Feasibility of simple machine learning approaches to support detection of non-glaucomatous visual fields in future automated glaucoma clinics.

Objectives: To assess the performance of feed-forward back-propagation artificial neural networks (ANNs) in detecting field defects caused by pituitary disease from among a glaucomatous population.

Methods: 24-2 Humphrey Visual Field reports were gathered from 121 pituitary patients and 907 glaucomatous patients. Optical character recognition was used to extract the threshold values from PDF reports. Left and right eye visual fields were coupled for each patient in an array to create bilateral field representations. ANNs were created to detect chiasmal field defects. We also assessed the ability of ANNs to identify a single pituitary field among 907 glaucomatous distractors.

Results: Mean field thresholds across all locations were lower for pituitary patients (20.3 dB, SD = 5.2 dB) than for glaucoma patients (24.4 dB, SD = 5.0 dB) indicating a greater degree of field loss (p < 0.0001) in the pituitary group. However, substantial overlap between the groups meant that mean bilateral field loss was not a reliable indicator of aetiology. Representative ANNs showed good performance in the discrimination task with sensitivity and specificity routinely above 95%. Where a single pituitary field was hidden among 907 glaucomatous fields, it had one of the five highest indexes of suspicion on 91% of 2420 ANNs.

Conclusions: Traditional artificial neural networks perform well at detecting chiasmal field defects among a glaucoma cohort by inspecting bilateral field representations. Increasing automation of care means we will need robust methods of automatically diagnosing and managing disease. This work shows that machine learning can perform a useful role in diagnostic oversight in highly automated glaucoma clinics, enhancing patient safety.