Deep learning on ECGs for efficient diabetes screening in atrial fibrillation patients

https://doi.org/10.1007/s00392-024-02526-y

Yazan Mohsen (Köln)1, I. Vatsaraj (Baltimore)2, L. Steffens (Köln)1, H. Horlitz (Köln)1, L. Grüne (Köln)1, G. Zandieh (Baltimore)2, J. Mombaur (Köln)1, M. Horlitz (Köln)1, N. Trayanova (Baltimore)2, F. Stöckigt (Köln)1

1Krankenhaus Porz am Rhein gGmbH Klinik für Kardiologie, Elektrophysiologie u. Rhythmologie Köln, Deutschland; 2Johns Hopkins University Baltimore, USA

 

Introduction:
Diabetes Mellitus (DM) significantly impacts the cardiovascular system, notably altering the atrial substrate and increasing the prevalence of Atrial Fibrillation (AF). These effects are due to diabetes-induced structural, electrical, and autonomic changes in the atrium, leading to a higher AF risk. The early diagnosis and effective management of DM are crucial for treating AF and improving ablation outcomes. Despite the straightforward nature of DM diagnosis, large-scale screening presents challenges due to the invasive nature of traditional methods such as glucose tolerance tests and hemoglobin A1c, coupled with their availability constraints. Given the widespread accessibility of electrocardiograms (ECGs) in this patient population. our study explores the use of deep learning on 12-lead ECGs and clinical covariates such as sex, age, and Body Mass Index (BMI) for a non-invasive and efficient DM screening approach in AF patients.
 
Objective:
Developing a non-invasive and efficient screening tool for DM in AF patients using multimodal deep learning on 12 lead ECG and clinical covariates.
 
Method:
For this study, pre-procedural 12-lead ECGs in sinus rhythm were gathered from 210 patients scheduled for AF ablation. These ECGs were accompanied by key clinical covariates (age, sex, and BMI). The data was analyzed using an innovative multimodal deep learning framework that combined elements of Long Short-Term Memory and Convolutional Neural Networks with a cross-attention mechanism. This model was trained on the ECGs and the clinical covariates to predict DM. The model's performance was rigorously evaluated through a 5-fold stratified cross-validation process to confirm its effectiveness and reliability.
 
Results:
In the examined group 16 (7.6%) of the patients were diagnosed with DM. The deep learning model demonstrated 93% Accuracy, 95% Specificity in detecting DM with a 67% sensitivity.
 
Conclusion:
The multimodal deep learning model demonstrated high potential in DM-screening in AF patients. This approach can proficiently pinpoint patients less likely to be affected by DM, thereby optimizing the diagnostic workflow. Such a method allows for the concentration of more extensive diagnostic resources on individuals more likely to have DM, enhancing the overall efficiency of the screening process.
Diese Seite teilen