Deep neural network for glaucoma detection using contact lens sensor data

Glaucoma is a common disease occurring worldwide that is a significant public health challenge. Substantial efforts have been made to improve efficacy of non-invasive diagnostic methods based on retinal fundus images or optical coherence tomography (OCT). However, these common eye imaging techniques do not provide conclusive diagnostic information at the early stages of the disease. This paper presents our work concerning the design of a deep neural network model for glaucoma detection using data acquired by an alternative method: Triggerfish contact lens sensor (CLS). CLS device provides a way to monitor relative ocular volume changes during a 24-hour session that forms the basis for model inference. Specific patterns in this data may indicate the risk of glaucoma development or progression. Our model achieved ROC AUC of 0.89 estimated using cross-validation on a small and noisy dataset, consisting of 139 cases, outperforming other approaches used for this type of data. Apart from that, we achieved the shortest patients’ engagement time, effectively shortening the time required for a patient to wear the sensor in order to receive a diagnosis. The proposed solution could be used as a preliminary glaucoma screening procedure or provide supplementary insights for comprehensive diagnosis of the disease.