Autonomous AI-based System Compliant with Expert Consensus for AMD Treatment

Published on: 

Study data indicate the AI-based follow-up system could make an autonomous decision in 73% of the cases, of which 91.8% were in agreement with retina experts consensus.

New findings suggest the potential of a novel artificial intelligence-based system for autonomous follow-up of patients treated for neovascular age-related macular degeneration (nAMD).

The research demonstrated both its safety and compliance with expert consensus as being on par with decisions made in clinical practice, with particularly low rates of false-positive classification of choroidal neovascularization (CNV) activity.

“Over half of the eyes with AMD could be followed autonomously without additional risk to the patient,” wrote study author Ivan Potapenko, MD, PhD, Department of Ophthalmology, Rigshospitalet. “The current results are encouraging; however, a live implementation is needed to establish real-life performance.”

Management strategies for AMD typically require close monitoring and administration of anti-VEGF injections for extended periods. In combination with the demographic transition to an older population, an increased burden will affect ophthalmology services.

In order to maintain quality of care in the future, Potapenko and colleagues stressed the importance of novel approaches for more effective patient management. They presented a novel design for a follow-up system of patients with AMD based on a combination of temporally aware artificial intelligence (AI) and deterministic logic. The framework would autonomously suggest patient treatment in accordance with an observe-and-plan (O&P) regimen and handle several advanced concepts.

The AI system was designed to make treatment decisions through evaluation of clinical and imaging data from the current and previous examinations. It consisted of two main components: a deep learning-based model designed to detect CNV activity on optical coherence tomography (OCT) scans and a layer of deterministic logic that defines how its output should impact the management and treatment of the patient.

Investigators prospectively collected a data set of 200 real-world AMD follow-ups, including the treatment decision made by clinicians at our department. Each case was reevaluated by retina specialists and agreement between the AI-decision and the expert consensus was compared with the original treatment decision. From there, investigators attempted to determine the proportion of patients that could be autonomously followed by the proposed AI system.

The findings indicate unanimous agreement was reported in 46% of cases, higher in the passively observed eyes (56%) than in actively treated eyes (36%; P = .007). Data show the temporal AI model was superior at detecting disease activity compared with the model without temporal input (area under the curve [AUC], 0.900 [95% CI, 0.894 - 0.906] and 0.857 [95% CI, 0.846 - 0.867], respectively).

The new models with and without fundus photography had additionally higher AUC in sensitivity (AUC, 0.836 [95% CI, 0.827 - 0.845]) and 0.837 (95% CI, 0.828 - 0.846) versus 0.762 (95% CI, 0.746 - 0.778), respectively. Then, in validation against expert consensus, data show the comprehensive AI system made safe autonomous decisions in 67% (n = 134) and unsafe decisions in 6% (n = 12) of eyes, while a second opinion was asked for in 27% (n = 54) of cases.

Moreover, the AI system made the same decision as expert consensus in 89.2% (n = 157) of the cases, while the retina clinic agreement rate with expert consensus was 85.8% (n = 151; P = .42). The AI-based follow-up could make an autonomous decision in 73% of the cases, of which 91.8% of which were in agreement with expert consensus.

Investigators noted it was on par with the 87.7% agreement rate between decisions made in the clinic and expert consensus (P = .33).

“Even if not deployed to its full potential, the proposed algorithm could greatly ease the pressure on the public ophthalmology departments from an increasing number of AMD patients and be a part of an efficient system for follow-up and treatment in concert with the primary sector services,” Potapenko concluded.

The study, “Automated artificial intelligence-based system for clinical follow-up of patients with age-related macular degeneration,” was published in Acta Ophthalmologica.