OR WAIT null SECS
A convolutional neural network accurately predicted eyes that progressed from intermediate AMD to GA within 1 year using volumetric SD-OCT scans.
A convolutional neural network-based deep learning algorithm accurately predicted eyes that progressed from intermediate age-related macular degeneration (iAMD) to geographic atrophy (GA) within 1 year, according to a new analysis.1
The investigative team from Duke University added that simulations using the convolutional neural network for clinical trial recruitment of patients with risk for disease progression showed a greater yield in identifying progression to GA within the trial cohort.
“The ability to predict imminent GA progression could facilitate clinical trials aimed at preventing the condition and could guide clinical decision-making regarding screening frequency or treatment initiation,” wrote investigators, led by Eleonora M. Lad, MD, PhD, department of ophthalmology, Duke University Medical Center.
An advanced nonexudative form of AMD, GA is preceded by the intermediate stage of dry AMD, with minimal effects on visual acuity. Although GA rates are rising, owing to increases in the global median age, reports indicate a low likelihood of progressing from iAMD to GA, with per-year incidences estimated to be 0.75–3.67%.2
Without a way to identify those patients most likely to progress, Lad and colleagues indicated clinical studies suffer due to the need for long study periods and large patient cohorts.1 Predicting near-term GA progression could also be valuable in the event of potential future therapeutics or early initiation of currently available therapeutics, to target those who might benefit the most.
For this analysis, the investigative team aimed to create a fully automated and accurate deep-learning algorithm to predict the progression from iAMD to GA. They designed the algorithm based on volumetric spectral-domain optical coherence tomography (SD-OCT) scans, to predict progression during the year following the scan.
The retrospective cohort included individuals with iAMD at baseline, who either progressed or did not progress to GA during the following 13 months. These participants were recruited from centers in 4 states in the United States and included 3 independent data sets. Data set 1 included patients from the Age-Related Disease Study 2 (AREDS2) Ancillary SD-OCT study from July 2008 - August 2015; data sets 2 and 3 were collected from routine outpatient encounters at a tertiary referral center and associated satellite practices between July 2022 – February 2023.
Investigators then trained and cross-validated the position-aware convolutional neural network with proactive pseudointervention on Bioptigen SD-OCT volumes in data set 1 and Heidelberg Spectralis SD-OCT scans in data sets 2 and 3. The primary outcome was predicting progression to GA within 13 months, evaluating the area under the receiver-operator characteristic curves (AUROC) and the area under the precision-recall curve (AUPRC), including sensitivity and specificity.
A total of 417 patients were included in the analysis: 316 in data set 1 (mean age, 74 years; 185 [59%] female; 53 in data set 2 (mean age, 83 years; 32 [60%] female); 48 in data set 3 (mean age, 81 years; 32 [67%] female).
Upon analysis, for data set 1, the AUROC for the prediction of progression from iAMD to GA within 1 year was 0.94 (95% CI, 0.92 - 0.95) and the AUPRC was 0.90 (95% CI, 0.85 - 0.95). Additional performance values for the prediction of GA after 1 year showed a sensitivity of 0.88 (95% CI, 0.84 - 0.92) and a specificity of 0.90 (95% CI, 0.87 - 0.92) for data set 1.
Expert-annotated SD-OCT features added to the model produced an AUROC of 0.95 (95% CI, 0.92 - 0.95) for the 1-year prediction, showing no improvement compared to the fully autonomous model (+0.01; 95% CI, 0.02 - 0.03; P = .19).
On the independent validation data set, or data set 2, the model predicted progression to GA in 13 months with an AUROC of 0.94 (95% CI, 0.91 - 0.96), AUPRC of 0.92 (95% CI, 0.89 - 0.94), sensitivity of 0.91 (95% CI, 0.74 - 0.98), and specificity of 0.80 (95% CI, 0.63 - 0.91).
Further, in the analysis, investigators calculated the enrichment that could be achieved in patients progressing from iAMD to GA if models were used to screen and enroll 1000 patients for a speculative clinical trial. Depending on the baseline incidence of iAMD to GA progression, the model would lead to an 11.2 to 20.7-fold enrichment in progressing patients.
As the model would require an autonomous application to multiple image databases during clinical trial recruitment, Lad and colleagues also tested the performance in data set 3 using the same operating threshold. The high-specific operating point showed a specificity of 0.96 (95% CI, 0.95 - 0.99), a sensitivity of 0.60 (95% CI, 0.49 - 0.68), and an 8.3 to 12.2-fold enrichment, again depending on baseline disease progression.
Lad and colleagues noted this finding as a latter strength of the algorithm, being essential for its application to clinical trials and/or patient care.
“Investigators seeking to test new therapies to prevent the progression from iAMD to GA could apply the model to large databases of SD-OCT volumes and return a list of patients likely to undergo progression during the 1- to 2-year duration of the trial,” investigators wrote.