Purpose: Our previous study predicted genuine glaucomatous visual field (VF) impairment in the central 10° VF, excluding the effect of cataract, using visual acuity (VA) and global indexes of VF more accurately than pattern deviation (PD). This study aimed to improve the accuracy by using pointwise total deviation (TD) values with the machine-learning method of random forest model (RFM) and to investigate whether incorporating optical coherence tomography-measured ganglion cell-inner plexiform layer (GCIPL) thickness is useful.
Methods: This retrospective study included 89 eyes with open-angle glaucoma that underwent successful cataract surgery (with or without iStent implantation or ab interno trabeculotomy). Postoperative TD in each of the 68 VF points was predicted using preoperative (1) PD, (2) VA and VF with a linear regression model (LM), and (3) VA and VF with RFM, and averaged as predicted mean TD (mTDpost). Further prediction was made by incorporating the preoperative GCIPL into the best model.
Results: The mean absolute error (MAE) between the actual and predicted mTDpost with RFM (1.25 ± 1.03 dB) was significantly smaller than that with PD (3.20 ± 4.06 dB, p < 0.01) and LM (1.42 ± 1.06 dB, p < 0.05). The MAEs with the model incorporating GCIPL into RFM (1.24 ± 1.04 dB) and RFM were not significantly different.
Conclusions: Accurate prediction of genuine glaucomatous VF impairment was achieved using pointwise TD with RFM. No merit was observed by incorporating the GCIPL into this model.
Translational relevance: This pointwise RFM could clinically reduce cataract effect on VF.