Time series kernel similarities for predicting Paroxysmal Atrial Fibrillation from ECGs

Filippo Maria Bianchi, Lorenzo Livi, Alberto Ferrante, Jelena Milosevic, Miroslaw Malek

Abstract

This paper introduces a machine learning framework for predicting the onset of Paroxysmal Atrial Fibrillation (PAF) up to 15 minutes in advance. Utilizing ECG signals represented as multivariate time series (MTS), the approach leverages specialized kernel similarity measures—Learned Pattern Similarity (LPS) and Time Series Cluster Kernel (TCK)—to handle missing data and noise inherent in wearable sensor data. The framework achieves state-of-the-art accuracy on the PhysioNet PAF Prediction Challenge dataset, demonstrating robust performance without the need for data imputation.

Introduction

Atrial fibrillation is the most common heart arrhythmia, but the paroxysmal (PAF) variety is particularly difficult to manage because episodes are intermittent and often asymptomatic. While detecting an active episode is valuable, predicting an onset before it occurs is critical for clinical intervention and patient alerting.

The primary technical challenge in analyzing ECG data from wearable sensors is the presence of noise and missing segments. Traditional machine learning models often require "imputation" to fill these gaps, which can introduce significant bias. This research proposes a method that bypasses imputation by using kernel similarities that are inherently robust to missing values.

Methodology

The authors propose a multi-stage pipeline to transform raw ECG data into actionable predictions.

The proposed multi-stage pipeline for ECG-based PAF prediction.
  1. Feature Extraction & Selection: Initially, 95 numerical descriptors are extracted from ECG heartbeats, including R-R intervals and P-wave characteristics. These are pruned to 22 high-impact features using Pearson correlation to reduce redundancy and computational load.
  2. Multivariate Time Series (MTS) Representation: The ECG is transformed into an MTS where each variable tracks a specific feature over time.
  3. Kernel Similarities: This core innovation uses specialized kernels:
  4. Learned Pattern Similarity (LPS): Uses randomized regression trees to identify local patterns.
  5. Time Series Cluster Kernel (TCK): An unsupervised approach using an ensemble of Gaussian Mixture Models. Both methods naturally handle missing data and are highly parallelizable.

  6. Classification: Predictions are made either directly from the kernel similarity matrix (e.g., k-Nearest Neighbors) or by treating similarity values as feature vectors for a classifier like an SVM.

Dataset

The study utilizes the PhysioNet PAF Prediction Challenge dataset, consisting of 106 ECG recordings (30 minutes each). The data includes "control" subjects, "far-from-onset" recordings, and "close-to-onset" recordings (immediately preceding a PAF event).

Results and Discussion

The framework achieved a peak classification accuracy of 71.4% using kNN with the LPS kernel. This performance is on par with or slightly better than the top-performing entries from the original PhysioNet challenge.

Similarity matrices for LPS and TCK kernels, illustrating the distinction between 'far-from-onset' and 'close-to-onset' ECG segments.

Key Insights

  • Imputation-Free Learning: By using TCK and LPS kernels, the model learns from the patterns of available data rather than relying on artificial imputed values.
  • Lead Time Stability: A "sliding window" experiment demonstrated that the model maintains stable accuracy even when analyzing data from 15 minutes prior to an event. This suggests the heart enters a detectable "pre-PAF" state well in advance.
  • Robustness: Representing patients by their similarity to known cases (embedding space approach) provides strong regularization, making the model more robust to sensor noise.

Conclusion

This research demonstrates that specialized time series kernels can effectively predict PAF onset while handling the practical challenges of real-world medical data. While the results are promising, the small size of the benchmark dataset (106 records) suggests that validation on larger, more diverse populations is necessary. Future work will also focus on improving fiducial point detection in noisy environments and optimizing the pipeline for real-time implementation on low-power wearable devices.

Summary updated

Abstract

We tackle the problem of classifying Electrocardiography (ECG) signals with the aim of predicting the onset of Paroxysmal Atrial Fibrillation (PAF). Atrial fibrillation is the most common type of arrhythmia, but in many cases PAF episodes are asymptomatic. Therefore, in order to help diagnosing PAF, it is important to design procedures for detecting and, more importantly, predicting PAF episodes. We propose a method for predicting PAF events whose first step consists of a feature extraction procedure that represents each ECG as a multi-variate time series. Successively, we design a classification framework based on kernel similarities for multi-variate time series, capable of handling missing data. We consider different approaches to perform classification in the original space of the multi-variate time series and in an embedding space, defined by the kernel similarity measure. We achieve a classification accuracy comparable with state of the art methods, with the additional advantage of detecting the PAF onset up to 15 minutes in advance.
BibTeX
@inproceedings{bianchi2018time,
  author = {Filippo Maria Bianchi and Lorenzo Livi and Alberto Ferrante and Jelena Milosevic and Miroslaw Malek},
  title = {{Time series kernel similarities for predicting Paroxysmal Atrial Fibrillation from ECGs}},
  booktitle = {IJCNN},
  year = {2018},
  url = {https://openreview.net/forum?id=29_dKvH1Dn},
}
Figure