Machine Learning Proven Beneficial for Diagnosing, Monitoring wAMD

Peter M. Maloca, MD

Support vector machine learning (SVML) can be a useful tool in monitoring wet age-related macular degeneration (wAMD), proving to be as effective as retina specialists in detecting wAMD activity features.

A team of European investigators conducted a retrospective pilot study in order to demonstrate the ways to implement and use SVML algorithms in a small, three-dimensional sample. They tested 588 consecutive pairs of optical coherence tomography (OCT) volumes selected randomly from 70 randomly-chosen patients with wAMD. Study authors noted that patients were treated with ranibizumab and had a mean age of 80.3 years.

Investigators employed 4 independent, diagnosis-blinded retina specialists to indicate whether wAMD was activity was present between 100 pairs of consecutive OCT volumes in the rest of the 40 patients for comparison with the SVML algorithm. They also compared a non-complex baseline algorithm that used only retinal thickness. The SVML algorithm was assessed using inter-observer variability and receiver operating analyses.

“For me it is very exciting to see how we cross the threshold from the ‘analog data recording and storing’ to what I call ‘dead data graveyards’ towards my vision of a self-measuring and self-communicating retina analysis system that works worldwide and reduces the physical barriers and democratizes the technology for the benefit of mankind,” Peter M. Maloca, MD, told MD Magazine®.

The algorithm produced either ‘activity’ or ‘no activity’ binary outputs, which was then compared to the retina specialists’ observations. The retina specialists spent a mean of 18.5 seconds per pair of consecutive OCT volumes before making an ‘activity’ or ‘no activity’ decision, the study authors wrote. The SVML algorithm completed this step in 16.1 seconds.

The investigators wrote that the signs of wAMD activity were evaluated by the clinic and described as the “ground truth,” where 40% of pairs of eyes showed activity and 60% showed no activity. The SVML algorithm showed 41% activity and 59% activity, compared to the 4 retina specialists who showed 38%, 27%, 49%, and 30% activity and 62%, 73%, 51%, and 70% no activity, respectively.

Despite the success of the proposed SVML algorithm, machine-based learning has not been frequently reported or discussed in the literature and is likely unfamiliar to the retina specialist community, the study authors wrote. SVML could one day eliminate the need for a retina specialist to read all OCT images, which would advance efficiency, that capability might be the main reason that the community of retina specialists would be restrained in applying the algorithm to clinical practice at first, the study authors suggested.

“We are on the threshold of hyper-medicine, where we as subjects become part of the data loop,” Maloca continued. “The future physician will be shaped and successfully work for his patients if he intensively uses such new technologies as a companion that raises his level and he finally chooses together with his patient the best option for a therapy and monitoring. The training to become a physician will change fundamentally.”

The study authors also wrote that the need for telehealth is pushed forward by the “mounting pressure to serve more patients with chronic eye disease in an increasingly resource-constrained environment.”

However, they said, despite barriers, telehealth could be a way to simplify procedures, deliver services more efficiently and control costs better.

This study,“Feasibility of support vector machine learning in age-related macular degeneration using small sample yielding sparse optical coherence tomography data,” was published online in Acta Ophthalmologica.