Abstract
Purpose :
To investigate retinomics changes in mental disorders and develop algorithms to detect depression, anxiety, schizophrenia, and bipolar disorder via fundus photography.
Methods :
We employed cross-sectional data from institutions in both UK and China, with participants spatially partitioned across 7 recruitment centers and randomly divided into training and validation sets. Retinal images were obtained from age-matched individuals diagnosed with depression, anxiety, schizophrenia, bipolar disorder, and healthy volunteers using a non-mydriatic fundus camera. A validated semantic segmentation model based on ResNet101-UNet was employed to extract 99 retinomics features, encompassing retinal arterial and venous trajectories, as well as optic nerve indices. Multivariate regression analyses were used to study associations between retinomics and mental diseases. Different models were constructed after testing algorithms for feature selection, considering both retinomics-only and a combination of retinomics and clinical risk factors. The discriminative ability was assessed in both training and validation datasets using Receiver Operating Characteristic curves and AUC metrics.
Results :
A total of 517 individuals with mental disorders (517 eyes, mean[SD] age, 54.9[8.1]years; 60.6% female) and 142 healthy controls (142 eyes, mean[SD] age, 58.2[18.2]years; 40.1% female) were included after quality control. In terms of optic disc characteristics, the mental disorder group exhibited a larger tilt angle of optic disc (OR=1.46, 95%CI=1.13-1.90, per 0.01°), particularly in the depression (OR=1.53, 95%CI=1.11-2.12, per 0.01°) and anxiety group (OR=1.51, 95%CI=1.08-2.10, per 0.01°). Accounting for vessel tortuosity, the mental disorder group showed a significantly larger curvature of veins (OR=1.66, 95% CI=1.14-2.41) in all depression, anxiety, and schizophrenia groups (OR=1.69, 95% CI=1.08-2.63; OR=1.71, 95% CI=1.07-2.75; OR=2.41, 95% CI=1.35-4.31, respectively). The LASSO-based models, incorporating retinomics and clinical factors, achieved accuracies of 83.8%, 86.9%, 81.0%, and 77.7% for classifying depression, schizophrenia, bipolar, and anxiety in the training set, with corresponding accuracies of 78.2%, 79.5%, 84.8%, and 78.8% in the validation set.
Conclusions :
Considering the relationship between retinal/cerebral vasculatures and neurons, retinomics emerges as potential effective markers for mental disorders.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.