Dynamic prefrontal coupling coordinates adaptive decision-making

Abstract

Adaptive decision-making requires flexibly maintaining or changing behavior in response to uncertainty. While the dorsomedial (dmPFC) and dorsolateral (dIPFC) prefrontal cortex are each essential for this ability, how they coordinate to drive adaptation remains unknown. Using intracranial EEG recordings from human participants performing a dynamic reward task, we identified distinct, frequency-specific computations: dmPFC high-gamma activity encoded uncertainty before stay decisions but transitioned to prediction error before switches, while theta activity shifted from uncertainty to value representation. In contrast, dIPFC theta activity signaled both value and uncertainty before stays, but predominantly value before switches. Crucially, these regions coordinated through two temporally specific coupling mechanisms that predicted behavioral changes: theta-theta amplitude coupling during feedback processing and theta-gamma phase coupling before decisions. Both coupling mechanisms strengthened before switches, suggesting that changing behavior requires greater dmPFC-dIPFC integration than maintaining. These findings reveal how the dorsal prefrontal cortex employs frequency-specific computations and precise temporal coordination to guide adaptive behavior.

Keywords: adaptation; dorsolateral prefrontal cortex; dorsomedial prefrontal cortex; inter-regional coupling; stay; switch; uncertainty; value.

Fig. 1.. Experimental setup, task design, computational modeling, and behavioral results.

(a) Intracranial EEG recordings from 14 epilepsy patients, with 190 channels in the dorsolateral prefrontal cortex (dIPFC) and 83 channels in the dorsomedial prefrontal cortex (dmPFC). (b) Three-armed restless bandit task. Patients chose one of three options (face, car, house), followed by reward or no-reward feedback. Upper right panel: example trials showing how reward probability for each bandit varied in a random walk manner (on each correct trial, there was a 10% chance that the reward probability for each target would either increase or decrease by 0.2, with these probabilities bounded between 0.1 and 0.9, see Methods). Lower right panel: Response time (RT) distribution across all patients (model-free results and more time information see TableS2 & Table S3). (c) Volatile Kalman filter model implementation and model fits. The model computes trial-by-trial value and uncertainty estimates for each arm. Relative value (RV) and relative uncertainty (RU) were calculated as the ratio of the chosen option’s value/uncertainty to the sum across all options. Right panels show model fits from three representative patients: colored dots indicate actual choices for specific options (e.g., car), black lines show the model-predicted choice probabilities. Bayesian model comparison results revealed that the volatile Kalman filter with relative value and relative uncertainty incorporated into the softmax function outperformed other models (model validation analyses see Supplementary Notes1, Fig. S1, model comparison results see Table S4) (d) Choice behavior of the next trial as a function of the current chosen option’s relative value and uncertainty. Left box plots showing the distributions of relative value and right box showing uncertainty for stay versus switch trials. Higher relative value and uncertainty promote stay decisions, while lower relative value and uncertainty lead to switches (n=10 points/condition shown using stratified sampling for visualization purposes, statistical significance assessed using linear mixed-effects models across all trials, see Methods).

Fig. 2.. Differential roles of dIPFC and dmPFC in feedback processing and choice selection.

(a, c) Time-frequency analyses (t-value maps) of local field potentials in the dIPFC and dmPFC during post-feedback processing (2000ms after outcome onset). Both regions show significant differences in neural activity between reward and non-reward feedback conditions. (b, d) The dIPFC exhibits distinct activity patterns associated with stay and switch decisions before selection, while the dmPFC shows no significant differences between these decision types.

Fig. 3. Neural representations of prediction error (PE) in dIPFC and dmPFC preceding stay/switch decisions

(a-c) PE representation preceding stay decisions: (a) dmPFC high-gamma band (70–150 Hz), (b) dmPFC theta band (4–9 Hz), and (c) dIPFC theta band showing minimal PE representation, with only a brief significant window in dIPFC. (d-f) PE representation preceding switch decisions: (d) dmPFC high-gamma band, (e) dmPFC theta band, and (f) dIPFC theta band, all showing robust PE representation.

Fig. 4. Choice-dependent neural representation of value and uncertainty in prefrontal regions

Neural representations of relative value (RV) and relative uncertainty (RU) during stay and switch decisions. (a-c) Neural activity preceding stay decisions: (a) dmPFC high-gamma band (70–150 Hz) and (b) dmPFC theta band (4–9 Hz) exclusively represent relative uncertainty, while (c) dIPFC theta band represents both variables with predominant uncertainty representation. (d-f) Neural activity preceding switch decisions: (d) dmPFC high-gamma shows no significant representation, (e) dmPFC theta band selectively represents relative value, and (f) dIPFC theta band represents also selectively represents value.

Fig.5.. Dynamic inter-regional coupling between dmPFC and dIPFC predicts decision changes

Left panel: theta-theta amplitude-amplitude coupling (AAC) (a) Comparison of AAC strength between post-feedback and pre-selection stages, showing stronger coupling during post-feedback period (b) Post-feedback stage AAC significantly predicts subsequent switch decisions (higher AAC associated with increased switch probability) (c) Pre-selection stage AAC shows no significant predictive relationship with subsequent decisions. Right panel: phase-amplitude coupling (PAC) between dIPFC theta phase and dmPFC high-gamma amplitude (d) Comparison of PAC strength between stages, showing stronger coupling during pre-selection stage (e) post-feedback stage PAC shows no significant predictive relationship with subsequent decisions (f) Pre-selection stage PAC significantly predicts subsequent switch decisions (higher PAC associated with increased switch probability).

Figure 6.. Decision-specific patterns of inter-regional coupling and their relationships with computational variables

Analysis of coupling-variable relationships during post-feedback and pre-selection periods across different decision contexts. Left panels, post-feedback stage amplitu-deamplitude coupling (AAC) (a) Combined trials: AAC exhibits significant negative correlations with both prediction error (PE) and relative uncertainty (RU) (b) Stay-decision trials: AAC shows selective negative correlation with RU (c) Switch-decision trials: AAC shows selective negative correlation with PE. Right panels: the pre-selection stage phase-amplitude coupling (PAC) (d) Combined trials: PAC shows significant negative correlations with both PE and relative value (RV) (e) Stay-decision trials: No significant correlations observed (f) Switch-decision trials: PAC exhibits selective negative correlation with PE.