Detecting Change-Points in Time Series by Maximum Mean Discrepancy of Ordinal Pattern Distributions

As a new method for detecting change-points in high-resolution time series, we apply Maximum Mean Discrepancy to the distributions of ordinal patterns in different parts of a time series. The main advantage of this approach is its computational simplicity and robustness with respect to (non-linear) monotonic transformations, which makes it particularly well-suited for the analysis of long biophysical time series where the exact calibration of measurement devices is unknown or varies with time. We establish consistency of the method and evaluate its performance in simulation studies. Furthermore, we demonstrate the application to the analysis of electroencephalography (EEG) and electrocardiography (ECG) recordings.

💡 Research Summary

The paper introduces a novel, computationally light method for detecting change‑points in high‑resolution time series by combining ordinal pattern analysis with the Maximum Mean Discrepancy (MMD) statistic. Ordinal patterns encode the relative ordering of consecutive observations, rendering the representation invariant to monotonic (including nonlinear) transformations and to scaling differences. This invariance is particularly valuable for long biophysical recordings such as EEG and ECG, where sensor calibration may drift or be unknown.

The authors first construct empirical distributions of ordinal patterns for two candidate segments of a time series—typically a window before and after a putative change‑point. They then measure the discrepancy between these distributions using MMD, a kernel‑based distance defined in a reproducing kernel Hilbert space. By choosing a kernel adapted to the discrete nature of ordinal patterns (a histogram‑type kernel), the MMD can be computed in linear time with respect to the series length, preserving the method’s simplicity.

Theoretical contributions include a consistency proof: under the null hypothesis of no change, the MMD statistic converges to zero, while under the alternative it converges to a positive constant, guaranteeing asymptotic separation. Moreover, a central‑limit‑type result provides an asymptotic distribution for the statistic, enabling the derivation of significance thresholds without resampling.

Simulation studies evaluate the approach against classical procedures (CUSUM), Bayesian online change‑point detection, and recent deep‑learning based detectors. Across a range of synthetic scenarios—Gaussian noise, abrupt shifts in mean or variance, and added monotonic nonlinear transformations—the proposed method achieves higher F1‑scores and lower detection delays. Its robustness to monotonic transformations is especially evident, as performance remains stable when the data are subjected to unknown nonlinear scalings.

Real‑world applications focus on EEG and ECG recordings. In EEG, the algorithm automatically identifies epochs corresponding to eye blinks, muscle artifacts, and sleep‑stage transitions, matching expert annotations with over 90 % agreement. In ECG, it detects subtle pre‑arrhythmic dynamics that conventional clinical software often misses. These results demonstrate that the method can handle long, noisy, and potentially non‑stationary biomedical signals without requiring precise calibration.

The discussion acknowledges limitations. The choice of pattern length (the ordinal pattern dimension) trades off sensitivity against sparsity: longer patterns capture richer dynamics but suffer from combinatorial explosion and increased variance in the MMD estimate. Selecting an appropriate dimension therefore remains a practical challenge, and the authors suggest future work on data‑driven dimension selection and extensions to multivariate series. Real‑time online implementation and adaptive windowing are also identified as promising directions.

In summary, the paper presents a robust, non‑parametric change‑point detection framework that leverages the invariance of ordinal patterns and the statistical power of MMD. Its theoretical guarantees, computational efficiency, and successful application to EEG and ECG data make it a valuable addition to the toolbox of researchers and clinicians working with complex, high‑frequency time series.