Data-Driven Spatio-Temporal Filters for Crosstalk Reduction in High-Density Surface EMG

Objective: Spatial filters have been extensively studied to improve selectivity of information extraction from high density surface electromyography (HD-EMG). However, variations in volume conductors and experimental conditions require tailored solutions, that predetermined filters with fixed weights cannot guarantee to be optimal. This work proposes the development of data-driven filters, adapted to the specific subject and experimental conditions, with a focus on crosstalk. Methods: A general framework to develop filters is introduced. The data are partitioned into patches, which the filter linearly combines across both electrode channels and time samples (obtaining a spatio-temporal filter). With the purpose to reduce crosstalk, different objective functions are proposed: maximal local correlation with the input and minimal cross-correlation across the output channels; minimum entropy of the filter output. Both optimization strategies promote sparse and spatially localized outputs, thereby reducing the effect of distant sources generating crosstalk. Results: Simulated monopolar HD-EMGs from two close muscles have been used to test the filters on a benchmark (n =60: 10 simulated subjects, 3 fat layer thicknesses, 2 levels of additive noise). Moreover, experimental data from selective contractions of the flexor carpi radialis and the pronator teres were used as a preliminary validation in the field (8 subjects). All proposed filters reduced the spread of the recorded signal, better focusing on the sources below the detection point. The estimation of the envelopes of the target muscles (i.e., the muscles located under the detection points of the filters) and the signal-to-crosstalk ratio were statistically improved in simulations when using the filters instead of the raw data. These positive outcomes were confirmed by trends observed in the experimental tests; however, the limited size of the dataset resulted in only a small number of statistically significant improvements. Conclusion: Defining a filter by optimizing an objective function allows to target a property of interest while adapting to the specific recording condition. This adaptive, data-driven approach offers a versatile tool for information extraction from HD-EMGs and may open new avenues for the investigation of neuromuscular signals.