Investigating correlations between mental disorders and fundus imaging data using deep learning: A study from the UK Biobank.

Purpose: To automatically identify mental and behavioral disorders using fundus imaging data and interpret the potential associations between mental diseases and fundus biomarkers.

Methods: In this study, a deep learning-based multimodality training method is introduced primarily to explore the relationship between fundus imaging features and mental disorders. The methodology is carefully trained and evaluated using a dataset containing fundus images and optical coherence tomography (OCT) measured features from 1,494 participants in the UK Biobank database. The participants took part in the assessment center proceedings from December 2009 to June 2013, during which fundus images and OCT scans were collected. These participants later received diagnoses related to mental disorders between October 2013 and September 2021. A five-fold cross-validation strategy was employed to select the optimal hyperparameters, followed by training on the entire training set to obtain the best-fitted model. The best-fitted model was subsequently tested on the testing set.

Results: The multimodality model demonstrated an overall Area under the ROC Curve (AUC) value of 0.8490 (95% CI, 0.8477-0.8526), sensitivity of 0.7702 (95% CI, 0.7698-0.7785), and specificity of 0.8552 (95% CI, 0.8546-0.8564) on the fundus images and OCT measures. The Random Forest classifier and Linear Classifier, when applied to the OCT measures, achieved final AUC of 0.8121 (95% CI, 0.8118-0.8126) and 0.8094 (95% CI, 0.7936-0.8102), respectively, indicating a negative correlation between average retinal nerve fiber layer and average ganglion cell-inner plexiform layer thickness and mental disorders.

Conclusions: Preliminary results demonstrated this method can reveal a correlation between fundus imaging and mental disorders, suggesting a promising avenue for noninvasive early detection and intervention.