Machine learning may predict optimal treatment regimen for patients with nAMD

At ARVO 2022, Carlos Quezada Ruiz, MD, senior medical director at Genentech, presented his talk entitled, “Predicting optimal treatment regimen for patients with neovascular age-related macular degeneration using machine learning.”

This transcript has been lightly edited for clarity.

This year at ARVO, we are presenting our data on a pilot study looking at machine learning and its ability to predict treatment regimen for patients with neovascular AMD. We believe that personalized health care and personalized treatment for patients with neovascular MD is the way that we are going to be treating patients and the retina in the future.

And at Genentech Roche, we are committed to better understanding how machine learning fits into this picture and how it can help us evolve into getting to that personalized approach for patient treatment and patient counseling as well, which is so important.

In this work, we evaluated 324 patients who were enrolled in the AVENUE and the STAIRWAY trial, and who received ranibizumab or faricimab for the treatment of neovascular AMD. We evaluated the baseline characteristics of those 324 patients, which included age, sex baseline BCVA, CSD treatment regimen arm. We also included image-derived characteristics from the spectral domain OCTs of all of those patients. We then implemented different machine learning algorithms. In the results, we actually realize that one of them, XG boost, was the one who came with the with the highest R square. This is basically the highest prediction rate for treatment interval for these patients.

Now, what this means for patients is basically that if this algorithm is able to work, in the future we may be able to predict at baseline during that initial visit how a patient might respond over time with regards to BCVA response, but also what would be the most appropriate treatment regimen, meaning the least frequent regimen that would give the best BCVA outcomes for that individual patient.

And while we recognize this is a pilot study of only a small proportion of patients, we do think that we need bigger datasets and more diverse datasets as well to better replicate these results and be able to extrapolate it to a broader population as well.