Skip to main content

  • Unsupervised AI System May Predict Rapid VF Loss

    By Lynda Seminara
    Selected by Russell N. Van Gelder, MD, PhD
    Add to My Bookmarks

    Journal Highlights

    Ophthalmology, December 2022

    Download PDF

    Recent advances in models of artificial intelligence (AI), coupled with greater availability of data, represent promis­ing objective tools to assess visual field (VF) data. Models of unsupervised learning do not require large clinically annotated datasets, nor do they need human expertise. Even so, it is crucial to assess these models with input from human experts to ensure their clinical relevance and utility. Yousefi et al. pro­posed an objective approach, based on machine-identified patterns of VF loss, to identify potential rapid glaucoma progression and subsequent vision loss. Their system was capable of identifying patterns of VF loss, which may pave the way for reproducible nomenclature to characterize early signs of visual defects and rapid glaucoma progression.

    For this cross-sectional longitudinal study, the researchers included 2,231 abnormal VFs from 205 eyes (176 patients) in the Ocular Hypertension Treatment Study (OHTS), which included follow-up of approximately 16 years.

    The VFs were assessed by an un­supervised deep archetypal analysis algorithm and by OHTS-certified VF readers. The 18 machine-identified patterns of glaucomatous damage were compared with the expert-identified patterns. To determine the extent and severity of glaucoma in eyes within each pattern cluster, the authors calcu­lated the average mean deviation (MD) for each cluster. Based on longitudinal data, the patterns of VF loss with strong correlation to glaucoma progression were documented. The main outcome measure was machine-expert agree­ment on the patterns of VF loss that signify rapid progression.

    According to the analysis, the aver­age VF MD at conversion to glaucoma was –2.7 dB (SD, 2.4 dB), and the average MD of eyes at their last visit was –5.2 dB (SD, 5.5 dB). Fifty (24.4%) of 205 eyes had an MD rate of –1 dB/year or worse and were deemed rapid progressors. The mean rate of MD decline in nonprogressing eyes was .2 dB/year (SD, +2.1). Thirteen of the 18 machine-identified patterns were simi­larly identified by the experts. The most common expert-identified patterns were partial arcuate, paracentral, and nasal-step defects. The most prevalent machine-identified patterns were tem­poral wedge, partial arcuate, nasal-step, and paracentral defects. After adjust­ment for covariates, one machine-iden­tified pattern was predictive of rapid VF decline. This pattern was present in 52% of fast-progressing eyes and 9% of nonprogressing eyes.

    With further refinements and larger datasets reflecting a wide range of severity, the authors believe that the system may aid in glaucoma manage­ment. 

    The original article can be found here.