[Paper Review] Kalman-based Spectro-Temporal ECG Analysis using Deep Convolutional Networks for Atrial Fibrillation Detection

Zheng Zhao, Simo Särkkä+1 more|arXiv (Cornell University)|Dec 12, 2018

ECG Monitoring and AnalysisMedicine1 references3 citations

TL;DR

This paper proposes a Kalman-based spectro-temporal ECG analysis framework using deep convolutional networks for atrial fibrillation (AF) detection. By modeling time-varying Fourier coefficients via Kalman filtering and introducing a stochastic oscillator model for computational efficiency, the method achieves an F1 score of 80.2% on the PhysioNet/CinC 2017 dataset, outperforming traditional spectral methods and demonstrating state-of-the-art performance with improved temporal feature retention and scalability for long signals.

ABSTRACT

In this article, we propose a novel ECG classification framework for atrial fibrillation (AF) detection using spectro-temporal representation (i.e., time varying spectrum) and deep convolutional networks. In the first step we use a Bayesian spectro-temporal representation based on the estimation of time-varying coefficients of Fourier series using Kalman filter and smoother. Next, we derive an alternative model based on a stochastic oscillator differential equation to accelerate the estimation of the spectro-temporal representation in lengthy signals. Finally, after comparative evaluations of different convolutional architectures, we propose an efficient deep convolutional neural network to classify the 2D spectro-temporal ECG data. The ECG spectro-temporal data are classified into four different classes: AF, non-AF normal rhythm (Normal), non-AF abnormal rhythm (Other), and noisy segments (Noisy). The performance of the proposed methods is evaluated and scored with the PhysioNet/Computing in Cardiology (CinC) 2017 dataset. The experimental results show that the proposed method achieves the overall F1 score of 80.2%, which is in line with the state-of-the-art algorithms.

Motivation & Objective

• To develop a robust spectro-temporal representation of ECG signals that preserves temporal dynamics better than traditional methods like STFT and CWT.
• To address the computational inefficiency of Kalman-based spectro-temporal estimation in long ECG signals by introducing a stochastic oscillator model with a stationary Kalman filter.
• To evaluate and compare multiple deep learning architectures (e.g., DenseNet, ResNet, RF) on spectro-temporal ECG data for AF detection.
• To achieve state-of-the-art performance in AF detection using end-to-end deep learning on spectro-temporal features derived from Bayesian state-space modeling.
• To improve scalability and accuracy of AF detection in real-world ECG data, particularly for short or noisy recordings, through optimized feature representation and classification.

Proposed method

• Model ECG signals using time-varying Fourier series coefficients in a state-space framework, where the coefficients evolve stochastically over time.
• Apply Kalman filtering and smoothing to estimate the time-varying spectrum, incorporating Gaussian process priors on Fourier coefficients for Bayesian inference.
• Introduce a stochastic oscillator differential equation model to replace the standard Fourier-based state model, enabling faster computation via a stationary Kalman filter and smoother.
• Derive a closed-form solution for the steady-state covariance matrix (P∞) to drastically reduce computational cost during repeated spectro-temporal estimation on multiple ECG segments.
• Transform the estimated spectro-temporal representations into 2D feature maps and feed them into deep convolutional neural networks (e.g., DenseNet-18, ResNet-18) for end-to-end classification.
• Use an averaging strategy across multiple ECG segments to enhance robustness and improve generalization, especially for short recordings.

Experimental results

Research questions

• RQ1 Can Kalman-based spectro-temporal estimation outperform traditional methods like STFT, CWT, and autoregressive modeling in capturing dynamic ECG features for AF detection?
• RQ2 Does the proposed stochastic oscillator model with stationary Kalman filter significantly reduce computation time without sacrificing estimation accuracy?
• RQ3 Which deep learning architecture performs best when classifying spectro-temporal ECG representations into AF, normal, other abnormal, and noisy classes?
• RQ4 How do different spectro-temporal representations affect the activation patterns and feature learning in convolutional layers of deep networks?
• RQ5 To what extent does the averaging of spectro-temporal features improve classification robustness in short or noisy ECG recordings?

Key findings

• The proposed FourierKS and OscKS spectro-temporal estimation methods achieved the highest F1 scores across all tested classifiers, outperforming STFT, CWT, and BurgAR methods.
• The OscKS method reduced computational cost by precomputing the steady-state covariance matrix (P∞), enabling faster estimation without loss of accuracy.
• DenseNet-18 (Dense18+) achieved the best overall performance across all spectro-temporal representations, with an F1 score of 80.2% on the PhysioNet/CinC 2017 dataset.
• Feature map visualization showed that FourierKS and CWT produced more active and diverse responses in the first convolutional layer, particularly in high-energy and structural regions.
• AF detection performance was highest for AF class (F1 > 80%) and lowest for Noisy class (F1 < 60%), with Other class performance being moderate due to class imbalance and diversity.
• The method demonstrated robustness to short ECG segments when combined with averaging, though performance was limited by insufficient data length in some recordings.