Graded decisions in the human brain

Decision-makers objectively commit to a definitive choice, yet at the subjective level, human decisions appear to be associated with a degree of uncertainty. Whether decisions are definitive (i.e., concluding in all-or-none choices), or whether the underlying representations are graded, remains unclear. To answer this question, we recorded intracranial neural signals directly from the brain while human subjects made perceptual decisions. The recordings revealed that broadband gamma activity reflecting each individual’s decision-making process, ramped up gradually while being graded by the accumulated decision evidence. Crucially, this grading effect persisted throughout the decision process without ever reaching a definite bound at the time of choice. This effect was most prominent in the parietal cortex, a brain region traditionally implicated in decision-making. These results provide neural evidence for a graded decision process in humans and an analog framework for flexible choice behavior.

), which show significant modulation for both saccade and button press choices.b, SC&BP-graded electrodes (Fig. 5a), in which the signals were significantly modulated by the DV at the time of choice for both saccade and button press choices.c, SC-modulated and BP-modulated electrodes (Fig. 3b), which show significant modulation for saccade and button press choices, respectively.d, SC-graded and BP-graded electrodes (Fig. 5b), in which the signals were significantly modulated by the DV at the time of choice for saccade and button press choices, respectively.e, The proportion of electrodes located within individual Brodmann areas.f, DV-graded electrodes (in d) during congruent and reversed task contexts.Specifically, the proportion of SC-graded and BP-graded electrodes in congruent (reversed) sessions was 2.4% (1.2%) and 4.1% (3.2%), respectively.Subject S8 also provided access to deep brain targets.Electrodes from all 13 sessions (8 subjects) are included.The numbers above each bar indicate the number of sessions (numerator) and number of subjects (denominator).was derived from the presented evidence (black dots indicating time and laterally of click sounds) in the same way as many previous studies 15,70 .b, DV at the time of choice (DV@choice).We define the DV at the time of choice as the DV value 100 ms preceding movement onset.c, Model-free analysis.We averaged the broadband gamma ( ) from effector-modulated electrodes (black traces), and calculated the mean value of the averaged over a 100 ms window preceding movement onset for each trial (y@choice).d, Model-based analysis.We applied a linear regression between the of the effector-modulated electrodes (black traces) and the corresponding value of the modeled DV (red trace).This provided a set of weights that enabled us to predict the modeled DV from the .Next, we computed the correlation between y@choice and DV@choice, and determined its significance using a randomization test.To avoid the potential mixing of decision difficulty and response type, we have separated saccade and button press trials in the model-free and model-based analysis.The analysis of the congruent and reversed data was also performed separately.We used saccade trials as an example in this graphical figure.⇤: BP-modulated regions were significantly graded by the DV at the time of saccade choices (top histogram) with an average R = 0.12 (t(12) = 3.4, p = 5.6 ⇥ 10 3 , two-tailed t-test); SC-modulated regions were significantly graded by the DV at the time of button press choices (bottom histogram) with an average R = 0.11 (t(12) = 3.9, p = 2.0 ⇥ 10 3 , two-tailed t-test).b Graded effect in effector-selective regions.Effectorselective regions were defined as those exclusively modulated by the effector indicated in the subtitle.Same analyses and format as a, but for the BP-selective and SC-selective electrodes.The BP-selective regions were not graded (ns, p > 0.05) by the DV at the time of saccade choices (top histogram) with an average R = 0.07 (t(12) = 2.0, p = 0.07; two-tailed t-test).Similarly, The SC-selective regions were not graded by the DV at the time of button press choices (bottom histogram) with an average R = 0.0 (t(9) = 0.0, p = 1.0; two-tailed t-test).We used the same format and analysis as in Fig. 3b. .Decision-related signals in the peripheral nervous system.a, Electromyographic (EMG) activity of the hand pressing the button (mean ± s.e.m., n = 13) as a function of time.b, Horizontal eye gaze (mean ± s.e.m., n = 13) as a function of time.To better visualize the DV-graded effect, in each session, the EMG and eye gaze signals were normalized by subtracting the trial-averaged signal for saccade and button press choice, respectively.We explored the DV-graded effect in the effector peripheral systems that execute each choice.On each trial, the EMG amplitude was averaged between 50 ms preceding and 50 ms following a button press.The amplitude of eye gaze was taken as the maximal value within the period from 200 ms preceding to 200 ms following saccade onset.We found that for valid trials, hand EMG and eye gaze were significantly (p < 0.05) graded by the DV at the time of choice using the same analysis as Fig. 3b.⇤: t(12) = 4.0, p = 0.0017 for hand EMG in button press choices; t(12) = 3.1, p = 0.010 for eye gaze in saccade choices; two-tailed t-tests.

Figure S1 .Figure S2 .
Figure S1.Subjects information.a, Electrode coverage for all subjects.The black dots show the locations of all 799 electrodes (prior visual inspection) inferred from CT scans.b, Electrode locations were pooled across the subjects and projected onto a template brain.For visualization purposes, we projected the left hemisphere electrodes of subjects S2, S4, S6, and S7 onto the right hemisphere.Solid/hollow circles indicate the location of all 571 electrodes (after visual inspection) across all subjects engaged in the congruent/reversed tasks, respectively.c, Information of individual subjects.8 humans (with mean±s.d.age and IQ equal to 39±15 and 94.5±17.2,respectively) with intractable epilepsy participated in our study.d Number (mean±s.e.m.) of performed trials in congruent and reversed tasks.The black circles indicate the average number of trials for each session.Specifically, subjects performed an average of 218 (241) congruent trials for the saccade (button press) choice, while subjects performed an average of 205 (252) reversed trials.e, Auditory-related electrodes that were excluded.Electrodes with significant (p < 0.05, corrected using false discovery rate, one-tailed randomization tests) broadband gamma increase for both the decision task and passive listening task were defined as auditory-related electrodes.We extracted 79/571 auditory-related electrodes from eight subjects, and these electrodes were excluded.Together, 492 electrodes in eight subjects were included for further analyses.

Figure
Figure S3.Effector-modulated and DV-graded Brodmann areas.a, SC&BP-modulated electrodes (Fig.3a), which show significant modulation for both saccade and button press choices.b, SC&BP-graded electrodes (Fig.5a), in which the signals were significantly modulated by the DV at the time of choice for both saccade and button press choices.c, SC-modulated and BP-modulated electrodes (Fig.3b), which show significant modulation for saccade and button press choices, respectively.d, SC-graded and BP-graded electrodes (Fig.5b), in which the signals were significantly modulated by the DV at the time of choice for saccade and button press choices, respectively.
FigureS4.A graphical rendering of the analyses.a, Signal processing.The decision variable (DV, red trace) was derived from the presented evidence (black dots indicating time and laterally of click sounds) in the same way as many previous studies15, 70  .b, DV at the time of choice (DV@choice).We define the DV at the time of choice as the DV value 100 ms preceding movement onset.c, Model-free analysis.We averaged the broadband gamma ( ) from effector-modulated electrodes (black traces), and calculated the mean value of the averaged over a 100 ms window preceding movement onset for each trial (y@choice).d, Model-based analysis.We applied a linear regression between the of the effector-modulated electrodes (black traces) and the corresponding value of the modeled DV (red trace).This provided a set of weights that enabled us to predict the modeled DV from the .Next, we computed the correlation between y@choice and DV@choice, and determined its significance using a randomization test.To avoid the potential mixing of decision difficulty and response type, we have separated saccade and button press trials in the model-free and model-based analysis.The analysis of the congruent and reversed data was also performed separately.We used saccade trials as an example in this graphical figure.

Figure S6 .Figure S7 .
Figure S5.Effector-related broadband gamma ( ) signals index developing decisions.a, Decision signals in effector-modulated regions.Left panel: Mean (n = 13 sessions) activity of SC-modulated (top panel) and BPmodulated (bottom panel) regions (Fig. 3b) for the three different levels of decision evidence (HI, ME, LO).Middle panel: Mean (n = 13) of least squares fitted activity of SC-modulated (top panel) and BP-modulated (bottom panel) regions.Right panel: Mean (n = 13) value of the DV as a function of time.The signals are aligned to the stimulus onset (dashed line).The top (bottom) panel shows activity (left, middle panels) and DV (right panel) for trials that resulted in a saccade (button press) choice.b, Spearman's correlation R between the time course of activity and the time course of DV, throughout each decision period (stimulus onset to the time of choices).The individual values in the histogram represent individual trials, and are presented separately for saccade (top histogram) and button press (bottom histogram) choices.The triangle denotes the mean R value.⇤: p = 8.4 ⇥ 10 21 and p = 8.1 ⇥ 10 40 for saccade and button press, respectively (two-tailed t-tests).