Skip to main content

A feasibility study for objective evaluation of visual acuity based on pattern-reversal visual evoked potentials and other related visual parameters with machine learning algorithm

Abstract

Background

To develop machine learning models for objectively evaluating visual acuity (VA) based on pattern-reversal visual evoked potentials (PRVEPs) and other related visual parameters.

Methods

Twenty-four volunteers were recruited and forty-eight eyes were divided into four groups of 1.0, 0.8, 0.6, and 0.4 (decimal vision). The relationship between VA, peak time, or amplitude of P100 recorded at 5.7°, 2.6°, 1°, 34′, 15′, and 7′ check sizes were analyzed using repeated-measures analysis of variance. Correlations between VA and P100, contrast sensitivity (CS), refractive error, wavefront aberrations, and visual field were analyzed by rank correlation. Based on meaningful P100 peak time, P100 amplitude, and other related visual parameters, four machine learning algorithms and an ensemble classification algorithm were used to construct objective assessment models for VA. Receiver operating characteristic (ROC) curves were used to compare the efficacy of different models by repeated sampling comparisons and ten-fold cross-validation.

Results

The main effects of P100 peak time and amplitude between different VA and check sizes were statistically significant (all P < 0.05). Except amplitude at 2.6° and 5.7°, VA was negatively correlated with peak time and positively correlated with amplitude. The peak time initially shortened with increasing check size and gradually lengthened after the minimum value was reached at 1°. At the 1° check size, there were statistically significant differences when comparing the peak times between the vision groups with each other (all P < 0.05), and the amplitudes of the vision reduction groups were significantly lower than that of the 1.0 vision group (all P < 0.01). The correlations between peak time, amplitude, and visual acuity were all highest at 1° (rs = − 0.740, 0.438). VA positively correlated with CS and spherical equivalent (all P < 0.001). There was a negative correlation between VA and coma aberrations (P < 0.05). For different binarization classifications of VA, the classifier models with the best assessment efficacy all had the mean area under the ROC curves (AUC) above 0.95 for 500 replicate samples and above 0.84 for ten-fold cross-validation.

Conclusions

Machine learning models established by meaning visual parameters related to visual acuity can assist in the objective evaluation of VA.

Peer Review reports

Background

Visual acuity (VA) examination, one of the most basic tests for visual function, is divided into subjective and objective examination methods. The subjective ways are simple and easy to perform using various visual acuity charts. However, these methods require a high degree of patient cooperation. Infants, children with cerebral visual impairment, pseudo-blindness and hysterical patients need an objective VA assessment method, a classic common challenge for ophthalmologists. Likewise, this assessment technology is also required for the VA identification of flight personnel.

Pattern-reversal visual-evoked potentials (PRVEPs) are a cluster of bioelectrical signals generated in the visual cortex located in the brain’s occipital lobe after graphic visual stimulus. They are waveforms recorded by regular graphic stimulation of the subject and processed by applying computerized averaging and superimposition techniques. The P100 wave is the first positive wave to appear in PRVEPs. It includes peak time and amplitude evaluation elements, with low variability across subjects and a strong relationship with VA. PRVEPs are mainly applied in patients who are unable or unwilling to complete a subjective VA examination and those with cognitive deficits. They are widely used in the assessment of optic neuropathy [1,2,3] and functional visual loss [4], in the identification of malingering [5], and in the early diagnosis and prognostication of multiple sclerosis [6,7,8]. Over the years, several reports have demonstrated that using the visual-evoked potential technique to detect the completeness of visual pathways has been widely recognized in ophthalmology [9, 10]. However, most prior studies on the objective assessment of VA have focus on normal VA or best-corrected VA; there is little information in the literature on the effect of reduced VA caused by refractive error alone on the PRVEPs waveform [10, 11]. When visual acuity was greater than Snellen 20/100, there was a good correlation between subjective VA and PRVEP-estimated VA and dioptric blur would diminish the PRVEP responses to smaller pattern elements more than to larger ones [12]. In addition, it has been shown that the correlation between PRVEP-estimated VA and subjective VA in patients with organic ocular pathology is less than that in subjects with refractive error only [12]. VA is also related to visual parameters such as contrast sensitivity (CS), refractive error, wavefront aberrations, and visual field [13,14,15,16]. Therefore, we studied the correlation of these parameters with VA and incorporated them into machine modeling with the P100 of PRVEPs. As a subfield of artificial intelligence, machine learning has been increasingly applied to medical practice since it can extract deeper features from raw data and combine several predictors in a highly interactive manner [17]. In recent years, machine learning algorithms have become more prevalent in the diagnosis and prognostication of many ophthalmic diseases [18,19,20].

This study focused on the relationship between VA and visual parameters, including PRVEPs, CS, refractive error, wavefront aberrations, and visual field. We were using machine learning algorithms to explore a feasible objective examination and assessment method for VA to provide theoretical support and an experimental basis for improving the quality of vision assessment for special operation staff, such as pilots.

Methods and subjects

Subjects

Twenty-four male subjects from Air Force Medical University were recruited; their ages ranged from 20 to 35 (mean 27.75 ± 4.33) years. The inclusion criteria for selecting participants were as follows: spherical degree (D) − 2.00 D ~ + 2.50 D, cylindrical degree − 1.00 D ~ + 1.00 D, and best-corrected VA ≥ 1.0 (decimal vision). The exclusion criteria were a history of refractive surgery, central nervous system disease, glaucoma, diabetic retinopathy, and other organic ocular diseases. The experiments were conducted in the same laboratory under the same luminance environment, and all subjects rested and adapted to the light environment for 10 min before the examination. An automated, computerized refractive examiner (KR-8100PA, Topcon, Japan) was used to detect the subjects’ refractive condition. First, the refractive error of the subjects was examined. After meeting the criteria, twenty-four subjects were randomly divided into four groups: 1.0/1.0, 0.8/0.8, 0.6/0.6, and 0.4/0.4 (decimal vision). Six subjects and twelve eyes were in each group. Second, subjects were checked for monocular VA at 5 m using a standard logarithmic visual acuity E chart with 96% contrast in a 200 cd/m2 luminance lightbox. Finally, the subjects were corrected to the corresponding VA using the subjective insertion method with appropriate lenses.

Pattern visual evoked potential

The VEP stimulated by reversal checkerboard patterns was recorded using the Visual Electrophysiology System (LS-D1, ChongQing Sunkingdom, China). The field size of the checkerboard stimulus was 375 × 300 mm, the contrast of the checkerboard was 96%, and the average luminance of the screen was 120 cd/m2. Each subject sat 1 m from a black-and-white checkerboard monitor in a moderately lit room, with their eyes at the same height as the center of the screen. The reversal rate was two reversals per second. Monocular PRVEPs of both eyes (first the right, then the left) were recorded using gold disc scalp electrodes with an average of 100 stimulations. According to the International Society for Clinical Electrophysiology of Vision standard [21], the active electrode was placed at Oz with the reference electrode at Fz, and the ground electrode at the right earlobe, all with inter-electrode impedance < 5 kΩ. Six sizes of checkerboard patterns (5.7°, 2.6°, 1°, 34′, 15′ and 7′ check sizes) were tested in each eye, with 1 to 2 min of rest between each examination with the eyes closed. Throughout the test, the examiner closely monitored the participants’ fixation on a red marker at the center of the screen. Each eye was examined at least two times at each check size, and similar graphs were considered reliable. The entire PRVEPs examination was 20 ~ 30 min for each subject.

Other visual parameters

CS examinations were performed by MetroVision (MonPack one, Perenchies, France) under photopic conditions (80 cd/m2) while the patient was sitting 2 m away from the screen, with one eye of the corrective lenses. Spatial frequencies of 0.6, 1.1, 2.2, 3.4, 7.1, and 14.2 cycle per degree (cpd) were tested. The refractive error was calculated from the refractive error of the bare eye and the corrective lenses. Spherical equivalent (SE) was calculated using the standard formula (spherical equivalent = spherical power + [cylinder power/2]). Wavefront aberrations for a 4 mm pupil were tested by the OPD-Scan III aberrometer (Nidek, Gamagori, Japan). To avoid the influence of lenses on the examination, we recorded the root-mean-square (RMS) values for the total higher-order aberrations (HoA), total coma aberrations (CA), total trefoil aberrations (TA), and total spherical aberrations (SA) in the naked eye. The visual field was determined using a Humphrey Field Analyzer 750i (Zeiss Humphrey Systems, Dublin, CA). The central 30 − 2 threshold protocol was performed in each eye with the corrective lenses. The absolute values of mean deviation (MD) and pattern standard deviation (PSD) were recorded for statistical analysis.

Statistical analysis

SPSS version 21.0 (IBM, USA) was used for statistical analysis. Measurement data were expressed as mean ± standard deviation (\(\stackrel{-}{x}\)± s). Kolmogorov-smirnov test was conducted on the data, and the results showed that the data conformed to the normal distribution. Statistical analyses were performed using repeated-measures analysis of variance (ANOVA) and rank correlation with a significance level of 5%. The univariate ANOVA method was used if the sphericity test was satisfied. Otherwise, the multivariate ANOVA method was used. Further two-by-two comparisons were carried out using the least significant difference test. Subjective VA was binarized using 1.0, 0.8, and 0.6 VA as cutoff points, and the mean area under the ROC curves (AUC) was used to compare the performances of different models based on meaningful objective visual parameters.

Machine learning modeling

Machine learning models are computer programs used to recognize patterns in data or make predictions. The models are created from machine learning algorithms, which are trained using labeled, unlabeled, or mixed data. Different machine learning algorithms are suited to different goals, such as classification or prediction modeling, so data analysts use different algorithms as the basis for different models.

Based on the Scikit Learn machine learning algorithm package in Python, the P100 and other visual parameters related to VA were used as input variables, and the binarized VA divided by different VA cutoffs (1.0, 0.8, and 0.6) were used as output variables to construct classification models by support vector machine (SVM), random forest (RF), K-nearest neighbor (KNN), multilayer perceptron (MLP), and Voting. Hybrid Classifier uses a combination of different classifiers to make a common decision. Voting is a common method of classifier ensemble, aggregating the predictions of each basic classifier and using the one with the most votes as the final prediction. Voting has the advantage of achieving better results than individual classifiers by complementing the strengths of each basic classifier. In this study, a voting classifier was constructed as shown in Fig. 1. It utilizes SVM, RF, KNN, and MLP as the basic classifiers to obtain the final prediction based on the voting results. The classification performance of the models was compared. On one hand, 70% of the 48-sample data were selected as the training set and 30% as the test set, and the training test was repeatedly sampled 500 times to compare the mean AUC of each classification model. On the other hand, ten-fold cross-validation was performed on the 48-sample data to compare the accuracy, sensitivity, specificity, and other parameters of different classification models. In the ten-fold cross-validation, the 48-sample data were randomly divided into 10 parts, and when one part was used as the validation set, the other 9 parts were used as the training set, and the same binary classification of “good” and “bad” visual acuity was performed. The 10 samples were cross-trained and validated 10 times, and the results of the ten-fold cross-validation of each model were obtained. The accuracy, sensitivity and specificity of the classification were compared.

Fig. 1
figure 1

Hybrid classifier based on voting, which utilized support vector machine (SVM), random forest (RF), K-nearest neighbor (KNN), and multilayer perceptron (MLP) as the basic classifiers

Results

P100 peak time versus VA at different check sizes

There was a statistically significant difference in the main effect in P100 peak time between the different VA groups (F = 14.969, P < 0.001) and the different check sizes (F = 61.386, P < 0.001) and no statistically significant difference in the interaction between VA and check size (F = 0.967, P > 0.05). Overall, the peak time tended to increase with a decrease in VA, and the peak time of the 1.0 vision group was the shortest (101.15 ± 8.69). The peak time decreased with increasing check size, reaching the minimum value at 1° (97.73 ± 3.55). Further statistical tests revealed statistically significant differences when comparing the peak time between the vision groups with each other at 1° (all P < 0.05). It can be seen that compared with 1.0 vision, the P100 peak time of 0.4 vision at 7′ and 2.6°, that of 0.8 vision at 15′, those of the vision reduction groups at 34′, those of 0.6 vision at 2.6° and 5.7° were increased significantly (all P < 0.05) (Fig. 2). Figure 3 showed the waveforms of PRVEPs at 7′ and 1° check sizes of different VA.

Fig. 2
figure 2

Group graph of comparison of P100 peak time of various VA at different check sizes (n = 48). *P < 0.05, **P < 0.01, ***P < 0.001. The P100 peak time of different VA at 7′ (A), 15′ (B), 34′ (C), 1° (D), 2.6° (E), and 5.7° (F)

Fig. 3
figure 3

PRVEP waveforms of different vision. A: PRVEP waveforms at 7’ check size of different vision; B: PRVEP waveforms at 1° check size of different vision

P100 amplitude versus VA at different check sizes

There was a statistically significant main effect difference in P100 amplitude between the different VA groups (F = 3.153, P = 0.034) and the different check sizes (F = 27.155, P < 0.001) and a statistically significant difference in the interaction between VA and the check size (F = 1.955, P = 0.024). Overall, the amplitude tended to decline with a decrease in VA. The amplitude of the 1.0 vision group (7.98 ± 2.52) was the highest among all groups and the amplitude of the 15′ (7.86 ± 2.92) was the highest among all check sizes. At the 7′ and 1° check sizes, the amplitude of the vision reduction groups decreased significantly compared to that of the 1.0 vision group (all P < 0.05). At the 15′ and 34′ check sizes, the amplitude of the 0.4 vision group decreased significantly compared to that of the 1.0 vision group (all P < 0.05) (Fig. 4).

Fig. 4
figure 4

Group graph of comparison of P100 amplitude of various VA at different check sizes (n = 48). *P < 0.05, **P < 0.01, ***P < 0.001. The P100 amplitude of different VA at 7′ (A), 15′ (B), 34′ (C), 1° (D), 2.6° (E), and 5.7° (F)

Correlation analysis of P100 peak time and amplitude versus VA

There was a negative correlation between P100 peak time and VA at all check size (all P < 0.05). Namely, the higher the VA, the shorter the peak time at the same check sizes, with the highest correlation at 1° check size (rs = − 0.740) (Fig. 5- A, B, C). Except for the 2.6° and 5.7° check sizes, there was a positive correlation between the P100 amplitude and VA at all check sizes (all P < 0.05), which means that the higher the VA, the higher the magnitude at the same check size, especially at the 1° check size (rs = 0.438) (Fig. 5- D, E, F).

Fig. 5
figure 5

Group graph of correlation of P100 peak time and amplitude versus VA at different check sizes (n = 48). The relationships between VA and P100 peak time at 7′, 15′ (A), 34′, 1° (B), 2.6°, and 5.7° (C) and P100 amplitude at 7′, 15′ (D), 34′, 1° (E), 2.6°, and 5.7° (F). Lines represent least-squares linear model fits to the data. rs, Spearman rank correlation coefficient

Correlation analysis of CS, refractive error, wavefront aberrations and visual field versus VA

Figure 6 (A, B, C) shows correlation coefficients relating VA to CS in the spatial frequencies ranging from 0.6 cpd to 14.2 cpd. There was a positive correlation between VA and CS with all spatial frequencies (all P < 0.001) and high spatial frequencies (7.1 cpd and 14.2 cpd) generally had a stronger relationship to VA (rs = 0.815, 0.800) than other low spatial frequencies. Figure 6- D showed that SE and VA had a statistically significant correlation (rs = 0.815, P < 0.001). There was a negative correlation between VA and CA (rs = − 0.299, P = 0.039), but not with HoA, TA, and SA (Fig. 6- E, F). There was no significant correlation between VA and the absolute values of both MD and PSD for the visual field (Fig. 6- G).

Fig. 6
figure 6

Group graph of correlation of CS, SE, wavefront aberrations and visual field versus VA (n = 48). The relationships between VA and CS of 0.6 cpd, 1.1 cpd (A), 2.2 cpd, 3.4 cpd (B), 7.1 cpd, and 14.2 cpd (C). The relationships between VA and SE (D), HoA, CA (E), TA, SA (F), MD, and PSD (G). Lines represent least-squares linear model fits to the data. rs, Spearman rank correlation coefficient. CS, contrast sensitivity; SE, spherical equivalent; VA, visual acuity; HoA, higher-order aberrations; CA, coma aberrations; TA, trefoil aberrations; SA, spherical aberrations; MD, mean deviation; PSD, pattern standard deviation

Construction of classification models for objective evaluation of VA

Several classification models were constructed using machine learning algorithms such as SVM, RF, KNN, MLP, and Voting. The P100, CS, SE, and CA related to VA were input variables and binarized VA divided by different VA cutoff points were output variables. The means and standard deviations of the AUC values of the five models were obtained by training and testing with 500 replicate sampling (Table 1). For different visual acuities, the individual classifier models with the best assessment efficacy all had mean AUC values above 0.95. Similarly, the Voting classifier achieves the same assessment performance.

We also compared the classification performances of the ten-fold cross-validation of the five models for 48 samples (Table 2; Fig. 7). In the classification models of distinguishing between 1.0 and 0.8, 0.6, and 0.4 at the VA level, the SVM classifier performed well with an accuracy of 89.58%, sensitivity of 91.67%, specificity of 83.33%, and AUC value of 0.8750. In the classification models of distinguishing between 1.0 and 0.8 from 0.6 to 0.4 of VA, the Voting classifier performed well with an accuracy of 93.75%, sensitivity of 91.67%, specificity of 95.83%, and AUC value of 0.9375. In the classification models of distinguishing between 1.0, 0.8, and 0.6 from 0.4 at the VA level, the Voting classifier performed best with an accuracy of 89.58%, sensitivity of 75.00%, specificity of 94.44%, and AUC value of 0.8472.

Table 1 Comparison of AUC for 500 replicate samples of machine learning classification models for different VA objective assessment (\(\stackrel{-}{x}\)± s, n = 48)
Table 2 Comparison of the effectiveness of ten-fold cross-validation of machine learning classification models for different VA objective assessments (n = 48)
Fig. 7
figure 7

Group graph of ROC curves of ten-fold cross-validation of machine learning classification models (n = 48). The ROC curves of distinguishing 1.0 from 0.8, 0.6, and 0.4 (A), 1.0 and 0.8 from 0.6 and 0.4 (B), 1.0, 0.8, and 0.6 from 0.4 (C). The 95% CI of the AUC value has been demonstrated in the figures. AUC, area under the receiver operating characteristic curves; SVM, support vector machine; RF, random forest; KNN, K-nearest neighbor; MLP, multilayer perceptron

Discussion

The results of this study showed that with the decrease of VA, the P100 peak time tended to lengthen and the P100 amplitude tended to decrease. The peak time reflects the optic nerve conduction function of the subject eye: the longer the peak time, the slower the nerve conduction speed. The amplitude reflects the number of neurons that generate electrical activity in the human eye responding to picture stimulus [22]: the lower the amplitude, the fewer the neurons stimulated. As a result, when vision decreases due to refractive error, the image projected onto the retina becomes blurred and the optic nerve pathway becomes less responsive to stimuli, eventually leading to longer peak times and decreased amplitude. The stimulus check size was “non-linear” to the peak time and amplitude. The shortest peak time was observed at 1° and the peak time increased gradually with the change of the check size. This finding was similar to those of Chen and her colleagues [11], who observed that when the VA is less than 0.2, the peak time is shortest at the 1° check size. Kurita-Tashima [23] found a curvilinear relationship between the P100 peak time and the check size in normal-vision subjects, with a minimum peak time at 35′. Steele et al. [12] studied the relationship between six check sizes from 96′ to 2′ and the amplitude and found that the maximum amplitude occurred at the 12′ check size. Consistent with their research, we also observed that the amplitude of the 15′ check size was higher than other check sizes. A possible explanation for these results may be that the edges of the stimulus pattern become blurred when the stimulus check size is small, resulting in a delay in peak time and a decrease in amplitude. Interestingly, we found that the correlations between P100 peak time, amplitude, and VA were highest at the 1° check size. It seems that greater efficacy of assessing visual acuity using peak time and amplitude of PRVEPs with 1° check size. This may be because the tests included subjects with different visual acuities and the 1° check size was moderate and appropriate for the central retinal area of these visual acuity individuals. Sun et al. [5] revealed a good correlation between VA and medium checkerboard size, especially at the 50′ check size, by examining PRVEPs to identify malingering. It has also been reported that a 15′ or 1° check size is better for studying visual-evoked potentials in hereditary optic neuropathy [24]. These differences may be partly explained by the various selection of the subject groups and the stimulus parameters. In addition, the International Society for Clinical Electrophysiology of Vision recommends 1° and 15′ as common stimulation check sizes for PRVEPs examinations [21].

The following methods have been reported to use PRVEPs to assess objective VA [25]: (1) converting the minimum check size of the recordable waveform into VA; (2) determining the spatial frequency that produces the optimal peak time and amplitude; (3) making a linear regression equation of amplitude-check size, extrapolating the amplitude to 0 µv, and taking the cutoff value; (4) evaluating the peak time and amplitude of the binocular VEP together; and (5) determining the absolute value of P100 parameters. A review study by Hamilton et al. [10] considered that direct conversion of the perspective of PRVEPs threshold stimuli to subjective VA is inaccurate. For one thing, subjective VA tests use a fixed target, whereas PRVEPs are dynamic and continuous visual stimuli. For another, the higher cognitive cortex is involved in the subjective VA tests, but the PRVEPs are only cellular activities of the primary visual cortex. Jeon et al. [26] found a significant positive correlation between P100 amplitude and VA in normal and amblyopic individuals. By plotting ROC curves, they determined that an absolute amplitude value > 5.77 µv helped distinguish between visual disability and malingering. However, due to individual variability in examining PRVEPs and the influence of various check parameters (such as pattern contrast, pattern size, average luminance, signal filtering, patient age, and pupil size), there may be some error in inferring VA from the absolute value of the P100 peak time or amplitude for a single check size. In addition, wavefront aberration, CS, refractive error, and visual field are also closely related to VA. The minimum angle of resolution (MAR) varied linearly with the magnitude of wavefront aberration [14]. Correcting wavefront aberration can improve mesopic CS and improve the quality of long-distance vision [27, 28]. Higher-order aberrations were associated with diminished visual acuity and perception in highly aberrated eyes [29]. Far visual acuity (logMAR) significantly correlates negatively with peak CS [13]. There was a correlation between the visual field mean deviation and VA in patients with glaucoma [16]. This study similarly found a correlation between CS, SE, and CA and visual acuity. Some of the discrepancies in results may be due to differences in subject populations.

Yperman et al. [30] found that the machine learning model based on visual-evoked potentials showed good predictive performance in analyzing disease progression in multiple sclerosis. Bach et al. [31] assessed objective VA with 89 machine learning algorithms based on steady-state brief-onset low-contrast checkerboard stimulation evoked potentials. They found that nearly half of the machine algorithms obtained higher agreement between subjective and objective VA than traditional heuristics and were more widely used and testable. They concluded that the machine learning approach is a useful alternative to traditional VA assessment. In contrast, we built five machine learning models based on combining the P100 values of multiple check sizes and other visual parameters (CS, SE, and CA) correlating with visual acuity to assess VA objectively. The performance of machine learning models was evaluated internally by ROC curves with the random split and the ten-fold crossover. The results of the study showed that in the internal validation results of machine learning models distinguishing different VA, the average AUC values of the 500 random split validations of the models with better evaluation performance were above 0.95 (Table 1), and the AUC values of ten-fold cross-validations were above 0.84 (Fig. 7). Both the single machine learning model and the ensemble classifier method (Voting) showed better evaluation performance. Voting exhibited evaluation performance that was superior or approximate to that of a single model. However, this study only performed an internal validation and comparison of the evaluation performances of different models. Which model has a more powerful generalization capability and which pattern performs better in objective evaluation of VA based on multiple visual parameters or a single visual parameter needs to be compared by the external validation of the model in subsequent studies. Nevertheless, the results of the internal validation of the models obtained in our current study indicated that the performance of objective assessment for VA through the parameters of P100, CS, SE, and CA was satisfactory.

Our study had several limitations that need to be considered. Firstly, the study sample size was small and only four VA points were selected. The machine learning model could not evaluate VA outside the modeling data, which means that a new VA evaluation model must be established when assessing the other VAs is needed. Secondly, the subjects in this trial were all male, and it has been shown that the standard values of the P100 peak time and amplitude of PRVEPs differ between genders [32]. Finally, our current experiment focuses on visual acuity assessment in subjects with vision loss due to refractive error, and further research is needed for visual acuity assessment in visual system disorders. Consequently, it is necessary to increase and enrich the groups and expand the range of VA to improve the availability of the test results and the reliability of the method.

Conclusions

In conclusion, the correlations between P100 peak time, amplitude, and VA were highest at the 1° check size. There were positive correlations between CS, SE, and VA and a negative correlation between CA and VA. Furthermore, the classification models with AUC values above 0.84 were available in various VA cutoff point divisions. Machine learning classification models based on PRVEPs and other visual parameters appear to be an effective approach to the intelligent assistance of VA evaluation.

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Abbreviations

ANOVA:

analysis of variance

AUC:

area under the receiver operating characteristic curves

CA:

coma aberrations

cpd:

cycle per degree

CS:

contrast sensitivity

D:

degree

HoA:

higher-order aberrations

KNN:

K-nearest neighbor

MD:

mean deviation

MLP:

Multilayer perceptron

PRVEPs:

pattern-reversal visual evoked potentials

PSD:

pattern standard deviation

RF:

Random forest

RMS:

root-mean-square

ROC:

receiver operating characteristic

SA:

spherical aberrations

SE:

spherical equivalent

SVM:

Support vector machine

TA:

trefoil aberrations

VA:

visual acuity.

References

  1. Elkholy SH, El-Jaafary SI, Kotb MS, El GA, Elbhy BA. Trans-orbital sonography versus visual evoked potentials in acute demyelinating optic neuritis. MULT SCLER RELAT DIS. 2020;40:101934. https://doi.org/10.1016/j.msard.2020.101934.

    Article  Google Scholar 

  2. Ringelstein M, Kleiter I, Ayzenberg I, Borisow N, Paul F, Ruprecht K, Kraemer M, Cohn E, Wildemann B, Jarius S, et al. Visual evoked potentials in neuromyelitis optica and its spectrum disorders. MULT SCLER J. 2014;20(5):617–20. https://doi.org/10.1177/1352458513503053.

    Article  Google Scholar 

  3. Neto SP, Alvarenga RM, Vasconcelos CC, Alvarenga MP, Pinto LC, Pinto VL. Evaluation of pattern-reversal visual evoked potential in patients with neuromyelitis optica. MULT SCLER J. 2013;19(2):173–8. https://doi.org/10.1177/1352458512447597.

    Article  Google Scholar 

  4. Xu S, Meyer D, Yoser S, Mathews D, Elfervig JL. Pattern visual evoked potential in the diagnosis of functional visual loss. Ophthalmology. 2001;108(1):76–80. https://doi.org/10.1016/s0161-6420(00)00478-4.

    Article  CAS  PubMed  Google Scholar 

  5. Sun I, Lee J, Huang H, Kuo H. Pattern visual evoked potentials for identifying malingering. DOC OPHTHALMOL. 2015;130(3):221–9. https://doi.org/10.1007/s10633-015-9483-0.

    Article  PubMed  Google Scholar 

  6. Vidal-Jordana A, Rovira A, Arrambide G, Otero-Romero S, Rio J, Comabella M, Nos C, Castillo J, Galan I, Cabello S, et al. Optic nerve topography in multiple sclerosis diagnosis: the utility of visual evoked potentials. NEUROLOGY. 2021;96(4):e482–90. https://doi.org/10.1212/WNL.0000000000011339.

    Article  PubMed  PubMed Central  Google Scholar 

  7. Behbehani R, Ali A, Al-Omairah H, Rousseff RT. Optimization of spectral domain optical coherence tomography and visual evoked potentials to identify unilateral optic neuritis. MULT SCLER RELAT DIS. 2020;41:101988. https://doi.org/10.1016/j.msard.2020.101988.

    Article  Google Scholar 

  8. Leocani L, Guerrieri S, Comi G. Visual evoked potentials as a biomarker in multiple sclerosis and Associated Optic Neuritis. J NEURO-OPHTHALMOL. 2018;38(3):350–7. https://doi.org/10.1097/WNO.0000000000000704.

    Article  Google Scholar 

  9. Zheng X, Xu G, Zhang K, Liang R, Yan W, Tian P, Jia Y, Zhang S, Du C. Assessment of Human Visual Acuity using visual evoked potential: a review. SENSORS-BASEL. 2020;20(19):5542. https://doi.org/10.3390/s20195542.

    Article  PubMed  PubMed Central  Google Scholar 

  10. Hamilton R, Bach M, Heinrich SP, Hoffmann MB, Odom JV, McCulloch DL, Thompson DA. VEP estimation of visual acuity: a systematic review. DOC OPHTHALMOL. 2021;142(1):25–74. https://doi.org/10.1007/s10633-020-09770-3.

    Article  PubMed  Google Scholar 

  11. Chen X, Li Q, Liu X, Yang L, Xia W, Tao L. Visual acuity evaluated by pattern-reversal visual-evoked potential is affected by check size/visual angle. NEUROSCI BULL. 2012;28(6):737–45. https://doi.org/10.1007/s12264-012-1292-9.

    Article  PubMed  PubMed Central  Google Scholar 

  12. Steele M, Seiple WH, Carr RE, Klug R. The clinical utility of visual-evoked potential acuity testing. AM J OPHTHALMOL. 1989;108(5):572.

    Article  CAS  PubMed  Google Scholar 

  13. Thurman SM, Davey PG, McCray KL, Paronian V, Seitz AR. Predicting individual contrast sensitivity functions from acuity and letter contrast sensitivity measurements. J Vis. 2016;16(15):15. https://doi.org/10.1167/16.15.15.

    Article  PubMed  PubMed Central  Google Scholar 

  14. Cheng X, Bradley A, Ravikumar S, Thibos LN. Visual impact of Zernike and Seidel Forms of monochromatic aberrations. OPTOMETRY Vis SCI. 2010;87(5):300–12. https://doi.org/10.1097/OPX.0b013e3181d95217.

    Article  Google Scholar 

  15. Ginsburg AP. Contrast sensitivity: determining the visual quality and function of cataract, intraocular lenses and refractive surgery. CURR OPIN OPHTHALMOL. 2006;17(1):19–26. https://doi.org/10.1097/01.icu.0000192520.48411.fa.

    Article  PubMed  Google Scholar 

  16. Hawkins AS, Szlyk JP, Ardickas Z, Alexander KR, Wilensky JT. Comparison of contrast sensitivity, visual acuity, and Humphrey visual field testing in patients with glaucoma. J GLAUCOMA. 2003;12(2):134–8.

    Article  PubMed  Google Scholar 

  17. Xiao Y, Hu Y, Quan W, Zhang B, Wu Y, Wu Q, Liu B, Zeng X, Lin Z, Fang Y, et al. Machine learning-based prediction of anatomical outcome after idiopathic macular hole surgery. ANN TRANSL MED. 2021;9(10):830. https://doi.org/10.21037/atm-20-8065.

    Article  PubMed  PubMed Central  Google Scholar 

  18. Diao T, Kushzad F, Patel MD, Bindiganavale MP, Wasi M, Kochenderfer MJ, Moss HE. Comparison of machine learning approaches to improve diagnosis of Optic Neuropathy using photopic negative response measured using a Handheld device. FRONT MED-LAUSANNE. 2021;8. https://doi.org/10.3389/fmed.2021.771713.

  19. Rohm M, Tresp V, Müller M, Kern C, Manakov I, Weiss M, Sim DA, Priglinger S, Keane PA, Kortuem K. Predicting Visual Acuity by using machine learning in patients treated for Neovascular Age-Related Macular Degeneration. Ophthalmology. 2018;125(7):1028–36. https://doi.org/10.1016/j.ophtha.2017.12.034.

    Article  PubMed  Google Scholar 

  20. Wei L, He W, Wang J, Zhang K, Du Y, Qi J, Meng J, Qiu X, Cai L, Fan Q, et al. An optical coherence tomography-based deep learning algorithm for visual acuity prediction of highly myopic eyes after cataract surgery. FRONT CELL DEV BIOL. 2021;9. https://doi.org/10.3389/fcell.2021.652848.

  21. Odom JV, Bach M, Brigell M, Holder GE, McCulloch DL, Mizota A, Tormene AP. ISCEV standard for clinical visual evoked potentials: (2016 update). DOC OPHTHALMOL. 2016;133(1):1–9. https://doi.org/10.1007/s10633-016-9553-y.

    Article  PubMed  Google Scholar 

  22. Jia F, Liang Y, Zhang X, Luo F, Xiong Y, Cheng L, Liu J. Characteristics of Pattern Visual evoked potential in two eyes with varying visual acuity in One Eye and forensic application. CURR MED SCI. 2018;38(2):342–8. https://doi.org/10.1007/s11596-018-1885-2.

    Article  PubMed  Google Scholar 

  23. Kurita-Tashima S, Tobimatsu S, Nakayama-Hiromatsu M, Kato M. Effect of check size on the pattern reversal visual evoked potential. Electroencephalogr Clin Neurophysiol. 1991;80(3):161–6. https://doi.org/10.1016/0168-5597(91)90118-h.

    Article  CAS  PubMed  Google Scholar 

  24. Sacai PY, Salomão SR, Carelli V, Pereira JM, Belfort R, Sadun AA, Berezovsky A. Visual evoked potentials findings in non-affected subjects from a large brazilian pedigree of 11778 Leber’s hereditary optic neuropathy. DOC OPHTHALMOL. 2010;121(2):147–54. https://doi.org/10.1007/s10633-010-9241-2.

    Article  PubMed  Google Scholar 

  25. Gundogan FC, Sobaci G, Bayer A. Pattern visual evoked potentials in the Assessment of Visual Acuity in Malingering. Ophthalmology. 2007;114(12):2332–7. https://doi.org/10.1016/j.ophtha.2007.04.026.

    Article  PubMed  Google Scholar 

  26. Jeon J, Oh S, Kyung S. Assessment of visual disability using visual evoked potentials. BMC OPHTHALMOL. 2012;12(1):36. https://doi.org/10.1186/1471-2415-12-36.

    Article  PubMed  PubMed Central  Google Scholar 

  27. Denoyer A, Le Lez M, Majzoub S, Pisella P. Quality of vision after cataract surgery after Tecnis Z9000 intraocular lens implantation. J CATARACT REFR SURG. 2007;33(2):210–6. https://doi.org/10.1016/j.jcrs.2006.10.035.

    Article  Google Scholar 

  28. Yoon G, Jeong TM, Cox IG, Williams DR. Vision improvement by correcting higher-order aberrations with phase plates in normal eyes. J REFRACT SURG. 2004;20(5):523–S527. https://doi.org/10.3928/1081-597X-20040901-22.

    Article  Google Scholar 

  29. Rocha KM, Vabre L, Chateau N, Krueger RR. Enhanced visual acuity and image perception following correction of highly aberrated eyes using an adaptive Optics Visual Simulator. J REFRACT SURG. 2010;26(1):52–6. https://doi.org/10.3928/1081597X-20101215-08.

    Article  PubMed  Google Scholar 

  30. Yperman J, Becker T, Valkenborg D, Popescu V, Hellings N, Wijmeersch BV, Peeters LM. Machine learning analysis of motor evoked potential time series to predict disability progression in multiple sclerosis. BMC NEUROL. 2020;20(1). https://doi.org/10.1186/s12883-020-01672-w.

  31. Bach M, Heinrich SP. Acuity VEP: improved with machine learning. DOC OPHTHALMOL. 2019;139(2):113–22. https://doi.org/10.1007/s10633-019-09701-x.

    Article  PubMed  Google Scholar 

  32. Dotto PDF, Berezovsky A, Sacai PY, Rocha DM, Salomão SR. Gender-based normative values for pattern-reversal and flash visually evoked potentials under binocular and monocular stimulation in healthy adults. DOC OPHTHALMOL. 2017;135(1):53–67. https://doi.org/10.1007/s10633-017-9594-x.

    Article  PubMed  Google Scholar 

Download references

Acknowledgements

Not applicable.

Funding

This study was supported by Aviation Medicine Project of Xijing Hospital (2023) and Medicine Promotion Project of Air Force Medical University (No. 2018HKZL04, No. 2020ZTB01 and No. 2021JSTS28).

Author information

Authors and Affiliations

Authors

Contributions

Design of the study: CJZ, CT, WXC, ZZM; development of the algorithm: ZT; gathering the data: CJZ, LCC, LSH, SYT, ZT; performing the data analysis: CJZ, LCC, ZT, WYS, DGR; drafting the first version of the manuscript: CJZ, LCC; revision and approval of the manuscript: all authors.

Corresponding authors

Correspondence to Tao Chen, Xiao Cheng Wang or Zuo Ming Zhang.

Ethics declarations

Ethics approval and consent to participate

Ethics approval was granted by the Ethics Committee of Xijing Hospital, Air Force Medical University (No. KY20213088-1). All procedures in this study adhered to the tenets of the Declaration of Helsinki. Informed consent was obtained from all individual participants included in the study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Chen, J.Z., Li, C.C., Li, S.H. et al. A feasibility study for objective evaluation of visual acuity based on pattern-reversal visual evoked potentials and other related visual parameters with machine learning algorithm. BMC Ophthalmol 23, 293 (2023). https://doi.org/10.1186/s12886-023-03044-7

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s12886-023-03044-7

Keywords