Communication and control in biological systems is mediated by the timing of discharges -spikes- from excitable cells such as neurons and muscle fibers. Each spike is associated to a characteristic waveform that can be captured by sensors. The waveform's characteristics depend on the cell's biophysical properties and the recording modality. Depending on the technique, e.g., electrical recordings with electrodes, optical imaging, ultrasound, the observed signals are mixtures of waveforms emitted from active cells/sources (multiunit data/signals). Recovering the timing and identity of these sources (multiunit or spike decoding) is central to neuroscience, clinical diagnostics, and neural interfacing, yet it remains challenging due to waveform superposition, non-stationarity, limited training labels, and the computational demands of high-density recordings. This review provides a unified methodological perspective on spike decoding by formalizing the problem as a sparse source separation task under a convolutive mixing model. Rather than organizing the literature by application domain, we group and critically compare methods by their underlying principles: classical spike sorting, Bayesian and probabilistic inference, blind source separation, and data-driven approaches, including deep learning and hybrid schemes. For each class of methods, we present the core mathematical formulation and algorithmic strategies and discuss assumptions and limitations. Our synthesis highlights parallels in signal processing across physical recording modalities and clarifies when and why particular approaches succeed or fail. By bridging previously compartmentalized literature, this survey aims to accelerate crosspollination of ideas between application areas and to provide a roadmap for selecting, adapting, and advancing decoding methods across diverse multiunit recording modalities.
Ge et al. (Thu,) studied this question.