June 2021
Volume 62, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2021
Deep-learning based sequence modelling of the dark adaptation curve for noise reduction and parameter prediction in patients with AMD
Author Affiliations & Notes
  • Tharindu de silva
    National Eye Institute, Bethesda, Maryland, United States
  • Peyton Grisso
    National Eye Institute, Bethesda, Maryland, United States
  • Souvick Mukherjee
    National Eye Institute, Bethesda, Maryland, United States
  • Brett G Jeffrey
    National Eye Institute, Bethesda, Maryland, United States
  • Henry Wiley
    National Eye Institute, Bethesda, Maryland, United States
  • Tiarnan D L Keenan
    National Eye Institute, Bethesda, Maryland, United States
  • Emily Y Chew
    National Eye Institute, Bethesda, Maryland, United States
  • Catherine A. Cukras
    National Eye Institute, Bethesda, Maryland, United States
  • Footnotes
    Commercial Relationships   Tharindu de silva, None; Peyton Grisso, None; Souvick Mukherjee, None; Brett Jeffrey, None; Henry Wiley, None; Tiarnan Keenan, None; Emily Chew, None; Catherine A. Cukras, None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2021, Vol.62, 2824. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Tharindu de silva, Peyton Grisso, Souvick Mukherjee, Brett G Jeffrey, Henry Wiley, Tiarnan D L Keenan, Emily Y Chew, Catherine A. Cukras; Deep-learning based sequence modelling of the dark adaptation curve for noise reduction and parameter prediction in patients with AMD. Invest. Ophthalmol. Vis. Sci. 2021;62(8):2824.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Dark adaptation (DA) has shown to reveal early functional changes in eyes with age-related macular degeneration (AMD). This work aims to translate recent advances in deep learning sequence modelling to robustly estimate DA curves from sparse, noisy measurements and forecast late phase of the curve to reliably estimate parameters in shorter test times.

Methods : The data was collected in a clinical study involving AMD patients (NCT01352975). The DA test involved 82% focal bleach and recordings of log thresholds using a 3-down/1-up modified staircase threshold estimate procedure. 1496 DA curves were acquired from multiple annual study visits from 207 patients. During analysis, the sparse data points were interpolated within a 1.5 log unit reduction (ΔLU) and 3.1 ΔLU range using isotonic regression to obtain a monotonically-decreasing fit to the data. A recurrent neural network with long-short term memory (LSTM) autoencoder model was then devised to estimate the DA curve minimizing the curve fluctuations due to noise. The same model was then trained to forecast the latter part of the curve using the sequence of measurements up to different early portions of the curve ranging from 2 ΔLU – 2.9 ΔLU. The models were developed using 1271 train, 150 validation, 75 test curves. In each model, rod intercept time (RIT) was measured as the time required to reach 3 ΔLU and evaluated by comparing the model predicted RIT to ground truth.

Results : RIT measurements in the acquired data set varied between 3.6–39.9 min. When the entire curve was estimated from the model, the model successfully mitigated noisy fluctuations and the measured RIT had error (mean ± stdev) = 0.3±0.4 min (range=0.0–2.8 min) compared to ground truth. When the curve was forecasted at 2.5 ΔLU the predicted RIT had error = 1.2±1.3 min whereas at 2 ΔLU the predicted RIT error degraded to 3.1±3.1 min. Even at 2 ΔLU, 78.7% cases reported error < 5min.

Conclusions : LSTM autoencoder modelling provides a useful method to estimate and predict DA curve utilizing patterns of variation and mitigating noise. The ability to forecast from early phase of the DA curve suggests that the trends in the curve could help estimate the parameters from patients who are unable to complete the test and could lead to the development of shorter tests and identification of additional curve parameters to serve as useful endpoints.

This is a 2021 ARVO Annual Meeting abstract.

×
×

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.

×