Abstract
生体の断層イメージを測定する核磁気共鳴(NMR)のシグナルや、Belousov-Zhabotinsky反応(BZ)などの振動反応の酸化還元ポテンシャルには、測定環境によりSN比が急激に悪くなる時系列データが見られ、そのノイズ除去が問題となっている。また、生体における時系列データの一つである心電図においても、ノイズによる心電図解析の妨げが問題となっている。そこで本研究では、フーリエ変換、ベイズモデル、トレンドモデルの3手法を用いたノイズ除去を行い、それぞれのモデルによる結果の比較、検討を行った。数値実験には、NMRのシグナルに代表的なMultiple nonstationary frequenciesの近似関数と心電図の2種類のデータを用いた。その結果、Multiple nonstationary frequenciesの近似関数では、ベイズモデルが最も有効であった。心電図では、フーリエ変換により、基線の揺れを抑えることができ、ベイズモデルやトレンドモデルを用いて平滑化を行った後に、フーリエ変換によってノイズを除去する方法が有効であることが示唆された。
Introduction
In fields ranging from medical imaging (NMR) to chemical kinetics (Belousov-Zhabotinsky reactions), researchers often encounter time-series data where the Signal-to-Noise Ratio (SNR) drops precipitously due to environmental factors. A particularly messy example is portable ECG monitoring. Unlike clinical ECGs taken while a patient is lying still, portable monitors used during daily activities suffer from two distinct types of noise: high-frequency transmission spikes caused by wireless glitches and low-frequency baseline wander caused by physical movement and breathing. This study evaluates three mathematical filters—Fast Fourier Transform (FFT), Bayesian models, and Trend models—to determine the optimal approach for cleaning such signals without destroying critical diagnostic features like the QRS complex.
Methodology
The researchers compared three distinct mathematical frameworks:
- Fast Fourier Transform (FFT): A standard approach that decomposes a signal into its constituent frequencies, allowing for noise removal via low-pass or high-pass filtering.
- Bayesian Model: Based on Tanabe’s method, this approach treats noise reduction as a probabilistic estimation problem. It utilizes the Akaike Bayesian Information Criterion (ABIC) to find the optimal balance between data fitting and smoothness.
- Trend Model: This uses a state-space representation and the Kalman filter/smoother to model the signal as a long-term trend plus white noise.
Experimental Setup
The study utilized two primary datasets for evaluation:
- Synthetic Nonstationary Data: A mathematical function representing "Multiple nonstationary frequencies," common in NMR and BZ signals, tested at varying SNR levels.
- Real-world Exercise ECG: Data recorded from a healthy individual while climbing stairs, intentionally containing both baseline drift and transmission spikes.
Results and Discussion
Synthetic Signal Performance
The Bayesian model was the clear winner for synthetic nonstationary signals. In the most difficult test case (SNR=5), the Bayesian model achieved a relative error of 3.56, significantly outperforming FFT (4.19) and the Trend model (4.17). The Bayesian approach’s use of ABIC allows for a more automated and robust selection of smoothness compared to the manual tuning required for Trend models.
ECG Noise Reduction
For portable ECG data, no single method was a "silver bullet." FFT was excellent at stabilizing the baseline wander (the slow wavy motion of the graph), while Bayesian and Trend models were superior at smoothing out the jagged transmission noise spikes.
The authors suggest a multi-stage hybrid pipeline: first, use a Bayesian or Trend model to smooth out high-frequency spikes, then apply an FFT-based high-pass filter to flatten the baseline. This preserves the critical QRS peaks needed for medical diagnosis while removing the artifacts of movement and wireless transmission.
Conclusion
While Bayesian models offer superior robustness for nonstationary signals, real-world medical data often requires a hybrid approach to tackle complex noise profiles. Future research aims to disentangle whether model performance is driven by the nature of the underlying signal or the specific characteristics of the noise to further refine these filtering techniques.
Summary updated