Graded decisions in the human brain

Decision-making, a fundamental cognitive process, involves deliberation over time. While objectively a definitive choice is made, subjectively, humans experience uncertainty. The neural implementation of this deliberation has been a topic of debate, with models ranging from the traditional "all-or-none" bound-based drift-diffusion model to more recent graded models. The all-or-none model posits that a decision is made when neural activity representing the decision reaches a fixed threshold or bound. However, this model has faced criticism, primarily due to its inability to fully account for behavioral and neural data observed in situations with time constraints. Alternative models, such as those incorporating collapsing bounds or multidimensional attractor networks, propose that decision-related activity is graded by accumulated evidence, reaching a stable state influenced by various factors like time constraints and accuracy requirements. These alternative models are difficult to distinguish experimentally because computational models often predict similar choice behavior despite invoking different neural mechanisms. Previous studies using non-invasive methods or recordings from specific brain regions in non-human primates have provided limited insight. This research aimed to directly address the nature of the decision process in humans by recording high-fidelity neural activity from multiple brain regions during perceptual decision-making.

The dominant model in decision-making research is the drift-diffusion model, which suggests a process of evidence accumulation towards a decision boundary. This model has been successful in explaining various behavioral phenomena. However, recent critiques highlight that the fixed boundary assumption doesn't fully account for decision-making under time pressure or dynamic conditions. Alternative models like those with collapsing boundaries and multidimensional attractor networks have emerged to address these limitations. These models incorporate the influence of factors like time constraints, urgency, and reward expectations on decision processes. However, distinguishing between these models experimentally has been challenging due to the overlapping behavioral predictions of different computational models. Most prior research was conducted with non-invasive techniques like EEG or single-unit recordings in specific brain regions of non-human primates. This study aims to address these limitations by employing high-fidelity intracranial recordings from multiple brain areas in humans.

Intracranial local field potentials (LFPs) were recorded from eight human subjects (five male, three female, ages 15–57) during a perceptual decision task. Subjects had electrodes implanted for clinical reasons (epilepsy localization). The task involved listening to binaurally presented click sounds (Poisson-distributed) and deciding whether more clicks were heard in the left or right ear. Two response types were used: saccades (eye movements) and button presses. Two task contexts were employed: a congruent context (left clicks -> left saccade/right button press) and a reversed context (left clicks -> right button press/left saccade), allowing for a manipulation of the mapping between stimulus and action. A decision variable (DV) was calculated based on the cumulative evidence (number of clicks in each ear), reflecting the subjects' decisions. Eye gaze and electromyographic (EMG) activity were also recorded to monitor the responses. Data were preprocessed to remove artifacts and noise. Broadband gamma (γ) activity (70–170 Hz), a surrogate of multi-unit spiking activity, was analyzed. Effector-modulated regions (showing γ activity modulation around choice time) were identified. The study employed a model-free analysis (correlating γ activity with the DV at choice time) and a model-based analysis (regressing γ activity on a modeled DV with a fixed bound to test the "all-or-none" hypothesis). The spatial distribution of graded effects was mapped to Brodmann areas to identify regions contributing to the graded responses. Statistical analyses included Spearman's correlation, t-tests, and ANOVAs.

The study's key finding is that broadband gamma (γ) activity in multiple brain regions, prominently the parietal cortex (BA40), remained graded by the decision variable (DV) at the time of choice. This graded effect was observed across response types (saccades and button presses), task contexts (congruent and reversed), and analytical approaches (model-free and model-based). The γ activity ramped up gradually during the decision process, consistent with evidence accumulation models. Importantly, the γ activity did not abruptly reach a bound at the moment of choice but rather maintained a graded relationship with the DV, indicating the analog nature of the decision process. The strength of the correlation between γ activity and DV was statistically significant (p < 0.01) in the effector-modulated regions. The parietal cortex (BA40), along with frontal eye fields (BA8) and premotor/supplementary motor areas (BA6), showed the most prominent graded effect. Furthermore, the probability of subjects making a choice was significantly modulated by the level of the DV (F(2,36) > 27.9, p < 4.8 × 10⁻⁷ for both response types). The regression analysis explicitly rejecting the bound hypothesis showed a similar effect, further confirming the graded nature of the decision process. Modest but significant grading was also observed in efferent motor activity (EMG and saccade amplitude).

The findings provide strong neural evidence for a graded decision process in humans. This contradicts the traditional view of decision-making as an "all-or-none" process reaching a fixed bound. The graded nature of neural activity at choice time suggests that decisions are made probabilistically, with the strength of the neural signal reflecting the certainty or level of supporting evidence. This aligns with embodied cognition theories, suggesting that decision formation and action planning are intertwined within the same neural circuits. The multi-regional representation of evidence accumulation, particularly in parietal, frontal, and premotor areas, highlights the coordinated function of these regions in decision-making. The robust graded effect observed even with varying stimulus-response mappings underscores the flexibility and adaptability of human decision processes. The graded neural activity also reflects the probabilistic nature of decision-making, indicating a direct link between neural representation and behavioral outcomes.

This study offers compelling evidence for a graded decision process in the human brain, challenging the traditional "all-or-none" bound-based models. The observed graded relationship between broadband gamma activity and the decision variable, especially prominent in the parietal cortex, supports a more flexible, analog representation of decision evidence. Future research could explore the specific factors influencing the graded representation (e.g., urgency, reward, confidence) and examine these processes in more naturalistic decision-making settings. Investigating deeper brain structures and expanding the scope of recorded brain regions will provide more comprehensive insights into decision-making.

The study has limitations inherent to laboratory-based decision tasks. The exact factor represented by the graded gamma activity (DV, confidence, etc.) cannot be definitively determined. The generalizability to real-world scenarios remains to be confirmed. Only a subset of the recorded brain regions showed DV modulation at choice time, suggesting that other regions may also be involved. Further investigation with more extensive neural recordings, including deep brain structures and naturalistic settings, is necessary to improve our understanding of the decision-making process.