[Paper Review] End-to-end Deep Learning from Raw Sensor Data: Atrial Fibrillation Detection using Wearables

Igor Gotlibovych, Stuart L. Crawford+7 more|arXiv (Cornell University)|Jul 27, 2018

ECG Monitoring and AnalysisMedicine37 references21 citations

TL;DR

This paper proposes an end-to-end deep learning model using a convolutional-recurrent neural network with long short-term memory (LSTM) to detect atrial fibrillation (AFib) directly from raw photoplethysmography (PPG) signals collected by wearable devices. The model achieves state-of-the-art performance with an AUC of 0.9999 and false positive/negative rates below 0.2%, outperforming traditional methods that rely on handcrafted features like heart rate extraction.

ABSTRACT

We present a convolutional-recurrent neural network architecture with long short-term memory for real-time processing and classification of digital sensor data. The network implicitly performs typical signal processing tasks such as filtering and peak detection, and learns time-resolved embeddings of the input signal. We use a prototype multi-sensor wearable device to collect over 180h of photoplethysmography (PPG) data sampled at 20Hz, of which 36h are during atrial fibrillation (AFib). We use end-to-end learning to achieve state-of-the-art results in detecting AFib from raw PPG data. For classification labels output every 0.8s, we demonstrate an area under ROC curve of 0.9999, with false positive and false negative rates both below $2 imes 10^{-3}$. This constitutes a significant improvement on previous results utilising domain-specific feature engineering, such as heart rate extraction, and brings large-scale atrial fibrillation screenings within imminent reach.

Motivation & Objective

• To develop a real-time, end-to-end deep learning system that detects atrial fibrillation from raw PPG signals without relying on domain-specific feature engineering.
• To overcome the limitations of traditional ECG-based AFib detection, such as high cost, limited availability, and the need for expert interpretation.
• To enable large-scale, continuous, and non-invasive AFib screening using low-cost wearable sensors combined with deep learning.
• To demonstrate that neural networks can implicitly learn signal processing tasks such as filtering and peak detection from raw sensor data.

Proposed method

• A convolutional-recurrent neural network with LSTM units is trained end-to-end on raw PPG signals sampled at 20 Hz to classify heart rhythms in real time.
• The model learns time-resolved embeddings of the input signal through hierarchical feature extraction, implicitly performing filtering and peak detection.
• Intermediate layer activations are visualized using hierarchical clustering and optimal leaf ordering to interpret learned neuron functions.
• Heart rhythm embeddings from the final hidden layer are projected into 2D space using a method that preserves the logistic regression decision boundary for interpretability.
• The network is trained on over 180 hours of PPG data, including 36 hours of AFib episodes, collected from patients undergoing cardioversion and healthy volunteers during sleep.
• Synthetic data augmentation is used to improve detection of short AFib episodes, particularly those lasting 20–200 seconds.

Experimental results

Research questions

• RQ1 Can a deep learning model trained end-to-end on raw PPG signals achieve superior AFib detection performance compared to traditional feature-based methods?
• RQ2 How well can a neural network learn implicit signal processing operations such as filtering and peak detection without explicit engineering?
• RQ3 What is the minimum detectable duration of an isolated AFib episode using this end-to-end approach?
• RQ4 Can latent space representations of PPG signals enable personalized, unsupervised detection of rhythm anomalies beyond AFib?
• RQ5 How generalizable is the model across diverse demographics and real-world conditions such as motion and varying signal quality?

Key findings

• The model achieves an area under the ROC curve (AUC) of 0.9999, indicating near-perfect discrimination between AFib and normal sinus rhythm.
• The false positive rate is below 0.2% (2×10⁻³), and the false negative rate is also below 0.2%, demonstrating high sensitivity and specificity.
• The model detects AFib episodes as short as 20–200 seconds, with performance varying by patient and signal-to-noise ratio.
• Latent space embeddings from the network's hidden layer reveal patient-specific patterns in heart rhythm, enabling potential for personalized anomaly detection.
• The use of synthetic data during training improves detection of short AFib episodes, suggesting a path to robustness in real-world deployment.
• Visualizations of intermediate neurons and embeddings confirm that the network learns meaningful, interpretable representations of PPG morphology and dynamics.