Fig. 4: Language responses in the anaesthetized human hippocampus.

a, The response (spike rate, Hz) of an example unit, aligned to word onset (vertical line). Data are mean ± s.e.m. b, Neuronal responses as a function of word frequency shown for each neuron. Each datapoint corresponds to a neuron and word, colour-coded according to the patient. The dashed line shows the linear fit for that patient. c,d, The distribution of the Pearson’s correlation coefficient for predicted versus actual firing rates for all (c) and unique (d) words, per patient. All correlations are greater than zero, indicating that the models are better than chance at fitting neural responses. e, The distribution of words within semantic categories sorted by frequency within the stimulus set, per patient. f, The percentage of units selective for each semantic category, compared with all other semantic categories, per patient. The horizontal pink lines represent values from awake data. g, The number of categories decoded by individual units, per patient. Neurons generally participate in coding of multiple categories. h–j, Similar to e–g, respectively, but for part of speech categories. k,l, The decoding accuracy of a SVM across units for semantic (k) and part of speech (l) categories, shown per patient. The grey lines represent chance; the boxes represent the 25th–75th percentiles, notched at the median. The whiskers are the extrema within 150% of the IQR past the box. The plus symbols indicate outliers. n = 4 patients, 375 units. m, Correlation coefficients for predicted versus actual firing rates as a function of word Index. Index = 0, current word. Negative indices indicate past words; positive indices indicate future words. Black, anaesthetized (n = 4 patients, 375 units); pink, awake (n = 10 patients, 356 units). n,o, The correlation coefficient (Spearman) for each neuron between surprisal and firing rate in anaesthetized (n) and awake (o) patients.