Surface electromyography (sEMG) datasets are widely used to develop and benchmark machine learning models for gesture recognition. Many studies investigating classification performance under different signal resolutions assume that publicly available datasets preserve the quantization characteristics of the original acquisition hardware, often reported as 16-, 24-, or higher-bit analog-to-digital conversion. However, this assumption is rarely verified. Signal processing operations such as filtering, normalization, interpolation, or floating-point conversion may alter or even destroy the original discrete quantization structure, potentially introducing arbitrary intermediate values that were not present in the raw measurements.
This work introduces the concept of quantization integrity to evaluate whether public sEMG datasets preserve the discrete amplitude structure associated with their original measurement process. A data-driven methodology is proposed based on minimum nonzero sample difference (Δmin), zero-difference proportion between consecutive samples, signal cardinality, and histogram analysis of difference magnitudes. These metrics enable the identification of preserved quantization levels as well as numerical densification caused by post-processing.
The motivation for this analysis arises from ongoing research on wearable multichannel sEMG systems, where low- and medium-resolution acquisition (e.g., 8–12 bits) offers practical implementation advantages over high-resolution systems (16–24 bits) in terms of hardware complexity, bandwidth, and power consumption. Meaningful performance comparisons across resolutions require reliable high-resolution references. The presented framework helps determine whether public sEMG datasets can serve as trustworthy high-resolution baselines for controlled resolution degradation experiments and reproducible gesture recognition studies.