# Larger Real-World OCT Reference Database Improves Accuracy of Glaucoma Flagging Using Summary Metrics

## TL;DR

A larger real-world OCT reference database improves the accuracy of identifying glaucoma by reducing random errors compared to a smaller commercial database.

## Contribution

Demonstrates that a larger real-world OCT reference database reduces flagging discrepancies and improves glaucoma detection accuracy.

## Key findings

- 16% of ON-G eyes had different flags for g-cpRNFL between the two databases.
- 7% of ON-G eyes had different flags for g-GCL+ between the two databases.
- The difference in flagging is attributed to reduced random error in the larger real-world database.

## Abstract

To compare the healthy control and glaucoma eyes flagged as outside normal limits (yellow or red) by a commercial reference database (C-RDB) and a larger real-world (RW)–RDB.

The C-RDB consisted of 398 eyes/individuals. Based only on optical coherence tomography (OCT) reports, the RW-RDB consisted of 4830 eyes/individuals selected from optometry practices using a reading center method. The fifth and first percentile quantile regression lines (QRLs) versus age were calculated for both RDBs for common OCT metrics, including global circumpapillary retinal nerve fiber (g-cpRNFL) and global ganglion cell layer plus inner plexiform layer (g-GCL+) thickness. The test dataset contained 175 healthy control (H) eyes and 183 eyes with OCT defects consistent with optic neuropathy–glaucoma (ON-G). These eyes were flagged as yellow (red) if they fell below the fifth (first) percentile QRL. The QRLs were also compared to a Gaussian model and Monte Carlo simulations.

The C-RDB and RW-RDB did not flag an identical set of eyes as red or yellow. In fact, 16% (g-cpRNFL) and 7% (g-GCL+) of the 183 ON-G eyes had a different color flag. The results of the model and simulations support the hypothesis that both RDBs are sampled from essentially the same underlying “normal” population. Thus, the difference between them is largely due to less random error in the larger sample.

There is a difference in the eyes flagged, and this difference is largely due to the greater size of the RW-RDB.

These findings support the clinical value of expanding reference databases to improve diagnostic accuracy of glaucoma flagging.