A multicentric study examining a deep-learning-based computer model for classifying bipolar disorder using retinal vascular images

Abstract

Objectives

Due to easy accessibility, the retina is considered a window to the brain. Recent studies have reported retinal vascular abnormalities in bipolar disorder. Deep learning analysis, an advanced computational approach, has been implemented in medicine for diagnosis using retinal vascular images. However, deep learning analysis has not been implemented to classify patients with bipolar disorder and healthy individuals using retinal vascular images to date.

Methods

A total of 383 subjects participated in the study (188 patients with Bipolar Disorder (BD) and 195 Healthy Volunteers(HV)). 327 subjects were classified as the training dataset, and the remaining 56 subjects were classified as the test dataset. We acquired retinal fundus images using a non-mydriatic fundus camera. We applied an optimized convolutional neural network (CNN) model for the current analysis. The model was trained using the training dataset and its transfer learning ability was tested in the test dataset.

Results

The overall metrics of the CNN for the training dataset are 88.0 % sensitivity, 85.7 % specificity, 96.2 % positive predictive value, and 83.7 % negative predictive value. The model parameters were close to each other, suggesting a balanced model. The model achieved high performance in the test dataset; accuracy = 85.7 %, sensitivity = 83.3 %; specificity = 92.9 %, and positive predictive value = 83.3 %.

Conclusions

Diagnostic markers are the need of the hour in psychiatry. The CNN model differentiated BD patients and HV with greater accuracy. Replicating the findings in a non-overlapping, independent, test dataset suggests the potential for transfer learning and clinical utility.