Glaucoma, a chronic neurodegenerative disease affecting the retina and optic nerve, and a leading cause of blindness, is characterized by a progressive, irreversible loss of vision. As currently available treatment paradigms focus primarily on a predisposing factor, elevated intraocular pressure (IOP), and do not allow for repair of the retina and optic nerve once the disease has progressed and damage has occurred, technologies enabling an early diagnosis of glaucoma are needed urgently. Consequently, such new diagnostic modalities enabling early therapeutic intervention would significantly improve treatment outcomes. Current methods of glaucoma diagnosis are based on psychophysical techniques and the assessment of structural changes to the retina and optic nerve (Bussel et al., 2014). Standard automated perimetry testing, including the widely used Humphrey visual field testing, currently represents the most commonly utilized technique for glaucoma diagnosis and monitoring of disease progression and therapy outcomes (Ernest et al., 2012; Fidalgo et al., 2015). Recent efforts to employ machine-learning (ML) approaches to improve the analysis of behavioral psychophysical testing approaches produced moderate improvements over conventional analysis algorithms (Saeedi et al., 2021). However, significant damage to the retina and optic nerve, including loss of retinal ganglion cells (RGCs) has often already occurred before changes can be detected with standard automated perimetry testing (Turalba and Grosskreutz, 2010).

Recently, automated retinal image analysis (ARIA) systems have been developed for the diagnosis of complex diseases such as diabetic retinopathy and glaucoma (Sim et al., 2015; Lee et al., 2017). The development of these ARIA systems involved ML-based methods to detect structural changes determined with optical coherence tomography (OCT) imaging resulting in high analytical accuracy in automatically classifying disease phenotypes based on structural characteristics (Zhu et al., 2014; Asaoka et al., 2016; An et al., 2019). Despite such significant progress, early detection of glaucoma is still a challenge (Brandao et al., 2017), given the highly significant limitations of early detection of glaucoma based on structural methods. Systems employing the analysis of structural changes for glaucoma diagnosis are based on measuring retinal nerve fiber layer (RNFL) thickness in OCT images of the retina, which is highly variable and weakly correlated with RGC counts despite RNFL thickness being a surrogate marker of RGC degeneration and optic nerve fiber loss, hallmarks of glaucoma pathogenesis (Ledolter et al., 2015). Further, RGC loss often occurs early during pathogenesis in the absence of measurable RNFL thinning, prompting an urgent clinical need for methods with higher sensitivity, such as functional measures including ERG (Harwerth et al., 2002; Fortune et al., 2003; Takagi et al., 2012; Ledolter et al., 2015; Brandao et al., 2017). In contrast, functional measures such as visual field and ERG are sensitive to subtle changes in RGC function and RGC damage, which suggest a significant potential to enable early detection of glaucoma, even in the absence of elevated IOP, as seen in patients with normotensive glaucoma (Fortune et al., 2003; Aldebasi et al., 2004; Brandao et al., 2017). Therefore, this study aims to investigate such potential considering ERG signals.

Consequently, interventions could be initiated before irreversible damage occurs, allowing for the optimization of treatment strategies based on the improvement of RGC function (Ventura and Porciatti, 2006). This is of high clinical importance as determining the efficacy of therapies aimed at lowering IOP in open-angle glaucoma (Palmberg, 2002; Leske et al., 2007) requires early validation of therapy success (An et al., 2019), but will also be of importance for the development of novel alternative and complementary glaucoma therapies based on neuroprotective strategies (Rohowetz et al., 2018). Recently, in a study conducted by Tang et al. (2020) photopic negative response (PhNR) was used to assess the short-term changes in inner retinal function following intraocular pressure (IOP) decrease in glaucoma using eyedrops. Hui et al. (2020) showed that Nicotinamide supplementation helps improve the function of the inner retina in glaucoma.

Recent advances in the acquisition of complex neuroscience data have a significant innovative potential to progress toward more effective diagnostic systems (Kononenko, 2001). The adequate, timely, and clinically relevant analysis of such data has potentially high clinical impact (Lisboa, 2002). However, while such data sets can be readily acquired and technologies to further improve and simplify data acquisition continue to emerge (McPadden et al., 2019), critical barriers to implement the effective use of such novel data in clinical diagnostics and therapy delivery remain (Lee and Yoon, 2017). While the analysis of complex biomedical data is often part of medical diagnostics, current expert analysis standards and algorithms are limited by pattern recognition in few dimensions, which results in less than optimal identification or even exclusion of potentially relevant diagnostic features (Hannun et al., 2019). Machine learning could significantly augment medical diagnostics and increase their efficacy by analyzing aspects of complex and multi-dimensional biomedical data that are either not being considered adequately or that are not accessible to current analysis methods (Holzinger, 2014). Such machine-learning based diagnostic approaches have been developed and are being actively used for the detection of cardiovascular diseases (Al'Aref et al., 2019), and cancer (Cruz and Wishart, 2006).

ERG data are one such type of complex and multi-dimensional biomedical data that are potentially relevant to the diagnosis of glaucoma, but are currently not considered during routine clinical practice or in clinical research. Historically, this is due to multiple barriers related to clinical ERG data acquisition, such as limitations in reproducibility, high costs of both equipment and of individual tests, long test duration and complex test administration resulting in reduced patient acceptance and compliance, and the need for highly trained experts to administer tests. With the advent of novel ERG technologies, most of these barriers related to clinical ERG data acquisition have been removed (Nakamura et al., 2016; Asakawa et al., 2017; Kato et al., 2017; Hobby et al., 2018; Liu et al., 2018; Man et al., 2020), opening up the possibility to effectively use ERG data for glaucoma diagnostics, calling the necessity for the development of novel approaches (e.g., M-L-based ones) that is capable to quickly and thoroughly analyze such data.

Machine learning is based on statistical techniques to learn from data and develop predictive models (Jordan and Mitchell, 2015). Recently, there has been a surge of interest in machine learning as significant advancements in computational hardware (Shi et al., 2016) facilitate the development of novel machine learning approaches as solutions to problems in various disciplines from financial forecasting to public transportation and healthcare (Trafalis and Ince, 2000; Omrani, 2015; Ahmad et al., 2018). There are several predictive techniques in machine learning with various complexities, ranging from simple linear models to advanced non-linear models such as those based on deep learning algorithms (Shailaja et al., 2018; Khan et al., 2021; Saxe et al., 2021). Currently, available ERG analysis methods, such as those developed by Hood et al. (2000); Ventura and Porciatti (2006), have contributed to a significantly improved understanding of the relationship between ERG signals and vision loss. These methods are limited to frequency domain analysis (Miguel-Jiménez et al., 2010; Luo et al., 2011; Palmowski-Wolfe et al., 2011; Ledolter et al., 2013) and the analyses of differences in amplitude and latency of ERG (Fortune et al., 2002; Thienprasiddhi et al., 2003; Stiefelmeyer et al., 2004; Chu et al., 2007; Todorova and Palmowski-Wolfe, 2011; Ho et al., 2012; Hori et al., 2012). In addition, these methods are often time-consuming, labor-intensive, and focused on parameters developed to address a small subset of mostly genetic diseases of the eye affecting predominantly pediatric patient populations (Frishman et al., 2000; Graham et al., 2000; Dale et al., 2010). To achieve higher accuracy and a more detailed understanding of disease progression and of the impact of therapeutic intervention, more sophisticated features such as those obtained from wavelet analysis are required (Forte et al., 2008; Barraco et al., 2014). Additionally, currently available methods are often not suitable for analyzing large data sets and databases, rendering them incapable of taking advantage of complex and rich datasets (Consejo et al., 2019; Armstrong and Lorch, 2020). These drawbacks prompted others (Bowd et al., 2014; Yousefi et al., 2015; Atalay et al., 2016; Verma et al., 2017) and us to design and develop novel methods capable of handling complex and large datasets and ultimately to provide a unique approach for diagnosing early-stage glaucoma. However, it should be noted that early detection of glaucoma is not possible with currently available techniques during the early stages of glaucoma pathogenesis, when cellular changes occur that do not result in structural damage or visual impairment yet. Such early-onset factors predisposing to glaucoma development include processes preceding the onset of ocular hypertension, for example, the onset of iris pigment dispersion preceding IOP elevation in the DBA/2 mouse model. However, and more importantly, we identified cellular changes resulting in altered ERG signals, such as changes in oscillatory potentials, that currently cannot be detected with other functional or structural measures.

Boquete and colleagues developed a method to automate glaucoma diagnosis based on ERG signals using neural networks and structural pattern analysis (Boquete et al., 2012). They utilized thirteen features (morphological and transitional characteristics) for training the model and achieved a testing accuracy of 80.7% (Boquete et al., 2012). This study was limited to basic morphological characteristics of mfERG recordings (Boquete et al., 2012). Miguel-Jiménez et al. (2015) also employed neural networks for ERG-based glaucoma diagnosis but used continuous wavelet transformed coefficients and achieved a binary classification accuracy of 86.90% (Miguel-Jiménez et al., 2015). Although a higher accuracy was achieved, this analysis was limited to wavelet features only (Miguel-Jiménez et al., 2015). Nevertheless, both studies showed that machine learning-based methods trained even on compact data sets provide powerful tools to analyze ERG signals and provide potentially new information relevant for the early detection of glaucoma. Sarossy and colleagues investigated the relationship between a compact set of features and glaucoma that can be analyzed with machine learning approaches; however, the study was limited to the analysis of the photopic negative response (PhNR) and five additional features (Sarossy et al., 2021).

The goal of the present study was to comprehensively assess the capability of machine-learning-based methods to detect early-stage glaucoma using time-series ERG signals. In particular, the following points are addressed during method development:

A machine learning-based approach was developed and trained using the balanced ERG data previously published by Grillo et al. (2018). Although a compact dataset of 60 observations and 540 signals was used in this study, the current framework was able to consistently detect features (Figures 6, 9) that are known to be medically relevant such as OP, STR, Flicker reported in various studies (Tyler, 1981; Saszik et al., 2002; de Lara et al., 2014, 2015; Porciatti, 2015; Grillo et al., 2018). In particular, studies conducted by Wilsey and Fortune (2016); Hermas (2019); Beykin et al. (2021) investigating the variability of PhNR in glaucomatous and healthy subjects in PERG and fERG have found that PhNR to be an important biomarkers for detection of glaucoma. It is worth noting that in fERG analysis (ERG protocol for this study), pSTR, nSTR, PhNR are extracted from STR.

Therefore, we were able to demonstrate that the proposed framework for early-stage glaucoma diagnosis can be reproducibly evaluated and validated even on such a compact database. Furthermore, we would like to note that there are other investigations that successfully applied ML-based method in different fields, including biomedical (Seo et al., 2020) and material science (Zhang and Ling, 2018) using compact datasets. The procedure employed for the development of the predictive modeling framework is summarized below.

The dataset was divided into two parts; 80% of the data was used for training and validation, and the remaining 20% was set aside for testing. The hold-out testing strategy ensured that the test data was never a part of the training process (Yadav and Shukla, 2016). Dimensionality reduction was performed using feature extraction and feature selection. MRMR feature selection algorithm was used to identify the important predictors. K-fold (K = 10) cross-validation was used for training and hyper-parameter tuning (Duan et al., 2003). The cross-validation technique significantly reduces bias when working with small datasets (Varma and Simon, 2006). The loss function was the objective minimization function for both classification regressions during hyper-parameter optimization. The hyper-parameters associated with corresponding ML algorithms (Feurer and Hutter, 2019), as shown in Table 1, were optimized through nested cross-validation. Next, the trained model with optimized hyper-parameters was evaluated using test data that was not a part of training. To further ensure that the machine learning models compared in this investigation were not over-fitted, given the compact dataset used in the present study, the behavior of training and testing error vs. training cycles was monitored. Different techniques, including Tree, Discriminant, SVM, Naive Bayes, Tree Ensemble, and KNN, were applied, and their performances were assessed. The performance of each technique was assessed based on the accuracy (discussed in section 2.9) is tabulated in Table 2. Considering binary and multiclass classifications, it can be seen that the Ensemble-based technique (bagged tree) was consistently outperforming other techniques. Additionally, other performance metrics for ensemble bagged trees (discussed in section 2.9) are summarized in Table 3.

Our goal was to determine the feasibility of applying ML-based methods to the analysis of ERG signals for glaucoma detection at different stages of the disease. In the present study, we systematically applied machine-learning-based methods for the first time to detect glaucoma and predict RGC loss based on ERG signals. The present study utilized ERGs measured in mice rather than from human patients, because the use of data from a preclinical model allowed us to validate “ground truth” data sets with a range of complimentary and alternative experimental strategies, which is not possible in human clinical studies. These include histology, biochemical, and immunochemical assays, as well as optomotor reflex measurements. We were able to determine for the first time that advanced features (wavelet-based features) are capable of detecting subtle changes in the ERG signal and perform multiclass classification based on the progression level of the disease with 80% accuracy. In particular, we found that Shannon Entropy Values for Maximal Overlap Discrete Wavelet Packet Transform (MOD-PWT) and AR coefficients represent important features capable of detecting early-stage glaucoma. Among the nine available ERG signals, Flicker, STR, OP, and Rod-Cone appear integral for such successful detection. This is in agreement with the results published in Lei et al. (2006). However, given that these features are highly correlated, the ML-based algorithm picks only one for each set of highly correlated features to reduce the model complexity as shown in Figure 6.

In addition, the method proposed here performs ERG analysis in a wavelet domain instead of a frequency domain, which allows to capture subtle changes in the signals. In addition, various intricate features such as multiscale wavelet variance estimates, Shannon entropy, and autoregressive coefficients are incorporated in the method, compared to basic features such as differences in amplitude and latency in previous studies (Hood et al., 2000; Fortune et al., 2002; Thienprasiddhi et al., 2003; Stiefelmeyer et al., 2004; Ventura and Porciatti, 2006; Chu et al., 2007; Miguel-Jiménez et al., 2010; Luo et al., 2011; Palmowski-Wolfe et al., 2011; Todorova and Palmowski-Wolfe, 2011; Ho et al., 2012; Hori et al., 2012; Ledolter et al., 2013; Consejo et al., 2019). The results strongly suggest that such advanced features in the wavelet domain are necessary for detection of early-stage glaucoma. Moreover, in contrast to the recent study that leverages ML-based technique to analyze ERG using solely the photopic negative response (PhNR) component (Armstrong and Lorch, 2020), the current method uses all ERG components in the analysis to fully utilize the capability of the ML-based technique to crunch large data sets and draw complicated relationships. Therefore, the proposed framework is not limited to a small subset of genetic eye diseases like previous studies (Fortune et al., 2002; Thienprasiddhi et al., 2003; Stiefelmeyer et al., 2004; Chu et al., 2007; Miguel-Jiménez et al., 2010; Luo et al., 2011; Palmowski-Wolfe et al., 2011; Todorova and Palmowski-Wolfe, 2011; Ho et al., 2012; Hori et al., 2012; Ledolter et al., 2013; Consejo et al., 2019); instead, it is capable of mapping ERG signals to various eye diseases.

Results obtained in the present study strongly suggest that the methods employed can reproducibly identify dominant features for classification and regression from STR, Oscillatory potentials (OPs), and other ERG tests consistent with the results reported in previously published work on the sensitivity of and OPs and flicker to subtle changes in RGC function and viability (Tyler, 1981; Brandao et al., 2017). Further, our approach identified additional dominant distinguishing features such as Shannon Entropy Values for Maximal Overlap Discrete Wavelet Packet Transform (MOD-PWT) and AR coefficients, which are not distinguishable by traditional methods used in Grillo et al. (2018). This strongly suggests that the current machine-learning-based algorithm has significant potential in distinguishing subtle changes in ERG signals corresponding to different stages of glaucoma disease development. This capability of the technique could be used as a foundational step to create a reliable framework for the early detection of glaucoma and to monitor efficacy of therapeutic intervention in both clinical practice and novel drug development for glaucoma. In addition, the inclusion of various ERG protocols in this framework, such as cones, rods and cones, STR, and oscillatory potentials, represent responses from different cell types in the eye. Therefore, ERG response can be mapped to diseases specific to those cell types. It should be noted that this study was based on mice and with 12 h of dark adaptation. The promising results obtained here suggest the great potential for this method to help detect early stage, pre-symptomatic glaucoma. However, an additional study on adaptation requirements would be required before extending this framework to humans.

The datasets generated for this study are available on request to the corresponding author. Requests to access these datasets should be directed to mehdizadeha@umkc.edu.

MG contributed in machine learning framework development, formal analysis, investigation, validation, visualization, and writing—original draft. LR contributed in writing—review and editing. PK contributed in providing the data, conceptualization, supervision, and writing—review and editing. AM contributed in conceptualization, supervision, and writing—review and editing. All authors contributed to the article and approved the submitted version.

Research reported in this publication was supported by the Felix and Carmen Sabates Missouri Endowed Chair in Vision Research, the Vision Research Foundation of Kansas City, and in part by National Eye Institute grant EY031248 of the National Institutes of Health (PK). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. The publication cost was covered by PK.

The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

PK, AM, and MG have a patent-pending based on this study. The remaining author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.