Development and Evaluation of an Artificial Intelligence Model to Set Target IOP for Glaucoma.

PURPOSE

Develop an artificial intelligence (AI) model to predict a personalized target intraocular pressure (IOP) for eyes with glaucoma. Compare the impact of achieving the AI-predicted, clinician-defined, and society guideline-based target IOP on rates of visual field (VF) worsening.

DESIGN

Development and evaluation of an AI treatment algorithm

SUBJECTS

A dataset of 14,871 eyes with a defined target IOP (set by glaucoma specialists during routine clinical care) and baseline OCT, VF, and clinical measurements. A non-overlapping progression dataset of 10,559 eyes with longitudinal VF testing (≥ 5 reliable tests).

METHODS

We trained and tested a machine-learning model on the dataset of eyes with baseline structural, functional, and clinical data to predict a target IOP for each eye in the progression dataset. Linear models estimated the effect of mean target difference (measured IOP - target IOP) on the rate of VF worsening, defined by mean deviation (MD) slope. The effect of deviations from AI targets was compared to deviations from clinician-set targets and targets from society-based guidelines by the Canadian Ophthalmological Society (COS). Linear models also estimated the effect of mean absolute IOP on the rate of VF worsening. The effect of mean target difference (AI, clinician, or COS) was compared to the effect of mean absolute IOP.

MAIN OUTCOME MEASURES

Effect of 1 mm Hg increase in mean target difference (AI, clinician, or COS) on the rate of MD worsening (dB/year).

RESULTS

The AI model achieved a mean absolute error of 2.28 mm Hg for predicting target IOP. AI and clinician target differences had similar effects on VF outcomes (0.032 dB/year vs 0.026 dB/year faster rate of MD worsening per 1 mm Hg increase, respectively, p = 0.09). AI and clinician target difference had greater effects than COS target difference (0.011 dB/year faster per 1 mm Hg increase, p < 0.001) and mean absolute IOP (0.001 dB/year faster per 1 mm Hg increase, p < 0.001).

CONCLUSIONS

AI models can set target IOP with comparable performance to glaucoma specialists and are superior to utilizing mean IOP or society-based guidelines to set targets. Further work is needed to assess the clinical impact of AI-based target IOP guidance for patients managed by non-glaucoma specialists.