A post-hoc analysis of intravitreal aflibercept-treated nAMD patients from ARIES & ALTAIR: predicting treatment intervals and frequency for aflibercept treat-and-extend therapy regimen using machine learning

Abstract

Purpose: To predict potential treatment need during treat-and-extend (T&E) anti-vascular endothelial growth factor (VEGF) treatment in neovascular age-related macular degeneration (nAMD) using an artificial intelligence (AI) model trained using transfer learning.

Methods: ARIES and ALTAIR were randomized controlled Phase 3b/4 trials assessing intravitreal aflibercept (IVT-AFL) in patients with nAMD. Following treatment initiation with three monthly injections of IVT-AFL, treatment intervals were re-assessed continuously during the study based on prespecified criteria. In this post- hoc analysis, spectral domain optical coherence tomography (SD-OCT) scans from Week (Wk) 8 and Wk 16 visits from patients treated with T&E regimens of 2 mg IVT-AFL over 2 years were utilized to predict individual treatment intervals and frequency. Automated image segmentation of the SD-OCT scans was performed, predictive models of treatment intervals and frequency were developed using machine learning or logistic regression methods, and their performance was evaluated using a fivefold cross-validation. A transfer learning technique was used to adapt existing AI models previously trained on a pro-re-nata therapy regimen to the T&E dataset.

Results: In total, 205 ARIES and 112 ALTAIR patient datasets were used for training and evaluation. The following results were achieved with an AI model trained using transfer learning (for ARIES) and logistic regression (for ALTAIR). For prediction of the first treatment interval (short [< 12 weeks] or long [≥ 12 weeks]) following treatment initiation, at Visit 4 (Wk 16), the AI model achieved an area under the receiver operating characteristic curve (AUC) of 0.87 and 0.78 for ARIES and ALTAIR, respectively. For assessment of the individual frequency of IVT-AFL in the first and second study years, the model achieved an AUC of 0.84 and 0.79, respectively, for ARIES, and 0.79 and 0.78, respectively, for ALTAIR. For prediction of the last intended individual treatment interval at the end of Year 2, the AI model achieved an AUC of 0.74 and 0.77 for ARIES and ALTAIR, respectively.

Conclusion: AI trained using transfer learning can be used to predict potential treatment needs for anti-VEGF treatment in nAMD based on SD-OCT scans at Wk 8 and Wk 16, supporting medical decisions on interval adjustments and optimizing individualized IVT-AFL treatment regimens.

Keywords: Anti-VEGF therapy; Artificial intelligence; Neovascular age-related macular degeneration; Transfer learning; Treat-and-extend.

Fig. 1

B-scan preprocessing steps. a) Original B-scan b) Normalization of pixel distances and FoV c) AI layer segmentations d) Detection of ROI e) CLAHE application f) Slice alignment g) Resized image. AI, artificial intelligence; BM, Bruch’s membrane; CLAHE, contrast-limited adaptive histogram equalization; FoV, field of view; ILM, inner limiting membrane; OCB, outer choroidal boundary; ROI, region of interest; RPE, retinal pigment epithelium

Fig. 2

Experiment 1: Receiver operating characteristic curve showing runs from 5 × fivefold cross-validation following transfer learning (ARIES). AUC, area under the ROC curve; ROC, receiver operating characteristic; std, standard deviation

Fig. 3

Experiment 2: Receiver operating characteristic curves showing runs from 5 × fivefold cross-validation for Year 1 and Year 2 following transfer learning (ARIES). AUC, area under the ROC curve; ROC, receiver operating characteristic; std, standard deviation

Fig. 4

Experiment 3: Receiver operating characteristic curve showing runs from 5 × fivefold cross-validation following transfer learning, threshold 12 weeks (ARIES). AUC, area under the ROC curve; ROC, receiver operating characteristic; std, standard deviation