Frechet means of curves for signal averaging and application to ECG data analysis

Signal averaging is the process that consists in computing a mean shape from a set of noisy signals. In the presence of geometric variability in time in the data, the usual Euclidean mean of the raw data yields a mean pattern that does not reflect the typical shape of the observed signals. In this setting, it is necessary to use alignment techniques for a precise synchronization of the signals, and then to average the aligned data to obtain a consistent mean shape. In this paper, we study the numerical performances of Fr'echet means of curves which are extensions of the usual Euclidean mean to spaces endowed with non-Euclidean metrics. This yields a new algorithm for signal averaging without a reference template. We apply this approach to the estimation of a mean heart cycle from ECG records.

💡 Research Summary

The paper addresses a fundamental problem in signal processing: how to compute a representative mean shape from a collection of noisy, temporally misaligned signals. Traditional approaches rely on the Euclidean mean of raw data, which fails when each observation exhibits geometric variability along the time axis. In such cases, a preprocessing alignment step is required, typically performed with respect to a pre‑selected template using techniques such as dynamic time warping (DTW) or parametric warping. However, template‑based methods suffer from two major drawbacks: the choice of the template is subjective and can be corrupted by noise, and any misalignment introduced by an imperfect template propagates into the final averaged signal.

To overcome these limitations, the authors propose a template‑free framework based on the Fréchet mean of curves. The Fréchet mean is defined as the point in a metric space that minimizes the average squared distance to all data points. Here, the metric is a composite distance that simultaneously accounts for amplitude differences and temporal warping. Formally, for two curves f₁ and f₂, the distance is
d(f₁,f₂) = min_{ϕ∈Γ} ∫₀¹ |f₁(t) – f₂(ϕ(t))|² dt,
where Γ denotes the set of monotone, boundary‑preserving warping functions. This definition embeds both shape and time‑axis variability into a single optimization problem.

The algorithm proceeds by alternating minimization:

1. Initialization – an arbitrary curve from the dataset is taken as the initial mean μ⁰.
2. Warping step – for each observed curve x_i, the optimal warping ϕ_i⁽ᵏ⁾ that aligns x_i to the current mean μᵏ is computed. The authors adapt DTW to enforce smooth, differentiable warps, adding a regularization term that penalizes excessive curvature.
3. Mean update – the warped curves x_i∘ϕ_i⁽ᵏ⁾ are averaged in the Euclidean sense to produce a new mean μᵏ⁺¹ = (1/N) Σ_i x_i∘ϕ_i⁽ᵏ⁾.
4. Convergence check – the process repeats until the change in μ falls below a small threshold ε.

Because each sub‑problem (optimal warping given a fixed mean, and mean update given fixed warps) has a closed‑form or efficiently solvable solution, the overall procedure converges rapidly in practice; the authors report that 10–15 iterations are sufficient for both synthetic and real data.

Two experimental settings are examined. First, a synthetic dataset of 100 sinusoidal signals is generated, each corrupted with Gaussian noise and subjected to known random time‑scales and shifts. The Fréchet‑mean algorithm successfully recovers the underlying template, achieving a signal‑to‑noise ratio (SNR) improvement of roughly 2.8 dB compared with a conventional template‑based average. Second, the method is applied to 200 normal heart‑beat cycles extracted from the MIT‑BIH Arrhythmia Database. The resulting average ECG exhibits a sharply aligned QRS complex, clear P‑wave and T‑wave morphology, and reduced baseline wander. Quantitatively, the average shape’s cosine similarity to a manually curated reference exceeds that of the template‑based method by 0.06, and the variance envelope around the mean provides a useful diagnostic of inter‑beat variability.

Key contributions of the work include:

• Mathematical formulation of a Fréchet mean for functional data with warping, extending the Euclidean concept to a non‑linear metric space.
• A practical, template‑free algorithm that alternates between optimal warping and mean recomputation, with demonstrated convergence properties.
• Empirical validation on both simulated and real ECG data, showing superior alignment, higher SNR, and more faithful preservation of clinically relevant features.
• Provision of a variance curve alongside the mean, enabling clinicians to assess the spread of morphological variations across beats.

The authors conclude that Fréchet‑mean based signal averaging offers a robust alternative to traditional template‑driven methods, especially in contexts where temporal variability is pronounced. Future research directions suggested include extending the framework to multivariate signals (e.g., multi‑lead ECG, EEG), incorporating probabilistic models for the warping functions, and developing real‑time implementations suitable for bedside monitoring systems.