June 2023
Volume 64, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2023
Improving Glaucoma Trials Using Deep Learning to Forecast VF Variability
Author Affiliations & Notes
  • Jithin Yohannan
    Ophthalmology, Johns Hopkins University, Baltimore, Maryland, United States
    Malone Center for Engineering in Healthcare, Johns Hopkins University, Baltimore, Maryland, United States
  • Ruolin Wang
    Malone Center for Engineering in Healthcare, Johns Hopkins University, Baltimore, Maryland, United States
  • Kaihua Hou
    Malone Center for Engineering in Healthcare, Johns Hopkins University, Baltimore, Maryland, United States
  • Patrick Herbert
    Malone Center for Engineering in Healthcare, Johns Hopkins University, Baltimore, Maryland, United States
  • Gregory Hager
    Malone Center for Engineering in Healthcare, Johns Hopkins University, Baltimore, Maryland, United States
  • Mathias Unberath
    Malone Center for Engineering in Healthcare, Johns Hopkins University, Baltimore, Maryland, United States
  • Pradeep Y Ramulu
    Ophthalmology, Johns Hopkins University, Baltimore, Maryland, United States
  • Chris Bradley
    Ophthalmology, Johns Hopkins University, Baltimore, Maryland, United States
  • Footnotes
    Commercial Relationships   Jithin Yohannan Topcon, Ivantis, Abbvie, Genetech, Code C (Consultant/Contractor); Ruolin Wang None; Kaihua Hou None; Patrick Herbert None; Gregory Hager None; Mathias Unberath None; Pradeep Ramulu None; Chris Bradley None
  • Footnotes
    Support  NIH 1K23EY032204-02, Unrestricted Funding from Research to Prevent Blindness
Investigative Ophthalmology & Visual Science June 2023, Vol.64, 343. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Jithin Yohannan, Ruolin Wang, Kaihua Hou, Patrick Herbert, Gregory Hager, Mathias Unberath, Pradeep Y Ramulu, Chris Bradley; Improving Glaucoma Trials Using Deep Learning to Forecast VF Variability. Invest. Ophthalmol. Vis. Sci. 2023;64(8):343.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : The FDA has required that neuroprotective drugs must show differences in VF loss between treatment and control groups. However, high VF variability makes this challenging. We develop a deep learning model (DLM), to forecast eyes that will have low future VF variability using data from as little as one baseline VF, OCT and clinical visit. We conduct simulations to study the impact of using this DLM on sample size.

Methods : We included one eye per patient with a baseline set of reliable VF, OCT and clinical measures (demographics, IOP, visual acuity) and 5 subsequent reliable VFs. We used these data to perform the simulations and develop the DLM described below.

For sample size simulations, the outcome was mean deviation (MD) slope (derived from linear mixed effects models) between treatment and control groups. We estimated sample size for three different groups of eyes: all eyes (AE), low variability eyes (LVE), and DLM-predicted (DLP) low variability eyes. LVE were defined as the subset of AE with a standard deviation of MD slope resiuals in the bottom 25th percentile. DLP was defined as the subset of AE that was predicted to be low variability by our DLM models.

DLP1 eyes were identified using a vision transformer based DLM that took baseline VF/OCT/Clinical data as input and output the probability LVE. DLP2 eyes were identified in a similar manner with but a second VF (after baseline) was also included in DLM2 input. We compared DLM performance to logtistic regression and neural network models. Data were split 60/10/30 into train/val/test. Simulations were performed only on the test set.

We estimated the sample size necessary to detect treatment effects of 20%-50% with a power of 80%. Power was defined as the percentage of simulated clinical trials where the MD slope was significantly different from control. Clinical trials were simulated with visits every 3 months for a course of 10 visits total.

Results : 2,817 eyes were included in the dataset used to conduct the simulations and create the DLM. DLM1 and DLM2 achieved an AUC of 0.732 (95% CI: 0.68, 0.76) and 0.82 (95% CI: 0.78, 0.85) and performed better than non-DLM models (Table 1, p<0.05). When compared to including AE, including DLP eyes reduced sample size requirements for various treatment effects (Fig 1).

Conclusions : DLMs can forecast LVE. This can reduce sample size requirements and has the potential to reduce the burden of future glaucoma clinical trials.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

 

×
×

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.

×