Assistant Professor University of Toronto Toronto, Ontario, Canada
Background: Inherited arrhythmia syndromes, such as Arrhythmogenic Right Ventricular Cardiomyopathy (ARVC) and Long QT Syndrome (LQTS), can be challenging to diagnose in the undifferentiated patient. Our objective was to develop a single ML model capable of distinguishing between ARVC, LQTS, and healthy controls based on ECG alone.
METHODS AND RESULTS: Digital ECG records were gathered from the Canadian Hearts in Rhythm Organization (HiRO) Inherited Arrhythmia Registries, including individuals with ARVC (gene-positive and gene-negative), LQTS (Type 1 and Type 2) and unaffected controls/unaffected family members. Various convolutional neural network (CNN) architectures (21) were tested across multiple classification tasks. Records were categorized by site into internal, validation, and blinded test sets to prevent overlap in ECGs or recording methodology.
1423 ECG records were included, including 351 ARVC exams (225 genotype-positive; 126 genotype-negative), 452 LQTS exams (340 Type 1; 112 Type 2), and 620 control exams. ARVC patients had a median age of 39.5 years (range 5.7-82.0) and recorded a median of 2 ECGs/patient. LQTS patients had a median age of 34.9 years (range 3.8-94.9), recorded a median of 1 ECG/patient, and had a mean QT interval of 435±53 msec.
The best model was a 15-layer CNN with a multi-head attention mechanism, achieving excellent differentiation between LQTS and controls (area under curve [AUC]=0.92, sensitivity=0.97), ARVC and controls (AUC=0.93, sensitivity=0.96), and moderate performance in the combined classification of ARVC, LQTS, and controls (AUC=0.84, sensitivity=0.88; Table 1). The model also achieved excellent performance in distinguishing ARVC from LQTS (AUC=0.93, sensitivity=0.93) and differentiating ARVC, LQTS Type 1 and Type 2 (AUC=0.96, sensitivity=0.95). Other models also achieved strong performance, including popular architectures ResNet-32 (mean AUC across tasks = 0.89) and AlexNet (mean AUC across tasks = 0.88).
Conclusion: The CNN model demonstrated strong performance in distinguishing between ARVC, LQTS, and healthy controls, highlighting its potential for early diagnosis and differentiation of inherited arrhythmia syndromes.
