• Machine Learning for Diagnosis in Automated Glaucoma Clinics

    By Lynda Seminara
    Selected By: Deepak P. Edward, MD

    Journal Highlights

    Eye
    2019;33(7):1133-1139

    Download PDF

    The demand for global glaucoma services is growing rapidly, driven by factors such as longer life expectancy, earlier diagnosis, and the plethora of available treatments. Machine learning is a powerful tool for analyzing patient data, and various computer algorithms can distinguish between glaucomatous signs and neurological field defects. In a glaucomatous population, Thomas et al. explored the utility of feed-forward back-propagation artificial neural networks (ANNs) for detecting field defects caused by pituitary disease. They found that by inspecting bilateral field representations, traditional ANNs are efficient in detecting chiasmal field de­fects. The findings suggest that machine learning may play a key “diagnostic oversight” role in the future.

    For this study, the researchers obtained 24-2 Humphrey visual field reports for 121 patients with pituitary disease and 907 patients with glau­comatous findings, and they used optical character recognition to extract threshold values from the reports. For each patient, left- and right-eye visual fields were coupled in an array to create bilateral field representations. ANNs were created to detect chiasmal field defects. The authors then assessed the networks’ ability to identify a single pituitary field among the 907 glauco­matous distractors.

    Results of the analyses showed that mean field thresholds in all locations were lower for the pituitary group (20.3 dB, standard deviation [SD] = 5.2 dB) than for the glaucoma group (24.4 dB, SD = 5.0 dB), suggesting that the degree of field loss is greater in patients with pituitary disease (p < .0001). However, the substantial overlap in results im­plies that mean bilateral field loss alone is not a reliable indicator of etiology.

    Overall, ANNs performed well in the discrimination task; sensitivity and specificity exceeded 95%. When a single pituitary field was hidden among 907 glaucomatous fields, it had one of the five top indexes of suspicion in 91% of the ANNs.

    The advent of telemedicine means that better methods are needed to au­tomatically identify not only glaucoma but also any accompanying pathology or masquerading conditions, the authors said. They added that ANNs such as theirs may have a role to play in the automated care of the future.

    The original article can be found here.