Effect of Encoding Method on the Distribution of Cardiac Arrhythmias

This paper presents the evaluation of the effect of the method of ECG signal encoding, based on nonlinear characteristics such as information entropy and Lempel-Ziv complexity, on the distribution of cardiac arrhythmias. Initially proposed a procedure electrocardiographic gating to compensate for errors inherent in the process of filtering segments. For the evaluation of distributions and determine which of the different encoding methods produces greater separation between different kinds of arrhythmias studied (AFIB, AFL, SVTA, VT, Normal’s), use a function based on the dispersion of the elements on the centroid of its class, the result being that the best encoding for the entire system is through the method of threshold value for a ternary code with E = 1 / 12.

💡 Research Summary

The paper investigates how different ECG signal encoding schemes influence the separability of cardiac arrhythmia classes when nonlinear descriptors—Shannon entropy and Lempel‑Ziv (LZ) complexity—are used as features. The authors first address a practical problem in ECG preprocessing: applying conventional low‑pass and high‑pass digital filters introduces an initial‑error segment (due to unknown filter states) and a group‑delay segment (caused by phase lag). They demonstrate that for a 721‑sample segment, up to 65 samples can be lost or corrupted, which would distort subsequent feature extraction. To remedy this, they propose an “ECG gating” method that pads the original segment with additional samples at the beginning (I) and end (F) sufficient to cover the filter’s maximum order. After filtering, only the central original portion is retained, thereby eliminating both initial‑error and delay artifacts at the cost of increased computational load.

Feature extraction relies on two well‑known nonlinear measures. Shannon entropy is computed from the probability distribution of symbols in the encoded sequence, normalized by the alphabet size. LZ complexity is obtained by scanning the symbol sequence and counting the number of new substrings that appear for the first time; this count is then normalized by n·logα n, where n is the sequence length and α the alphabet size. Both metrics capture different aspects of signal randomness and structural regularity.

The core of the study is the comparison of four encoding strategies: (1) slope‑based binary, (2) slope‑based ternary, (3) threshold‑based binary, and (4) threshold‑based ternary. The slope method assigns symbols according to the sign of the first‑difference between consecutive samples (positive, negative, or zero). The threshold method compares each sample to a reference value derived from the segment mean Y and a deviation factor E; for ternary coding two thresholds (Ta and Tb) are defined, yielding symbols {1,0,−1}. The authors test several E values (e.g., 1/10, 1/12, 1/14…) to explore how the granularity of the threshold influences feature dispersion.

To quantify dispersion, they introduce a centroid‑based distance metric λ. For each class i, the centroid Φi is the mean of its feature vectors. For each element eij, the distance lij to its own centroid is compared with distances to all other class centroids. λi counts how many elements are farther from their own centroid than from at least one other centroid; λi = 0 indicates perfect clustering, λi = ni indicates complete overlap. Normalized λiⁿ, total dataset dispersion λD, normalized total λDN, and average per‑class dispersion λDP are derived to enable fair comparison across methods with differing class sizes.

The experimental dataset consists of 2,124 two‑second ECG segments (720 samples each) drawn from the MIT‑BIH arrhythmia database, distributed as 800 Normal, 325 AFIB, 324 AFL, 325 SVTA, and 350 VT. After encoding each segment with the four strategies and computing entropy and LZ complexity, the λ metrics are evaluated. The results show that the ternary threshold encoding with deviation E = 1/12 yields the highest separation: λDN = 0.4035 and λDP = 0.4358. Binary threshold and slope‑based encodings produce lower λ values, indicating more class overlap. Detailed λ tables for all encoding‑E combinations are provided in an online supplement.

The paper’s contributions are threefold: (1) a practical gating technique that removes filter‑induced artifacts, improving the fidelity of downstream nonlinear feature extraction; (2) a systematic evaluation framework based on centroid distances that objectively ranks encoding schemes for arrhythmia discrimination; (3) empirical evidence that a ternary threshold code with a modest deviation (E = 1/12) maximizes class separability when using entropy and LZ complexity as descriptors.

Limitations include reliance on a single public database, which may limit generalizability to other acquisition settings, and the λ metric’s focus on Euclidean distances in a two‑dimensional feature space, which may not capture more complex, nonlinear class boundaries. Future work could extend the analysis to multi‑lead ECG, incorporate additional nonlinear measures (e.g., sample entropy, fractal dimensions), and integrate the λ‑based assessment with machine‑learning classifiers to validate whether the identified optimal encoding indeed improves automated arrhythmia detection performance.