Predicting visual acuity from patient data with a machine-learning multistage system


(Image Credit: AdobeStock/Lee)

(Image Credit: AdobeStock/Lee)

German investigators are working on a better way to predict the visual acuity (VA) levels in patients with vision-threatening diseases, in this case, age-related macular degeneration (AMD), diabetic macular edema (DME), and retinal vein occlusion (RVO).

Their machine learning multistage system predicts the expected disease progression of patients and their VA in the 3 diseases,1 according to lead author Tobias Schlosser, PhD, Junior Professorship of Media Computing, Chemnitz University of Technology, Chemnitz, Germany, who, along with his colleagues, described how they constructed this multistage system.

In commenting on the rationale for this research, they said, “… in real-world settings, patients often suffer from loss of vision on time scales of years despite therapy, whereas the prediction of the VA and the earliest possible detection of deterioration under real-life conditions is challenging due to heterogeneous and incomplete data.”

Constructing the system

In their report, they described the workflow for the development of a research-compatible body of data that brought together different IT systems of the Department of Ophthalmology of Klinikum Chemnitz gGmbH, Chemnitz. The text and images used for analysis were obtained from over 49,000 patients.

The workflow facilitates processing of medical patient data to enable optical coherence tomography (OCT) biomarker classification, VA prediction, and general statistical evaluation and visualization. For this purpose, they explained, the developed patient progression visualization and modeling dashboard enables visualization, annotation, and assessment of patient progression with a focus on VA.

Using their proposed multistage system, they classified the VA progression into 3 groups of patients who were treated: “winners,” “stabilizers,” and “losers.”

The extensive body of data, they reported, found that for exudative AMD, for example, a high percentage of patients, 60%, lost a significant amount of vision over the course of 3 to 5 years. For DME, they reported a “weakly significant deterioration of VA”; for RVO, there was no significant decrease in the VA.

A more specific analysis can show the effect of medical co-existing factors such as other diseases, for example, DME with an epiretinal membrane.

At the point, however, the data are too weak to derive reliable correlations for statistical surveys of comorbidities in combination with different observation time windows.

Schlosser and colleagues further explained that for the VA-based treatment progression modeling, incomplete OCT documentations are completed by classifying the OCT B-scans, which allows the classification of OCT scans when only single OCT slices are provided. Based on the obtained OCT slice classifications, a scan-wise OCT classification of the OCT biomarkers, external limiting membrane, ellipsoid zone, foveal depression, retinal pigment epithelium, scars, and subretinal fibrosis, resulted in an overall classification accuracy of over 98% in the F1-score.

Finally, the completed OCT documentations are combined with additional medical data, defining our ophthalmic feature vectors for predicting VA. “We achieved a final prediction accuracy of 69% in macro average F1-score with 77.2% true positives, while our main ophthalmologist shows a macro average F1-score of 57.8% with 82.2% true positives. To further validate these results, we evaluated the annotations of 8 different ophthalmic doctors given randomized subsets of our test set, which resulted in an overall macro average F1-score of 50.0±10.7% and with 70.1±5.9% true positives.”

However, this is just the beginning. The investigators explained that the effect of OCT biomarkers is not fully understood. “Further investigations have to be conducted, for which additional OCT biomarkers as well as their influence for the visual acuity modelling process have to be evaluated,” they concluded.

Future fine-tuning will include the ability to determine the optimal times for changes in therapies and treatment options such as laser coagulation, pars plana vitrectomy, or phacoemulsification with posterior chamber lens implantation, among others.

  1. Schlosser T, Beuth F, Meyer T, et al. Visual acuity prediction on real-life patient data using a machine learning basedmultistage system. Sci Rep. 2024;14:5532;
© 2024 MJH Life Sciences

All rights reserved.