Features
Two powerful modules for AF detection and temporal phenotyping
Atrial Fibrillation Detection
ArNet2 Deep Learning Model
Overview
ArNet2 is a deep recurrent neural network that detects AF events from RR interval sequences. The model analyzes 60-beat windows and provides beat-by-beat AF classification with high accuracy and strong generalization across demographics and geography.
How It Works
Input Processing
- Accepts RR intervals from ECG recordings
- Supported formats: WFDB, CSV, EDF
- Preprocessing: 200 Hz resampling, bandpass filter [0.67-100 Hz]
- Quality check: bSQI ≥ 0.8
Two-Stage Analysis
- Stage 1: ResNet extracts embedded features (5 ResNet blocks)
- Stage 2: GRU units perform temporal analysis (4 GRU layers)
- Output: AF probability per 60-beat window
- Window size: 60 beats (non-overlapping)
Clinical Output
- AF Burden percentage
- Episode timestamps
- Confidence scores
- Quality indicators
ArNet2 Architecture
Stage 1: Residual Neural Network
- 5 ResNet blocks
- 1D convolutional layers
- Batch normalization
- Dropout regularization
- Extracts embedded features
Stage 2: Recurrent Neural Network
- 4 GRU (Gated Recurrent Units)
- Temporal sequence encoding
- AF burden-specific pathways
- Outputs AF probability
- Dataset: 2,147 patients, 51,386 hours
- Optimization: Adam optimizer
- Loss: Weighted binary cross-entropy
- Validation: 5-fold cross-validation
- Hardware: NVIDIA A100 GPU
Validated Performance
(internal test)
(patient-level)
validated
hours
Reference: Biton et al., NPJ Digital Medicine (2023)
Circadian AF Phenotyping
5 Temporal Patterns
Overview
Hierarchical clustering analysis identifies 5 distinct temporal patterns of paroxysmal AF based on 24-hour burden profiles. These chronophenotypes may have prognostic significance for risk stratification and treatment planning.
The 5 Phenotypes
Type I: Nocturnal
Peak Burden: 00:00-08:00 hours
Characteristics:
- Highest AF activity during sleep
- Associated with vagal/parasympathetic dominance
- May respond to beta-blocker therapy
Type II: Evening-to-Morning
Peak Burden: 18:00-04:00 hours
Characteristics:
- Extended nocturnal AF burden
- Gradual onset in evening hours
- May span evening through early morning
Type III: Daytime
Peak Burden: 08:00-18:00 hours
Characteristics:
- AF during waking hours
- May be adrenergically mediated
- Physical activity-related
Type IV: Persistent
Peak Burden: All hours (>80% burden)
Characteristics:
- Continuous or near-continuous AF
- High overall burden throughout 24h
- May require rhythm control strategies
Type V: Non-AF
Peak Burden: None (<30s total AF)
Characteristics:
- Minimal or no AF detected
- Normal sinus rhythm predominant
- May have rare PACs or PVCs only
Clinical Significance
Temporal patterns may inform:
Reference: Brimer et al., ML4H (2024)
Try These Features with Sample Data
Experience both AF detection and phenotyping in our interactive demo
Launch Demo