Predictive Modeling of Long-Term Glaucoma Progression Based on Initial Ophthalmic Data and Optic Nerve Head Characteristics

Abstract

Purpose: The purpose of this study was to develop a model, based on initial optic nerve head (ONH) characteristics, predictive of long-term rapid retinal nerve fiber layer (RNFL) thinning in patients with open-angle glaucoma (OAG).

Methods: This study evaluated 712 eyes with OAG that had been followed up for >5 years with annual evaluation of RNFL thickness. Baseline ophthalmic features were incorporated into the machine learning models for prediction of faster RNFL thinning. The model was trained and tested using a random forest (RF) method, and was interpreted using Shapley additive explanations. Factors associated with faster rate of RNFL thinning were statistically evaluated using a decision tree.

Results: The RF model showed that greater lamina cribrosa (LC) curvature, higher intraocular pressure (IOP), visual field mean deviation converging towards -5 dB, and thinner peripapillary choroid at baseline were the four most significant features predicting faster RNFL thinning. Partial interaction between the features showed that larger LC curvature was a strong factor for faster RNFL thinning when it exceeded approximately 12.0. When the LC curvature was ≤12, higher initial IOP and thinner peripapillary choroid played a role in the rapid RNFL thinning. Based on the decision tree, higher IOP (>26.5 mm Hg), greater laminar curvature (>13.95), and thinner peripapillary choroid (≤117.5 µm) were the 3 most important determinants affecting the rate of RNFL thinning.

Conclusions: Baseline ophthalmic data and ONH characteristics of patients with OAG were predictive of eyes at risk of faster progression. Combinations of important characteristics, such as IOP, LC curvature, and choroidal thickness, could stratify eyes into groups with different rates of RNFL thinning.

Translational relevance: This work lays the foundations for developing prediction models to estimate glaucoma prognosis based on initial ONH characteristics.

Figure 1.

Interpretation of the final model based on baseline patient variables. (A) Feature importance plot based on mean SHAP values. (B) Interpretation of the importance of features using the SHAP plot. The red and blue colors indicate feature values of high and low levels, respectively. For example, a larger mean LCCI had a strong and negative impact on the rate of RNFL thinning (i.e. faster RNFL thinning). LCCI, lamina cribrosa curve index; IOP, intraocular pressure; VF, visual field; MD, mean deviation; CT, choroidal thickness; AXL, axial length; PSD, pattern standard deviation.

Figure 2.

Partial dependence plots showing SHAP values versus feature values for the four most important variables. (A) Highest IOP during initial 6 months, (B) mean LCCI, (C) VF MD, and (D) global CT. IOP, intraocular pressure; LCCI, lamina cribrosa curve index; VF, visual field; MD, mean deviation; CT, choroidal thickness.

Figure 3.

Partial interaction plots showing interactions between (A) initial IOP and mean LCCI and (B) global CT and mean LCCI. Overall, larger mean LCCI and larger initial IOP A, and larger mean LCCI and smaller global CT B together contributed to faster rates of RNFL thinning. However, in the region where mean LCCI was >12, the rate of RNFL thinning was mainly dependent on the mean LCCI, whereas higher IOP A and thinner global CT B were main contributors when the mean LCCI was <12. IOP, intraocular pressure; LCCI, lamina cribrosa curve index; CT, choroidal thickness.

Figure 4.

Decision tree model stratifying groups with faster or slower glaucoma progression based on factors best explaining the rate of RNFL thinning. IOP, intraocular pressure; LCCI, lamina cribrosa curve index; CT, choroidal thickness.