b) 50 µs.

Export to PPT

3.4 The influence of electrode-RGC distance on RGC activation

Fig. 8 delineates the interplay between electrode-RGC distance, pulse duration, and pulse waveform in modulating OFF RGC activation thresholds. Increasing electrode-RGC distance universally elevated current thresholds, yet biphasic anodic-first (BA) and cathodic-first (BC) waveforms demonstrated superior efficacy in avoiding unwanted passing axon fiber activation at larger separations. At 120 µm, only BA (50 µs) achieved focal activation, with axonal thresholds 1.6-fold higher than AIS thresholds. At 240 µm, BA maintained selectivity for both durations (axonal thresholds 1.44-fold and 2.22-fold higher for 0.5 ms and 50 µs, respectively), while BC (50 µs) exhibited moderate selectivity (1.27-fold). At 480 µm, BA and BC further improved, with BA axonal thresholds reaching 2.6-fold (50 µs) and BC axonal thresholds 2.13-fold (50 µs) higher than AIS thresholds. Monophasic cathodic (MC) waveforms posed persistent risks of axonal activation, particularly at longer durations, while monophasic anodic (MA) waveforms showed negligible AIS-axon discrimination across all conditions. These findings demonstrate BA as the most robust waveform for focal activation across distances, with BC becoming viable at 240 µm or greater. Strategic parameter selection, prioritizing BA or BC at extended electrode separations, thus enables focal activation while minimizing unwanted passing axon fiber activation, critical for optimizing retinal implant efficacy.

Close

Fig. 8.

Comparison of current thresholds (Y-axis) in (µA) for RGC activation. The effect of four pulse waveforms is shown at two pulse durations (X-axis) (0.5 ms, left; 50 µs, right) and three distances from the stimulating electrode: (a) 120 µm, (b) 240 µm, and (c) 480 µm..

Export to PPT

4. Discussion

This study employed computational modeling of morphologically realistic RGCs) to systematically evaluate epiretinal stimulation strategies that optimize focal activation while mitigating unwanted passing axon fiber activation. Across the full matrix of pulse widths (0.5 ms, 50 µs), waveforms (monophasic cathodic (MC), monophasic anodic (MA), biphasic charge-balanced cathodic-first (BC), and biphasic charge-balanced anodic-first (BA)), and electrode–RGC separations (120, 240, 480 µm), focal activation was achieved exclusively with BA pulses, and BC pulses became viable at separations of 240 µm or greater. Our findings demonstrate that parametric optimization of pulse duration, waveform, and electrode-RGC distance enables focal RGC stimulation. Achieving focal activation required optimizing the interplay relationship between these three factors. While symmetric balanced biphasic cathodic-first (BC) waveforms failed to achieve focal activation at 120 µm (Fig. 6), biphasic charge-balanced anodic-first (BA) waveforms enabled robust discrimination at the same distance with 50 µs pulses, exhibiting a 1.6-fold higher axonal threshold than AIS thresholds (Fig. 7). Increasing electrode-RGC separation enhanced BA’s efficacy, yielding more than 2.5-fold axonal threshold elevation at 480 µm (50 µs). BC waveforms became viable at 240 µm or greater, achieving 2.13-fold higher axonal threshold than AIS thresholds at 480 µm (50 µs). Conversely, monophasic cathodic (MC) waveforms consistently activated axons first across all parameters, with axonal and AIS thresholds converging at larger distances. Monophasic anodic (MA) waveforms showed negligible discrimination. Notably, MC exhibited the lowest absolute current thresholds across all conditions despite its poor selectivity.

Stimulating RGCs while minimizing inadvertent activation of passing axon fibers remains a critical challenge for effective visual prostheses. Research demonstrates two primary strategies to address this challenge (Beyeler et al., 2019; Chang et al., 2019; Esler et al., 2018; Freeman et al., 2010; Jensen et al., 2005; Rattay & Resatz, 2004; Tong, Hejazi, et al., 2020; Vilkhu et al., 2021; Weitz et al., 2015; Werginz et al., 2022): electrode geometry/orientation optimization (Esler et al., 2018; Rattay & Resatz, 2004; Vilkhu et al., 2021) and pulse waveform/duration manipulation (Beyeler et al., 2019; Chang et al., 2019; Werginz et al., 2022), each with distinct mechanisms and implications. Computational and experimental studies strongly support the efficacy of using elongated electrodes aligned parallel to axon fibers (Esler et al., 2018; Rattay & Resatz, 2004; Tong, Hejazi, et al., 2020). Specifically, Rattay and Resatz’s pioneering work demonstrated that such electrodes (e.g., slot electrodes) increase the activation threshold for distant RGCs by approximately 2.3 times compared to local targets near the electrode (Rattay & Resatz, 2004). This principle was confirmed by Esler et al. (2018), who showed using multielectrode arrays (100 µm diameter, 200 µm spacing), showing improved RGC focal activation with simultaneous stimulation of multiple parallel-aligned disc electrodes. Focal activation increased from one to four electrodes, with four electrodes achieving 78% AIS activation before passing axons fired (Esler et al., 2018). Tong et al. (2020) study, using electrode geometries similar to Rattay and Resatz study, found that parallel alignment to axons enabled selective RGC stimulation, with distant thresholds up to 1.64-fold higher using 33 µs anodic-first biphasic pulses. Perpendicular alignment showed minimal threshold differences (Tong, Hejazi, et al., 2020). A recent experimental study found that paired electrodes straddling an axon and delivering equal-and-opposite currents boost soma-over-axon selectivity by 1.44-fold higher compared with single-electrode stimulation (Vilkhu et al., 2021). Despite promising results in avoiding unwanted axonal activation, fabrication of these narrow, long electrodes for chronic use remains untested, and their high predicted charge densities raise safety concerns. Modeling limitations, such as neglecting active properties and geometry, and reliance on unrealistic assumptions, impact validity. Additionally, the method’s dependence on unvalidated real-time axon tracking alignment poses a major challenge.

Beyond electrode geometry, shaping the stimulus waveform itself can further sharpen RGC selectivity (Corna et al., 2024; Freeman et al., 2010; Weitz et al., 2015). Experimental work on rabbits showed that low-frequency sinusoids (<25 Hz) activate somas via indirect pathways while avoiding axons. The axonal thresholds were higher by 7–10 fold than soma levels compared to ∼3× with biphasic pulses (Freeman et al., 2010). Weitz et al. (2015) demonstrated that 20 Hz sinusoidal waveforms achieved a remarkable somatic-to-axon activation selectivity ratio of 16:1 (Weitz et al., 2015). Confirming minimal selectivity at higher frequencies, recent work shows axonal activation begins at 60 Hz, where axonal thresholds are only 1.6 times somatic thresholds, and this ratio decreases as frequency increases (Corna et al., 2024). Collectively, these studies suggest that low-frequency sinusoids can achieve targeted somatic activation. However, this method often necessitates higher charge densities, which may exceed safe injection limits (Freeman et al., 2010; Weitz et al., 2015). The considerable variability in reported selectivity ratios also emphasizes the need for further research to refine and standardize the technique.

The manipulation of pulse duration offers a temporal strategy broadly categorized into long and short pulses (Beyeler, 2019; Chang et al., 2019; Jensen et al., 2005; Paknahad et al., 2020; Tong et al., 2019; Tong, Hejazi, et al., 2020; Tong, Meffin, et al., 2020; Weitz et al., 2015; Werginz et al., 2022). The use of long pulses (>25 ms) bypasses direct axon activation and yields up to a threefold improvement in selectivity (Weitz et al., 2015). This method potentially harnesses the retina’s intrinsic processing capabilities. However, significant challenges accompany it (Tong, Hejazi, et al., 2020): network-mediated responses often produce temporally imprecise spike patterns (multiple spikes, longer/unpredictable latencies), hindering the reproduction of natural temporal resolution. The safety of prolonged pulses exceeding a few milliseconds remains largely unverified. Furthermore, increased charge density requirements (>1 mC/cm² for >25 ms pulses) pose risks of exceeding safe charge injection limits and elevate power consumption. The functional integrity of the degenerate retinal network and its suitability for mediating responses are also uncertain, compounded by reports of post-degeneration reorganization and potential links to perceptual phenomena like image fading reported by patients. In contrast, short pulses (<100 µs) promote direct soma stimulation while minimizing both axon activation and network involvement (Tong, Hejazi, et al., 2020). Short pulses (≤120 µs) provide up to threefold more focal activation than long pulses (≥0.5 ms) in both wild-type and RD retinas, though RD may have fewer responsive RGCs (Beyeler, 2019; Chang et al., 2019; Jensen et al., 2005; Tong, Hejazi, et al., 2020; Tong, Meffin, et al., 2020; Werginz et al., 2022). Chang et al. (2019) confirmed that ≤120 µs biphasic pulses, cathodic-first or anodic-first, selectively stimulate somas over axons; anodic-first is most focal despite higher currents. Pulses 0.5 ms or larger activate both axons and somas non-selectively. Comparably, Tong et al. 2020 discovered that shorter pulses, 33 µs and 100 µs, were superior to longer pulses, 0.5 ms, in terms of preventing the activation of passing axon fiber (Tong, Hejazi, et al., 2020). Additionally, the study conducted by Jensen et al. 2005 discovered that short-duration pulsed stimuli less than 100 µs can directly activate RGCs, resulting in a three-fold increase in selectivity for activating somas over axons (Jensen et al., 2005). A recent computational modeling with morphologically realistic multicompartment mouse ON-alpha RGCs model showed that a 10 µs biphasic epiretinal pulse raised axonal activation thresholds (mean ratio: 4.1), whereas a 200 µs pulse yielded uniform thresholds, while monophasic stimuli performed markedly worse (Werginz et al., 2022).

Our findings support the possibility that preferential activation of RGCs (avoiding passing axon fibers activation) can be achieved through varying pulse durations, pulse waveforms, and electrode-RGC distance. Our results demonstrate that short pulses (50 µs) enable focal activation, producing a greater than 2.5-fold difference in current threshold between the AIS and the axon, thereby avoiding unwanted passing axon activation. Long pulses (0.5 ms), in contrast, failed to achieve this preferential activation, consistent with prior studies (Chang et al., 2019; Jensen et al., 2005; Tong, Hejazi, et al., 2020; Werginz et al., 2022). Critically, this effect also depends on pulse waveform: biphasic anodic-first pulses proved most effective at avoiding passing axon activation, despite requiring higher thresholds, aligning with findings in (Chang et al., 2019; Tong, Hejazi, et al., 2020). Conversely, biphasic cathodic-first pulses, the most common pulse used in retinal prostheses, became effective at larger electrode-RGC distances and offer the advantage of lower current requirements than anodic-first pulses. Taken together, these results emphasize that achieving focal activation requires carefully balancing pulse duration, waveform, and electrode-RGC distance, rather than focusing on just one factor.

The main benefit of epiretinal stimulation, bypassing the RGCs, also represents its main weakness: it bypasses the inherent preprocessing that the inner retinal network carries out. The stimulation does not involve bipolar, amacrine, or horizontal cells and therefore has no processing to form a detailed visual image; however, it still does not form a rich natural visual experience but makes the observer see elementary phosphenes. Even though superior stimulation methods can partially utilize the remaining components of the network, they are incapable of recreating the high-level, multi-layered processing of the intact retina (Palanker et al., 2025). This gap constitutes an essential functional gap between epiretinal solutions and upstream retinal circuitry interventions.

The surgical placement of the electrode array fundamentally dictates which neural elements are activated and the propensity for unintended axonal stimulation. While epiretinal implants provide direct access to RGC somas, their placement adjacent to passing axon bundles makes unintended axonal activation a fundamental challenge, resulting in the perception of elongated phosphenes (Beyeler et al., 2019; Nanduri et al., 2012). Subretinal implants mainly excite the bipolar cells and utilize the remaining network of the retina to activate the orthodromic response that avoids the direct stimulation of the axons (Stingl et al., 2013; Zrenner et al., 2011). Suprachoroidal implants generate the diffuse electric fields, which, in general, do not selectively activate axonal bundles but demand more currents and relatively lower spatial resolution (Ayton et al., 2014, 2020). Therefore, although subretinal and suprachoroidal methods inherently reduce axonal activation, the epiretinal method, notwithstanding this problem, has the potential to be higher in resolution with the use of more sophisticated stimulation schemes, like those discussed in this paper and in other studies.

While these results demonstrate promising strategies for avoiding unwanted passing axon activation, some issues warrant consideration. Although our parameter optimization enhances spatial resolution, the clinical application of these efforts requires long-term material stability. It has been established through chronic studies that materials like liquid crystal polymers and iridium oxide exhibit stable functions over time, with surface modifications like a coating with graphene being even more biocompatible (Eggenberger et al., 2021; Jeong et al., 2016; Nguyen et al., 2021). Although biphasic charge-balanced anodic-first (BA) pulses proved the most robust waveform across distances, their higher current requirements raise concerns about exceeding the safe charge injection limit for platinum electrodes (350 µC/cm²) (Cogan, 2008), particularly at greater distances. Notably, in our study at the farthest distance tested (480 µm), BA pulses with 50 µs duration produced a charge density of ∼183 µC/cm², remaining safely within this limit. Nevertheless, a trade-off between achieving focal activation and adhering to charge safety limits must be carefully weighed.

A variety of RGC types will be investigated to generalize the optimal stimulation parameters in future work. Moreover, an effect of superimposed retinal tissue layers will also be added, which was not in consideration in the present model, to more closely represent the in vivo condition. Adaptive stimulation systems might be able to change their parameters dynamically in real time depending on the individual retinal physiology of a patient. With the incorporation of neural feedback and other novel imaging tools, these would enable truly individualized therapeutic solutions, hence optimizing therapeutic outcomes.

5. Conclusions

The impact of pulse duration, waveform, and electrode-RGC distance on preventing undesired activation of passing axon fibers during epiretinal stimulation was clearly demonstrated. Using a computational model of a realistic morphological RGC, we investigated different pulse durations, waveforms, and electrode-RGC distances. This approach advances prior research by revealing how these factors interact synergistically, a relationship previously underexplored. Shorter BA pulses proved most effective across all distances, requiring axon activation thresholds approximately three times higher than AIS thresholds. BC pulses became viable at larger distances. Conversely, MA pulses required nearly identical current thresholds for both AIS and axon across all distances, making them less effective at preventing undesired axon activation. MC pulses performed poorly in avoiding unwanted axon activation, with the axon activating first at every distance. Crucially, our findings demonstrate that increasing electrode-cell distance enhances selectivity, but only when paired with an optimal waveform, revealing that focal activation depends primarily on pulse shape rather than distance alone.