Left ventricular systolic dysfunction screening in muscular dystrophies using deep learning-based electrocardiogram interpretation

Background: 

Routine echocardiographic monitoring is recommended in muscular dystrophy patients to detect left ventricular systolic dysfunction (LVSD) but is often challenging due to physical limitations. This study evaluates whether artificial intelligence-based electrocardiogram interpretation (AI-ECG) can detect and predict LVSD in muscular dystrophy patients. 

Methods: 

Patients aged >16 years who underwent an ECG and echocardiogram within 90 days at the University Medical Center Utrecht were included. Patients with Duchenne (DMD), Becker (BMD), limb-girdle muscular dystrophy (LGMD). myotonic dystrophy (MD), and female DMD/BMD carriers, were identified. A convolutional neural network (CNN) was trained on a derivation cohort of patients without muscular dystrophy to detect LVSD and tested on muscular dystrophy patients. A Cox proportional hazards model assessed AI-ECG's predictive value for new-onset LVSD. 

Results: 

The derivation cohort included 53,874 ECG-echocardiogram pairs from 30,978 patients, while the muscular dystrophy test set comprised 390 ECG-echo pairs from 390 patients. LVSD prevalence varied from 81.3 % in DMD to 13.4 % in MD. The model achieved an AUROC of 0.83 (0.79–0.87) in the muscular dystrophy test set, with sensitivity 0.87 (0.81–0.93), specificity 0.58 (0.52–0.63), NPV 0.91 (0.86–0.95), and PPV 0.49 (0.43–0.56). AI-ECG predicted new-onset LVSD with an AUROC of 0.72 (0.66–0.78), with AI-ECG probability being a significant predictor. 

Conclusions: 

AI-ECG can detect LVSD in muscular dystrophy patients, offering a non-invasive, accessible tool for risk stratification and an alternative to routine echocardiography. It may also predict new-onset LVSD, enabling earlier intervention. Further research should explore external validation, pediatric application, and integration within the clinical care plan.