Machine Learning-Enabled Longitudinal Volumetric Fluid Assessment in the Phase III VISTA Clinical Trial

Abstract

Purpose : To assess the feasibility of machine learning enhanced volumetric fluid characterization in diabetic macular edema (DME) and compare the longitudinal fluid dynamics between the treatment groups in the Phase III VISTA DME clinical trial.

Methods : In the Phase III VISTA clinical trial, patients were randomly assigned to receive either 2mg IAI every 4 weeks (2q4), 2mg IAI every 8 weeks following 5 initial monthly doses (2q8), or laser photocoagulation (rescue with IAI was allowed beginning at week 24). Utilizing a machine learning-enhanced fluid feature extraction platform, OCT images taken monthly for each patient were analyzed from Baseline to Week 24, at Week 52, and at Week 100. For each visit, intraretinal fluid (IRF) and subretinal fluid (SRF) metrics were extracted through automated fluid segmentation with manual correction, as needed. IRF/SRF total macular volume, central subfield volume (CSV), and central subfield fluid indices (i.e. percentage of retina occupied by fluid) were analyzed.

Results : Four hundred forty-three patients were included in the study with successful OCT feature extraction, with 148 in the 2q4 group, 148 in the 2q8 group, and 147 in the laser group. At baseline, all groups demonstrated similar IRF and SRF volumetric parameters (p>0.05). At Week 100, the pooled IAI group showed significantly less mean panmacular IRF volume than the laser group 0.135±0.39mm³ vs 0.275±0.47mm³; p=0.006, respectively) and lower CSV IRF fluid index (0.06±0.13% vs 0.12±0.18%; p=0.002), despite 41% of laser patients receiving IAI between weeks 24 and 100. Laser and pooled IAI showed significant improvement in SRF from Baseline to Week 100 (both p<0.01).

The 2q4 group showed significantly greater mean change in all IRF parameters at week 100 compared to 2q8, including panmacular IRF volume (-0.137±0.34mm³ vs -0.064±0.16mm³; p=0.034), IRF CSV (-0.012±0.034mm³ vs -0.003±0.02mm³ respectively; p=0.006), and CSV IRF index (-0.04±0.09% vs -0.01±0.07%; p=0.0019).

Conclusions : Utilizing a machine learning-enhanced extraction platform, longitudinal volumetric fluid assessment was feasible and demonstrated greater reduction in IRF with IAI at week 100 compared to laser photocoagulation. The 2q4 group demonstrated a greater reduction in IRF at week 100 compared to the 2q8 group.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.