Artificial intelligence for the detection, prediction, and management of atrial fibrillation

Figures

• 

  Fig. 1. Distinction and overlap between artificial intelligence (AI), machine learning (ML), and neural networks/deep learning (DL). Only DL operates directly on raw waveforms, while other AI methods rely on markers or features. AI example: explicit programming to determine whether left ventricular hypertrophy (LVH) is present or not based on R and S amplitudes. ML example: (imperfect) discrimination between two classes using a support vector machine. DL example: Discrimination between sinus rhythm (SR) and atrial fibrillation (AF) using a simple neural network operating directly on a raw electrocardiogram (ECG) rhythm strip

• 

  Fig. 2. Top: Model performance is similarly displayed by the receiver-operator characteristics (ROC) curve and the precision-recall curve (PRC) when classes are balanced. Bottom: Poor model performance combined with imbalanced classes results in a good-looking ROC, but the PRC curve reveals the poor performance. For PRC curves, positive predictive value is often termed precision, and sensitivity termed recall. (AF atrial fibrillation, AUC area under the curve)

• 

  Fig. 3. Examples of sinus rhythm and atrial fibrillation recorded using a single lead KardiaMobile (AliveCor) electrocardiogram (ECG) device (top) and a smartphone-based FibriCheck photoplethysmography (PPG, bottom). Recordings are from different individuals