In new research published in JAMA Ophthalmology, investigators from Duke University developed a deep learning algorithm to determine a more accurate prediction for the progression of intermediate age-related macular degeneration (iAMD) to geographic atrophy (GA).
Let’s start with some background.
With no current mechanism in place for determining patients at risk of progressing from iAMD to GA—an increasingly prevalent global disease—clinical studies on GA prevention are often impeded by a need for long study periods and large patient cohorts, according to investigators.
What would the benefits be to this prediction?
The researchers noted that, in a potential future where therapeutics are able to prevent iAMD to GA progression, predicting this occurrence would be “valuable in targeting treatment to the patients who stand to benefit.”
And for the immediate future, more frequent monitoring of iAMD patients with a high risk for GA could enable them to take advantage of currently available therapeutics at an earlier stage.
Gotcha. Now talk about this algorithm
Dubbed “DeepGAze,” the fully automated and accurate convolutional neural network-based deep learning algorithm was designed to predict progression from iAMD to GA within 1 year based on spectral-domain optical coherence tomography (SD-OCT) imaging studies.
To note, the technology is similar to those used to monitor other retinal diseases such as diabetic retinopathy (DR).
And its capabilities?
The investigators’ goal was for the deep learning-based algorithm to be:
- Generalized for multiple SD-OCT devices, including current standard-of-care models
- Validated on data taken from routine patient care
- Capable of making predictions on a clinically meaningful timeframe
- Automated for screening large patient databases
- Able to function with no need for human intervention
Now talk about the study.
DeepGAze examined three datasets (listed below) of SD-OCT imaging studies from 417 iAMD patients obtained from four centers across the United States.
- Data set 1
- Imaging of 316 patients (mean standard deviation [SD] age = 74; 59% female)
- Imaging retrieved from the Age-Related Eye Disease Study 2 (AREDS2) and Ancillary Spectral-Domain Optical Coherence Tomography Study (A2A)
- Timing: July 2008 - August 2015
- Data set 2
- Imaging of 53 patients (mean SD age = 83; 60% female)
- Imaging retrieved from routine clinical care
- Timing: January 2013 - January 2023
- Data set 3
- Imaging of 48 patients (mean SD age = 81; 67% female)
- Imaging retrieved from routine clinical care
- Timing: January 2013 - January 2023
What was measured?
Investigators evaluated the prediction of progression to GA within 13 months with the following:
- Area under the receiver-operator curve (AUROC)
- Area under the precision-recall curve (AUPRC)
- Sensitivity
- Specificity
- Positive/negative predictive value
- Accuracy
To note, AUROC is used to measure the accuracy of diagnostic tests while AURPRC is a performance method for imbalance data.
How were the scans analyzed?
DeepGAze was trained and cross-validated on two forms of SD-OCT:
- Data set 1 → Bioptigen SD-OCT volumes
- Data sets 2 & 3 → Heidelberg Spectralis SD-OCT scans
And the findings?
For data set 1, the AUROC for predicted progression from iAMD to GA within 1 year was calculated to be 0.94 (94%).
To note, after 1 year, the accuracy of this predicted progression was 0.88 (88%).
For data set 2, the AUROC predicted progression also reached 0.94 (94%) within 13 months; however, after 13 months, the AUROC increased slightly to 0.97 (97%).
How did AUROC perform beyond 1 year?
Per the study, the AUROC remained high at 0.88 (88%); however, DeepGAze’s accuracy diminished rapidly beyond 24 months.
And how did this compare to human-selected imagery?
Comparable, give or take 0.01 (1%).
A model using human-annotated SD-OCT images produced an AUROC of 0.95 (95%).
Any limitations?
The authors noted the relatively small number of cases for deep learning, which would also limit the nature of the predicted GA progression. Others included:
- Manual assembly of the data set
- Evolving definition of GA
- Class-imbalanced data set
And overall?
The ability of DeepGAze to predict GA progression within 1 year was accurate 94% of the time.
Lastly… what’s the significance of this?
According to the study authors, this algorithm’s value is key because it could:
- Facilitate enrollment for iAMD clinical trials using large databases to identify patients at a high risk for GA.
- Pending the availability of an effective GA therapy: help eyecare professionals (ECPs) determine which GA patients might benefit the most from treatment.
- Also pending an approved GA therapy: identify GA patients requiring more frequent monitoring in order to initiate treatment earlier.