Artificial intelligence-based algorithms analysing retinal images to accurately identify diabetic CAN

Aims: To determine if Artificial Intelligence (AI) using retinal images can provide an efficient diagnostic method for cardiac autonomic neuropathy (CAN) in people with diabetes.

Methods: A single centre, observational study in a sub-section of people (n=229) with type 1 and type 2 diabetes from the Cardiovascular Disease in Patients with Diabetes: The Silesia Diabetes-Heart Project (Poland). CAN was diagnosed using the Toronto criteria where cardiovascular autonomic function tests are the gold standard. Retinal images underwent analysis using Multiple Instance Learning and primarily the ResNet 18 was backbone model. The models underwent training and validation on an unseen image set.

Results: 2275 retinal images were analysed using the model, which using a binary classification of CAN correctly identified 93% or cases with CAN and 89% of no CAN cases. The model achieved an area under the receiver operating characteristic curve (AUCROC) of 0.87 (95% CI 0.74-0.97). The ResNet 18 model was able to accurately determine 78% of patients with definite/severe CAN (dsCAN) and 93% of cases without dsCAN with an AUCROC of 0.94 (95% CI 0.86-1.00)

Conclusions: AI based algorithms using retinal images were able to accurately identify patients with CAN and severe CAN with a good sensitivity and specificity.

Comments. CAN is a common yet one of the most underdiagnosed complications of diabetes. It is an independent risk factor for cardiovascular mortality. The difficulty in accessing CARTs and their limitations are just some of the factors that lead to the underdiagnosis. It is therefore imperative that new methods of screening of this complication are sought. In this study the team have utilised novel AI-based algorithms to examine retinal images to determine if they can identify CAN as well as CAN severity. Retinal images are easily acquired and standard of care of people with diabetes for retinopathy screening. AI-based algorithms have been used for retinopathy and already to predict the presence of diabetic polyneuropathy however this is a trendy new kid on the block in the field of autonomic neuropathy. The team have developed a novel AI-based deep learning algorithm to detect CAN and classify severity with the binary algorithm achieving good sensitivity with excellent specificity for CAN detection. Furthermore, an AI-based deep learning algorithm was able to detect severe CAN presence with a high sensitivity and specificity. These subtle structural CAN related changes are not obvious to humans viewing the images (and not captured by the diagnostic criteria of retinopathy) and the use of AI in this situation has shown how much more information we may be able to unveil from fundus images. As retinal screening is readily available in the Western work, AI analysis of fundus images to detect CAN may be a viable option in routine clinical practice to identify those patients at the highest CV risk. Further work is needed to perform an external validation of this AI algorithm in larger multi-centre prospective studies of different populations.

Shazli Azmi

Reference. Nabrdalik K, Irlik K, Meng Y, Kwiendacz H, Piaśnik J, Hendel M, Ignacy P, Kulpa J, Kegler K, Herba M, Boczek S, Hashim EB, Gao Z, Gumprecht J, Zheng Y, Lip GYH, Alam U. Artificial intelligence-based classification of cardiac autonomic neuropathy from retinal fundus images in patients with diabetes: The Silesia Diabetes Heart Study. Cardiovasc Diabetol. 2024 Aug 10;23(1):296. doi: 10.1186/s12933-024-02367-z. PMID: 39127709; PMCID: PMC11316981.

https://cardiab.biomedcentral.com/articles/10.1186/s12933-024-02367-z