Development of An Automated Electroretinography Analysis Approach

Abstract

Purpose : Electroretinography (ERG) is used to assess retinal function in the clinic and animal models of ocular disease. However, there can be inter-observer variability in data analysis, and analysis is often time-intensive due to the number of flash intensities, animals, and timepoints in a study. We developed an automated approach to perform non-subjective and repeatable feature identification (“marking”) of the ERG waveform, and compared it to manual marking.

Methods : The automated approach pre-processed measured waveforms through a wavelet filter to remove noise, low-pass filtered (60 Hz) the output, and marked the b-wave (highest positive peak). The a-wave was marked as the initial large negative response after the flash stimuli. For oscillatory potential (OP) markings, the waveform was bandpass filtered (60-235 Hz) and the initial OP wavelet (OP1) was marked either by using a second derivative inflection point criterion or taking advantage of the a-wave implicit time. The remaining OP waves (up to OP5) were then sequentially labelled. We tested our approach on ERG data sets from 2 animal cohorts. 1) ERGs with 8 dark-adapted and 3 light-adapted stimuli in Brown-Norway rats from -6 to 1.5 log cd-s/m². 2) ERGs with a 5 step dark-adapted protocol (-3 to 2 log cd-s/m²) on Long Evans rats divided into control or diabetic (DM) groups. DM was induced with streptozotocin (STZ;100 mg/kg) and ERGs were recorded at 4 weeks of DM.

Results : The automated approach showed good correlation with manual markings of the major ERG features, including amplitue of the a-wave (r² = 0.99) and b-wave (r² = 0.99) and the respective timing of the a-wave (r² = 0.96) and b-wave (r² = 0.90). In the second cohort, we did not find a significant difference between the automated and manual marking approaches (p=0.65), with both approaches detecting differences between control and DM groups, e.g. longer implicit times in DM animals (Fig 1). The automated program took 45 seconds to complete ERG marking, reducing the analysis time 40 fold.

Conclusions : These results provide initial verification of our automated ERG analysis. Next steps include expanding our datasets to ensure accuracy in various animal models of ocular pathologies and human ERGs. Further, our approach allows us to include advanced ERG analysis tools, including PII and PIII modeling, and power analysis of the OPs.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.