Source-linked AI summary
Defecting or not defecting: how to "read" human behavior during cooperative games by EEG measurements
F. De Vico Fallani, V. Nicosia, R. Sinatra, L. Astolfi, F. Cincotti, D. Mattia, C. Wilke, A. Doud, V. Latora, B. He, F. Babiloni
TL;DR
The paper asks whether neural signatures of social interaction can be measured directly across two interacting brains. It constructs hyper-brain networks from simultaneous EEG recordings during the Iterated Prisoner’s Dilemma and finds that defection produces distinctive, more separated connectivity patterns that can be classified before the players act.
Problem
Social-interaction studies rarely measure signals from both players simultaneously, making direct assessment of inter-brain connectivity difficult.
Method
The study combines EEG hyperscanning, cortical localization, Partial Directed Coherence, and graph analysis to construct hyper-brain networks from 26 couples playing the Iterated Prisoner’s Dilemma.
Results
Up to 90% accuracy was achieved in predicting the DD strategy before button presses, while DD networks showed reduced inter-brain connectivity and greater separation than CC and TT networks.
Takeaways & Limitations
Connectivity changes in hyper-brain networks can provide an advance neural signature of non-cooperative behavior during social decision-making.
Takeaways & Limitations
The networks used only 12 predefined cortical ROIs, limiting graph-analysis power and restricting coverage to those regions.
Abstract
from arXiv · showhide
Understanding the neural mechanisms responsible for human social interactions is difficult, since the brain activities of two or more individuals have to be examined simultaneously and correlated with the observed social patterns. We introduce the concept of hyper-brain network, a connectivity pattern representing at once the information flow among the cortical regions of a single brain as well as the relations among the areas of two distinct brains. Graph analysis of hyper-brain networks constructed from the EEG scanning of 26 couples of individuals playing the Iterated Prisoner's Dilemma reveals the possibility to predict non-cooperative interactions during the decision-making phase. The hyper-brain networks of two-defector couples have significantly less inter-brain links and overall higher modularity - i.e. the tendency to form two separate subgraphs - than couples playing cooperative or tit-for-tat strategies. The decision to defect can be "read" in advance by evaluating the changes of connectivity pattern in the hyper-brain network.
Introduction
The study addresses the difficulty of directly measuring interacting brains during social decision-making. It combines simultaneous EEG recordings and hyper-brain network analysis to examine cooperation and defection in the Iterated Prisoner’s Dilemma.
- Single-player measurements infer inter-brain connectivity indirectly rather than measuring interactions between players directly.
- The study combines simultaneous EEG hyperscanning, cortical localization, and Partial Directed Coherence to estimate information propagation within and between brains.
- The Iterated Prisoner’s Dilemma allows each player to choose cooperation or defection and potentially respond to previous non-cooperative behavior.
- Analysis of 26 couples found different hyper-brain network structures for cooperative and selfish interactions.
- Two-defector couples showed stronger separation between the brains, whereas cooperative and tit-for-tat couples had more intertwined regions.
Results
The results compare hyper-brain network topology across Prisoner’s Dilemma strategies and identify connectivity patterns associated with defection. Lower efficiency, higher divisibility and modularity, and altered frontal connectivity characterize DD networks, enabling advance classification of defection.
- Standard graph averages across 26 couples could not discriminate the six strategies because between-couple variability produced large standard deviations.
- Each couple’s six strategy-specific hyper-brain networks were compared using efficiency, divisibility, and modularity.
- 50% of couples had minimum efficiency in DD networks, compared with 11.6% for CC and 19.2% for TT in the Theta band.
- DD networks had longer paths and fewer inter-brain links, while CC and TT networks were more interconnected.
- Task significantly affected total node strength across all frequency bands, with ROI effects significant except in Alpha and task-by-ROI interactions significant in Beta and Gamma.
- In Beta, Brodmann area 10_L had lower total strength for DD than CC and TT, with post-hoc p-values of 0.000153 and 0.000002.
- DD networks showed lower efficiency, higher divisibility, and higher modularity than the other strategies.
- A nonlinear classifier predicted DD strategy before button presses with up to 90% accuracy.
Discussion
The study uses simultaneous EEG hyper-scanning and hyper-brain network analysis to link inter-brain connectivity with cooperative and non-cooperative decisions. Defection is associated with reduced cross-brain connectivity, especially in prefrontal regions, and can be predicted before players communicate their choices.
- Connectivity patterns: Across all analyzed frequency bands, DD strategies show significantly lower connectivity between the two brains than TT and CC trials.Cooperative and tit-for-tat networks are more tightly connected and intermingled, whereas DD networks show fewer inter-brain links.
- Connectivity patterns: Prefrontal cortex regions, including Brodmann area 10 and the anterior cingulate cortex, principally account for the observed decrease in inter-brain connectivity.The effect is particularly evident in Beta and Gamma bands, where these regions contain local representations of intended decisions before they are communicated.
- Method: Simultaneous EEG hyper-scanning records the brain activity of both players during an Iterated Prisoner’s Dilemma game.EEG offers high temporal resolution compared with hemodynamic measurements, supporting analysis during decision-making.
- Prediction: A non-linear classifier discriminates DD strategies with up to 90% accuracy before players press the keyboard buttons.The authors suggest that predicting CC and TT may require a larger dataset than the 26 couples studied.
- Method: The study applies complex network theory to functional brain connectivity and its correlation with observed social behavior.Hyper-brain networks represent information flow within each brain and relations between the two brains.
- Limitations: The main methodological limitation is that each hyper-brain network contains only 12 predefined ROIs, restricting the analyzed cortical networks.This restriction is linked to the data requirements of the Partial Directed Coherence approach and its multivariate autoregressive models.
Material and methods
The study records and classifies repeated Prisoner’s Dilemma behavior, then constructs directed hyper-brain networks from paired cortical signals to compare strategies using graph measures.
- Experimental design: Each couple repeatedly plays the Prisoner’s Dilemma, with outcomes and player behavior classified across cooperation, defection, and Tit-for-Tat strategies.The experiment also obtained informed consent and ethics approval, and recorded EEG from participants using separate 64-channel systems at 200 Hz.
- EEG and cortical activity: Cortical activity is estimated from scalp EEG using an average realistic head model and six regions of interest for each subject.The cortical estimates are obtained by solving an electromagnetic linear inverse problem.
- Hyper-brain network: A hyper-brain network merges the six cortical signals from each participant into 12 signals and estimates directed functional connectivity with Partial Directed Coherence.The spectral causality criterion excludes connections whose intensities do not exceed a predetermined significance threshold.
- Graph representation: The networks are represented as directed weighted graphs whose arc weights encode directed connections and whose node degrees and strengths characterize connectivity.The weighted adjacency matrix allows wij to differ from wji, reflecting directionality.
- Graph measures: Efficiency evaluates communication through shortest paths, while divisibility and modularity quantify separation into two node sets corresponding to the players’ brains.Higher modularity indicates a better partition into the two brain communities; disconnected node pairs do not contribute to efficiency.
- Normalization and comparison: For each couple, strategy-specific efficiency, divisibility, and modularity are standardized with Z-scores computed across strategies before averaging across couples.The within-couple averages and standard deviations are calculated over all strategies, and strategy-level Z-scores are then averaged over couples.
Figure legends
The figures define the experiment timeline and visualize hyper-brain connectivity and graph measures across strategies, frequency bands, and brain regions.
- Figure 1: Each trial separates strategy communication from the subsequent decision period, with the first 1 second of EEG recordings designated as the period of interest.Players choose cooperation or defection, receive a 4-second strategy-and-score report, and then make the next decision.
- Figure 2: Figure 2 displays inter-brain connectivity for CC, DD, and TT strategies in the Alpha band using six regions of interest per brain.Only links between the two brains are shown, with directed-connection size and color representing PDC values.
- Figure 3: Figure 3 compares the percentages of couples with minimal or maximal efficiency, divisibility, and modularity for pure strategies in the Theta band.CC, DD, and TT are color-coded, while mixed strategies are shown in white; analogous diagrams appear for other frequency bands.
- Figure 4: Figure 4 plots average Z-scores for efficiency, divisibility, and modularity across CC, DD, and TT strategies and frequency bands.The panels pair divisibility with efficiency or modularity for comparison across couples.
- Figure 5: Figure 5 shows Beta-band total ROI strength for 52 subjects across CC, DD, and TT tasks, with confidence intervals and significance markers.The x-axis lists ROIs, the y-axis gives total strength, and stars identify strategy differences at p<0.001.