Fig. 6: AC population decoder analysis can explain hearing-loss-related behavioral deficits.

a Schematic of AC population AM depth encoding with a linear population readout procedure. Hypothetical population responses for individual trials of a Safe (gray; unmodulated) and Warn (black; modulated) stimulus. Spikes were counted across the entire stimulus duration (1 sec) such that spike firing responses from N neurons to T trials of S stimuli (“Warn” and “Safe”) formed a population “response vector”. A proportion of trials (“leave-one-out” procedure) from each neuron were randomly sampled (without replacement) and fitted to a linear hyperplane that was determined by a support vector machine (SVM) procedure (”train” set). Symbols represent “support vectors”, which are points used to create the linear boundary. Cross-validated classification performance was assessed on the remaining trials (“test” set). Performance metrics included the proportion of correctly classified Warn trials (“Hits”) and misclassified Safe trials (“False Alarms”). Similar to the psychometric and individual unit neurometric analyses, we converted population decoder performance metrics into d′ values. This procedure was conducted across 500 iterations with a new randomly drawn train and test set for each iteration. b Population decoder performance (d′) for two example individual recording sessions as a function of AM depth from Ctrl and HL neuron populations. Neural d′ values were fit with a logistic function (solid line), and threshold is defined as the AM depth where the fit crosses d′ = 1. Sample size indicates the number of single and/or multi-units included within that session. Shaded vertical bars indicate the behavioral threshold for that session. c Within-session correlations of behavioral threshold plotted as a function of multi-and single-unit population thresholds for that session. Dotted lines indicate the fitted linear regression for Ctrl and HL thresholds, and the solid line indicates the linear regression for both groups combined (gray). Pearson’s r (two-tailed) and statistical significance of each fit are noted in the bottom right corner of each plot. Behavioral sessions were included if they had at least 10 single and/or multi-units for the decoder analysis. d Average population decoder performance as a function of AM depth from Ctrl and HL single unit populations pooled across recording sessions (159 total sessions; 88 Ctrl sessions, 71 HL sessions). Sessions were included if there were ≥8 trials per depth.