AN AI ENABLED DIGITAL TWIN FOR SIMULATING ANTI-VEGF TREATMENT STRATEGIES IN REAL-WORLD NAMD CARE
Author(s)
Ashwin Kumar Rai, MS1, Nick Boucher, BSc2, Fabio Ishii, BS3, Devika Bhandary, MS4, Ariel Berger, MPH5.
1Director of Data Science & Advanced Analytics, Thermo Fisher Scientific, Overland Park, KS, USA, 2Thermo Fisher Scientific, Ottawa, ON, Canada, 3Thermo Fisher Scientific, Rio De Janeiro, Brazil, 4Thermo Fisher Scientific, London, United Kingdom, 5Thermo Fisher Scientific, Waltham, MA, USA.
1Director of Data Science & Advanced Analytics, Thermo Fisher Scientific, Overland Park, KS, USA, 2Thermo Fisher Scientific, Ottawa, ON, Canada, 3Thermo Fisher Scientific, Rio De Janeiro, Brazil, 4Thermo Fisher Scientific, London, United Kingdom, 5Thermo Fisher Scientific, Waltham, MA, USA.
OBJECTIVES: Real-world neovascular age-related macular degeneration (nAMD) care requires strategies that maintain visual and anatomic outcomes while reducing injection burden. This study developed an AI/ML-enabled patient-to-population digital twin to quantify real-world anti-VEGF treatment trajectories, identify unmet need, and simulate population-level impact of alternative injection-burden reduction strategies.
METHODS: A longitudinal real-world AMD dataset comprising 1,015,175 encounters, 61,220 patients, and 75,000 patient-eye trajectories from January 2023 through April 2026 was analyzed. Variables included demographics, laterality, visual acuity (VA), central subfield thickness (CST), retinal-fluid markers, macular edema, injection status, anti-VEGF brand, insurance, geography, population density, and practice characteristics. Patient-eye timelines were constructed to derive baseline disease state, treatment history, VA/CST trends, fluid activity, injection intervals, switching, and cumulative injection burden. Supervised machine-learning models were trained to predict future VA, CST, fluid activity, switching probability, treatment interval, and vision-loss risk. Population projections scaled simulated patient-eye outcomes across the observed cohort.
RESULTS: The cohort had a median of 12 encounters and 9 injections per patient-eye, with mean annualized treatment burden of 5.8 injections. Retinal-fluid or macular-edema activity was observed in 78.0% of patient-eyes, and 59.2% received ≥8 injections. Among injected patient-eyes, 55.7% received more than one anti-VEGF brand, indicating diverse switching pathways. Mean VA change was -0.5 letters and mean CST change was -51.1 microns over follow-up. The twin produced patient profiles, unmet-need segments, switch-pathway maps, burden estimates, and scenario projections. As one illustrative output, In a digital twin simulation scenario, a 30% uptake of an injection-burden reduction strategy among high-burden, persistent-activity patient-eyes projected 14,603 fewer annual injections, equivalent to 195 fewer injections per 1,000 patient-eyes.
CONCLUSIONS: This AI/ML-enabled nAMD digital twin links patient-eye trajectories with population-level treatment strategy simulation and may support HEOR and RWE evidence generation by quantifying unmet need, treatment burden, switching pathways, and simulated impact of alternative anti-VEGF strategies.
METHODS: A longitudinal real-world AMD dataset comprising 1,015,175 encounters, 61,220 patients, and 75,000 patient-eye trajectories from January 2023 through April 2026 was analyzed. Variables included demographics, laterality, visual acuity (VA), central subfield thickness (CST), retinal-fluid markers, macular edema, injection status, anti-VEGF brand, insurance, geography, population density, and practice characteristics. Patient-eye timelines were constructed to derive baseline disease state, treatment history, VA/CST trends, fluid activity, injection intervals, switching, and cumulative injection burden. Supervised machine-learning models were trained to predict future VA, CST, fluid activity, switching probability, treatment interval, and vision-loss risk. Population projections scaled simulated patient-eye outcomes across the observed cohort.
RESULTS: The cohort had a median of 12 encounters and 9 injections per patient-eye, with mean annualized treatment burden of 5.8 injections. Retinal-fluid or macular-edema activity was observed in 78.0% of patient-eyes, and 59.2% received ≥8 injections. Among injected patient-eyes, 55.7% received more than one anti-VEGF brand, indicating diverse switching pathways. Mean VA change was -0.5 letters and mean CST change was -51.1 microns over follow-up. The twin produced patient profiles, unmet-need segments, switch-pathway maps, burden estimates, and scenario projections. As one illustrative output, In a digital twin simulation scenario, a 30% uptake of an injection-burden reduction strategy among high-burden, persistent-activity patient-eyes projected 14,603 fewer annual injections, equivalent to 195 fewer injections per 1,000 patient-eyes.
CONCLUSIONS: This AI/ML-enabled nAMD digital twin links patient-eye trajectories with population-level treatment strategy simulation and may support HEOR and RWE evidence generation by quantifying unmet need, treatment burden, switching pathways, and simulated impact of alternative anti-VEGF strategies.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
MT16
Topic
Medical Technologies, Methodological & Statistical Research, Real World Data & Information Systems
Topic Subcategory
Digital Health
Disease
Sensory System Disorders (Ear, Eye, Dental, Skin)