Purchase this article with an account.
Suman Sedai, Bhavna Antony, Hiroshi Ishikawa, Gadi Wollstein, Joel S Schuman; Forecasting Visual Field parameters at the Future visits using machine learning regression. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1465. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop and evaluate a machine learning regression method to forecast visual field mean deviation (MD) and visual field index (VFI) and compare with conventional trend based analysis (TBA) approach.
Optical coherence tomograph (OCT) scans were acquired from both eyes on 634 glaucoma patients, 404 glaucoma suspects and 49 healthy controls using commercial OCT device (Cirrus HD-OCT, 200x200 Optic Disc Cubes; Zeiss, Dublin, CA) over at least 4 visits (visit interval was 6 months). All subjects had visual field (VF) tests at each visit (Humphrey VF, SITA 24-2 test; Zeiss). A support vector regression (SVR) based forecasting model was developed utilizing clinical test results and OCT measurements at multiple visits including VF mean deviation, VF visual field index, circumpapillary RNFL thickness (both global and sectoral parameters), cup-to-disc ratio, cup volume, rim area, intraocular pressure, age and follow-up duration. Three consecutive visits were used for forecasting the MD and VFI at the 4th visit. The mean absolute error (MAE) was calculated to assess the forecasting performance and was statistically compared with TBA. TBA used 4 consecutive visits to linearly regress the MD and VFI against the duration of the follow-ups to forecast MD and VFI for the 5th visit.
Proposed machine learning regression based model outperformed TBA on MAE for both MD and VFI forecasting tasks regardless of the subject group; MAEs for MD forecasting task were (1.29 vs 0.69 DB, 1.17 vs 0.91 DB, and 1.27 vs 1.05 DB for healthy, glaucoma suspects, and glaucoma subjects, respectively, all p < 0.01). MAEs for VFI forecasting were (2.30 vs 2.05 %, 2.32 vs 1.64 %, and 3.65 vs 2.95 % for healthy, glaucoma suspects, and glaucoma subjects, respectively, all p<0.05). There was no significant difference in MAE of the both models among subgroups.
The proposed method showed improved accuracy in forecasting the MD and VFI in the future visit while using fewer number of visits baseline information compared to the conventional TBA. This may facilitate the personalized clinical management of glaucoma.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
This PDF is available to Subscribers Only