Purchase this article with an account.
Aushim Kokroo, Hiroshi Ishikawa, Mengfei Wu, Yu-Ying Liu, James Rehg, Gadi Wollstein, Joel S Schuman; Prediction Performance of a Trained Two-Dimensional Continuous Time Hidden Markov Model for Glaucoma Progression. Invest. Ophthalmol. Vis. Sci. 2018;59(9):4076.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
We previously described the two-dimensional continuous time hidden Markov model (2D CT-HMM) to model glaucoma progression using structural and functional measurements simultaneously. The purpose of this study was to validate the glaucoma progression prediction performance of a previously trained model on data collected from a different cohort.
A 2D CT-HMM was trained using optical coherence tomography (OCT; Cirrus HD-OCT, Zeiss, Dublin, CA) mean circumpapillary retinal nerve fiber layer (cRNFL) thickness and visual field index (VFI; Humphrey Field Analyzer, Zeiss) obtained from 107 eyes of 107 subjects, including glaucoma and glaucoma suspect. Average observation period was 4.2 years (7.1 visits). Approximately 1 year of longitudinal data were collected from a separate cohort. 78 eyes of 39 subjects, glaucoma and glaucoma suspect, with an average of 2.2 ± 0.4 visits were included. After matching the distribution of OCT and VF data on the training cohort, 19 eyes from 14 subjects were selected. The previously trained model was tested on these cases. One visit was used as an input to the model to predict the state at the next visit at least 6 months later, with 4 possible state changes (stable, OCT, VF, or OCT+VF progression). The percentage of correct prediction against the actual recorded state was reported as the prediction accuracy.
Baseline age of the test cohort was 58.4 ± 13.9 years, VFI 93.6 ± 8.3, mean cRNFL thickness 74.0 ± 10.9μm. Figure 1 shows the trained model. The size of the circle (state) shows the number of subjects passing through the state. The grayscale of the state indicates the length of time spent there, increasing white to black. Lines indicate state changes, with the blue line being the most likely. This information is also shown in numerical form. The inset shows an example of model use. The calculated prediction accuracy of this pre-trained 2D CT-HMM on test data was 52.6%.
Although the glaucoma progression prediction performance of the trained 2D CT-HMM was slightly lower than that previously reported, it is acceptable given the training and testing cohorts were different, and it exceeds the random chance of making a correct prediction, 25%. Furthermore, unlike conventional methods, this model requires only one visit as an input, which makes it a potentially useful tool in the clinical prediction of glaucoma progression.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.
This PDF is available to Subscribers Only