The Use of Machine Learning to Forecast Progression to Advanced AMD

There is a need for more comprehensive prediction models for advanced age-related macular degeneration (AMD) that consider a wider range of risk factors. Researchers tested a prediction model and applied a machine learning algorithm that autonomously identified the most significant clinical, genetic, and lifestyle risk factors for AMD.

The training set, obtained from the Rotterdam Study I (RS-I), included 3,838 patients aged 55 years or older. Median follow-up was 10.8 years, and there were 108 incident cases of advanced AMD. The test set, obtained from the ALIENOR study, included 362 participants aged 73 years or older. Median follow-up was 6.5 years, and there were 33 incident cases of advanced AMD.

The following factors were retained by the prediction model:

  • Age
  • Phenotypic predictors, per the presence of intermediate drusen, hyper-pigmentation in one or both eyes, and age-related eye disease study simplified score
  • A summary genetic risk score, per 49 single nucleotide polymorphisms
  • Smoking
  • Diet quality
  • Education
  • Pulse pressure

In the RS-I group, the cross-validated area under the receiver operating characteristic curve (AUC) estimation was: at five years, 0.92; at 10 years, 0.92; and at 15 years, 0.91. In the ALIENOR cohort, at five years, the AUC was 0.92. The researchers noted that when it came to calibration, the prediction model underestimated the cumulative incidence of advanced AMD in high-risk groups; this was particularly evident in the ALIENOR cohort.

They concluded that their prediction model achieved “high discrimination abilities” and was a step toward precision medicine for patients with AMD.