Direct stimulation of anterior insula and ventromedial prefrontal cortex disrupts economic choices

Risky decisions depend on the desirability of potential outcomes, yet the causal neural mechanisms remain unclear. Prior correlational evidence from fMRI and intracranial electrophysiology implicates the vmPFC and anterior insula (aINS) in shaping risk preferences: pre-stimulus activity in vmPFC tends to promote risk-taking by overweighting gains, whereas activity in aINS tends to temper risk-taking by overweighting losses. Both regions have also been linked to decision confidence, and confidence can be biased by monetary prospects, suggesting potential interactions between value and confidence signals. Causal evidence in humans is limited due to challenges in noninvasively targeting deep structures (vmPFC, aINS), inconsistencies in lesion studies (heterogeneous lesion extents/locations), and a paucity of IES studies combining stimulation with tasks that dissociate decision components. Building on prior observations that IES of ventral aINS can elicit disgust or depressed affect and dorsal aINS can evoke ecstatic sensations or ingestive behaviors, the present study used the high spatial specificity of IES during an accept/reject economic choice task to test whether vmPFC and aINS subregions causally modulate risk-taking and confidence and to delineate the computational mechanisms (e.g., sensitivity to gains vs. losses).

Correlational neuroimaging and intracranial EEG studies associate vmPFC activity with reward valuation, risk-seeking, and confidence, and aINS activity with avoidance, loss anticipation, negative affect, and reduced confidence. Confidence judgments are biased by monetary prospects, indicating intertwined value and confidence codes. Noninvasive stimulation studies have struggled to engage these deep targets effectively; lesion studies report mixed effects on risk-taking due to variable lesion profiles. IES studies in humans rarely targeted vmPFC/aINS with cognitive paradigms, though non-human primate work shows orbitofrontal value codes are causally related to choice. Prior insula stimulation reports dissociable affective/somatic phenomena along dorsal–ventral axes (disgust/depressed affect vs. ecstatic/ingestive behaviors), hinting at functional subregions relevant for approach–avoidance and risk processing.

Participants: Fifteen adults with drug-resistant focal epilepsy (mean age 34.9 ± 2.7 years; 7 females; 5 left-handed) undergoing stereo-EEG (SEEG) implantation at Grenoble University Hospital participated; 19 healthy participants performed the same task for comparison. Inclusion required at least one electrode pair in aINS and/or vmPFC and ability to perform the task. Ethical approvals and informed consent obtained. Electrode implantation and localization: Robot-assisted SEEG (ROSA) with 15–18 semi-rigid depth electrodes per patient (0.8 mm diameter, 10–18 contacts; Dixi Medical). Electrode locations were verified by co-registering pre-op T1 MRI with post-op CT (IntrAnat Electrodes). aINS subregions defined anatomically via Destrieux atlas: ventral aINS (valns: short insular gyri and anterior circular sulcus; parcels 18, 47) and dorsal aINS (dalns: superior circular sulcus; parcel 49; sites ensured anterior to central sulcus). vmPFC defined per MarsAtlas and sulcal anatomy (superior rostral sulcus, ROS-S; suprarostral sulcus, SU-ROS) with manual identification. Intracranial electrical stimulation (IES): Bipolar stimulation between contiguous contacts (a stimulation site). Parameters: 50 Hz, 0.5 ms pulse width, 1–3 mA (highest safe intensity below clinical symptoms), delivered for 5 s. For some sites, intensity reduced (1–2 mA) if clinical effects occurred at higher intensities. Stimulation initiated ~1 s before trial onset; inter-stimulation interval ≥1 min; iEEG monitored for afterdischarges/seizures. Behavioral task and sessions: Each session targeted one site and intensity, comprising 28 trials alternating stimulation (n=14) and no-stimulation (n=14), with first trial randomized. Choice phase: participants accepted or rejected an offer specifying gain (1–5 €), loss (1–5 €), and challenge difficulty (target window set to ~50% success). Accepting played for the offered stakes; rejecting played for minimal stakes (±€0.10). Confidence rating: before the challenge, participants rated prospective confidence (“Do you think you will win?”) on a 0–100 VAS with randomized starting cursor. Challenge: timed response to a moving ball entering a target window; ball visible first 500 ms of 1 s trajectory; adaptive target window maintained success ~50% by adjusting tolerated error. Feedback provided each trial; total earnings displayed per session. Training and calibration: Three-step training familiarized participants with the challenge, integrated choice attributes, and confidence reporting. Individual performance variability estimated to set initial tolerated error for ~55% theoretical success on first trial of each session. Data acquisition: Across 15 participants, 54 stimulation sites were tested: aINS n=38 sites (13 participants), vmPFC n=16 sites (9 participants). Five sessions excluded due to invariant choices; 54 retained. Statistical analysis: Generalized linear mixed-effects models (GLME; Matlab fitglme) with full random-effects at participant and site levels assessed: (i) effects of gains, losses, difficulty, and confidence on choice in non-stim trials; (ii) spatial gradients relating MNI coordinates (x,y,z) of IES to acceptance changes; (iii) interactions between stimulation condition and clusters/subregions (k-means clusters; anatomical parcellations) on acceptance and confidence. Participant ablation analyses evaluated robustness. Multiple comparisons controlled via Benjamini–Hochberg FDR. Computational modeling: Expected utility model fit to additional no-stimulation data (128–192 trials) from 10/15 patients using VBA toolbox to estimate parameters: gain weight (kg), loss weight (kl), bias (k0), temporal drift (kt), and subjective precision (σ) mapping challenge difficulty to success probability. For main sessions, model re-fit on with- vs without-stimulation trials allowing kg and kl to vary, with other parameters fixed to prior posterior means. GLME tested stimulation × cluster/subregion effects on posterior kg and kl.

• Baseline choice behavior (no IES): Acceptance increased with gain and decreased with loss magnitude (βgain=0.11; βloss=-0.09; F7,72=4.26, p<0.001; logistic mixed-effects). Challenge difficulty did not significantly affect choices (βdiff=7.10, F7,72=0.33, p=0.741). Confidence positively related to acceptance probability (βconf=0.39, F7,58=2.23, p=0.026).
• Spatial dependence of IES effects: In aINS, acceptance changes varied along anteroposterior (y) axis (βy=0.04, t134=2.15, p=0.039) and trended along ventrodorsal (z) axis (βz=0.04, t134=1.86, p=0.072). In vmPFC, ventrodorsal (z) coordinate significantly predicted choice changes (βz=-0.10, t112=3.71, p=0.003). K-means revealed two clusters per region (aINS: posteroventral n=15, anterodorsal n=23; vmPFC: ventral n=9, dorsal n=7).
• Opponent effects on risk-taking by subregion (GLME interactions): • aINS: Significant stimulation × cluster interaction on acceptance (β=-0.041, F1,72=11.6, p=0.001). Post-hoc: ventral aINS IES decreased risk-taking (F1,72=6.7, p=0.012); dorsal aINS IES increased risk-taking (F1,72=5.4, p=0.023). • vmPFC: Significant stimulation × cluster interaction (β=-0.038, F1,28=10.1, p=0.004). Post-hoc: ventral vmPFC IES increased acceptance (F1,28=6.5, p=0.016); dorsal vmPFC tended to decrease risk-taking (F1,28=4.0, p=0.056).
• Computational mechanisms: Stimulation × cluster interactions on loss sensitivity kl observed in both regions: • aINS: β=0.536, F1,72=15.3, p<0.001. Ventral aINS IES increased kl (F1,72=7.6, p=0.008; greater sensitivity to losses); dorsal aINS IES decreased kl (F1,72=6.3, p=0.014). • vmPFC: β=0.541, F1,28=13.6, p<0.001. Ventral vmPFC IES decreased kl (F1,28=8.9, p=0.006); dorsal vmPFC IES increased kl (F1,28=5.2, p=0.030). No significant effects on gain sensitivity (kg; Supplementary Fig. 6).
• Anatomical parcellation replication: Using Destrieux atlas subregions in aINS, interaction on acceptance (β=0.04, F1,72=12.4, p<0.001) and on kl (β=-0.47, F1,72=11.9, p=0.001), with dorsal aINS increasing risk-taking and decreasing kl, ventral aINS decreasing risk-taking and increasing kl. Results replicated with Julich Brain Atlas (β=0.038, F1,64=4.6, p=0.036). For vmPFC, inadequate sampling in dorsal SU-ROS (n=3) precluded full test, but in ventral ROS-S (n=11 sites) IES increased risky choices (β=-0.039, F1,20=4.5, p=0.046) and decreased kl (β=0.59, F1,20=5.0, p=0.037).
• Confidence: IES over aINS significantly reduced prospective confidence (main effect: F1,70=10.3, p=0.002), with no dorsal–ventral dissociation (interaction ns). In vmPFC, no significant effects of IES on confidence (main or interaction). Effects on confidence were not explained by changes in task performance.

The study provides causal evidence that vmPFC and anterior insula implement opponent, spatially dissociable mechanisms in economic choice. Along a dorso–ventral axis, ventral aINS stimulation increased sensitivity to potential losses and reduced risk-taking, whereas dorsal aINS stimulation decreased sensitivity to losses and promoted risk-taking. In vmPFC, ventral stimulation reduced loss sensitivity and increased risk-taking; dorsal stimulation showed the opposite trend. These findings align with a broader opponent-view framework wherein vmPFC supports appetitive valuation and approach, and aINS supports aversive valuation and avoidance, but refine it by revealing subregional gradients with opposing effects within each structure. The computational analyses indicate that modulation of subjective loss weighting, rather than gain weighting, underlies stimulation-induced choice changes. Metacognitively, IES of aINS lowered prospective confidence irrespective of subregion, providing causal support for the insula’s role in metacognitive judgments and extending correlational findings linking aINS activity to lower confidence. In contrast, vmPFC stimulation did not alter confidence despite prior associations, suggesting that confidence computations may be implemented outside the stimulated vmPFC subfields, require different timing relative to stimulation, or are more variable/noisy, reducing detectability. The causal patterns bridge invasive electrophysiology and stimulation literatures and are consistent with animal studies demonstrating insular subcircuits for approach–avoidance. They suggest that local stimulation can bias decision policies by altering loss sensitivity and possibly by modulating connectivity within heterogeneous dorsal–ventral networks in aINS and vmPFC.

Intracranial electrical stimulation over vmPFC and anterior insula causally disrupts economic choices and confidence via subregion-specific mechanisms. Ventral aINS stimulation reduces risk-taking by increasing loss sensitivity, whereas dorsal aINS and ventral vmPFC stimulation increase risk-taking by reducing loss sensitivity. aINS stimulation also robustly lowers prospective confidence. These results unveil dorso–ventral functional parcellations within aINS and vmPFC that differentially bias decision-making and metacognition. The findings have translational relevance for neuropsychiatric disorders involving maladaptive risk processing and suggest targets and strategies for closed-loop brain stimulation interventions. Future work should increase within-patient sampling, trials, and electrophysiological recordings to parse local versus network mechanisms and to more fully characterize dorsal vmPFC contributions to confidence and loss/gain weighting.

• Patient population: Findings derive from individuals with drug-resistant epilepsy, which may limit generalizability despite behavior comparable to healthy controls and convergent imaging evidence.
• Sampling constraints: Limited and uneven spatial sampling, particularly in dorsal vmPFC (SU-ROS, n=3), reduced power to test full dorso–ventral dissociation in vmPFC.
• Trial numbers: Relatively few trials per stimulation condition constrain model parameter estimation and may contribute to discrepancies with prior electrophysiological findings.
• Temporal factors: Confidence ratings occurred later relative to stimulation onset, potentially reducing sensitivity to vmPFC effects on metacognition.
• Electrophysiological mechanisms: Stimulation artifacts and limited trials precluded analysis of local/network electrophysiology or functional connectivity changes mediating behavioral effects.