July 2019
Volume 60, Issue 9
Open Access
ARVO Annual Meeting Abstract  |   July 2019
Machine learning to identify multifocal ERG deficits in patients taking hydroxychloroquine.
Author Affiliations & Notes
  • Tom Wright
    Kensington Vision & Research Centre, Toronto, Ontario, Canada
    Ophthalmology and Vision Sciences, University of Toronto, Toronto, Ontario, Canada
  • Peng Yan
    Kensington Vision & Research Centre, Toronto, Ontario, Canada
    Ophthalmology and Vision Sciences, University of Toronto, Toronto, Ontario, Canada
  • Michael Easterbrook
    Ophthalmology and Vision Sciences, University of Toronto, Toronto, Ontario, Canada
  • Footnotes
    Commercial Relationships   Tom Wright, None; Peng Yan, None; Michael Easterbrook, None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science July 2019, Vol.60, 5959. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Tom Wright, Peng Yan, Michael Easterbrook; Machine learning to identify multifocal ERG deficits in patients taking hydroxychloroquine.. Invest. Ophthalmol. Vis. Sci. 2019;60(9):5959.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose : Machine learning algorithms are able to identify and analyse complex patterns in data. The multifocal electroretinogram (mfERG) provides both spatial and temporal information about macular function. In this paper we describe an approach using machine learning to detect hydroxychloroquine (hcq) associated retinal toxicity using the mfERG.

Methods : A two stage approach was used for mfERG analysis. First, a neural network was trained to encode temporal and amplitude information present in the mfERG waveforms. A second neural network is used to identify spatial information in the mfERG result. The first stage neural network was trained using 43432 waveforms from 356 subjects, 30% of the waveforms were randomly excluded from the training set for model validation. The second stage network was trained to identify patients with indicators of hydroxychloroquine toxicity based on previously published mfERG criteria. mfERG results referred for hcq toxicity screening (n=241) were manually labeled as demonstrating toxicity using a previously published ring ratio analysis. The second stage neural network was trained using 10-Fold cross validation.

Results : The first stage neural network encoding time and amplitude domain waveform information was able to closely match values calculated using a previously published algorithm (time: Pearsons r= 0.97, amplitude: r = 0.96, p < 0.001). Patients with mfERG toxicity were identified with 66% specificity and 70% sensitivity.

Conclusions : Our results show that neural networks can be designed to analyze the spatial and temporal information encoded in the multifocal electroretinogram and are able to approach the performance of current standard procedures. Deep learning strategies and ongoing training of neural networks with other clinical indicators of retinal toxicity will improve early detection of hydroxychloroquine toxicity, prevent vision loss and improve patient care.

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

×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×