Photoplethysmography based atrial fibrillation detection: a review

Tânia Pereira, Nate Tran, Kais Gadhoumi, Michele M. Pelter, Duc H. Do, Randall J. Lee, Rene Colorado, Karl Meisel, Xiao HuNpj Digit. Medicine 2020 · OpenReview

TL;DR

This review surveys the landscape of using photoplethysmography (PPG) signals from wearable devices for atrial fibrillation (AF) detection. It highlights the potential of PPG as a low-cost, non-intrusive alternative to traditional ECG for widespread AF screening, while also detailing the signal processing, machine learning, and deep learning approaches used, along with their current limitations and future directions.

What Problem Does This Paper Address?

Atrial fibrillation (AF) is a major risk factor for stroke and other cardiovascular complications, leading to significant morbidity and mortality. Early detection is crucial for effective prevention and management. Traditional ECG monitoring can be intermittent and costly, making it difficult to capture paroxysmal AF episodes. The paper addresses the challenge of developing accessible, continuous, and cost-effective methods for AF detection, particularly in ambulatory settings, by exploring the potential of photoplethysmography (PPG) signals captured by ubiquitous wearable devices.

Methods

The review categorizes PPG-based AF detection methods into three main groups:

  • Traditional Statistical Analysis: These methods rely on extracting statistical features from PPG signals (often focusing on RR intervals, similar to ECG analysis) and setting thresholds to classify AF. This approach is generally simpler but can be less robust to variations.
  • Machine Learning (ML): ML approaches involve extracting a set of pre-selected features (morphological, temporal, spectral) from PPG signals and then using classifiers like k-nearest neighbors (KNN), support vector machines (SVM), or decision trees (DT) to distinguish between AF and normal sinus rhythm (NSR). This requires more domain expertise for feature engineering.
  • Deep Learning (DL): DL methods, particularly Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), are highlighted as a powerful emerging trend. DL models can automatically learn feature representations from raw PPG data, reducing the need for manual feature engineering. Hybrid approaches combining time-series and image representations of PPG signals are also explored.

Datasets

The paper references the use of publicly available ECG databases such as MIT-BIH atrial fibrillation, MIT-BIH normal sinus rhythm, and MIT-BIH arrhythmia database for training and validation in some studies. However, many studies also utilize data collected from real-world wearable devices (e.g., smartwatches, smartphones, wristbands) from patients in both inpatient and outpatient settings. The paper does not introduce any new benchmark datasets but reviews studies that have used various existing datasets and real-world collected data.

Key Insights and Ideas

  • PPG's Promise: PPG, the optical sensing technology behind many wearables, offers a compelling low-cost and non-intrusive avenue for continuous cardiac rhythm monitoring, potentially democratizing AF screening.
  • Beyond Heart Rate: While early PPG analysis focused on heart rate, researchers are increasingly exploring the richer physiological information embedded in the PPG waveform itself, including pulse morphology, to detect complex arrhythmias like AF.
  • DL's Ascendancy: Deep learning models are showing significant promise in AF detection from PPG, outperforming traditional methods by automatically learning complex patterns and reducing reliance on manual feature engineering.
  • The Artifact Challenge: Motion artifacts and poor sensor contact remain significant hurdles for PPG-based AF detection, often leading to data rejection and impacting overall performance. Developing robust artifact detection and correction mechanisms is critical.
  • Clinical Translation Gap: Despite promising results in research settings, translating these PPG-based AF detection algorithms into clinically validated, widely adopted tools faces challenges related to data annotation, computational resources for real-time inference on wearables, and the need for robust, generalizable models.

Key Results

The review summarizes performance metrics from numerous studies across statistical, ML, and DL approaches. While specific numbers vary widely based on dataset, methodology, and patient population, several studies report high accuracy, sensitivity, and specificity:

  • Statistical Methods: Studies using statistical approaches often report accuracies in the range of 90-96%.
  • Machine Learning: ML-based studies frequently achieve accuracies above 90%, with some reaching as high as 98%.
  • Deep Learning: DL approaches have demonstrated even higher performance, with some studies reporting accuracies exceeding 98% and AUC values close to 0.99. For instance, one study achieved an accuracy of 0.9819 with an AUC of 0.9967 using a CNN.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Other Cardiac Arrhythmias: Current PPG-based AF detection methods often struggle to differentiate AF from other cardiac arrhythmias, which can present similar PPG signal characteristics. Future work should focus on developing algorithms that can robustly distinguish between AF and other rhythm disturbances.
  • Corrupted Signals (Motion Artifacts): Motion artifacts are a major limitation. While discarding corrupted segments is common, it leads to data loss. Future research should focus on developing more sophisticated methods for artifact detection, correction, and signal recovery, potentially using techniques like Generative Adversarial Networks (GANs).
  • Data Annotation: Manual annotation of PPG data is time-consuming and prone to inter-rater variability. Active learning strategies and improved annotation guidelines are needed to create larger, more reliable labeled datasets.
  • Computational Resources: Deploying complex DL models on resource-constrained wearable devices for real-time inference presents a significant engineering challenge. Research into efficient model architectures and hardware optimization is crucial.
  • Interpretability: The "black-box" nature of many DL models can be a barrier to clinical adoption. Developing explainable AI (XAI) methods to understand how these models make predictions is essential for building trust with clinicians.
  • Generalizability: Many studies are based on data from a single medical center. Future work should focus on developing models trained on diverse, multi-center datasets to ensure generalizability across different populations and settings.
  • Clinical Validation: Large-scale, prospective clinical trials are needed to rigorously validate the performance and clinical utility of PPG-based AF detection systems before widespread adoption.

Summary updated

Abstract

Atrial fibrillation (AF) is a cardiac rhythm disorder associated with increased morbidity and mortality. It is the leading risk factor for cardioembolic stroke and its early detection is crucial in both primary and secondary stroke prevention. Continuous monitoring of cardiac rhythm is today possible thanks to consumer-grade wearable devices, enabling transformative diagnostic and patient management tools. Such monitoring is possible using low-cost easy-to-implement optical sensors that today equip the majority of wearables. These sensors record blood volume variations—a technology known as photoplethysmography (PPG)—from which the heart rate and other physiological parameters can be extracted to inform about user activity, fitness, sleep, and health. Recently, new wearable devices were introduced as being capable of AF detection, evidenced by large prospective trials in some cases. Such devices would allow for early screening of AF and initiation of therapy to prevent stroke. This review is a summary of a body of work on AF detection using PPG. A thorough account of the signal processing, machine learning, and deep learning approaches used in these studies is presented, followed by a discussion of their limitations and challenges towards clinical applications.
BibTeX
@inproceedings{pereira2020photoplethysmography,
  author = {Tânia Pereira and Nate Tran and Kais Gadhoumi and Michele M. Pelter and Duc H. Do and Randall J. Lee and Rene Colorado and Karl Meisel and Xiao Hu},
  title = {{Photoplethysmography based atrial fibrillation detection: a review}},
  booktitle = {Npj Digit. Medicine},
  year = {2020},
  url = {https://openreview.net/forum?id=Q73vdHKAd0},
}
Figure