Fast computation of the statistical significance test for spatio-temporal receptive field estimates obtained using spike-triggered averaging of binary pseudo-random sequences

Background

Spatio-temporal receptive fields (STRFs) of visual neurons are often estimated using spike-triggered averaging (STA) with binary pseudo-random stimulus sequences. An exact analytical test—called the STA-BPRS test—has been developed to determine the statistical significance of each pixel in the STRF estimate. However, computing this test can take minutes to days, or even longer, for certain neurons.

New method

Here, the STA-BPRS test is accelerated by approximating the null distribution of STRF pixel estimates with a Normal distribution. This methodological refinement significantly reduces computation time, making large-scale data analysis feasible.

Results

The approximate test is systematically validated on real mouse retinal ganglion cell data and synthetic spike train data, demonstrating that it yields identical significance thresholds to the exact test. For neurons where, exact computation would be prohibitively long (e.g., hundreds of years), the approximate test completes in seconds or minutes.

Comparison with existing methods

Few approaches address pixel-by-pixel significance in STA-based STRF estimates. While subspace methods like spike-triggered covariance exist for STRF estimation, they typically do not provide direct voxel-wise or pixel-wise p-values. The proposed method specifically accelerates an exact distribution-based test.

Conclusions and impact

The proposed Normal approximation drastically reduces computation time, enabling high-throughput analysis of STRF mapping from spike data. This advancement may foster broader adoption of precise statistical tests of STRFs in large-scale, high-density electrophysiological recordings. Moreover, fast detection of significant STRF features could facilitate closed-loop experiments where stimuli dynamically adapt to changing STRF structures.