Slide 1: Post Hoc Analysis of Aflibercept-Treated nAMD Patients from ARIES & ALTAIR
A Post Hoc Analysis of Intravitreal Aflibercept–Treated nAMD Patients from ARIES & ALTAIR: Predicting Patient-Individualized Treatment Interval for Aflibercept Treat-and- Extend Therapy Regimen by Adapting AI Algorithms Trained on Pro Re Nata Data 23rd Euretina Congress, Amsterdam, September 5-8, 20
Slide 2: Disclosures
Disclosures Disclosures: Presenting author Matthias Gutfleisch: Study participation: Amgen, Apellis, Iveric Bio, Kodiak, Neurotech, Novartis, Regeneron, Roche; Shareholder: deepeye Medical GmbH; Travel support: DORC; Consultant: Bayer. Disclosures: Co-author group PS: Employee: Bayer; HY: Employee:
Slide 3: Purpose and Methods – Data
Purpose • Prediction of potential treatment need/expected therapy response and neovascular age-related macular degeneration (nAMD) disease course using artificial intelligence (AI) • ARIES1 and ALTAIR2 • Randomized, controlled, Phase 3b/4 trials • Treat & Extend (T&E) regimens in newly diagnosed nAM
Slide 4: Methods – criteria for interval adaptation
Methods – criteria for interval adaptation Criteria for interval adaptation for ARIES & ALTAIR interventional studies (see Table) *Patients could receive more frequent treatment if identified as injection-intensive and were excluded from the per-protocol set; †Assessed by OCT; ‡Loss of ≥5 ETDRS lett
Slide 5: Methods – AI pipeline
Methods – AI pipeline • Evaluation and adaptation of the existing AI models in the deepeye® research tool1–3 • AI architecture • Input: SD-OCT images from Weeks 8 and 16 from ARIES & ALTAIR studies • Two AI networks: Biomarker segmentation (i) and prediction model (ii) • Use of AI model trained on S
Slide 6: Methods – experiments
Methods – experiments Experiment 1: • Prediction of treatment frequency and interval + - • Experiment 1: Prediction of potential adequate first injection interval: <3 vs ≥3 ≤ . interval extensions in the first four visits after initiation* • Experiment 2: Prediction of injection frequency in first a
Slide 7: Results of Experiment 3 – interval after second year
Results of Experiment 3 – interval after second year • ARIES: Interval?. • Sensitivity: Patients identified as needing a ‘short’ interval: From study data: 116 patients; from model: 93 patients (80% sensitivity) • Specificity: Patients identified as needing a ‘long’ interval: From study data: 105 pa
Slide 8: Results – additional experiments
Results – additional experiments • Experiment 1: <3 interval extensions in the four visits after treatment initiation* (starting from Week 16) • Experiment 2 (first year): ≥8 injections • Experiment 2 (second year): ≥5 injections First year Second year Experiment 1: Predict first potential adequate
Slide 9: Conclusions
Conclusions • AI models successfully adapted from PRN to T&E prediction • New AI algorithm accurately assigns a percentage between 71% and 76% of patients to the <12 weeks or ≥12 weeks interval extension groups ( Experiment 3) • Further experiments achieved a good* prediction accuracy, between 73%
Slide 10: Thank you for your attention
Thank you for your attention