A Noise-Robust Model-Based Approach to T-Wave Amplitude Measurement and Alternans Detection

TL;DR

This study tackles the challenge of reliable T-wave alternans (TWA) analysis in noisy, ambulatory ECG data by proposing a model-based framework that couples T-wave amplitude modeling with two detection schemes: surrogate data analysis (SDA) and a Markov-model state transition matrix (STM). By generating a synthetic ECG database grounded in real morphology and incorporating respiration and multiple noise types, the authors evaluate T-wave amplitude estimation, TWA level estimation via Modified Moving Average (MMA), and TWA detection under varying SNRs. They demonstrate that a polynomial T-wave model improves amplitude estimation in noisy conditions, while STM-based detection achieves higher F1-scores than SDA at low SNRs and benefits from longer analysis windows (more beats). The combination of model-based estimation and STM detection provides a robust framework for TWA analysis in wearable and ambulatory ECG monitoring, with implications for improved risk stratification in real-world settings. Future work includes validating on clinical data and extending the model to capture more complex alternation patterns beyond binary TWA states.

Abstract

T-wave alternans (TWA) is a potential marker for sudden cardiac death, but its reliable analysis is often constrained to noise-free environments, limiting its utility in real-world settings. We explore model-based T-wave estimation to mitigate the impact of noise on TWA level. Detection was performed using a previous surrogate-based method as a benchmark and a new method based on a Markov model state transition matrix (STM). These were combined with a Modified Moving Average (MMA) method and polynomial T-wave modeling to enhance noise robustness. Methods were tested across a wide range of signal-to-noise ratios (SNRs), from -5 to 30 dB, and different noise types: baseline wander (BW), muscle artifacts (MA), electrode movement (EM), and respiratory modulation. Synthetic ECGs with known TWA levels were used: 0 uV for TWA-free and 30-72 uV for TWA-present signals. T-wave modeling improved estimation accuracy under noisy conditions. With EM noise at SNRs of -5 and 5 dB, mean absolute error (MAE) dropped from 62 to 49 uV and 27 to 25 uV, respectively (Mann-Whitney-U test, p < 0.05) with modeling applied. Similar improvements were seen with MA noise: MAE dropped from 100 to 70 uV and 26 to 23 uV. In detection, the STM method achieved an F1-score of 0.92, outperforming the surrogate-based method (F1 = 0.81), though both struggled under EM noise at -5 dB. Importantly, beyond SNR, detection performance depended on the number of beats analyzed. These findings show that combining model-based estimation with STM detection significantly improves TWA analysis under noise, supporting its application in ambulatory and wearable ECG monitoring.

Figures (8)

• Figure 1: Example of a signal with respiratory modulation and muscle artifact noise at varying SNR levels (-5 dB to 20 dB).
• Figure 2: Example of polynomial modeling of T-waves. On the left a recording corrupted by electrode movement noise with an SNR of 10 dB is shown. The highlighted T-waves in this noisy signal are then modeled using an 8th-degree polynomial, as shown on the right.
• Figure 3: Example of corner cases for off-diagonal probabilities from the STM defined in \ref{['eq:stm']}, illustrating the expected STM patterns for different types of amplitude sequences: (green) perfect beat-wise T-wave alternans; (red) slowly varying periodic patterns such as respiration effects; (orange) Brownian motion (Wiener process), resulting in equal transition probabilities; and (blue) white noise, where frequent state switching leads to a higher probability of transitions than self-repetitions. These cases serve as reference cases for interpreting STMs derived from real or synthetic ECG.
• Figure 4: Distribution heatmaps of STM low-to-high ($p_{\text{LH}}$) and high-to-low ($p_{\text{HL}}$) transition probabilities for synthetic ECG with and without TWA in different SNR levels (0, 5, 10, and 15 dB), per row. In TWA-present cases, both off-diagonal probabilities converge to 1 as SNR increases, indicating increasingly reliable detection of TWA. TWA-free cases exhibit more dispersed distributions centered around 0.3 to 0.7, reflecting randomness in the absence of true TWA.
• Figure 5: Results of T-wave amplitude measurement with added noise, without (a) and with (b) respiratory modulation. Mean absolute error (MAE) with 95% confidence intervals was computed between the ground truth and both raw and modeled T-wave amplitudes (modeling results shown as dashed lines), across different noise types and SNR levels. Stars (color-coded by noise type) indicate cases with statistically significant improvements ($p<0.05$) of the modeled over the raw approach.