Purchase this article with an account.
Vuong Nguyen, Kelvin Teo, Judy Simpson, Vincent Daien, Daniel Barthelmes, Mark C Gillies, Richard Walton; A Probabilistic Forecast of 12 Month Visual Outcomes of nAMD. Invest. Ophthalmol. Vis. Sci. 2018;59(9):829.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To assess the potential of a large registry of patients with neovascular age-related macular degeneration (nAMD) to predict 12 month visual acuity (VA) outcomes and communicate uncertainty in our predictions in terms of probabilities.
Data were obtained from the Fight Retinal Blindness! registry, currently containing 12 month data for at least 4000 eyes. We used Bayesian regression models and random forest machine learning to predict 12 month VA using baseline characteristics including age, angiographic lesion type and size, and visual acuity. Predictions were expressed in terms of probabilities rather than single point estimates. Predictive performance was evaluated using cross-validation and diagnosed both calibration and sharpness.
An example of a probabilistic prediction from the random forest model is shown in Figures 1 and 2. Communicating the likelihood of achieving certain outcomes can be expressed using the survival function (complementary cumulative distribution function). For example, the patient described in Figure 1 has a 75% chance of not losing further vision and a 46% chance of achieving a VA≥70 letters at 12 months.Probability integral transform histograms suggested the predictions from the Bayesian regression model were overdispersed, while the random forest predictions were better calibrated. Predictions from random forest were also sharper; the average width of the 50% prediction intervals from cross-validation was 14.5 for random forest and 21.3 for random forest.
Integrating a predictive model into data registries to show likely outcomes for patients at baseline can be a valuable tool for patients and clinicians. Communicating these outcomes via probabilities rather than single point predictions can account for uncertainties and the possibility of negative outcomes. The random forest machine learning algorithm outperformed the Bayesian posterior in our data. Further investigations into machine learning algorithms for predicting 12 month outcomes and updating models with the 3 month VA will be explored.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.
Figure 1. Predictive probability density from random forest of a 72 year old patient with minimally classic lesion of size 3000µm and baseline VA of 60 letters. This patient went on to achieve 80 letters at 12 months.
Figure 2. Survival function for the patient described in Figure 1.
This PDF is available to Subscribers Only