Advertisement
Canadian Journal of Cardiology

DETECTION OF CONGENITAL LONG QT SYNDROME WITH ARTIFICIAL INTELLIGENCE

      Background

      Congenital long QT syndrome (LQTS) is a channelopathy associated with syncope, polymorphic VT, and rarely, sudden arrhythmic death. While often detected by QT prolongation on the resting ECG, half of patients have a normal or borderline QT interval, necessitating further investigations (i.e. exercise treadmill and genetic testing). We aimed to develop a deep neural network for identification of congenital LQTS and differentiate LQTS genotypes using both the resting single-lead and 12-lead ECGs.

      Methods and Results

      Patients with LQTS type 1, type 2, and unaffected family members were prospectively enrolled in a national LQTS registry. For each patient’s ECG, normalized voltages for the 8 non-augmented leads were down-sampled to 102.4 Hz over 10 seconds. Lead I was evaluated independently, followed by a combined analysis using the 8 non-augmented leads. For each task, we trained convolutional neural networks (Scikit-Learn v1.0.1, Keras v2.6.0, TensorFlow v2.6.0, Figure 1). Data was split into a training (80%) and testing dataset (20%). Model performance was assessed on an independent dataset to which the models were blinded during training. 149 LQTS (Type 1 n=105, Type 2 n=44) and 74 control patients were included, with a mean age of 47±18 years, 57% female. On resting ECG, the heart rate and corrected QT interval was 66±13 bpm and 468±36 msec in LQTS patients, and 74±15bpm and 423±24 msec in control patients. After training, the model was able to identify congenital LQTS with an area under the curve (AUC) of 0.71, sensitivity 53%, and specificity 83% using lead I alone. Using the full ECG (8 non-augmented leads combined), the model identified congenital LQTS with an AUC of 0.83, sensitivity 87%, and specificity 73% (Figure 2A-B). For patients with normal baseline QTc ( < 480 ms in females and < 470 ms in males, n=30), the model had comparable performance with an AUC of 0.85, sensitivity 87%, and specificity 80% (Figure 2C). Additional analyses demonstrated moderate accuracy in differentiating LQTS genotypes (AUC 0.76, sensitivity 57%, specificity 100%; Figure 2D).

      Conclusion

      Deep neural networks can be employed for the detection of congenital LQTS on ECG with moderate accuracy using a single-lead ECG alone, and good accuracy with a full 12-lead ECG (8 non-augmented leads). Future work includes developing a model to incorporate additional demographics and to effectively distinguish between LQTS. Deep neural networks are promising methodology for identifying LQTS and differentiating LQTS genotypes using ECG alone.
      Figure thumbnail fx1
      Figure thumbnail fx2