Detecting vision loss in glaucoma using eye movement scanpaths recorded during free viewing of movies - a proof of principle study

Abstract

 

Purpose

 

Vision is typically assessed by asking people to respond to synthetic stimuli whilst maintaining steady fixation. We test the hypothesis that vision loss in glaucoma can be detected by examining patterns of eye movement recorded whilst a person naturally watches a movie.

 

Methods

 

Thirty-two elderly people with normal vision (median age: 70, interquartile range [IQR] 64 to 75 years) and 44 patients with a clinical diagnosis of glaucoma (median age: 69, IQR 63 to 77 years) had standard vision examinations including automated perimetry. All participants then viewed three unmodified TV and film clips on a computer set up incorporating the Eyelink 1000 eyetracker (SR Research, Ontario, Canada). Scanpaths were plotted using purpose-written software that first filtered the data and then generated saccade density maps (see Figure). Maps were then subjected to a feature extraction analysis using kernel principal component analysis (KPCA). Features from the KPCA were then classified using a standard machine based classifier. The classification was trained and tested by a 10 fold cross validation which was repeated 100 times to estimate the confidence interval (CI) of classification hit rate and specificity.

 

Results

 

Patients had a range of glaucoma disease severity (median worse eye MD of -11.7 dB, IQR -17.1 to -5.9 dB). Average hit rate for correctly identifying a glaucoma patient at a fixed specificity of ~90% was 79% (95% CI: 67 to 93%). The area under the Receiver Operating Characteristic curve was 0.84 (95% CI: 0.79 to 0.92).

 

Conclusions

 

Huge data from scanpaths of eye movements recorded whilst people freely watch videos can be processed into maps that contain a signature of vision loss. In this proof of principle study we have demonstrated that a group of patients with a range of glaucomatous vision loss can be reasonably well separated from a group of peers with normal vision by considering these eye movement signatures alone. Estimates of ‘diagnostic precision’ for this approach are similar to what would be found in using a single result from a modern imaging device.

 

 

Each row in this schematic illustrates a different stage in the process of generating the saccadic map for one video viewed by one person. Each column represents a different time point in the process. (The final saccade map is shown in the bottom right corner of the schematic).

 

Keywords: 522 eye movements • 758 visual fields