Deep-learning fusion of OCT imaging and traditional risk factors to improve dementia detection in AlzEye

Abstract

Purpose : For dementia screening, the benefit of incorporating retinal information, in addition to traditional non-retinal risk factors (TRFs), has not yet been fully established. Here, we compare the performance of four models in screening for prevalent all-cause dementia using I. non-retinal TRFs, II. a combination of retinal layer thickness measurements and TRFs, III. a deep learning algorithm trained on OCT images, and IV. a multi-modal fusion model combining OCTs and TRFs.

Methods : The AlzEye study includes 353,157 patients who visited Moorfields Eye Hospital (MEH) from 2008 to 2018. We developed a multi-modal fusion model that combined both TRFs and retinal features. Traditional risk factors composed of I. TRFs - demographic (age, sex, ethnicity, deprivation index) and clinical features (hypertension, diabetes). Retinal features consisted of both II. OCT ETDRS layer thicknesses, and IV. automatically-learned features extracted by a convolutional neural network (CNN). Models were trained on data from 26,603 patients visiting four MEH hospitals, and validated on two test datasets; 2,946 independent patients visiting the same four hospitals (test set 1), and 2,966 independent patients visiting three distinct hospitals (test set 2).

Results : In predicting dementia, the Area Under the Receiver-Operator Curve (AUROC) of the non-retinal TRFs was 0.808, and 0.776 in test sets 1 and 2. The AUROC of the deep learning OCT image model was 0.777, and 0.710 respectively. The addition of the OCT layer thicknesses to the TRFs resulted in an AUROC of 0.812 and 0.784 (best) respectively. The addition of deep learning features to the TRFs resulted in an AUROC of 0.833 (best) and 0.773 in test sets 1 and 2.

Conclusions : We show in two retrospective independent UK validation sets that retinal data, in addition to traditional non-retinal risk factors, can boost performance in detecting all-cause dementia. In our validation sets both the interpretable retinal features defined by physicians, as well as the black-box retinal features extracted by deep learning, improved performance. This highlights the potential benefit of using retinal images for dementia screening in a community setting.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Fig 1. Linkage and dementia dataset curation approach for AlzEye.
Fig 2. Models.

 

Table 1. AUROC results with 95% confidence intervals for each model (1000 sample bootstrapping).

Fig 3. Violin plots displaying AUROC distributions.