Receiver operating characteristic (ROC) analysis of neural code efficacies

Abstract.

 Action potentials (APs) were intracellularly recorded from eccentric cells (which give rise to optic nerve fibers) in Limulus lateral eyes and their neural coding efficacies were determined over a wide range of light adaptation states and relative stimulus intensities. Extremely stringent data quality procedures were followed to ensure that the results are based on stable preparations. Waveforms which could be compared with those of comparable receptor potentials (RPs) were obtained by constructing plots of the reciprocals of successive interspike intervals, creating instantaneous spike frequency waveforms (ISFWs). Six candidate codes were then measured. They were: the area under the light-evoked ISFW, the mean value of the ISFW, the peak height of the ISFW, the slope of the onset of the ISFW, the duration required for the ISFW to drop from its peak by a given amount, and the duration required for the ISFW to end. Receiver operating characteristic (ROC) analyses were then applied to these coded ISFW values to provide objective indices of AP efficacies in the form of detectability (d′) measurements. Several reliable findings were obtained: (1) Both adaptation state and relative intensity affect efficacy. More specifically, the d′ values of all codes approach zero, representing chance detectability, when relatively weak flashes are presented in dark-adapted states. (2) Light adaptation produces a sensitivity-acuity tradeoff: as sensitivity decreases in more light-adapted states, detectabilities increase, indicating that ROC-characterized discrimination acuity increases. (3) The AP code efficacy order is similar to but not identical with the efficacy order previously found in photoreceptors: area = peak ? mean ? slope = duration-end = duration-drop. The previously measured photoreceptor RP efficacies were quantitatively compared with the present AP data from the optic nerve fiber level, with the following results: (1) All efficacies tend to decline at the more proximal neural level. (2) The decline is code dependent and transmission efficacy falls in this order: peak > area > mean > slope = duration-end =duration-drop. (3) The code which transmits best between photoreceptors and optic nerve fibers (i.e., the peak) differs from the code which has the highest efficacy at the photoreceptor level (i.e., the area); at the more proximal level, these two codes have indistinguishable efficacies. These findings support two conclusions: (1) They further demonstrate that arbitrary characterizations of stimulus-response relationships are very likely to be incomplete. This would be particularly important when, as is often the case in brain research, the mean code alone (i.e., the average firing rate) is used to characterize spike potentials. The present data show that the use of that code would have substantially underestimated detectability had it been used alone. (2) The fact that the code which most completely transmits information between cells is not the code which most completely represents information within a cell leads directly to a possible physiological basis for the existence of multiple neural codes. It particularly leads to the following extension of the Task Dependence Hypothesis of neural coding: Different codes may mediate performance in different behavioral tasks because different codes best serve different neural functions.