Risk and aversion coding in human habenula high gamma activity

Abstract

Neurons in the primate lateral habenula fire in response to punishments and are inhibited by rewards. Through its modulation of midbrain monoaminergic activity, the habenula is believed to play an important role in adaptive behavioural responses to punishment and underlie depressive symptoms and their alleviation with ketamine. However, its role in value-based decision-making in humans is poorly understood due to limitations with non-invasive imaging methods which measure metabolic, not neural, activity with poor temporal resolution.

Here, we overcome these limitations to more closely bridge the gap between species by recording local field potentials directly from the habenula in 12 human patients receiving deep brain stimulation treatment for bipolar disorder (n = 4), chronic pain (n = 3), depression (n = 3) and schizophrenia (n = 2). This allowed us to record neural activity during value-based decision-making tasks involving monetary rewards and losses.

High-frequency gamma (60–240 Hz) activity, a proxy for population-level spiking involved in cognitive computations, increased during the receipt of loss and decreased during receipt of reward. Furthermore, habenula high gamma also encoded risk during decision-making, being larger in amplitude for high compared to low risk. For both risk and aversion, differences between conditions peaked approximately between 400 and 750 ms after stimulus onset.

The findings not only demonstrate homologies with the primate habenula but also extend its role to human decision-making, showing its temporal dynamics and suggesting revisions to current models. The findings suggest that habenula high gamma could be used to optimize real-time closed-loop deep brain stimulation treatment for mood disturbances and impulsivity in psychiatric disorders.

Manssuer et al. record neural activity directly from the human habenula of patients receiving DBS treatment for psychiatric disorders. High-frequency activity in the habenula encodes punishment and risk, and could be a novel biomarker for closed-loop DBS therapy.

Introduction

The brain imbues stimuli with reward and punishment values to motivate adaptive behavioural responses, but when this process malfunctions it can lead to problems. A major goal of research disentangling the complex neural systems subserving value-based decision-making is the identification of cells, nuclei or regions that exclusively encode reward or aversion, thereby highlighting potential treatment targets.¹ While research in primates and rodents has shown that several regions such as the amygdala and orbitofrontal cortex encode both reward and aversion,^2,3 the lateral habenula (LHb) has emerged as a promising region specifically encoding aversion. Single neurons in the LHb are excited by punishments or the omission of rewards, inhibited by rewards and acquire similar responses to associated cues via negative prediction error learning.^4–7 The LHb receives inputs from forebrain regions via the striatum and hypothalamus and regulates monoamine secreting neurons in the brainstem. It is therefore an area of intense interest as a potential mediator of psychiatric symptoms, particularly depression, and its treatment with ketamine.^8–11

However, an understanding of the habenula’s role in human value-based decision-making is limited.^12–15 These studies are restricted to functional MRI (fMRI), which measures metabolic, not neural, activity and as a result has poor temporal resolution. As such, it can confound risk assessment activity from subsequent choice¹⁶ and cannot differentiate excitatory from inhibitory activity, which is important for contrasting habenula responses to reward and punishment given the inhibitory responses seen to reward.¹⁷ The use of deep brain stimulation (DBS) of the habenula in clinical trials^18,19 affords the rare opportunity to overcome these limitations by recording local field potentials (LFPs) during cognitive tasks with unparalleled signal to noise ratio, spatial and temporal precision.^20–22 Different frequency bands of the LFP provide a wealth of information and have distinct underlying generators and functions. Low frequencies are generated by postsynaptic potentials and orchestrate high gamma (60–240 Hz) activity, which is believed to reflect population-level spiking activity and to be involved in cognitive computations.^23–25 To date there have been two task-based LFP studies of human habenula function,^21,22 but neither were concerned with value-based decision-making and focused solely on slower oscillations (<40 Hz), which are believed to act more as carrier frequencies instead of reflecting local cognitive computations like high gamma.^20,23

In this study, we used high gamma LFPs to more closely bridge the study of value-based decision-making in the habenula between primates and humans. Value-based decision-making involves multiple computations such as valuation, motivation, learning and choice.²⁶ Some types of value-based decision-making are simple whereas others are complex.²⁶ We used two tasks to measure these different components while LFPs were recorded from patients implanted with habenula DBS electrodes for the treatment of refractory depression, bipolar disorder, schizophrenia and chronic pain. In the first experiment, we examined the LFP dynamics of habenula activity during a monetary incentive delay (MID) task, which involves the anticipation and receipt of rewards and losses which have the effect of speeding up reaction times. This task measures motivation, learning and outcome valuation and is a relatively implicit, low-level, Pavlovian type of valuation and behavioural output. We expected, and show, homologous patterns of high gamma modulation in the MID task to neurons coding reward and aversion in the primate LHb. In the second experiment, we examined habenula LFPs during a more high-level, explicit gambling task that manipulated the risk and uncertainty of receiving or losing money. We demonstrate that human habenula high gamma is not just involved in passive conditioning or receipt of outcomes but also specifically recruited in inferring the likelihood of future reward or loss based on probabilistic information and used to guide betting choices—a more complex type of decision-making behaviour. The findings represent an important step towards identifying biomarkers that can optimize real-time closed-loop DBS treatment for psychiatric disorders.

Materials and methods

Patients

The study took place in the neurosurgical service of Ruijin Hospital, Shanghai Jiao Tong University. In total, twelve patients took part (Table 1). Patients had an average age of 40 years old (SD = 17) and were mostly male (n = 10). All patients were undergoing experimental DBS of the habenula to treat refractory depression (n = 3), bipolar disorder (n = 4), schizophrenia (n = 2) or chronic pain (n = 3). Three patients did the risk task only, four patients did the MID task only and five patients did both tasks. Four patients did the risk task first followed by the MID task and one patient vice versa. On average, the MID task took 15.5 min and the risk task took 10.5 min. The ethics committee of Ruijin hospital, Shanghai Jiao Tong University School of Medicine approved all procedures used. All patients provided written informed consent in accordance with the Declaration of Helsinki.

Table 1.

Patient information

Patient	Diagnosis	Handedness	MID	Risk	Age	Gender	Hemisphere	MoCA	HAMD-17	HAMA-14	Medications at testing	Electrode model	Contacts analysed
1	Depression	Right	Yes	No	21	Male	Inaccuratea	–	24	27	–	Medtronic 3389	–
2	Schizophrenia	Right	Yes	No	21	Male	Left/right	–	–	–	Quetiapine; benzhexol	Medtronic 3389	L1–L2
3	Bipolar disorder	Right	Yes	Yes	46	Male	Left/right	26	23	24	Lamotrigine; quetiapine; magnesium valproate; amfebutamone; clonazepam	Medtronic 3389	L1–L2/R1–R2
4	Bipolar disorder	Right	Yes	Yes	30	Female	Left/right	28	21	20	Olanzapine; escitalopram oxalate	Medtronic 3389	L0–L1/R0–R1
5	Schizophrenia	Right	Yes	Yes	16	Male	Left/right	–	–	–	Quetiapine; risperidone; escitalopram oxalate	Medtronic 3389	L0–L1/R0–R1
6	Bipolar disorder	Right	No	Yes	48	Male	Left/right	23	23	24	Mirtazapine; lithium carbonate; sodium valproate sustained-release; clonazepam	Medtronic 3389	L0–L1/R0–R1
7	Depression	Right	Yes	No	28	Female	Left/right	29	28	31	Venlafaxine; mirtazapine; lithium carbonate; pregabalin; zopiclone	PINS L301	L1–L2/R1–R2
8	Depression	Right	Yes	No	35	Male	Left/right	27	25	33	Quetiapine; escitalopram oxalate; clonazepam	Sceneray 1200–30/40	R1–R2
9	Chronic pain	Right	No	Yes	71	Male	Right	–	4	0	Morphine; venlafaxine	Medtronic 3389	R0–R1
10	Chronic pain	Right	No	Yes	60	Male	Left	–	5	5	Pregabalin	Medtronic 3389	L1–L2
11	Chronic pain	Right	Yes	Yes	53	Male	Right	–	7	8	Pregabalin; tramadol; stilnox	Medtronic 3389	R1–R2
12	Bipolar disorder	Right	Yes	Yes	48	Male	Left/right	26	–	–	Duloxetine; olanzapine; clonazepam; tandospirone citrate	Medtronic 3389	L1–L2

MoCA = Montreal Cognitive Assessment Score; HAMD-17 = Hamilton Depression Score; HAMA-14 = Hamilton Anxiety Score.

All electrodes contacts were outside the habenula.

Monetary incentive delay task

In the MID task, patients saw one of three distinctive cues which signalled whether a reward or loss could be received depending on the speed and correctness of a simple visual discrimination response (Fig. 1). This task allows us to separate the period between the cue and response, when the subject anticipates the outcome, from when the outcome is received. On reward cue trials, a correct response led to a monetary reward, depicted by a 10 yuan note, and an incorrect or no response led to no reward/a neutral grey box. On loss cue trials, an incorrect or no response led to monetary loss, depicted by a 10 yuan note with a red cross overlaid, and a correct response led to no monetary loss/a neutral grey box. On neutral cue trials no money was won or lost and a grey box was shown regardless of correctness. To ensure equal numbers of trials across conditions and that patients were incentivized to respond quickly, the allotted time within which patients were allowed to respond started at 800 ms and was increased or decreased adaptively by 50 ms depending on the previous response being incorrect or correct, with a maximum time of 1000 ms and no fixed minimum. This means that roughly half of trials in each condition will be correct and incorrect/no response. Throughout the paper we define incorrect trials as both erroneous arrow classifications and non-responses unless otherwise stated. The 500 ms cue was followed by the 2000 ms anticipation phase during which the coloured square remained on the screen without the icon as a reminder of the type of outcome. A white arrow then appeared. Patients were instructed to press one of two buttons on a response box with their left or right thumb as quickly as possible in the direction of the arrow. This visual discrimination ensured patients maintained alertness and were attentive throughout the task. After the response, there was a blank screen of at least 500 ms followed by the outcome, which was presented for 2000 ms. The blank interval was intended to give any response-related activity time to dissipate so that it did not confound outcome activity. The total duration from arrow onset to outcome onset was always 1500 ms. The inter-trial interval (ITI)/fixation was 1500–2500 ms. Patients were told that they needed to respond as quickly as possible to either win or avoid losing money. Five patients performed 30 trials per condition and four performed 40 trials per condition. These were completed after 18 practice trials. The cumulative amount won/lost was presented in size 200 font below the money. Patients sat approximately 75 cm away from the screen. Both MID and risk tasks were programmed and run in MATLAB using Psychtoolbox 3.0 functions²⁷ and were displayed on an LG L1954 monitor that has a width/height of 380 × 300 mm and a resolution of 1280 × 1024 pixels.

Figure 1.

MID task. On each trial, patients saw a cue for 500 ms that was followed by a delay for 2000 ms, after which an arrow appeared requiring a speeded button press. Depending on the type of cue (for reward, loss or neutral) and whether the response was correct and quick enough, the patient was then presented with either a monetary gain, loss or neutral stimulus for 2000 ms after a short blank interval. See the ‘Materials and methods' section for more details.

Risk task

To assess the role of the habenula in coding risk, uncertainty and value-based choice, we used a card task (Fig. 2). In this task patients were presented with two playing cards on each side of the screen. One of the cards faced upwards and the second card faced downwards. The card numbers ranged from 1 to 10. The patients’ task was to press a button with their right thumb if they wanted to bet 1 yuan that the second card was higher than the first or press a button with their left thumb to abstain from betting. If the patient bet and the second card was higher than the first, then the patient won 1 yuan. However, if the patient bet and the second card was lower than the first, they lost 1 yuan. If patients chose not to bet, they neither won nor lost anything regardless of whether the second card was higher than the first. The two cards were initially presented for 2.1 s before a blue box appeared around the cards signalling to patients that they could make their response. This allowed us to separate decision-making activity from subsequent response activity. Patients were instructed to respond as quickly as possible and to try and win as much money as possible and had within 2 s to respond. After patients made their response, there was a 700 ms delay, to allow for any motor-related activity to dissipate, after which the second card was revealed. The number on the second card was chosen at random. This screen was presented for 1 s followed by how much the patient won or lost for 1.5 s. If the patient won, they were presented with the words ‘you win’ and a 1 yuan note surrounded by a green box. If the patient lost they were presented with the words ‘you lose’ surrounded with a red box with a red cross overlaid. If patients did not bet, they were presented with a screen saying ‘you win 0, you lose 0’. The cumulative total won or lost was presented beneath. Trials were separated by a jittered 1–1.5 s ITI in which a central fixation cross was presented. There were 78 trials in total. There were nine trials each for cards with a value between 2 and 9. There were three trials for cards with a value of 1 and 10. Patients were fully instructed on how to complete the task and completed 10 practice trials prior to the main task.

Figure 2.

Risk-taking task. On each trial patients saw two cards on the screen. One of the cards was faced upwards and the other downwards. Patients had to decide whether to bet 1 yuan that the card faced downwards was higher than the card faced upwards. Patients had 2.1 s to view the cards after which a rectangle appeared around them signalling that a response could be made. After patients made a button press, there was a delay of 700 ms after which the second card was revealed for 1 s and followed by the outcome for 1.5 s. The second card was randomly chosen. If patients bet and the second card was higher than the first, patients won 1 yuan. If patients bet and the second card was lower than the first, patients lost 1 yuan. If patients did not bet, they did not win or lose anything. The total cumulative amount won or lost throughout the task was presented below the 1 yuan won or lost on that trial. See the ‘Materials and methods' section for more details.

Behavioural data analyses

Reaction times were normalized using logarithmic transformation, z-scored to facilitate comparison across patients and outliers above or below 2.5 standard deviations (SD) from each patients mean were excluded from analyses. Reaction times were analysed with a linear mixed effects (LME) model using the fitlme function in MATLAB with the restricted maximum likelihood method. This method is advantageous over standard t-tests as it allows us to model trial-by-trial variation as a fixed effects factor, inter-subject variation as a random-effects factor and is not biased by the small differences in the number of trials between conditions. We confirmed that the assumptions for LME analysis were met. The residuals were shown to be normally distributed using histograms and qq-plots and heteroskedastic by plotting them as a function of the fitted values. As accuracy and betting behaviour are binomial, they were analysed using a generalized linear mixed effects model (GLME) with a logit link function using the fitglme function in MATLAB. The significance of each factor in the LME and GLME models are evaluated using a t-statistic.

Electrode implantation and localization

Patients were implanted with platinum–iridium alloy DBS electrodes encased in a poly-urethane sheath with four 1 mm contacts separated by 0.5 mm (either Medtronic 3389, PINS L301 or Sceneray 1200–30/40) under general anaesthesia using MRI-guided targeting (3T General Electric). The MRI was co-registered with the CT (General Electric) and with the Leksell stereotactic frame to obtain coordinate values.¹⁹ All patients had electrodes implanted bilaterally apart from the chronic pain patients who were implanted unilaterally. Lead DBS 2.5.3 was used to localize the DBS electrodes in MNI space.²⁸ The preoperative T₁-weighted anatomical MRI was co-registered with the postoperative CT using advanced normalization tools (ANTs) and subcortical refine. To improve co-registration, a T₂-weighted MRI was co-registered to the T₁-weighted MRI using the statistical parametric mapping (SPM) method. The CT was then normalized to the MNI_ICBM_2009b_NLIN_ASYM brain using ANTs. Accuracy of co-registration and normalization was inspected visually. Brain shift was corrected for using a course mask. The electrode paths were found automatically using the trac/core method, followed by manual adjustment.

Local field potential recording and preprocessing

The electrode leads were temporarily externalized for 1 week during which testing took place. During testing, LFP data were recorded using a BrainAmp MR amplifier (Brain Products) at a 500 Hz sample rate. The data were referenced online using a left mastoid electrode. The electro-oculogram (EOG) was recorded to confirm blinks and saccades did not contaminate the LFPs. Offline, the LFP data were preprocessed and analysed using MATLAB 2019b and FieldTrip.²⁹ The electrode contacts were bipolar re-referenced by subtracting adjacent contacts (L0–L1, L1–L2, L2–L3, R0–R1, R1–R2, R2–R3). This ensures that activity is restricted to the habenula as volume conducted activity from distant brain regions and reference-related activity is cancelled out. One contact from each hemisphere was selected for analysis if it was located in, or in contact with, the habenula with the highest probability. The contact with the next highest probability of being in the habenula, or closest to it, was used as the reference electrode (Fig. 3, Table 1 and Supplementary Fig. 1). This electrode selection procedure is commonly used.^21,22 Two-way IIR Butterworth zero-phase lag high-pass and notch filters were used to remove direct current (DC) bias below 1 Hz and powerline noise at 50 Hz and its harmonics, respectively. The data were z-normalized to facilitate comparison across patients and visually inspected (blind to conditions) to remove epochs or electrodes contaminated with artefactual activity. Across all analyses, the number of trials in each condition was equalized.

Figure 3.

Habenula recording sites. (A) Sagittal, coronal and axial views of the MNI standard brain with the location of the habenula highlighted. (B) 3D images of the habenula with the contact locations used in the analyses shown as spheres. Different colours represent different patients. A = anterior; I = inferior; L = left; R = right; P = posterior; S = superior.

High gamma amplitude analysis

To assess high gamma modulation, we band-pass filtered the signal between 60 and 240 Hz using a two-pass finite impulse response (FIR) zero-phase-lag filter and extracted the instantaneous amplitude/envelope by taking the absolute value of the Hilbert transform. The data were smoothed with a 250 ms moving average sliding window. After trials were averaged across conditions, all data (cue/anticipation/decision/outcome) were baseline corrected by subtracting the mean activity in the final 500 ms of the fixation cross period prior to cue onset in the MID task or decision phase onset in the risk task. Where possible, activity was then averaged across hemispheres to increase signal to noise ratio. We chose the 60–240 Hz range as this encompasses the full range of frequencies that have variously been referred to as high gamma in the literature while avoiding 50 Hz line noise and its harmonic at the Nyquist frequency (250 Hz).³⁰ High gamma activity was analysed using non-parametric cluster-based permutation testing as it allows for the control of multiple comparisons and does not require that the data are normally distributed.³¹ The permutation test works by repeatedly permuting the mapping between condition labels and time-series, calculating two-sided paired samples t-statistics for each data-point, clustering data-points that exceed a P-value of 0.05, and extracting the sum of t-values that form the largest cluster to build a distribution that can be used to compare with the non-permuted clusters. If the non-permuted cluster statistic is larger than 95% of clusters obtained after permuting the data (i.e. P < 0.05), it is considered significant. We used all 256 possible permutations. The same results were obtained when the cluster forming threshold was obtained non-parametrically either from the permutations or using Wilcoxon’s signed ranks test. We limited our analysis to between 0 and 1 s after stimulus onset based a priori on the non-human primate literature, which has shown punishment-related increases and reward-related decreases in single neuron action potentials in the LHb peaks within this time window.^4–7 As high gamma activity is believed to reflect population-level spiking we expected it would show similar modulation to rewards and punishments as shown in these non-human primate studies.^24,25 Our hypothesis was also informed by our previous demonstration of high gamma LFP activity coding reward and punishment in the amygdala and orbitofrontal cortex within this time window,³⁰ which are regions believed to provide input to the habenula.³² However, all effects were also significant when full epoch durations were analysed. We only report clusters as significant if the cluster-level significance exceeded a Bonferroni–Holm correction for the number of contrasts performed. All electrodes selected for analysis were either L0–L1, L1–L2, R0–R1 or R1–R2. L3 and R3 were always outside the habenula or were the least probable to be in the habenula in all patients. In order to demonstrate whether effects were specific to the habenula, we repeated all analyses using electrodes L2–L3 and R2–R3.

Data availability

The data supporting the results of this study are available upon request from the corresponding authors.

Results

Human habenula high gamma encodes reward and loss outcomes

Nine patients completed a MID task in which distinctive cues signalled whether a reward or loss could be received depending on the speed and correctness of a simple visual discrimination response (Fig. 1). The MID task is well validated and allows us to separate the period between the cue and response, when the subject anticipates the outcome, from when the outcome is received.³³ The reaction times of correct trials were analysed using an LME model with the fixed effects factors of condition (reward, loss, neutral) and arrow direction and the random effects factor of subject. The P-value threshold of 0.0167 was derived by Bonferroni correcting for the three tests performed. One-tailed P-values were used based a priori on previous studies showing faster reaction times on reward and loss trials relative to neutral using the same task.³⁰ On average there were 18.2 (SD = 5.2) correct reward trials, 15.4 (SD = 3.1) correct neutral trials and 19.6 (SD = 4) correct loss trials. Reaction times were significantly faster on reward trials [t(296) = 2.24, P = 0.0129, one-tailed] and loss trials [t(307) = 2.3, P = 0.011, one-tailed] relative to neutral (Fig. 4A and Table 2). There was no significant difference in reaction times between reward and loss trials ([t(335)=−0.05, P = 0.96, two-tailed]. A similar pattern was seen in the rates of correct relative to incorrect/missed responses. Differences in the frequency of correct relative to incorrect trials were analysed using a GLME model with the same factors and P-value thresholds as for the reaction time analysis. There was increased accuracy on loss trials relative to neutral trials [t(617) = 2.97, P = 0.003, one-tailed] and a trend for increased accuracy on reward trials relative to neutral trials [t(617) = 2, P = 0.023, one-tailed; Table 2]. There was no significant difference in accuracy between reward and loss trials [t(620) = 0.97, P = 0.33, two-tailed]. The patterns of reaction times are highly consistent with previous work³⁰ and show that the rewards and losses had a strong influence on motivational responding and therefore we should expect to see effects on LFPs if they are encoded by the human habenula.

Figure 4.

Human habenula high gamma encodes reward and loss outcomes. (A) Box-and-whisker plots showing mean (grey horizontal line) and median (black horizontal line) reaction times, range and interquartile range across conditions. Different coloured dots represent the mean of individual patients. (B) Time series of habenula high gamma in response to reward and loss in the outcome phase of the MID task. Shaded regions represent standard error and condition membership (see legend). Vertical dashed black line at t = 0 represents outcome stimulus onset time. The horizontal black line at the top of the plot represents the time intervals of significant clusters (P < 0.05 FWEC). (C) Mean activity within the significant time windows shown in B across patients. Different lines represent different patients. FWEC = family-wise error correction.

Table 2.

MID reaction times and accuracy rates

	Reward	Neutral	Loss
RT (Z) (SEM)	−0.14 (0.08)	0.08 (0.08)	−0.17 (0.06)
% Correct (SD)	52.2 (8.4)	45.3 (8.3)	56.6 (5.5)
% Incorrect (SD)	1.5 (2.6)	1.9 (2)	1.2 (2.6)
% No response (SD)	46.3 (10.2)	52.8 (8.5)	42.2 (5.3)

One patient was excluded from all LFP analyses due to inaccurate placing of electrodes. We first analysed high gamma activity in the outcome phase of the task seeking to test the hypothesis that the human habenula shows similar coding of rewards and punishments to primates. In primates, single neurons in the LHb increase in firing to punishments and decrease in firing to rewards and high gamma is believed to similarly reflect firing from large populations of neurons.^4–7,24,25 Therefore, contrasting loss and reward directly should be the most sensitive contrast. On average, there were 14.25 trials (SD = 2.7) in each condition, the same as previous LFP studies using the same task.³⁰ Consistent with our hypothesis, there was a significant increase in high gamma activity in response to loss when compared with reward, which showed a decrease (P = 0.004 cluster-based permutation test FWEC) peaking between 406 and 710 ms after stimulus onset (Fig. 4B). Within this time window, seven of eight patients showed increased high gamma activity in response to loss relative to reward (Fig. 4C). This effect was specific to the habenula. There were no significant differences in high gamma activity between reward and loss or any other contrast using contacts L2–L3 and R2–R3, which were the least likely to be in the habenula (Supplementary Fig. 3A).

The difference in activity between conditions was not driven by the correctness of responses as a control analysis showed no significant difference between correct and incorrect neutral trials. There were no significant differences between reward and neutral or loss and neutral. Previously, it was shown that the LHb response to neutral stimuli depends on the type of reinforcers that are used in the same block.⁶ In blocks where only an aversive reinforcer was used, there was a decrease in firing to neutral stimuli. In contrast, in blocks where only a rewarding reinforcer was used, there was an increase in firing to neutral. In our task, we presented rewards, losses and neutral all in one block. The adaptation of the neural response to the context in which the stimuli are presented may explain the selective response for the loss compared to reward contrast as the response to neutral may be more variable. This suggests that the habenula performs reference-dependent or relative coding of rewards and losses. This means that the coding of value adapts to the range of rewards and punishments that can be received in the task. Indeed, habenula neurons even show increased firing to a small reward in the context of a task where there is a possibility to receive a larger reward and the goal is to maximize reward.⁷ For this reason, we believe that the reward and loss contrast provides the most sensitive test of habenula function as these stimuli are the most potent and opposing stimuli presented. However, using non-parametric signed ranks tests on each condition separately, there was also significantly increased high gamma activity relative to baseline within the significant time window on loss trials (P = 0.012, one-tailed) and a decrease in high gamma on reward trials (P = 0.039, one-tailed). Seven of 8 patients showed increased high gamma activity in response to loss and 6 of 8 patients showed a decrease in response to reward. This provides support for the utility of high gamma as a biomarker for DBS treatment. The cue/anticipation phase showed the same pattern as in the outcome phase with greater high gamma for loss compared to reward but only trended in significance. Two clusters with P-values of 0.0508 and 0.063 did not reach significance (Supplementary Fig. 2).

Human habenula high gamma encodes risk during decision-making

Having established that human habenula high gamma encodes rewards and punishments in a similar manner to non-human primates, we next sought to examine whether such representations could be used to guide more complex, explicit, choice decision-making based on the probability of prospectively receiving rewards or losses. Eight patients participated in a card-based decision-making task that allowed us to dissociate risk and uncertainty (Fig. 2) and is well validated by similar prior studies.^34–36 Here, risk is defined as the probability of receiving a reward or loss and increases linearly from cards 1 to 10. In contrast, uncertainty is the sureness the patient will receive a particular outcome regardless of its value and shows an inverted U-shape from cards 1 to 10. For the analysis of behaviour and LFPs we divided trials into high and low risk, high and low uncertainty and bet and no bet and compared the two conditions. High-risk trials were those with a card value of 7–10 whereas low-risk trials were those with a card value of 1–4. We excluded the highest uncertainty cards (5 and 6) to maximize the difference in risk between conditions and sensitivity to differences in habenula activity while maintaining as many trials as possible. High-uncertainty trials were those with a value of 5, 6, 7 or 8 and low-uncertainty trials were those with a value of 1, 2, 3, 8, 9, or 10.

Based on prior research, we expected that patients would bet less as risk increased and respond slower as uncertainty increased.^34–36 Therefore, for these comparisons we used one-tailed tests. Differences in the frequency of binary decisions to bet or not bet were analysed using a GLME model with fixed-effects factors of risk and uncertainty and the random-effects factor of subject. A P-value threshold of 0.025 was derived by Bonferroni correcting for the two tests performed. For each patient, there were 36 high- and 42 low-uncertainty trials and 30 trials in the high- and low-risk conditions. There was a significant effect of risk [t(550)=−13.6, P < 0.0001, one-tailed] but not uncertainty [t(622) = 1.5, P = 0.14, two-tailed; Fig. 5A]. Reaction times were analysed using an LME model with the same factors and P-value thresholds as for betting choices. On average, in this analysis there were 29.25 (SD = 0.7) low- and high-risk trials, 41 (SD = 0.8) high- and 35 (SD = 0.9) low-uncertainty trials. There was a significant effect of uncertainty [t(607), 3.3, P = 0.006, one-tailed] but no effect of risk [t(537) = 0.9, P = 0.38, two-tailed; Fig. 5B]. These findings demonstrate a strong influence of risk on decision-making behaviour.

Figure 5.

Human habenula high gamma encodes risk during decision-making. (A) Bar graph showing proportion of gambles taken for each card number. Error bars represent ±SEM. (B) Reaction times as a function of card number. Error bars represent ±SEM. (C) Time series of habenula high gamma activity in the decision phase on high-risk compared to low-risk trials. Shaded regions represent standard error and condition membership (see legend). Vertical dashed line at t = 0 represents card stimulus onset time. The horizontal black line at the top of the plot represents the time intervals of significant clusters (P < 0.05 FWEC). (D) Mean activity within significant time points shown in C across patients. Each line is a different patient. (E) High gamma activity in the decision phase for the uncertainty contrast. (F) High gamma activity in the decision phase for the bet contrast.

For the LFP analysis, there were on average 25 trials (SD = 4.6) in the low- and high-risk conditions, 31 trials (SD = 6.8) in the low- and high-uncertainty conditions and 29 trials (SD = 5.6) in the bet and no bet conditions. There was significantly increased high gamma activity on high-risk trials relative to low-risk trials (P = 0.008, cluster-based permutation test FWEC) between 468 and 730 ms after stimulus onset (Fig. 5C). Within this time window, seven of eight patients showed increased high gamma activity in the high-risk condition relative to the low-risk condition (Fig. 5D). There was no significant effect of uncertainty or betting on high gamma (Fig. 5E and F). Using non-parametric signed ranks tests, there was significantly increased high gamma activity relative to baseline within the significant time window for high risk (P = 0.008, two-tailed) but no difference for low risk (P = 0.2, two-tailed). All patients showed increased high gamma in response to high risk and 6 of 8 patients showed a decrease to low risk. This provides additional support for the utility of high gamma as a biomarker for DBS treatment. There were no significant differences in high gamma between high and low risk or any other contrasts using contacts L2–L3 and R2–R3, which were the least likely to be in the habenula (Supplementary Fig. 3B). Our findings provide a significant advance in understanding the habenulas role in encoding risk in human decision-making, which to date is limited as studies have been restricted to inactivation in rats, in which probabilities must be learned via trial and error before decision-making,³⁷ or in humans with fMRI, which as mentioned previously has many limitations including mixing of activity during risk assessment with subsequent choice behaviour¹² which LFPs have the requisite temporal resolution to dissociate.

We tested hypotheses regarding outcome coding using the MID task as it has greater power for this analysis. This is because in the risk task it is easier to avoid loss outcomes and the outcome card value was chosen at random. This meant the number of loss outcomes was too small to compute a reliable average for four patients who all had only eight or fewer trials. In contrast, losses cannot be avoided in the MID task because response latency limits will increase until errors are made.

Discussion

The habenula is known as the anti-reward centre of the brain. An understanding of its role in reward and punishment is essential to understanding psychiatric disorders. However, this endeavour has largely been restricted to non-human primates and rodents or fMRI methods, which cannot inform us of the neurophysiological dynamics. In this study, we overcome these limitations to bridge the gap between human and non-human primate work. We used the rare opportunity to record LFPs with exceptional spatial and temporal resolution directly from the human habenula in patients receiving DBS treatment. Patients showed strong changes in motivational and decision-making behaviour with faster responses on reward and loss cue trials in the MID task and decreased betting for risky gambles in the risk task. LFPs showed a strong correspondence to behaviour and previous non-human primate work. High gamma activity, which is believed to reflect population-level spiking,^24,25 showed the same patterns of responses as single neurons recorded from the primate LHb, increasing during receipt of punishment and decreasing during receipt of reward.^4–7 A similar high gamma increase to high risk suggests that the habenula’s representation of punishment can be used for more complex, explicit, choice decision-making based on the probability of prospective rewards and losses.

To our knowledge, we are the first to report habenula neurophysiological responses to reward and punishment at the scale of the population. Whereas primate studies of the habenula focused on individual firing of between 42 and 76 cells,^4–7 high gamma has been estimated to reflect the pooled activity of around 500 000 cells which contact the DBS electrodes.²⁰ This suggests that reward/punishment coding of habenula cells is not due to selection bias and reflects part of a much larger coding ensemble. In the anticipation phase of the MID task, we did not find significant effects, although we observed the same trend of increased responses to loss and decreased responses to reward and neutral. The MID task is non-deterministic due to the necessity of generating incorrect trials for loss outcomes by adaptively adjusting response latency limits. Therefore, cues are only associated with reward and loss on approximately half of trials. Conditioned stimuli that predict the receipt of rewards and punishments with greater probability decrease and increase the firing of LHb neurons, respectively, to a greater extent than less-predictive stimuli. In contrast, responses to the outcomes show the opposite pattern. Rewards and punishments that are more unexpected decrease and increase LHb neuron firing more so than expected outcomes.^6,7 This could account for the differences in responses in the anticipation and outcome phase. In addition, LHb neurons can respond differently to cue and outcome⁶ and differential responses to reward and non-reward outcomes tend to remain after learning, suggesting that it is not just signalling negative prediction errors.⁵

We also demonstrate for the first time in humans habenula high gamma coding of risk during decision-making. This finding demonstrates that habenula high gamma activity is not simply involved in passive conditioning and receipt of outcomes. While a previous inactivation study in rats demonstrated a role for LHb in risk aversion, the probabilities of receiving rewards given different choices were learned through experience.⁴³ In contrast, here the stimuli presented during the decision phase had no direct association with reward or loss and therefore there is no learning component. This suggests that the coding of loss and reward omission in habenula high gamma can also be recruited by cognitive processes to guide future betting choices. This contrasts with the type of decision-making involved in the anticipation phase of the MID task, which influences the speed of an implicit response. The recruitment of habenula high gamma during risky decision-making may reflect propositional logic systems interacting with valuation systems. There was no difference in high gamma between bet and no bet trials, suggesting that it does not code the binary decision of the patient. This is consistent with the finding that LHb neurons more accurately tracked the value of conditioned stimuli during reversal learning than the actual behaviour of the animal.⁵ The increased high gamma response to high risk is consistent with studies showing decreased and increased firing of LHb neurons in response to conditioned stimuli that more accurately predicted reward and punishment outcome, respectively. The effect appeared to be driven more by an increase to high risk than a decrease to low risk, which is consistent with the finding that the LHb more precisely represents aversive stimuli.⁶

Although our discrete habenula recordings cannot delineate network interactions, they do inform us of the temporal dynamics which are absent from all fMRI studies. In both tasks, the high gamma responses to reward/loss and risk occurred approximately between 400 and 750 ms. The timing of high gamma in response to loss and reward in the MID task is similar to that observed in the amygdala and orbitofrontal cortex,³⁰ suggesting a potential relationship. Perhaps owing to the limitations and difficulty of studying the habenula with fMRI, requiring high-resolution imaging, custom acquisition parameters and preprocessing, the habenula has not yet been recognized as part of the human risk-taking network.^16,38–41 Although our findings require further corroboration, we believe that models of the neural basis of risk taking in humans should be expanded to take the habenula into account. Indeed, the nucleus accumbens and anterior cingulate cortex are recognized as part of the risk-taking network and are downstream targets of LHb.^{4,16,26,38–41}

The findings may have implications for the use of habenula DBS for the treatment of psychiatric disorders. To date, studies have used chronic stimulation using canonical frequencies, high-frequency stimulation and fine tuning.¹⁸ Our findings suggest that the habenula high gamma envelope may be a useful neurofeedback biomarker to trigger stimulation. This might help regulate habenula responses to negative thoughts and situations while at the same time retaining battery life and minimizing accumulation of damage to cells. Furthermore, this may need to be further tailored to the specific disorder. For example, in depression, DBS may be needed to suppress activity during a negative state, whereas in bipolar disorder it may also be needed to disrupt suppressions of high gamma that might accompany a manic state. Such closed-loop neuromodulation using amygdala high gamma to trigger ventral capsule/striatum stimulation has been used to successfully treat depression.^42,43 However, it has previously been shown that amygdala high gamma increases to both reward and loss.³⁰ Therefore, using habenula high gamma increases to trigger stimulation may confer further benefits to depression due to its more specific role in aversion. In addition to a disruptive effect, DBS may also have facilitatory effects when using certain stimulation settings.⁴⁴ This could be harnessed to use habenula DBS to increase risk assessment and reduce impulsivity, suggesting potential application to addiction. For example, a closed-loop approach could be used to determine the stimulation parameters that increase high gamma. These settings could then be activated by the patient when they feel cravings. In contrast, a disruptive effect of DBS may have the opposite effect. It will be necessary to investigate the short- and long-term effects of habenula DBS on risk-taking and impulsivity.

It is necessary to highlight limitations that are common to intracranial EEG studies including ours. Our habenula recordings are limited to neuropsychiatric patients whose neural systems underlying reward, aversion and risk processing could differ to those of healthy controls due to the disorder, its treatment with medication or the surgery. However, patients behaved in the same manner as in studies of the normal population, demonstrating that they were processing the tasks normally.^34–36,45 The patterns of LFP responses also conformed to what we expected would be a normal pattern based on primate work. The effects were also seen across almost all patients despite heterogeneity in disorders and medications used. It is difficult to make comparisons with fMRI studies as there may be confounding of differences in disease states with modality. We were only able to record from the habenula and therefore we do not know to what extent its functions are interdependent with other regions. Due to working with rare neurosurgical patients, the number of subjects was limited and we were also time-limited, meaning we could only collect a certain number of trials, although this is similar to that used previously.³⁰ However, this is mitigated by the excellent signal to noise ratio of intracranial LFP recordings²⁰ and pooling data across hemispheres.

In conclusion, we used the rare opportunity to record LFPs directly from the human habenula to study its role in reward processing and decision-making at a level of spatial and temporal resolution that hitherto was only possible in primates and rodents due to limitations with non-invasive methods. Our findings not only demonstrate homology with the primate habenula in terms of polarization of reward and aversion coding in high gamma activity but also extend its role to risk coding during decision-making. These signals could be used as potential neurophysiological biomarkers that could be targeted to optimize DBS treatment of psychiatric disorders.

Funding

Natural Science Foundation of China grant (81771482) to B.M.S.; SJTU Trans-med Awards Research (2019015) to B.M.S.; Shanghai Clinical Research Centre for Mental Health (19MC191100) to B.M.S.; Medical Research Council Senior Clinical Fellowship (MR/P008747/1) to V.V. All research at the Department of Psychiatry in the University of Cambridge is supported by the NIHR Cambridge Biomedical Research Centre (BRC-1215-20014) and NIHR Applied Research Centre. The views expressed are those of the author(s) and not necessarily those of the NIHR or the Department of Health and Social Care.

Competing interests

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as potential conflict of interest and report no conflicts of interest. The authors report no competing interests.

Supplementary material

Supplementary material is available at Brain online.