Artificial intelligence-based classification of cardiac autonomic neuropathy from retinal fundus images in patients with diabetes: The Silesia Diabetes Heart Study

Katarzyna Nabrdalik, Krzysztof Irlik, Yanda Meng, Hanna Kwiendacz, Julia Piaśnik, Mirela Hendel, Paweł Ignacy, Justyna Kulpa, Kamil Kegler, Mikołaj Herba, Sylwia Boczek, Effendy Bin Hashim, Zhuangzhi Gao, Janusz Gumprecht, Yalin Zheng, Gregory Y.h. Lip, Uazman AlamCardiovascular 2024 · DOI

Abstract: This study, part of the Silesia Diabetes-Heart Project (Clinical-Trials.gov Identifier: NCT05626413), investigates the efficacy of a deep-learning framework for the non-invasive detection of Cardiac Autonomic Neuropathy (CAN) using retinal fundus photography. By leveraging Multiple Instance Learning (MIL), the researchers developed a model capable of identifying CAN signatures in routine eye scans with high sensitivity, potentially transforming early screening for diabetic complications.

Introduction

Cardiac Autonomic Neuropathy (CAN) represents one of the most debilitating yet underdiagnosed complications of diabetes mellitus. Characterized by damage to the autonomic nerve fibers innervating the heart and blood vessels, CAN is associated with a three-fold increase in cardiovascular mortality and a significantly higher risk of silent myocardial ischemia.

Despite its clinical gravity, early-stage CAN often remains asymptomatic. Current diagnostic protocols rely on Cardiovascular Autonomic Reflex Tests (CARTs), which are time-intensive, require specialized equipment, and are difficult to implement in routine primary care. This research explores the "eye-heart" connection, hypothesizing that the retinal microvasculature—which shares embryological and physiological pathways with the autonomic nervous system—can serve as a proxy for systemic autonomic health.

Methodology

Multiple Instance Learning (MIL)

The technical cornerstone of this study is the implementation of Multiple Instance Learning. Unlike standard image classification, where each image is labeled individually, MIL treats a collection of images from a single patient as a "bag." This approach is particularly suited for retinal screening, where a patient may have multiple fundus views (fields), but only specific regions or views may contain the subtle pathological features indicative of CAN.

Model Architectures and Explainability

The researchers evaluated several convolutional neural network (CNN) backbones, including ResNet, EfficientNet, and MobileNet. To address the "black box" nature of deep learning, the team utilized Gradient-weighted Class Activation Mapping (Grad-CAM). This technique generates heatmaps that highlight the specific anatomical structures—such as the optic disc and retinal vasculature—that the model prioritized during classification, ensuring the AI's decisions align with clinical intuition.

Datasets and Experimental Setup

The study utilized data from 229 diabetic patients (109 with CAN, 120 without) enrolled in the Silesia Diabetes-Heart Project. A total of 2,275 non-mydriatic 5-field color fundus images were analyzed.

Patient-level data split for training, validation, and testing.

To ensure the robustness of the findings, the data was split at the patient level (60% training, 10% validation, 30% testing). This prevents "data leakage," a common pitfall where images from the same patient appear in both training and testing sets, leading to artificially inflated performance metrics.

Results and Performance

The AI model demonstrated exceptional diagnostic accuracy, particularly in identifying severe cases of the condition:

  • General CAN Detection: The model achieved a sensitivity of 93% and a specificity of 89%, with an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.87.
  • Severe CAN Detection: For patients with definite/severe CAN (dsCAN), the ResNet 18 backbone achieved an AUROC of 0.94.
  • Computational Efficiency: The model's inference time was recorded at approximately 0.19 seconds per patient, suggesting it could be seamlessly integrated into existing clinical workflows without adding significant time to patient visits.

Interestingly, the relatively lightweight ResNet 18 architecture outperformed deeper models. This suggests that for medical imaging tasks with modest dataset sizes, simpler models may be less prone to overfitting and more effective at capturing relevant clinical features.

Discussion and Clinical Implications

The study confirms that AI can detect CAN signatures even when traditional diabetic retinopathy markers are not yet apparent to human clinicians. However, the researchers noted that previous retinal laser therapy introduced significant "noise" into the images, which can degrade model performance.

While the results are promising, the authors acknowledge several limitations. The cohort was modestly sized and drawn from a single center in Poland, necessitating further validation in larger, multi-ethnic populations. Future iterations of the model may also benefit from "data fusion," combining retinal images with clinical metadata such as age, BMI, and duration of diabetes to further refine diagnostic precision.

Conclusion

This deep-learning approach offers a fast, non-invasive, and highly sensitive screening tool for Cardiac Autonomic Neuropathy. By integrating this AI into routine annual diabetic eye exams, clinicians could identify high-risk patients much earlier, allowing for timely intervention and potentially reducing the high cardiovascular mortality associated with this "silent" complication.

Summary updated

Abstract

This is a part of the Silesia Diabetes-Heart Project (Clinical-Trials.gov Identifier: NCT05626413).
BibTeX
@inproceedings{nabrdalik2024artificial,
  author = {Katarzyna Nabrdalik and Krzysztof Irlik and Yanda Meng and Hanna Kwiendacz and Julia Piaśnik and Mirela Hendel and Paweł Ignacy and Justyna Kulpa and Kamil Kegler and Mikołaj Herba and Sylwia Boczek and Effendy Bin Hashim and Zhuangzhi Gao and Janusz Gumprecht and Yalin Zheng and Gregory Y.H. Lip and Uazman Alam},
  title = {{Artificial intelligence-based classification of cardiac autonomic neuropathy from retinal fundus images in patients with diabetes: The Silesia Diabetes Heart Study}},
  booktitle = {Cardiovascular},
  year = {2024},
  doi = {10.1186/s12933-024-02367-z},
}
Figure