June 2022
Volume 63, Issue 7
Open Access
ARVO Annual Meeting Abstract  |   June 2022
Disease activity (DA) monitoring in neovascular age-related macular degeneration (nAMD) with artificial intelligence (AI): a feasibility study
Author Affiliations & Notes
  • Pearse Andrew Keane
    NIHR Biomedical Research Centre for Ophthalmology at Moorfields Eye Hospital, United Kingdom
    UCL Institute of Ophthalmology, London, United Kingdom
  • Jayashree Sahni
    Novartis Pharmaceuticals AG, Switzerland
  • Zufar Mulyukov
    Novartis Pharmaceuticals AG, Switzerland
  • Sandra Liakopoulos
    Department of Ophthalmology, Faculty of Medicine and University Hospital Cologne, Germany
    Department of Ophthalmology, Goethe University, Frankfurt, Germany
  • Katja Hatz
    Vista Klinik AG Binningen, Switzerland
    Universitat Basel Medizinische Fakultat, Basel, Basel-Stadt, Switzerland
  • Daniel Ting Shu Wei
    Yong Loo Lin School of Medicine, National University of Singapore, Singapore
  • Roberto Gallego Pinazo
    Oftalvist Clinic, Valencia, Spain
  • Daniel Lorand
    Novartis Pharmaceuticals AG, Switzerland
  • Footnotes
    Commercial Relationships   Pearse Keane Big Picture Medical, Code I (Personal Financial Interest), DeepMind, Code P (Patent), Novartis, DeepMind, Roche, Apellis, Bitfount, Heidelberg Engineering, Topcon, Allergan, Bayer, Code R (Recipient); Jayashree Sahni Novartis, Code E (Employment); Zufar Mulyukov Novartis, Code E (Employment); Sandra Liakopoulos Novartis, Appellis, Code C (Consultant/Contractor), Novartis, Appellis, Allergan, Alcon, Bayer, Heidelberg Engineering, Zeiss, Code R (Recipient); Katja Hatz Novartis, Roche, Allergan, Bayer, Code R (Recipient); Daniel Ting Shu Wei EyRIS Pte Ltd, Code P (Patent), Novartis, Code R (Recipient); Roberto Gallego Pinazo Novartis, Boehringer Ingelheim, Carl Zeiss Meditec, ORA Clinical, Roche, Heidelberg Engineering, Bloss Group, Fishawack, Code R (Recipient); Daniel Lorand Novartis, Code E (Employment)
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2022, Vol.63, 2997 – F0267. doi:
  • Views
  • Share
  • Tools
    • Alerts
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Pearse Andrew Keane, Jayashree Sahni, Zufar Mulyukov, Sandra Liakopoulos, Katja Hatz, Daniel Ting Shu Wei, Roberto Gallego Pinazo, Daniel Lorand; Disease activity (DA) monitoring in neovascular age-related macular degeneration (nAMD) with artificial intelligence (AI): a feasibility study. Invest. Ophthalmol. Vis. Sci. 2022;63(7):2997 – F0267.

      Download citation file:

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

  • Supplements

Purpose : DA assessed using optical coherence tomography (OCT) imaging and visual parameters is important for monitoring nAMD. We developed an AI-based DA model to support physicians who treat patients with nAMD with anti-vascular endothelial growth factor (VEGF) agents. We assessed the performance of the DA model in deriving a DA score in an adjudication study.

Methods : Measurements of OCT features and best-corrected visual acuity collected from patients evaluated with Heidelberg spectral domain (SD)-OCT during the phase 3 brolucizumab versus aflibercept HAWK (NCT02307682) & HARRIER (NCT02434328) (H&H) studies were used to develop an AI-based DA model. A retrospective review of selected DA assessments from the H&H dataset was conducted with an independent panel of 10 retina specialists (RS) to further train the model. DA was deemed to be present if the RS indicated treatment was needed and absent if no treatment was indicated. Three categories of DA assessment were defined using model-derived DA score and prediction uncertainty (Table 1).

Results : To train the DA model, 8970 longitudinal DA assessments from a total of 948 patients evaluated with Heidelberg SD-OCT by H&H study investigators were used. This model achieved a cross-validated accuracy of 0.84 (0.87 sensitivity, 0.84 specificity). The RS panel reviewed 486 selected DA assessments from 403 patients. Figure 1 shows the accuracy, sensitivity, and specificity achieved by RS, H&H investigators, and DA model for three categories of DA assessment. The DA model reached 0.94 accuracy and performed as well as RS and H&H investigators for ‘easy’ DA assessments. While the DA model outperformed H&H investigators for both ‘potential label noise’ (accuracy 0.67 vs 0.38) and ‘difficult’ DA assessments (accuracy 0.74 vs 0.58), the DA model underperformed the RS in these categories (mean RS accuracy 0.86 for ‘difficult’ and 0.82 for ‘potential label noise’ assessments).

Conclusions : In this study, we demonstrated encouraging performance of a model in detecting DA in nAMD patients treated with anti-VEGF agents. While further refinements and improvements are planned, these preliminary results suggest the potential of an AI-based DA algorithm in improving the consistency of treatment decisions based on DA assessments.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.


Table 1. Categories of DA assessments

Table 1. Categories of DA assessments


Figure 1. Performance

Figure 1. Performance


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.