Forecasting future Humphrey Visual Fields using deep learning

• DOI: 10.1371/journal.pone.0214875
• Publication Date: 2019-04-05T17:28:24Z
• Authors: Joanne C. Wen; Cecilia S. Lee; Pearse A. Keane; Sa Xiao; Ariel S. Rokem; Philip P. Chen; Yue Wu; Aaron Y. Lee
• Keywords: Decibel; Transfer of learning

Purpose: To determine if deep learning networks could be trained to forecast a future 24-2 Humphrey Visual Field (HVF). Participants: All patients who obtained HVF at the University of Washington. Methods: datapoints from consecutive HVFs 1998 2018 were extracted Washington database. Ten-fold cross validation with held out test set was used develop three main phases model development: architecture selection, dataset combination and time-interval training transfer learning, train artificial neural network capable generating point-wise visual field prediction. Results: More than 1.7 million perimetry points hundredth decibel 32,443 HVFs. The best performing 20 trainable parameters, CascadeNet-5, selected. overall MAE for 2.47 dB (95% CI: 2.45 2.48 dB). 100 fully models able successfully predict progressive loss in glaucomatous eyes up 5.5 years correlation 0.92 between MD predicted actual (p < 2.2 x 10 -16 ) an average difference 0.41 dB. Conclusions: Using unfiltered real-world datasets, show impressive ability not only learn spatio-temporal changes but also generate predictions years, given single HVF.