Integrating EEG-SVM Confidence and RT via cBCI Enhances Team Decisions in a VR Drone Task

Integrating EEG-SVM Confidence and RT via cBCI Enhances Team Decisions in a VR Drone Task | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Integrating EEG-SVM Confidence and RT via cBCI Enhances Team Decisions in a VR Drone Task Christopher Baker, Stephen Hinton, Akashdeep Nijjar, Riccardo Poli, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6985673/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Optimizing team decision-making by appropriately weighting individual contributions is a significant challenge. Collaborative Brain-Computer Interfaces (cBCIs) offer a novel approach by integrating neurophysiological and behavioral data. This study evaluated a cBCI system incorporating EEG-derived SVM decision confidence, response times (RT), and subjective confidence ratings to enhance team accuracy in a VR drone target-detection task, particularly under varying cognitive workload (Low vs. High). Seventeen participants performed the task; individual SVMs were trained, and team performance (N = 2–8 members) was simulated using diverse aggregation methods. Under High Workload, mixed cBCI methods (e.g., combining subjective and SVM confidence) significantly improved team accuracy, surpassing even the best individual's average performance (e.g., N = 8: 98.8% vs. 94.2%). This synergistic benefit was minimal under Low Workload due to ceiling effects in individual performance. These cBCI enhancements were evident for EEG data from both pre- and post-decision epochs. The findings demonstrate that cBCIs can markedly improve team decision-making in demanding contexts, facilitating a "superorganism" effect where team capabilities exceed those of the best individual. Physical sciences/Engineering Physical sciences/Mathematics and computing Biological sciences/Neuroscience Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Figure 13 Figure 14 Figure 15 Figure 16 Figure 17 Figure 18 Figure 19 Figure 20 Figure 21 Introduction Human beings have traditionally worked together in teams, harnessing their abilities as a collective to improve performance beyond the limits of the individual 1 . When people work in decision-making teams, inputs from individuals can be expressed, accumulated and condensed into a single, collective output from the team. This aggregation of individual decisions represents a form of collective cognition and can be expressed using several metaphors, such as swarm intelligence 2 , 3 , a ‘collective brain’ 4 , 5 or a superorganism 6 , 7 . This process of collating multiple judgements into a single collective decision can be formulated in several ways 8 , 9 . Where there is a finite number of options, such as a binary decision, we can use a majority vote process where each individual output functions as a single ‘vote’ that are counted to reach a team decision, i.e., which option accumulates the highest number of ‘votes’. However, this egalitarian approach ignores important differences between individual team members, for instance, some may be more experienced than others, others may have greater levels of skill and a stronger track record of performance. If we wish to reflect these differences in the collective decision, a process of weighting individual inputs can be applied alongside accumulation. As a secondary dimension, the aggregation of individual inputs also contains information about the degree of agreement within the collective. Some decisions may be unanimous whereas others will reflect disagreement within the team. The need for individual responses to be aggregated, weighed and assessed for coherence requires the existence of a supervisory agent 10 – 12 . This technological entity exists at a superordinate level to the team. The supervisory agent is capable of monitoring individual decisions, which are used as inputs to generate a collective decision and produce metadata about team decisions as collateral information, i.e., level of agreement within the team. The process of aggregating and weighing individual responses is defined in this superordinate, supervisory level. The purpose of the supervisory system is to improve the quality of team decision-making by applying dynamic adjustments to the process of aggregation 13 , 14 . However, guaranteeing that the addition of the supervisory agent will increase the number of “good” decisions is challenging 15 , 16 , partly because weightings can interact in unpredictable ways but also because the supervisory system is often not privy to the ground truth of the decision, if one exists. The supervisory agent can use several different metrics when assessing the quality of individual decisions. It can require every member of the team to explicitly provide a self-reported assessment of each decision. For example, each team member could rate their level of confidence after a decision has been made 17 , 18 . However, the repeated need to self-report, especially when decisions are made in quick succession, can be burdensome for the individual. Other metrics are implicit and do not require any overt response from the team member. Measures of behaviour, such as time to decide, can be meaningful indicators of decision quality, but their interpretation depends on the context of the decision-making task 19 – 21 . When tasks are complex, multifaceted and dependent on dynamic factors, a low decision time may be indicative of a hasty, poorly-planned response. In the case of simple decisions, a longer than average decision time can be interpreted as uncertainty on the part of the team member 19 , 22 – 25 . Neurophysiology (and psychophysiology) can also serve as implicit measures of decision-making quality from the individual 26 . These metrics can quantify the psychological state of the individual, e.g., to assess level of mental workload, fatigue; they can also be used to assess the behavioural intention of the team member 27 , 28 . The latter is called a collaborative Brain-Computer Interface (cBCI) 29 , which works via two stages of processing: (1) monitoring neurophysiology (EEG) during the decision-making process for each team member, and (2) using neurophysiological metrics (alongside other behavioural metrics) to ‘weigh’ the contribution of each individual decision to the collective decision of a team as a whole. In other words, a cBCI permits a process of assessment to be applied to each individual decision, which determines the weight of each individual decision as a contribution to a collective team decision. Collaborative BCI have been developed to improve team performance on a range of visual detection tasks. Early work 29 , 30 explored this concept using abstract visual target detection tasks and reported that groups generally outperformed individuals and that weighing responses by combining BCI with subjective ratings of decision confidence was an effective way to improve the performance of the team. This work was replicated using more naturalistic targets presented as still images 30 and extended by the inclusion of other behavioural markers such as eye tracking 31 . This work was developed to explore cBCI with more realistic visual search tasks that approximated real-world scenarios and were dynamic in nature 32 ; this paper demonstrated that cBCI outperformed other collective models, such as majority voting by integrating neurophysiological features from EEG with behavioural markers such as response time and reported confidence levels. In recent years, this work has been replicated using more complex decision-making tasks 33 and face recognition 34 . Others have developed cBCI systems that are capable of mutual learning within the team and responding to video feeds from Unmanned Aerial Vehicles 35 (UAV). The level of demand provides important context for cBCI performance. High workload tasks that are challenging are likely to lead to more variable human performance within a team compared to tasks that are less demanding. Furthermore, the benefits of accumulated performance may be more substantial for high workload tasks compared to the low workload tasks, especially when participants are performing decision-making tasks using partial or degraded visual information. The current paper will describe a cBCI system designed to assess the quality of individual decisions to be used as an input to the performance of human teams of different sizes. This system will amalgamate data from EEG and human performance to provide a weighting of each individual decision in the context of a team so that ‘good decisions’ make a larger contribution to the team decision compared to ‘bad decisions’, thus conferring an advantage on the performance of the team. The performance of this cBCI will be assessed in the context of high and low workload using a VR-based simulation of UAV target detection task. Results Individual Performance To verify the intended effects of the experimental manipulation, individual behavioral performance was analyzed by comparing outcomes between the Low and High Workload conditions. Key metrics including accuracy, response time (RT), and subjective confidence ratings were examined. Accuracy Analysis of decision accuracy revealed a significant impact of workload. Paired t-tests on subject-level mean accuracies indicated that participants performed significantly worse under High Workload compared to Low Workload, both when considering ReticleOn locked data (t(16) = -9.455, p < 0.0001) and ButtonPress locked data (t(16) = -8.581, p < 0.0001). Descriptively, mean accuracy assessed from ReticleOn epochs decreased from 93.2% (SD = 7.9%) in the Low Workload condition to 76.9% (SD = 11.8%) in the High Workload condition (Fig. 2 ). Response Times Response times (calculated for non-miss trials) were also significantly modulated by the workload manipulation. Participants exhibited significantly slower response times during High Workload trials compared to Low Workload trials. This was confirmed by paired t-tests on subject-level mean RTs for ReticleOn locked data (High: M = 1.460s, SD = 0.312s; Low: M = 0.862s, SD = 0.355s; t(16) = 10.395, p < 0.0001) (Fig. 3 ). Furthermore, response times for incorrect decisions were generally slower than for correct decisions, as visualized by the distributions in Fig. 4 . Subjective Confidence Furthermore, subjective confidence ratings reflected the change in task demand. Participants reported significantly lower confidence in their decisions (calculated for non-miss trials) during the High Workload blocks compared to the Low Workload blocks. Paired t-tests confirmed this effect for both ReticleOn locked data (High: M = 52.54, SD = 25.54; Low: M = 78.71, SD = 19.59; t(16) = -8.949, p < 0.0001) (Fig. 5 ). Participants also typically reported higher confidence for correct decisions compared to incorrect decisions, as illustrated by the distributions in Fig. 6 . Team Performance Analysis The performance of simulated teams, incorporating these BCI-SVM outputs alongside behavioural data, was then evaluated under different aggregation methods across varying group sizes (2, 4, 6, and 8). When interpreting the following team-based results, it is crucial to acknowledge that improvements in group performance over that of an average individual can arise from inherent mathematical and statistical principles. These include: 1. Error Cancellation (Wisdom of Crowds): Aggregating multiple, even noisy, independent or partially independent estimates (like individual decisions) can lead to a more accurate collective outcome as random errors tend to cancel each other out 36 . This principle underpins basic majority voting. 2. Increased Effective Sample Size: Each team decision, especially in weighted methods, effectively draws upon more data points (e.g., an individual's response, their RT, their BCI state) than a single individual's decision, potentially leading to more robust outcomes. 3. Exploitation of Complementary Error Profiles: This is a particularly important consideration for human-BCI synergy. If human decision-making and BCI classifications make errors on different types of trials or for different underlying reasons (i.e., their error profiles are complementary), then combining them can lead to a more accurate overall decision. In such cases, one system might be reliable and provide a "good decision" precisely when the other system is struggling or likely to fail, and vice-versa. A sophisticated integration strategy can capitalize on this complementarity. While the first two effects contribute to the general benefit of teamwork and information integration, a primary objective of this study is to determine if the specific cBCI methodologies employed offer advantages that extend significantly beyond these foundational aggregation benefits, particularly by leveraging the potential for complementary error profiles between human and BCI. To this end, our analyses focus on several key comparisons: 1. Surpassing the Average Individual: As a baseline confirmation of team benefit. 2. Outperforming Simple Aggregation Rules: Comparing advanced cBCI methods against standard majority votes and simple behavioral weighting (e.g., RT-only) to assess the added value of neurophysiological information and more sophisticated integration that might tap into complementary strengths. 3. Challenging the Best Individual: Critically, we evaluate whether cBCI methods can achieve accuracies surpassing the average performance of the best-performing individual member within each simulated team. Such an outcome would more strongly suggest a synergistic effect, where the integrated system creates a decision superior to what even the top individuals typically achieve. This could arise not just from better selection of confident decisions, but also from the system correctly weighting one information source (human or BCI) when the other is less reliable for a given trial. 4. Systematic Evaluation of Components: By examining a range of aggregation methods—from human-only approaches to those incorporating BCI data with and without filtering—we aim to elucidate the incremental contributions of different information sources and processing strategies, and how they might interact to exploit differing system strengths 37 . The following sections will detail team performance under various conditions, with these interpretive considerations in mind, paying particular attention to evidence for synergistic gains that may stem from the effective combination of human and BCI decision processes. Reticle On High Workload For the high workload reticle on epoch, the individual participant-specific SVM classifiers, which inform several BCI-driven aggregation methods, performed above chance on average (M = 0.673, SD = 0.072), although performance varied considerably across participants (range: 0.558–0.823). This indicates that while the BCI component provided useful information, its quality varied (see Fig. 7 ). Exploratory analysis of SVM feature importance revealed a diverse set of discriminative features, including both time-domain and frequency-domain measures; the top selected features for this condition. Analysis of raw agreement between simple human majority and BCI majority decisions showed that agreement increased steadily with group size, rising from approximately 63.6% for N = 2 to 82.4% for N = 8, suggesting a degree of concordance but also highlighting that human and BCI decisions were not perfectly redundant. Under High Workload conditions with EEG data epoched to the Reticle On event (Fig. 8 ), various team aggregation methods were compared. The standard Majority Human method (e.g., blue circle line in Fig. Y) achieved accuracies ranging from 81.8% (N = 2) to 95.2% (N = 8). Behavioral weighting methods included RT Weighted Human (e.g., light blue square line), which yielded accuracies from 83.4% (N = 2) to 94.7% (N = 8), and Subjective Confidence Weighted Human (e.g., dark blue triangle line), which produced accuracies from 87.0% (N = 2) to 97.3% (N = 8). The combined human-only method, RT + Subjective Confidence Human, resulted in accuracies from 86.8% (N = 2) to 96.8% (N = 8). The BCI-only aggregation, SVM Confidence Weighted BCI (e.g., light green square line), produced accuracies in the range of 72.8% (N = 2) to 87.6% (N = 8). The mixed methods, integrating human and BCI information, generally demonstrated the strongest performance. The Subjective Confidence + SVM Confidence Mixed method (e.g., magenta star line) showed the highest accuracies in this condition, increasing from 89.3% (N = 2) to 95.5% (N = 4), 97.9% (N = 6), and reaching 98.8% for eight-person teams. The comprehensive RT + Subjective Confidence + SVM Confidence Mixed method (e.g., firebrick H line) also achieved high accuracies, from 89.6% (N = 2) to 95.3% (N = 4), 97.6% (N = 6), and 98.6% (N = 8). The RT + SVM Confidence Mixed method (e.g., orangered pentagon line) yielded accuracies from 87.7% (N = 2) to 97.5% (N = 8). Notably, for larger team sizes, the Subjective Confidence + SVM Confidence Mixed method surpassed the average accuracy of the Best Individual Avg (black solid line; N = 6: 97.9% vs. 93.3%; N = 8: 98.8% vs. 94.2%), suggesting a synergistic benefit. Similar synergistic effects, where team accuracy exceeded that of the best individual member, were also observed for the RT + Subjective Confidence + SVM Confidence Mixed method (N = 6: 97.6% vs. 93.3%; N = 8: 98.6% vs. 94.2%) and the RT + SVM Confidence Mixed method (N = 6: 96.3% vs. 93.3%; N = 8: 97.5% vs. 94.2%). The Average Individual Avg (grey dashed line) was consistently around 81.5%. Reticle On Low Workload Under low workload conditions for the Reticle On epoch, individual SVM classifiers performed with an average accuracy of M = 0.710 (SD = 0.116), ranging from 0.497 to 0.891 (Fig. 9 ). While generally above chance, one participant's SVM performed below chance. The features most influential for SVM decisions in this low workload condition differed somewhat from the high workload scenario. Features such as Theta_Power_FC2, Theta_Power_F7, and Alpha_Power_Fp2 were among the most frequently selected by the SVMs. In terms of impact, Beta_Power_CP6 and Mean_CP1 exhibited high average importance scores when selected. Raw agreement between simple human majority and BCI majority decisions was higher than under high workload, increasing from approximately 74.1% for N = 2 to 90.3% for N = 8 Under Low Workload conditions with EEG data epoched to the Reticle On event (Fig. 10 ), overall team performance was considerably higher than under high workload, with many methods approaching ceiling levels, particularly for larger group sizes. The standard Majority Human method achieved accuracies ranging from 92.4% (N = 2) to 99.6% (N = 8). Human-only behavioral weighting methods generally performed very well: RT Weighted Human accuracies ranged from 94.7% (N = 2) to 99.8% (N = 8); Subjective Confidence Weighted Human from 94.5% (N = 2) to 99.6% (N = 8); and RT + Subjective Confidence Human from 95.0% (N = 2) to 99.7% (N = 8). The SVM Confidence Weighted BCI method performed lower, with accuracies from 76.5% (N = 2) to 90.9% (N = 8). The mixed methods integrating human and BCI information also achieved very high accuracies. Specifically, RT + SVM Confidence Mixed yielded accuracies from 95.1% (N = 2) to 99.8% (N = 8), and RT + Subjective Confidence + SVM Confidence Mixed produced similar results from 95.3% (N = 2) to 99.8% (N = 8). The Subjective Confidence + SVM Confidence Mixed method also performed strongly, with accuracies from 95.1% (N = 2) to 99.8% (N = 8). While these advanced aggregation methods, such as RT + SVM Confidence Mixed and RT + Subjective Confidence + SVM Confidence Mixed, slightly outperformed the Average Individual Avg (approximately 94.4%), their advantage over simpler, high-performing methods like RT Weighted Human or even the standard Majority Human vote was minimal in this low workload context. Furthermore, these cBCI methods generally did not surpass the average Best Individual Avg, which itself was extremely high (ranging from 98.0% for N = 2 to 99.9% for N = 8). The relative gains from the more complex cBCI aggregation strategies were thus substantially smaller and less distinct than those observed under high workload conditions. Button Press High Workload For the high workload button press epoch, the individual participant-specific SVM classifiers performed above chance on average (M = 0.692, SD = 0.101), with accuracies ranging from 0.587 to 0.863 across participants (Fig. 11 ). This average SVM performance was slightly higher than that observed for the ReticleOn epochs under high workload. Exploratory analysis of SVM feature importance for Button Press epochs revealed a distinct set of influential features compared to ReticleOn; the top selected features for this condition were Var_F10, Var_F7, and Theta_Power_Oz were the most frequently selected features (each 17.6%), while Var_F4 showed the highest average importance score (3.660). Analysis of raw agreement between simple human majority and BCI majority decisions showed agreement increasing from approximately 65.8% for N = 2 to 86.5% for N = 8, rates slightly higher than in the ReticleOn high workload condition. Under High Workload conditions with EEG data epoched to the Button Press event (Fig. 12 ), a similar pattern of team performance to the ReticleOn high workload analysis emerged. The standard Majority Human method achieved accuracies ranging from 81.8% (N = 2) to 94.7% (N = 8). Behavioral weighting methods, such as RT Weighted Human and Subjective Confidence Weighted Human, yielded accuracies from 83.4–94.4% and 86.9–97.0%, respectively. The combined human-only method, RT + Subjective Confidence Human, resulted in accuracies from 86.7% (N = 2) to 96.4% (N = 8). The BCI-only aggregation, SVM Confidence Weighted BCI, produced accuracies in the range of 75.3% (N = 2) to 91.1% (N = 8). The mixed methods integrating human and BCI information again demonstrated strong performance. The Subjective Confidence + SVM Confidence Mixed method showed the highest accuracies, increasing from 89.9% (N = 2) to 95.6% (N = 4), 97.8% (N = 6), and reaching 98.6% for eight-person teams. The RT + Subjective Confidence + SVM Confidence Mixed method also performed robustly, with accuracies from 89.8% (N = 2) to 95.4% (N = 4), 97.5% (N = 6), and 98.4% (N = 8). Similarly, the RT + SVM Confidence Mixed method yielded accuracies from 87.8% (N = 2) to 97.7% (N = 8). These mixed methods consistently outperformed simpler aggregation strategies and, for larger team sizes, surpassed the average accuracy of the Best Individual Avg; for instance, Subjective Confidence + SVM Confidence Mixed (N = 6: 97.8% vs. 93.1%; N = 8: 98.6% vs. 94.0%), RT + Subjective Confidence + SVM Confidence Mixed (N = 6: 97.5% vs. 93.1%; N = 8: 98.4% vs. 94.0%), and RT + SVM Confidence Mixed (N = 6: 96.5% vs. 93.1%; N = 8: 97.7% vs. 94.0%). The Average Individual Avg was consistently around 81.4%. Button Press Low Workload For the low workload button press epoch, the individual participant-specific SVM classifiers performed with an average accuracy of M = 0.700 (SD = 0.078), with accuracies ranging from 0.541 to 0.861 (Fig. 13 ). The features most influential for SVM decisions under these conditions were, Beta_Power_FC5 and Theta_Power_C4 were among the most frequently selected features (each 17.6%), while Var_O2 (average importance: 0.586) and Var_F10 (average importance: 0.368) showed high average importance scores. Raw agreement between simple human majority and BCI majority decisions was high, increasing from approximately 73.0% for N = 2 to 87.5% for N = 8. Team performance under Low Workload for ButtonPress epochs showed very high accuracies across most methods, consistent with the low task demand. The standard Majority Human method yielded accuracies from 92.3% (N = 2) to 99.6% (N = 8). Behavioral weighting methods generally performed strongly: RT Weighted Human accuracies ranged from 94.8% (N = 2) to 99.7% (N = 8); Subjective Confidence Weighted Human from 94.6% (N = 2) to 99.6% (N = 8); and RT + Subjective Confidence Human from 95.0% (N = 2) to 99.7% (N = 8). The SVM Confidence Weighted BCI method was less accurate, with values from 75.3% (N = 2) to 88.6% (N = 8). The mixed methods integrating human and BCI information also achieved very high accuracies. The RT + Subjective Confidence + SVM Confidence Mixed method produced accuracies from 95.4% (N = 2) to 99.8% (N = 8). Similarly, RT + SVM Confidence Mixed yielded accuracies from 95.2% (N = 2) to 99.8% (N = 8), and Subjective Confidence + SVM Confidence Mixed achieved accuracies from 95.0% (N = 2) to 99.7% (N = 8). However, similar to other low workload conditions, the incremental benefit of these complex combined methods, while achieving high absolute accuracies, was marginal compared to simpler high-performing aggregations like RT Weighted Human, or the Best Individual Avg (which ranged from 98.0% for N = 2 to 99.8% for N = 8). The Average Individual Avg was consistently around 94.5% (Fig. 14 ). Discussion The current study investigated the potential for a collaborative Brain-Computer Interface (cBCI), integrating EEG-derived decision confidence (SVM_Confidence) with behavioural metrics (RT, subjective confidence), to enhance team perceptual decision-making accuracy in a VR drone task under varying cognitive workload. Our primary finding demonstrated that under High Workload, specific mixed cBCI aggregation methods enabled simulated teams (N=2-8) to significantly surpass the average performance of their best individual member. This synergistic outcome, where the collective effectively transcended individual capabilities through efficient information fusion 29,32 , was critically dependent on high task demand; under Low Workload, where individual performance neared ceiling, the benefits of these advanced integration methods were minimal. This pronounced workload dependency highlights the neuroergonomic utility of cBCIs in challenging operational contexts where human performance variability is high 38,39 . Intriguingly, these cBCI benefits under high cognitive load were robust across EEG data epoched to both anticipatory pre-response ('ReticleOn') and peri-/post-response ('ButtonPress') events. This suggests that exploitable neural correlates of decision quality and relevant behavioral markers are present and integrable from both time windows, with the ReticleOn findings notably indicating a predictive capacity for team-level decision weighting 40–42 . Comparing aggregation strategies, mixed methods integrating human-derived information with BCI-derived SVM confidence consistently yielded the highest accuracies and demonstrated synergy. The 'Subjective Confidence + SVM Confidence Mixed' method frequently emerged as the top performer, suggesting a potent fusion of self-reported certainty and neurally-derived decision quality. While the comprehensive 'RT + Subjective Confidence + SVM Confidence Mixed' method was competitive, it did not consistently outperform the two-component SC+SVM mix in high-demand scenarios. Simpler aggregation methods (e.g., Majority vote, BCI-only) were generally less effective under high load, aligning with broader findings on the benefits of integrating multiple, diverse information streams in cBCI and team decision-making 43,44 . Exploratory analyses of trait impulsivity and risk-taking (BIS-11, BART) revealed no significant associations with individual performance metrics in this task, suggesting that workload and aggregation method were more dominant factors, or that the task demands were not optimally suited to elicit strong trait-based effects 45,46 . The current study has several limitations. The sample size (N=17) may limit the generalizability of individual difference analyses. The VR drone task, while designed for ecological relevance, is specific, and findings require replication across other domains and team compositions. Our BCI approach used SVM-derived confidence for weighting; future work could explore incorporating overall classifier reliability 47–49 . Furthermore, team performance was assessed via offline simulations, which, while enabling rigorous control, do not capture real-time team dynamics. In conclusion, this research provides compelling evidence that cBCIs integrating diverse human-derived and EEG-based information can markedly enhance team perceptual decision-making beyond best individual performance, particularly under high cognitive workload. This workload-dependent synergy, achievable using both anticipatory and post-response neural data, underscores the potential of neuroergonomically-informed cBCIs for supporting teams in demanding operational contexts. Future research should focus on replicating these findings with larger samples, investigating adaptive, AI-driven aggregation strategies 50,51 , and crucially, validating these cBCI frameworks in real-time, interactive team settings to further the development of systems that genuinely augment collective human intelligence 52,53 . Methods Participants Seventeen healthy adults (N=17; [10 Female]; Mean age ± SD = [24.8 ± 6.0]) participated in the study and were included in the final analysis dataset. Participants were recruited from the university undergraduate and postgraduate community. An initial cohort of 27 individuals was recruited; however, data from ten participants were excluded prior to the main analysis due to criteria established in preliminary quality control checks, including insufficient data quality after EEG preprocessing, or significant deviations in trial sequence alignment during task execution due to technical issues and the need for alignment of trials across all combinations of team group size. All included participants reported normal or corrected-to-normal vision and no history of neurological disorders or particular susceptibility to VR-induced motion sickness. Prior to the experiment, participants provided written informed consent and completed the Barratt Impulsiveness Scale (BIS-11) 54 and the Balloon Analogue Risk Task (BART) prior to commencing the study. Participants received monetary compensation for their participation. The experimental protocol was approved by UK MoDREC, App No: 2309/MODREC/24 Ref: RQ0000037929 and all procedures were conducted in accordance with the ethical standards outlined in the Declaration of Helsinki. Study Design The study employed a within-subject repeated measures design to evaluate the effectiveness of a collaborative Brain-Computer Interface (cBCI) system designed to enhance group decision-making in a dynamic virtual reality (VR) environment. The primary within-subject factor manipulated was cognitive workload, presented at two levels (Low vs. High) across distinct experimental blocks. Electroencephalographic (EEG) data were analyzed time-locked to two different events: the onset of a targeting reticle ('ReticleOn'), representing a pre-stimulus anticipatory period, and the participant's response ('ButtonPress'), capturing peri- and post-decisional neural activity. Team performance was assessed through offline simulations, evaluating decision accuracy for simulated groups of varying sizes (2, 4, 6, and 8 members) under different decision aggregation algorithms. Individual behavioral metrics, including accuracy, response time (RT), and subjective confidence ratings, served as primary dependent variables at the individual level. Additionally, personality traits related to impulsivity and risk-taking were assessed for exploratory analysis. VR Drone Task and Procedure Participants were seated in the laboratory while they viewed the virtual environment via a Varjo Aero HMD. The simulation, developed in Unreal Engine 5 and rendered on a high-performance PC, depicted the viewpoint of a quadcopter type drone flying over a simulated landscape designed to be semi realistic i.e.textured realistically for a temperate climate and populated with sparse foliage but tailored to not interfere with stimuli acquisition (See Fig. 15). Task stimuli consisted of 3D models appearing on the ground: Non-Targets (See Fig. 16) and Targets (See Fig. 17). Participants completed six 8-minute blocks, with each block containing 50 trials (30 Non-Targets, 20 Targets). Stimuli were evenly spaced 300m apart, but varied in lateral placement (-30m, -15m, 0m, +15m, +30m)] and rotation (0°, 90°, 180°, 270°) . Each trial followed a consistent sequence. Initially, the drone flew to a set virtual height (25m) and advanced across the landscape. As stimuli came into view a targeting reticle appeared over the stimuli ('ReticleOn') for 2500ms (See Fig. 18). Participants were tasked with discriminating the stimulus type and responding as quickly and accurately as possible by pressing the designated joystick button for Target or Non-Target. Participants were instructed to respond whilst the reticle was on-screen. Any responses made after the reticle disappeared were not counted, and those trials were classed as misses. The reticle initially appeared blue, then changed to red after 2000ms, to warn the participants that they only had 500ms to make a decision (See Fig. 19). Following this primary response ('ButtonPress'), a prompt appeared, requiring participants to rate their confidence in the preceding decision via tilting the joystick to select a value on a 0-100 scale (0 = Not Confident, 100 = Very Confident) (See Fig. 20). Cognitive workload was manipulated across different blocks of trials. The Low Workload condition had the virtual light level fixed as approximate for midday average daylight. Conversely, the High Workload condition reduced the light level by 50% , and the solar angle by 9° simulating a night condition (See Fig. 21). Participants first completed a practice block for each workload condition, followed by three blocks under each workload condition. Data Acquisition Continuous EEG data were recorded using a 32-channel LiveAmp system (Brain Products GmbH). Electrodes were arranged according to the international 10-20 configuration using an electrode cap (actiCAP snap electrode cap). Average referencing was utilised, the actiCAP utilised a dedicated ground electrode point at the front and centre of the head, between Fp1 and Fp2. Electrode impedances were kept below 30KΩ throughout the recording. The EEG signal was recorded at a sampling rate of 500Hz. Behavioral data, including response button presses (for RT calculation) and subsequent confidence ratings, were logged via joystick actions which generated LabStreamingLayer 55 (LSL) markers.. Event markers corresponding to critical task events (e.g., ReticleOn, StimulusOn, ButtonPress) were generated by the Unreal Engine environment using the LSL UE5 plugin and transmitted via LSL. Both EEG and event marker streams were simultaneously recorded and synchronized using LabRecorder and OpenSignals software, resulting in .xdf files for each session. Procedure Upon arrival at the laboratory, participants were briefed on the study's objectives and procedures. After providing written informed consent, they completed the BART and BIS-11 questionnaires. Participants were then fitted with the EEG cap and the VR HMD. Electrode impedances were checked and adjusted to be below. After baselining, participants undertook a training task of two blocks (one for each workload condition) then the VR drone task. They received instructions on the target/non-target discrimination, the response mapping (button presses), and the confidence rating scale. The main experiment consisted of six blocks, divided equally between the Low Workload and High Workload conditions (three blocks per condition). The entire experimental session, including setup, task execution, and debriefing, lasted approximately 150 minutes per participant. Upon completion, participants were debriefed and received their monetary compensation. EEG Signal Processing EEG data were processed offline using MNE-Python 56 and custom Python scripts. The processing pipeline included: Loading and Filtering: Data from each XDF file were loaded. A band-pass filter (0.1–30 Hz, FIR) and a 50 Hz notch filter were applied to the data. Trimming: The continuous recording was trimmed to the duration of the experimental task using the first and last LSL markers. Artifact Rejection (ICA): Independent Component Analysis (ICA) was employed to identify and remove stereotypical artifacts such as eye blinks and lateral eye movements. To enhance ICA quality, data from all sessions for a given participant were concatenated. FastICA was run on this concatenated data. Components reflecting ocular or other non-neural artifacts were manually identified by visual inspection of their topography and time course and were flagged for removal. The corresponding ICA demixing matrix was then applied to the individual session files to project out the artifactual components. Epoching: The cleaned continuous data were segmented into epochs relative to two primary events: (i) 'ReticleOn' epochs (-200 ms to +800 ms relative to marker onset) and (ii) 'ButtonPress' epochs (-500 ms to +300 ms relative to marker onset). Baseline Correction: Epochs were baseline-corrected using the pre-event interval: -200 ms to 0 ms for ReticleOn epochs, and -500 ms to -200 ms for ButtonPress epochs. Trial Validation: Specific trials identified as problematic during preliminary checks (e.g., Trial 18) were excluded during the creation of the final metadata associated with the epochs. BCI Feature Extraction and Classification A participant-specific BCI was developed using an SVM classifier to predict the likelihood of decision correctness based on EEG features. Feature Extraction: For each epoch (ReticleOn or ButtonPress), features were extracted per channel (32 channels). Time-domain features included the mean amplitude, maximum amplitude, and variance across the epoch. Frequency-domain features included the average Power Spectral Density (PSD) within the Theta (4–8 Hz), Alpha (8–13 Hz), and Beta (13–30 Hz) bands, estimated using Welch's method (1-40 Hz range). SVM Training: For each participant, an SVM classifier (sklearn.svm.SVC 57 ) was trained. Feature vectors were first standardized (StandardScaler). SelectKBest with mutual_info_classif scoring identified the top 5 features. To handle potential class imbalance (more correct than incorrect trials), the feature-selected training data was balanced using RandomUnderSampler. Hyperparameters (kernel type: 'rbf'/'linear'; C: 0.1-100; gamma: 'scale'/'auto') were optimized via 5-fold stratified cross-validation using GridSearchCV on this balanced training set, maximizing accuracy. Prediction and Confidence Score: The optimized SVM for each participant was then applied to predict the label (1=Correct Prediction, 0=Incorrect Prediction) for all of their valid, feature-selected (but unbalanced) trials. The classifier's decision function output for each trial, representing the signed distance from the separating hyperplane, was recorded as the SVM Confidence score. This score served as a BCI-derived measure of decision certainty. Team Simulation Procedure and Aggregation Methods Performance of simulated teams was evaluated offline. To ensure a comprehensive and robust analysis, an exhaustive combinatorial approach was employed. For each of the approximately 150 valid experimental trials within each workload condition (Low and High), team performance was simulated independently for every possible unique combination of participants drawn from the final N=17 cohort for each specified team size (m = 2, 4, 6, and 8). This meant that for any single trial, decisions were simulated for: all 136 unique two-person teams (the number of distinct combinations of 2 participants from 17), all 2,380 unique four-person teams (the number of distinct combinations of 4 participants from 17), all 12,376 unique six-person teams (the number of distinct combinations of 6 participants from 17), all 24,310 unique eight-person teams (the number of distinct combinations of 8 participants from 17). Given approximately 150 trials per workload condition, and two workload conditions (Low and High, totaling ~300 trials available for analysis after individual trial validation), the total number of simulated team decisions generated was substantial: For two-person teams, this involved 136 unique team combinations, each assessed over approximately 300 trials, resulting in approximately 40,800 simulated decisions. For four-person teams, this involved 2,380 unique team combinations, each assessed over approximately 300 trials, resulting in approximately 714,000 simulated decisions. For six-person teams, this involved 12,376 unique team combinations, each assessed over approximately 300 trials, resulting in approximately 3,712,800 simulated decisions. For eight-person teams, this involved 24,310 unique team combinations, each assessed over approximately 300 trials, resulting in approximately 7,293,000 simulated decisions. In total, this comprehensive simulation strategy yielded over 11.7 million individual team decision points for analysis across all team sizes and trials. For these per-trial, per-team-composition simulations, data were grouped by a unique trial identifier (Trial_Number) ensuring that for each simulated team decision, only data (e.g., individual response, RT, SVM confidence) from participants who had experienced that identical trial were included when constituting that specific team instance. Team decisions were aggregated using several methods (summarized in Table 1 and detailed below), which include various human-only, BCI-only, and mixed human-BCI approaches: Majority Human: A simple count of human responses (Target/Non-Target). Ties were typically resolved randomly or by a predefined rule (e.g., favouring Target classification). RT Weighted Human: Human responses were weighted by their corresponding normalized response time (a 0-1 scale where higher values indicate faster RT). The team decision favoured the response type (Target or Non-Target) with the higher sum of weights. Ties were resolved by favouring Target classification. Subjective Confidence Weighted Human: Human responses were weighted by their normalized subjective confidence (0-1 scale). The team decision favoured the response type (Target or Non-Target) with the higher sum of weights. Ties were resolved by favouring Target classification. RT + Subjective Confidence Human: For each team member, their normalized RT and normalized subjective confidence were averaged to form a single decision weight. This weight was assigned to their actual human response. The team decision favoured the response type with the higher total summed weight across members. Ties were resolved by favouring Target classification. SVM Confidence Weighted BCI: SVM-predicted labels (Target/Non-Target) from each team member were weighted by their corresponding normalized SVM confidence (a 0-1 scale derived from the absolute SVM decision function output). The team decision favoured the predicted label type with the higher sum of confidence weights. Ties were resolved by favouring Target classification. RT + SVM Confidence Mixed: This method integrated human response time and BCI-SVM confidence. For each individual, the evidence supporting a Target or Non-Target decision was calculated by giving equal 0.5 weighting to their RT-weighted human response and their SVM confidence-weighted BCI prediction. For example, the evidence for a particular choice was determined as half the human-derived score for that choice plus half the BCI-derived score for that choice. The team decision favoured the response type with the higher total summed evidence across members. Ties were resolved by favouring Target classification. Subjective Confidence + SVM Confidence Mixed: This method integrated human subjective confidence and BCI-SVM confidence. For each individual, the evidence supporting a Target or Non-Target decision was calculated by giving equal 0.5 weighting to their subjective confidence-weighted human response and their SVM confidence-weighted BCI prediction. For example, the evidence for a particular choice was determined as half the human-derived score for that choice plus half the BCI-derived score for that choice. The team decision favoured the response type with the higher total summed evidence across members. Ties were resolved by favouring Target classification. RT + Subjective Confidence + SVM Confidence Mixed: This comprehensive method integrated human RT, human subjective confidence, and BCI-SVM confidence. For each individual, a "human component" was derived from the average of their normalized RT and normalized subjective confidence. This human component and the individual's "BCI component" (normalized SVM confidence) were then each given a 0.5 weighting. These two weighted components (one reflecting the human response and one reflecting the BCI prediction) were summed to determine the individual's contribution to the team's evidence for a Target or Non-Target decision. The team decision favoured the response type with the higher total summed evidence. Ties were resolved by favouring Target classification. Individual Performance Baselines: Team performance for these aggregation methods was also compared against baseline accuracies derived from the average accuracy of the best, worst, and average human performer within each specific simulated group combination on each trial. The main outcome measure reported is the mean team accuracy, averaged across all unique team combinations and valid trials for each group size and decision method. Unless otherwise specified for a particular method (e.g., Majority Human), ties in accumulated evidence for weighted methods were resolved by favouring Target classification. Table 1: Calculation Summary of Key Team Decision Methods Method Label in Plot Core Logic Description Key Information Used per Team Member (Trial-Level) Worst Individual Avg Average of the lowest individual accuracy observed within each unique simulated team composition. Overall Human Accuracy (of the worst performer in each group iteration) Average Individual Avg Average of the mean individual accuracy observed within each unique simulated team composition. Overall Human Accuracy (mean of individuals in each group iteration) Best Individual Avg Average of the highest individual accuracy observed within each unique simulated team composition. Overall Human Accuracy (of the best performer in each group iteration) Majority Human Each member's binary human response (Target/Non-Target) contributes one vote. Team decision is the most frequent response. Ties typically resolved randomly or by pre-defined rule. Human Response (Target/Non-Target) RT Weighted Human Human responses are weighted by normalized RT (faster RT = higher weight). Team decision by sum of weights. Ties favour Target. Human Response, Normalized RT Subjective Confidence Weighted Human Human responses are weighted by normalized subjective confidence. Team decision by sum of weights. Ties favour Target. Human Response, Normalized Subjective Confidence RT + Subjective Confidence Human Human responses weighted by the average of normalized RT and normalized subjective confidence. Team decision by sum of weights. Ties favour Target. Human Response, Normalized RT, Normalized Subjective Confidence SVM Confidence Weighted BCI SVM-predicted labels are weighted by normalized SVM confidence. Team decision by sum of weights. Ties favour Target. SVM Predicted Label, Normalized SVM Confidence RT + SVM Confidence Mixed For each member, evidence for a decision (Target/Non-Target) is 0.5 * (Human Score from RT) + 0.5 * (BCI Score from SVM Confidence). Team decision by sum of total evidence. Ties favour Target. Human Response, Normalized RT, SVM Predicted Label, Normalized SVM Confidence Subjective Confidence + SVM Confidence Mixed For each member, evidence for a decision (Target/Non-Target) is 0.5 * (Human Score from Subjective Confidence) + 0.5 * (BCI Score from SVM Confidence). Team decision by sum of total evidence. Ties favour Target. Human Response, Normalized Subjective Confidence, SVM Predicted Label, Normalized SVM Confidence RT + Subjective Confidence + SVM Confidence Mixed For each member, a "human component" score is the average of Normalized RT and Normalized Subjective Confidence. Evidence for a decision (Target/Non-Target) is 0.5 * (Human Component Score) + 0.5 * (BCI Score from SVM Confidence). Team decision by sum of total evidence. Ties favour Target. Human Response, Normalized RT, Normalized Subjective Confidence, SVM Predicted Label, Normalized SVM Confidence Statistical Analysis Individual behavioral data (accuracy, RT, confidence) were analyzed to assess the impact of the Workload manipulation (Low vs. High). Depending on data distributions, paired t-tests or Wilcoxon signed-rank tests were used for continuous variables (RT, confidence), while Chi-square tests were used for accuracy (comparing counts of correct/incorrect decisions). For simulated team performance, mean accuracies for the proposed cBCI weighting method(s) were compared against baseline methods (Majority, Best/Average Individual) for each group size using paired t-tests or Wilcoxon tests. Corrections for multiple comparisons (e.g., Bonferroni) were applied where appropriate. Statistical significance was defined at an alpha level of p < 0.05. All statistical analyses were performed using Python 58 and its scientific computing libraries, primarily SciPy 59 , for significance testing. Data processing and manipulation were conducted using Pandas 60 and NumPy 61 . Visualizations were generated with Matplotlib 62 and Seaborn 63 . Hypotheses Individual Performance (Manipulation Check): We hypothesised that the High Workload condition would significantly impair individual performance compared to the Low Workload condition, manifesting as: a. Lower decision accuracy. b. Slower response times (RT). c. Lower subjective confidence ratings. Team accuracy using the cBCI method would be significantly higher than the average accuracy of the individual members comprising the team. Workload Interaction: We hypothesised that the performance benefit conferred by the cBCI method would be dependent on cognitive workload. Specifically: a. The accuracy gain provided by the cBCI method (relative to Majority vote or Average Individual performance) would be significantly greater under the High Workload condition compared to the Low Workload condition. b. (Stronger/More Specific version of 2b related to synergy): Under High Workload conditions, team accuracy using the cBCI method might surpass the average accuracy of the best-performing individual member within the team, particularly for larger group sizes. Epoch Timing: We hypothesised that while the underlying neural signals and individual SVM performance might differ between the anticipatory ('ReticleOn') and peri-/post-response ('ButtonPress') epochs, the overall pattern of cBCI-driven team performance enhancement (relative to baselines and across workload conditions) would be comparable between the two epoch types. Exploratory - Individual Differences: We exploratorily hypothesised that individual differences in trait impulsivity (measured by BIS-11) and risk-taking propensity (measured by BART) might be associated with individual performance metrics (accuracy, RT, confidence) within the VR drone task. Declarations Ethics approval and consent to participate The experimental protocol was approved by the UK Ministry of Defence Research Ethics Committee (MoDREC), Application Number: 2309/MODREC/24, Reference: RQ0000037929. All procedures were conducted in accordance with the ethical standards outlined in the Declaration of Helsinki. Written informed consent was obtained from all individual participants included in the study. Data Availability Statement The datasets generated and/or analysed during the current study are not publicly available due to restrictions imposed by the funding body (Defence Science and Technology Laboratory - Dstl). However, data are available from the corresponding author (CB) on reasonable request and subject to a data sharing agreement, if appropriate and in accordance with Dstl policy. Competing interests The authors declare that they have no competing interests. Funding This research was funded by the Defence Science and Technology Laboratory (DSTL) via RQ0000037929.The funders contributed to the conceptualisation of the broader project aims. The funders did not have a direct role in the specific design of this study, data collection, detailed analysis, interpretation of data from this specific study, or in the writing of this manuscript beyond the contributions of the DSTL-affiliated co-author (T.R.) as described in the Author Contributions section. Authors' contributions C.B.: Conceptualisation, Methodology, Software, Validation, Formal Analysis, Investigation, Data Curation, Writing – Original Draft, Writing – Review & Editing. S.H.: Conceptualisation, Methodology, Formal Analysis, Investigation, Writing – Review & Editing. S.F.: Conceptualisation, Methodology, Software, Formal Analysis, Investigation, Writing – Review & Editing. A.N.: Conceptualisation, Methodology, Writing – Review & Editing. R.P.: Conceptualisation, Methodology, Writing – Review & Editing. C.C.: Conceptualisation, Methodology, Writing – Review & Editing. T.R.: Writing – Review & Editing. References Salas, E., Reyes, D. L. & McDaniel, S. H. The science of teamwork: Progress, reflections, and the road ahead. Am. Psychol. 73 , 593–600 (2018). Heinrich, M. K., Wahby, M., Dorigo, M. & Hamann, H. Swarm Robotics. (2022) doi:10.7551/mitpress/13780.003.0009. Siean, A.-I., Gradinaru, B.-C., Gherman, O.-I., Danubianu, M. & Milici, L.-D. Opportunities and Challenges in Human-Swarm Interaction: Systematic Review and Research Implications. Int. J. Adv. Comput. Sci. Appl. 14 , (2023). Hung, W. Team-based complex problem solving: a collective cognition perspective. Educ. Technol. Res. Dev. 61 , 365–384 (2013). Canonico, L. B., Flathmann, C. & McNeese, N. Collectively Intelligent Teams: Integrating Team Cognition, Collective Intelligence, and AI for Future Teaming. Proc. Hum. Factors Ergon. Soc. Annu. Meet. 63 , 1466–1470 (2019). Poli, R. Super-Human and Super-AI Cognitive Augmentation of Human and Human-AI Teams Assisted by Brain Computer Interfaces. in Proceedings of the Genetic and Evolutionary Computation Conference 3 (Association for Computing Machinery, New York, NY, USA, 2023). doi:10.1145/3583131.3603554. van Ments, L. & Treur, J. Reflections on dynamics, adaptation and control: A cognitive architecture for mental models. Cogn. Syst. Res. 70 , 1–9 (2021). Xu, W., Chen, X., Dong, Y. & Chiclana, F. Impact of Decision Rules and Non-cooperative Behaviors on Minimum Consensus Cost in Group Decision Making. Group Decis. Negot. 30 , 1239–1260 (2021). Baltes, B. B., Dickson, M. W., Sherman, M. P., Bauer, C. C. & LaGanke, J. S. Computer-Mediated Communication and Group Decision Making: A Meta-Analysis. Organ. Behav. Hum. Decis. Process. 87 , 156–179 (2002). Sheridan, T. B. HCI in supervisory control: Twelve dilemmas. in Human error and system design and management (eds. Elzer, P. F., Kluwe, R. H. & Boussoffara, B.) 1–12 (Springer, London, 2000). doi:10.1007/BFb0110450. Parasuraman, R., Sheridan, T. B. & Wickens, C. D. A model for types and levels of human interaction with automation. IEEE Trans. Syst. Man Cybern. - Part Syst. Hum. 30 , 286–297 (2000). Wang, D. et al. From Human-Human Collaboration to Human-AI Collaboration: Designing AI Systems That Can Work Together with People. in Extended Abstracts of the 2020 CHI Conference on Human Factors in Computing Systems 1–6 (Association for Computing Machinery, New York, NY, USA, 2020). doi:10.1145/3334480.3381069. Kaklauskas, A. Intelligent Decision Support Systems. in Biometric and Intelligent Decision Making Support (ed. Kaklauskas, A.) 31–85 (Springer International Publishing, Cham, 2015). doi:10.1007/978-3-319-13659-2_2. Phillips-Wren, G. Intelligent Decision Support Systems. in Multicriteria Decision Aid and Artificial Intelligence 25–44 (John Wiley & Sons, Ltd, 2013). doi:10.1002/9781118522516.ch2. Smith, P. J., Geddes, N. D. & Beatty, R. HUMAN-CENTERED DESIGN OF DECISION-SUPPORT SYSTEMS. in Human-Computer Interaction (CRC Press, 2009). Herse, S., Vitale ,Jonathan & and Williams, M.-A. Using Agent Features to Influence User Trust, Decision Making and Task Outcome during Human-Agent Collaboration. Int. J. Human–Computer Interact. 39 , 1740–1761 (2023). Ma, S. et al. “Are You Really Sure?” Understanding the Effects of Human Self-Confidence Calibration in AI-Assisted Decision Making. in Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems 1–20 (Association for Computing Machinery, New York, NY, USA, 2024). doi:10.1145/3613904.3642671. Arshad, S. Z., Zhou, J., Bridon, C., Chen, F. & Wang, Y. Investigating User Confidence for Uncertainty Presentation in Predictive Decision Making. in Proceedings of the Annual Meeting of the Australian Special Interest Group for Computer Human Interaction 352–360 (Association for Computing Machinery, New York, NY, USA, 2015). doi:10.1145/2838739.2838753. Proaps, A. & Bliss, J. Team Performance as a Function of Task Difficulty in a Computer Game. Proc. Hum. Factors Ergon. Soc. Annu. Meet. 54 , 2413–2416 (2010). Caldwell, B. S. & Wang, E. Delays and User Performance in Human-Computer-Network Interaction Tasks. Hum. Factors 51 , 813–830 (2009). Longo, L. Experienced mental workload, perception of usability, their interaction and impact on task performance. PLOS ONE 13 , e0199661 (2018). MITTA, D. A. & and PACKEBUSH, S. J. Improving interface quality: an investigation of human-computer interaction task learning. Ergonomics 38 , 1307–1325 (1995). McFarlane, D. C. Coordinating the Interruption of People in Human- Computer Interaction. Fritz, T., Begel, A., Müller, S. C., Yigit-Elliott, S. & Züger, M. Using psycho-physiological measures to assess task difficulty in software development. in Proceedings of the 36th International Conference on Software Engineering 402–413 (Association for Computing Machinery, New York, NY, USA, 2014). doi:10.1145/2568225.2568266. Neerincx, M. A., van der Waa, J., Kaptein, F. & van Diggelen, J. Using Perceptual and Cognitive Explanations for Enhanced Human-Agent Team Performance. in Engineering Psychology and Cognitive Ergonomics (ed. Harris, D.) 204–214 (Springer International Publishing, Cham, 2018). doi:10.1007/978-3-319-91122-9_18. Dirican, A. C. & Göktürk, M. Psychophysiological measures of human cognitive states applied in human computer interaction. Procedia Comput. Sci. 3 , 1361–1367 (2011). Zhou, Y. et al. Cognitive Workload Recognition Using EEG Signals and Machine Learning: A Review. IEEE Trans. Cogn. Dev. Syst. 14 , 799–818 (2022). Hassan, J., Reza, Md. S., Ahmed, S. U., Anik, N. H. & Khan, Md. O. EEG workload estimation and classification: a systematic review. J. Neural Eng. (2024) doi:10.1088/1741-2552/ad705e. Poli, R., Valeriani, D. & Cinel, C. Collaborative Brain-Computer Interface for Aiding Decision-Making. PLoS ONE 9 , e102693 (2014). Valeriani, D., Poli, R. & Cinel, C. A collaborative Brain-Computer Interface for improving group detection of visual targets in complex natural environments. in 2015 7th International IEEE/EMBS Conference on Neural Engineering (NER) vols 2015-J 25–28 (IEEE, 2015). Valeriani, D., Poli, R. & Cinel, C. Enhancement of Group Perception via a Collaborative Brain–Computer Interface. IEEE Trans. Biomed. Eng. 64 , 1238–1248 (2017). Bhattacharyya, S., Valeriani, D., Cinel, C., Citi, L. & Poli, R. Anytime collaborative brain–computer interfaces for enhancing perceptual group decision-making. Sci. Rep. 11 , 17008 (2021). Valeriani, D., O’Flynn, L. C., Worthley, A., Sichani, A. H. & Simonyan, K. Multimodal collaborative brain-computer interfaces aid human-machine team decision-making in a pandemic scenario. J. Neural Eng. 19 , 056036 (2022). Salvatore, C., Valeriani, D., Piccialli, V. & Bianchi, L. Optimized Collaborative Brain-Computer Interfaces for Enhancing Face Recognition. IEEE Trans. Neural Syst. Rehabil. Eng. 30 , 1223–1232 (2022). Zuo, Z., Liu, C., Han, Q.-L. & Song, J. Unmanned Aerial Vehicles: Control Methods and Future Challenges. IEEECAA J. Autom. Sin. 9 , 601–614 (2022). Hamada, D., Nakayama, M. & Saiki, J. Wisdom of crowds and collective decision-making in a survival situation with complex information integration. Cogn. Res. Princ. Implic. 5 , 48 (2020). Thomas, E., Dyson, M. & Clerc, M. An analysis of performance evaluation for motor-imagery based BCI. J. Neural Eng. 10 , 031001 (2013). Dorneich, M. C., Mathan, S., Ververs, P. M. & Whitlow, S. D. An Evaluation of Real-Time Cognitive State Classification in a Harsh Operational Environment. Proc. Hum. Factors Ergon. Soc. Annu. Meet. 51 , 146–150 (2007). Bagheri, M. & Power, S. D. Simultaneous Classification of Both Mental Workload and Stress Level Suitable for an Online Passive Brain–Computer Interface. Sensors 22 , 535 (2022). Cinel, C., Fernandez-Vargas, J., Tremmel, C., Citi, L. & Poli, R. Enhancing performance with multisensory cues in a realistic target discrimination task. PLOS ONE 17 , e0272320 (2022). Kelley, B. Do You See It? Aspects of Effective Visual Cueing. in 2024 IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct) 658–659 (2024). doi:10.1109/ISMAR-Adjunct64951.2024.00198. Kelley, B. et al. Addressing Human Factors Related to Artificial Intelligence Integrated Visual Cueing. in 2024 IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct) 209–213 (2024). doi:10.1109/ISMAR-Adjunct64951.2024.00051. Kosch, T. et al. A Survey on Measuring Cognitive Workload in Human-Computer Interaction. ACM Comput Surv 55 , 283:1-283:39 (2023). Woolley, A. W., Chabris, C. F., Pentland, A., Hashmi, N. & Malone, T. W. Evidence for a Collective Intelligence Factor in the Performance of Human Groups. Science 330 , 686–688 (2010). Trumbo, M. et al. Individual Differences and the Science of Human Performance. in Foundations of Augmented Cognition. Directing the Future of Adaptive Systems (eds. Schmorrow, D. D. & Fidopiastis, C. M.) 46–54 (Springer, Berlin, Heidelberg, 2011). doi:10.1007/978-3-642-21852-1_6. Agrigoroaie, R. & Tapus, A. Cognitive Performance and Physiological Response Analysis. Int. J. Soc. Robot. 12 , 47–64 (2020). Chowdhury, M. R. Development of Multiple Class Human-Computer Interaction System using Machine Learning Algorithm for Eyeball Movement. (Khulna University of Engineering & Technology (KUET), Khulna, Bangladesht, 2019). Quitadamo, L. R. et al. Support vector machines to detect physiological patterns for EEG and EMG-based human–computer interaction: a review. J. Neural Eng. 14 , 011001 (2017). Vidaurre, C., Sannelli, C., Müller, K.-R. & Blankertz, B. Machine-Learning-Based Coadaptive Calibration for Brain-Computer Interfaces. Neural Comput. 23 , 791–816 (2011). Hauptman, A. I., Schelble, B. G., McNeese, N. J. & Madathil, K. C. Adapt and overcome: Perceptions of adaptive autonomous agents for human-AI teaming. Comput. Hum. Behav. 138 , 107451 (2023). Zhao, M., Simmons, R. & Admoni, H. The Role of Adaptation in Collective Human–AI Teaming. Top. Cogn. Sci. 17 , 291–323 (2025). Zheng, N. et al. Hybrid-augmented intelligence: collaboration and cognition. Front. Inf. Technol. Electron. Eng. 18 , 153–179 (2017). Cui, H. & Yasseri, T. AI-enhanced collective intelligence. Patterns 5 , (2024). Barratt, E. S. Impulsiveness Defined Within a Systems Model of Personality. in Advances in Personality Assessment (Routledge, 1985). Kothe, C. et al. The Lab Streaming Layer for Synchronized Multimodal Recording. bioRxiv 2024.02.13.580071 (2024) doi:10.1101/2024.02.13.580071. Gramfort, A. et al. MEG and EEG data analysis with MNE-python. Front. Neurosci. 7 , 267 (2013). Pedregosa, F. et al. Scikit-learn: Machine Learning in Python. J. Mach. Learn. Res. 12 , 2825–2830 (2011). van Rossum, G. & de Boer, J. Interactively testing remote servers using the Python programming language. CWI Q. 4 , 283–304 (1991). Virtanen, P. et al. SciPy 1.0: fundamental algorithms for scientific computing in Python. Nat. Methods 17 , 261–272 (2020). McKinney, W. Data Structures for Statistical Computing in Python. in 56–61 (Austin, Texas, 2010). doi:10.25080/Majora-92bf1922-00a. Harris, C. R. et al. Array programming with NumPy. Nature 585 , 357–362 (2020). Hunter, J. D. Matplotlib: A 2D Graphics Environment. Comput. Sci. Eng. 9 , 90–95 (2007). Waskom, M. L. seaborn: statistical data visualization. J. Open Source Softw. 6 , 3021 (2021). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6985673","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":487999756,"identity":"d79380b1-03c7-4d52-a2b7-81e69df92061","order_by":0,"name":"Christopher Baker","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAxUlEQVRIiWNgGAWjYFACHgbGBgYGORADDCSI1WJMupbEBqK18DfwHvw4o+Jeet+N3AMMP2oYEmc2ENAicYAvWXLDmeLcmTfyEhh7jjEkziZkiwEDj4Hkw7aE3A03cgwYeBsYEucRocX458N/CekGQC2Mf4nUYia5sSEhAaSFGWQLQYdJHOZLs5xxLMFw5pl3CYdljkkYE/Q+f3vv4Zs9NQnyfMdzDz58U2MjO+MAIWuYYYwDYERERCIAQcNHwSgYBaNgxAIA/MY/y1s30CAAAAAASUVORK5CYII=","orcid":"","institution":"Queen's University Belfast","correspondingAuthor":true,"prefix":"","firstName":"Christopher","middleName":"","lastName":"Baker","suffix":""},{"id":487999757,"identity":"3a83793d-1628-4d5b-bd84-eb3c21e75386","order_by":1,"name":"Stephen Hinton","email":"","orcid":"","institution":"Liverpool John Moores University","correspondingAuthor":false,"prefix":"","firstName":"Stephen","middleName":"","lastName":"Hinton","suffix":""},{"id":487999758,"identity":"31a9d02c-9ff7-4458-bc38-45260761fde4","order_by":2,"name":"Akashdeep Nijjar","email":"","orcid":"","institution":"University of Essex","correspondingAuthor":false,"prefix":"","firstName":"Akashdeep","middleName":"","lastName":"Nijjar","suffix":""},{"id":487999759,"identity":"0b87178a-209f-4de4-8e0d-473415ab2395","order_by":3,"name":"Riccardo Poli","email":"","orcid":"","institution":"University of Essex","correspondingAuthor":false,"prefix":"","firstName":"Riccardo","middleName":"","lastName":"Poli","suffix":""},{"id":487999760,"identity":"119c3c29-b8ff-49e4-8620-e32941e7a412","order_by":4,"name":"Caterina Cinel","email":"","orcid":"","institution":"University of Essex","correspondingAuthor":false,"prefix":"","firstName":"Caterina","middleName":"","lastName":"Cinel","suffix":""},{"id":487999761,"identity":"cd8129bb-f1c0-4877-a2df-91b541ff810b","order_by":5,"name":"Thomas Reed","email":"","orcid":"","institution":"Defence Science and Technology Laboratory","correspondingAuthor":false,"prefix":"","firstName":"Thomas","middleName":"","lastName":"Reed","suffix":""},{"id":487999762,"identity":"c74aa572-d669-4f73-b762-31b5ebe1dd54","order_by":6,"name":"Stephen Fairclough","email":"","orcid":"","institution":"Liverpool John Moores University","correspondingAuthor":false,"prefix":"","firstName":"Stephen","middleName":"","lastName":"Fairclough","suffix":""}],"badges":[],"createdAt":"2025-06-26 17:38:13","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6985673/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6985673/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":87382631,"identity":"8922e75a-4c6e-4dc4-a5ca-20d4c203fca5","added_by":"auto","created_at":"2025-07-23 08:39:09","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":71593,"visible":true,"origin":"","legend":"\u003cp\u003eSuperorganism Architecture\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/981893bb5d108359770d28bf.png"},{"id":87386143,"identity":"99c17d19-6b23-4eb0-8f3e-8466725f711b","added_by":"auto","created_at":"2025-07-23 08:55:10","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":62948,"visible":true,"origin":"","legend":"\u003cp\u003eMean individual decision accuracy (ReticleOn epochs) by Workload condition\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/ac8948add1977802c5c7078e.png"},{"id":87382670,"identity":"f1e1caab-dcb7-43dd-97f6-174d498184c7","added_by":"auto","created_at":"2025-07-23 08:39:12","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":66648,"visible":true,"origin":"","legend":"\u003cp\u003eMean individual response times (ReticleOn epochs) by Workload condition\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/2f571ba9e0e275da627170c4.png"},{"id":87382545,"identity":"bd193937-4bda-4390-87e0-0d529fb0a41d","added_by":"auto","created_at":"2025-07-23 08:39:04","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":68187,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of individual response times (ReticleOn epochs) by decision outcome\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/5a1c2210378f6339592f34f6.png"},{"id":87384120,"identity":"02a0820d-6dac-4d83-b8ef-3648af355cad","added_by":"auto","created_at":"2025-07-23 08:47:06","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":64552,"visible":true,"origin":"","legend":"\u003cp\u003eMean individual subjective confidence ratings (ReticleOn epochs) by Workload condition\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/184bbf43503ad0d4401784af.png"},{"id":87382732,"identity":"b4fbc317-e568-4d41-b78e-fd3887cbd637","added_by":"auto","created_at":"2025-07-23 08:39:14","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":74476,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of individual subjective confidence ratings (ReticleOn epochs) by decision outcome\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/039c8ef8a4814c7fcc97b569.png"},{"id":87382584,"identity":"58df5754-ba05-42da-b1d2-12ed9cf82259","added_by":"auto","created_at":"2025-07-23 08:39:07","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":485570,"visible":true,"origin":"","legend":"\u003cp\u003eIndividual SVM classifier accuracy for the ReticleOn, High Workload condition\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/4b3f5989103dd48c7432537e.png"},{"id":87382591,"identity":"44030e7b-0d0d-4080-83a2-2533dd419703","added_by":"auto","created_at":"2025-07-23 08:39:07","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":372123,"visible":true,"origin":"","legend":"\u003cp\u003eTeam decision accuracy by group size for the ReticleOn, High Workload condition\u003c/p\u003e","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/a04c5fa64a1fdd50dbfd542d.png"},{"id":87382607,"identity":"59e76920-27a8-4d92-b648-ef4975fcff5e","added_by":"auto","created_at":"2025-07-23 08:39:08","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":120354,"visible":true,"origin":"","legend":"\u003cp\u003eIndividual SVM classifier accuracy for the ReticleOn, Low Workload condition\u003c/p\u003e","description":"","filename":"floatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/d3f74909e49f6316ef8646e1.png"},{"id":87384133,"identity":"0a773312-dfb0-4715-a5d0-289f727730b0","added_by":"auto","created_at":"2025-07-23 08:47:09","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":316388,"visible":true,"origin":"","legend":"\u003cp\u003eTeam decision accuracy by group size for the ReticleOn, Low Workload condition\u003c/p\u003e","description":"","filename":"floatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/b2af27c2ade096e4762ef117.png"},{"id":87382520,"identity":"839ee0d6-0392-4142-9226-2bc153b1932c","added_by":"auto","created_at":"2025-07-23 08:39:02","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":115896,"visible":true,"origin":"","legend":"\u003cp\u003eIndividual SVM classifier accuracy for the ButtonPress, High Workload condition\u003c/p\u003e","description":"","filename":"floatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/58328417fecef63e73751821.png"},{"id":87382665,"identity":"36a7e361-052c-4380-a71b-4e05ee8d7b56","added_by":"auto","created_at":"2025-07-23 08:39:11","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":373337,"visible":true,"origin":"","legend":"\u003cp\u003eTeam decision accuracy by group size for the ButtonPress, High Workload condition\u003c/p\u003e","description":"","filename":"floatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/8ae91505908fe983798003c2.png"},{"id":87382518,"identity":"91520873-3398-4fba-90f5-0206f6d683d2","added_by":"auto","created_at":"2025-07-23 08:39:02","extension":"png","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":115508,"visible":true,"origin":"","legend":"\u003cp\u003eIndividual SVM classifier accuracy for the ButtonPress, Low Workload condition\u003c/p\u003e","description":"","filename":"floatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/3243d7cfe22cf75b6bb5b188.png"},{"id":87382657,"identity":"1ad6a0ac-3c9b-421d-8a08-5cdd1f58770b","added_by":"auto","created_at":"2025-07-23 08:39:11","extension":"png","order_by":14,"title":"Figure 14","display":"","copyAsset":false,"role":"figure","size":328774,"visible":true,"origin":"","legend":"\u003cp\u003eTeam decision accuracy by group size for the ButtonPress, Low Workload condition\u003c/p\u003e","description":"","filename":"floatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/ad96487b99219ffd5270b510.png"},{"id":87382637,"identity":"ac0449cf-3a99-4db7-9590-dc87d8980cd9","added_by":"auto","created_at":"2025-07-23 08:39:10","extension":"png","order_by":15,"title":"Figure 15","display":"","copyAsset":false,"role":"figure","size":617797,"visible":true,"origin":"","legend":"\u003cp\u003eView of the virtual landscape from the simulated drone\u003c/p\u003e","description":"","filename":"floatimage15.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/55e6e4458e281ef17ef3e7b7.png"},{"id":87382614,"identity":"cdfa08be-c516-40e4-8c03-c48438c2e29c","added_by":"auto","created_at":"2025-07-23 08:39:09","extension":"png","order_by":16,"title":"Figure 16","display":"","copyAsset":false,"role":"figure","size":792688,"visible":true,"origin":"","legend":"\u003cp\u003eExample of a Non-target stimulus\u003c/p\u003e","description":"","filename":"floatimage16.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/a96a480562690612cae36ada.png"},{"id":87384111,"identity":"12b07295-682f-4535-94d5-dce58bb80222","added_by":"auto","created_at":"2025-07-23 08:47:02","extension":"png","order_by":17,"title":"Figure 17","display":"","copyAsset":false,"role":"figure","size":749359,"visible":true,"origin":"","legend":"\u003cp\u003eExample of a Target stimulus\u003c/p\u003e","description":"","filename":"floatimage17.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/b38c26832318a271c44eb326.png"},{"id":87382554,"identity":"74e9305a-6337-4aee-942a-ba7df5e14c5e","added_by":"auto","created_at":"2025-07-23 08:39:05","extension":"png","order_by":18,"title":"Figure 18","display":"","copyAsset":false,"role":"figure","size":365397,"visible":true,"origin":"","legend":"\u003cp\u003eInitial appearance of the targeting reticle\u003c/p\u003e","description":"","filename":"floatimage18.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/a3214d089d5681c34a512e62.png"},{"id":87384118,"identity":"9c3c4912-2e8b-4bff-bfcb-8295c14ef652","added_by":"auto","created_at":"2025-07-23 08:47:06","extension":"png","order_by":19,"title":"Figure 19","display":"","copyAsset":false,"role":"figure","size":453176,"visible":true,"origin":"","legend":"\u003cp\u003eTargeting reticle changing color as a response deadline warnin\u003c/p\u003e","description":"","filename":"floatimage19.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/cb6ddb1611bd8b97611983ec.png"},{"id":87386916,"identity":"8024a7d2-cd25-4189-b885-15b75993efe2","added_by":"auto","created_at":"2025-07-23 09:03:12","extension":"png","order_by":20,"title":"Figure 20","display":"","copyAsset":false,"role":"figure","size":687175,"visible":true,"origin":"","legend":"\u003cp\u003eUser interface element for subjective confidence rating\u003c/p\u003e","description":"","filename":"floatimage20.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/94e336499d692db2038b60b2.png"},{"id":87382555,"identity":"caf670a2-dbb6-4f36-adb3-01e515ae14ef","added_by":"auto","created_at":"2025-07-23 08:39:05","extension":"png","order_by":21,"title":"Figure 21","display":"","copyAsset":false,"role":"figure","size":323006,"visible":true,"origin":"","legend":"\u003cp\u003eView from the simulated drone in the High Workload (low light) condition\u003c/p\u003e","description":"","filename":"floatimage21.png","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/cc22807b79fc1d087973a810.png"},{"id":98774757,"identity":"eb615a4f-798d-4d7a-bdd4-d8864b0f4530","added_by":"auto","created_at":"2025-12-22 12:13:51","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":7611582,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6985673/v1/5b1db4ab-f11c-44e8-b5cd-b1c0ddf2d9aa.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Integrating EEG-SVM Confidence and RT via cBCI Enhances Team Decisions in a VR Drone Task","fulltext":[{"header":"Introduction","content":"\u003cp\u003eHuman beings have traditionally worked together in teams, harnessing their abilities as a collective to improve performance beyond the limits of the individual\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. When people work in decision-making teams, inputs from individuals can be expressed, accumulated and condensed into a single, collective output from the team. This aggregation of individual decisions represents a form of collective cognition and can be expressed using several metaphors, such as swarm intelligence\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e, a \u0026lsquo;collective brain\u0026rsquo;\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e,\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e or a superorganism\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e,\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eThis process of collating multiple judgements into a single collective decision can be formulated in several ways\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e,\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e. Where there is a finite number of options, such as a binary decision, we can use a majority vote process where each individual output functions as a single \u0026lsquo;vote\u0026rsquo; that are counted to reach a team decision, i.e., which option accumulates the highest number of \u0026lsquo;votes\u0026rsquo;. However, this egalitarian approach ignores important differences between individual team members, for instance, some may be more experienced than others, others may have greater levels of skill and a stronger track record of performance. If we wish to reflect these differences in the collective decision, a process of weighting individual inputs can be applied alongside accumulation. As a secondary dimension, the aggregation of individual inputs also contains information about the degree of agreement within the collective. Some decisions may be unanimous whereas others will reflect disagreement within the team.\u003c/p\u003e\u003cp\u003eThe need for individual responses to be aggregated, weighed and assessed for coherence requires the existence of a supervisory agent\u003csup\u003e\u003cspan additionalcitationids=\"CR11\" citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. This technological entity exists at a superordinate level to the team. The supervisory agent is capable of monitoring individual decisions, which are used as inputs to generate a collective decision and produce metadata about team decisions as collateral information, i.e., level of agreement within the team. The process of aggregating and weighing individual responses is defined in this superordinate, supervisory level. The purpose of the supervisory system is to improve the quality of team decision-making by applying dynamic adjustments to the process of aggregation\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e,\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. However, guaranteeing that the addition of the supervisory agent will increase the number of \u0026ldquo;good\u0026rdquo; decisions is challenging\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e,\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e, partly because weightings can interact in unpredictable ways but also because the supervisory system is often not privy to the ground truth of the decision, if one exists.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe supervisory agent can use several different metrics when assessing the quality of individual decisions. It can require every member of the team to explicitly provide a self-reported assessment of each decision. For example, each team member could rate their level of confidence after a decision has been made\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e,\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. However, the repeated need to self-report, especially when decisions are made in quick succession, can be burdensome for the individual. Other metrics are implicit and do not require any overt response from the team member. Measures of behaviour, such as time to decide, can be meaningful indicators of decision quality, but their interpretation depends on the context of the decision-making task\u003csup\u003e\u003cspan additionalcitationids=\"CR20\" citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. When tasks are complex, multifaceted and dependent on dynamic factors, a low decision time may be indicative of a hasty, poorly-planned response. In the case of simple decisions, a longer than average decision time can be interpreted as uncertainty on the part of the team member\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e,\u003cspan additionalcitationids=\"CR23 CR24\" citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eNeurophysiology (and psychophysiology) can also serve as implicit measures of decision-making quality from the individual\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. These metrics can quantify the psychological state of the individual, e.g., to assess level of mental workload, fatigue; they can also be used to assess the behavioural intention of the team member\u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e,\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e. The latter is called a collaborative Brain-Computer Interface (cBCI)\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e, which works via two stages of processing: (1) monitoring neurophysiology (EEG) during the decision-making process for each team member, and (2) using neurophysiological metrics (alongside other behavioural metrics) to \u0026lsquo;weigh\u0026rsquo; the contribution of each individual decision to the collective decision of a team as a whole. In other words, a cBCI permits a process of assessment to be applied to each individual decision, which determines the weight of each individual decision as a contribution to a collective team decision.\u003c/p\u003e\u003cp\u003eCollaborative BCI have been developed to improve team performance on a range of visual detection tasks. Early work\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e,\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e explored this concept using abstract visual target detection tasks and reported that groups generally outperformed individuals and that weighing responses by combining BCI with subjective ratings of decision confidence was an effective way to improve the performance of the team. This work was replicated using more naturalistic targets presented as still images\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e and extended by the inclusion of other behavioural markers such as eye tracking\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e. This work was developed to explore cBCI with more realistic visual search tasks that approximated real-world scenarios and were dynamic in nature\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e; this paper demonstrated that cBCI outperformed other collective models, such as majority voting by integrating neurophysiological features from EEG with behavioural markers such as response time and reported confidence levels. In recent years, this work has been replicated using more complex decision-making tasks\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e and face recognition\u003csup\u003e\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e. Others have developed cBCI systems that are capable of mutual learning within the team and responding to video feeds from Unmanned Aerial Vehicles\u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e (UAV).\u003c/p\u003e\u003cp\u003eThe level of demand provides important context for cBCI performance. High workload tasks that are challenging are likely to lead to more variable human performance within a team compared to tasks that are less demanding. Furthermore, the benefits of accumulated performance may be more substantial for high workload tasks compared to the low workload tasks, especially when participants are performing decision-making tasks using partial or degraded visual information.\u003c/p\u003e\u003cp\u003eThe current paper will describe a cBCI system designed to assess the quality of individual decisions to be used as an input to the performance of human teams of different sizes. This system will amalgamate data from EEG and human performance to provide a weighting of each individual decision in the context of a team so that \u0026lsquo;good decisions\u0026rsquo; make a larger contribution to the team decision compared to \u0026lsquo;bad decisions\u0026rsquo;, thus conferring an advantage on the performance of the team. The performance of this cBCI will be assessed in the context of high and low workload using a VR-based simulation of UAV target detection task.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cb\u003eIndividual Performance\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo verify the intended effects of the experimental manipulation, individual behavioral performance was analyzed by comparing outcomes between the Low and High Workload conditions. Key metrics including accuracy, response time (RT), and subjective confidence ratings were examined.\u003c/p\u003e\u003cp\u003e\u003cb\u003eAccuracy\u003c/b\u003e\u003c/p\u003e\u003cp\u003eAnalysis of decision accuracy revealed a significant impact of workload. Paired t-tests on subject-level mean accuracies indicated that participants performed significantly worse under High Workload compared to Low Workload, both when considering ReticleOn locked data (t(16) = -9.455, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001) and ButtonPress locked data (t(16) = -8.581, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001). Descriptively, mean accuracy assessed from ReticleOn epochs decreased from 93.2% (SD\u0026thinsp;=\u0026thinsp;7.9%) in the Low Workload condition to 76.9% (SD\u0026thinsp;=\u0026thinsp;11.8%) in the High Workload condition (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eResponse Times\u003c/b\u003e\u003c/p\u003e\u003cp\u003eResponse times (calculated for non-miss trials) were also significantly modulated by the workload manipulation. Participants exhibited significantly slower response times during High Workload trials compared to Low Workload trials. This was confirmed by paired t-tests on subject-level mean RTs for ReticleOn locked data (High: M\u0026thinsp;=\u0026thinsp;1.460s, SD\u0026thinsp;=\u0026thinsp;0.312s; Low: M\u0026thinsp;=\u0026thinsp;0.862s, SD\u0026thinsp;=\u0026thinsp;0.355s; t(16)\u0026thinsp;=\u0026thinsp;10.395, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001) (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). Furthermore, response times for incorrect decisions were generally slower than for correct decisions, as visualized by the distributions in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eSubjective Confidence\u003c/b\u003e\u003c/p\u003e\u003cp\u003eFurthermore, subjective confidence ratings reflected the change in task demand. Participants reported significantly lower confidence in their decisions (calculated for non-miss trials) during the High Workload blocks compared to the Low Workload blocks. Paired t-tests confirmed this effect for both ReticleOn locked data (High: M\u0026thinsp;=\u0026thinsp;52.54, SD\u0026thinsp;=\u0026thinsp;25.54; Low: M\u0026thinsp;=\u0026thinsp;78.71, SD\u0026thinsp;=\u0026thinsp;19.59; t(16) = -8.949, p\u0026thinsp;\u0026lt;\u0026thinsp;0.0001) (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). Participants also typically reported higher confidence for correct decisions compared to incorrect decisions, as illustrated by the distributions in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eTeam Performance Analysis\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe performance of simulated teams, incorporating these BCI-SVM outputs alongside behavioural data, was then evaluated under different aggregation methods across varying group sizes (2, 4, 6, and 8).\u003c/p\u003e\u003cp\u003eWhen interpreting the following team-based results, it is crucial to acknowledge that improvements in group performance over that of an average individual can arise from inherent mathematical and statistical principles. These include:\u003c/p\u003e\u003cp\u003e1. Error Cancellation (Wisdom of Crowds): Aggregating multiple, even noisy, independent or partially independent estimates (like individual decisions) can lead to a more accurate collective outcome as random errors tend to cancel each other out\u003csup\u003e\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e. This principle underpins basic majority voting.\u003c/p\u003e\u003cp\u003e2. Increased Effective Sample Size: Each team decision, especially in weighted methods, effectively draws upon more data points (e.g., an individual's response, their RT, their BCI state) than a single individual's decision, potentially leading to more robust outcomes.\u003c/p\u003e\u003cp\u003e3. Exploitation of Complementary Error Profiles: This is a particularly important consideration for human-BCI synergy. If human decision-making and BCI classifications make errors on different types of trials or for different underlying reasons (i.e., their error profiles are complementary), then combining them can lead to a more accurate overall decision. In such cases, one system might be reliable and provide a \"good decision\" precisely when the other system is struggling or likely to fail, and vice-versa. A sophisticated integration strategy can capitalize on this complementarity.\u003c/p\u003e\u003cp\u003eWhile the first two effects contribute to the general benefit of teamwork and information integration, a primary objective of this study is to determine if the specific cBCI methodologies employed offer advantages that extend significantly beyond these foundational aggregation benefits, particularly by leveraging the potential for complementary error profiles between human and BCI. To this end, our analyses focus on several key comparisons:\u003c/p\u003e\u003cp\u003e1. Surpassing the Average Individual: As a baseline confirmation of team benefit.\u003c/p\u003e\u003cp\u003e2. Outperforming Simple Aggregation Rules: Comparing advanced cBCI methods against standard majority votes and simple behavioral weighting (e.g., RT-only) to assess the added value of neurophysiological information and more sophisticated integration that might tap into complementary strengths.\u003c/p\u003e\u003cp\u003e3. Challenging the Best Individual: Critically, we evaluate whether cBCI methods can achieve accuracies surpassing the average performance of the best-performing individual member within each simulated team. Such an outcome would more strongly suggest a synergistic effect, where the integrated system creates a decision superior to what even the top individuals typically achieve. This could arise not just from better selection of confident decisions, but also from the system correctly weighting one information source (human or BCI) when the other is less reliable for a given trial.\u003c/p\u003e\u003cp\u003e4. Systematic Evaluation of Components: By examining a range of aggregation methods\u0026mdash;from human-only approaches to those incorporating BCI data with and without filtering\u0026mdash;we aim to elucidate the incremental contributions of different information sources and processing strategies, and how they might interact to exploit differing system strengths\u003csup\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eThe following sections will detail team performance under various conditions, with these interpretive considerations in mind, paying particular attention to evidence for synergistic gains that may stem from the effective combination of human and BCI decision processes.\u003c/p\u003e\u003cp\u003e\u003cb\u003eReticle On High Workload\u003c/b\u003e\u003c/p\u003e\u003cp\u003eFor the high workload reticle on epoch, the individual participant-specific SVM classifiers, which inform several BCI-driven aggregation methods, performed above chance on average (M\u0026thinsp;=\u0026thinsp;0.673, SD\u0026thinsp;=\u0026thinsp;0.072), although performance varied considerably across participants (range: 0.558\u0026ndash;0.823). This indicates that while the BCI component provided useful information, its quality varied (see Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e). Exploratory analysis of SVM feature importance revealed a diverse set of discriminative features, including both time-domain and frequency-domain measures; the top selected features for this condition. Analysis of raw agreement between simple human majority and BCI majority decisions showed that agreement increased steadily with group size, rising from approximately 63.6% for N\u0026thinsp;=\u0026thinsp;2 to 82.4% for N\u0026thinsp;=\u0026thinsp;8, suggesting a degree of concordance but also highlighting that human and BCI decisions were not perfectly redundant.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eUnder High Workload conditions with EEG data epoched to the Reticle On event (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e), various team aggregation methods were compared. The standard Majority Human method (e.g., blue circle line in Fig. Y) achieved accuracies ranging from 81.8% (N\u0026thinsp;=\u0026thinsp;2) to 95.2% (N\u0026thinsp;=\u0026thinsp;8). Behavioral weighting methods included RT Weighted Human (e.g., light blue square line), which yielded accuracies from 83.4% (N\u0026thinsp;=\u0026thinsp;2) to 94.7% (N\u0026thinsp;=\u0026thinsp;8), and Subjective Confidence Weighted Human (e.g., dark blue triangle line), which produced accuracies from 87.0% (N\u0026thinsp;=\u0026thinsp;2) to 97.3% (N\u0026thinsp;=\u0026thinsp;8). The combined human-only method, RT\u0026thinsp;+\u0026thinsp;Subjective Confidence Human, resulted in accuracies from 86.8% (N\u0026thinsp;=\u0026thinsp;2) to 96.8% (N\u0026thinsp;=\u0026thinsp;8). The BCI-only aggregation, SVM Confidence Weighted BCI (e.g., light green square line), produced accuracies in the range of 72.8% (N\u0026thinsp;=\u0026thinsp;2) to 87.6% (N\u0026thinsp;=\u0026thinsp;8). The mixed methods, integrating human and BCI information, generally demonstrated the strongest performance. The Subjective Confidence\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed method (e.g., magenta star line) showed the highest accuracies in this condition, increasing from 89.3% (N\u0026thinsp;=\u0026thinsp;2) to 95.5% (N\u0026thinsp;=\u0026thinsp;4), 97.9% (N\u0026thinsp;=\u0026thinsp;6), and reaching 98.8% for eight-person teams. The comprehensive RT\u0026thinsp;+\u0026thinsp;Subjective Confidence\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed method (e.g., firebrick H line) also achieved high accuracies, from 89.6% (N\u0026thinsp;=\u0026thinsp;2) to 95.3% (N\u0026thinsp;=\u0026thinsp;4), 97.6% (N\u0026thinsp;=\u0026thinsp;6), and 98.6% (N\u0026thinsp;=\u0026thinsp;8). The RT\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed method (e.g., orangered pentagon line) yielded accuracies from 87.7% (N\u0026thinsp;=\u0026thinsp;2) to 97.5% (N\u0026thinsp;=\u0026thinsp;8). Notably, for larger team sizes, the Subjective Confidence\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed method surpassed the average accuracy of the Best Individual Avg (black solid line; N\u0026thinsp;=\u0026thinsp;6: 97.9% vs. 93.3%; N\u0026thinsp;=\u0026thinsp;8: 98.8% vs. 94.2%), suggesting a synergistic benefit. Similar synergistic effects, where team accuracy exceeded that of the best individual member, were also observed for the RT\u0026thinsp;+\u0026thinsp;Subjective Confidence\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed method (N\u0026thinsp;=\u0026thinsp;6: 97.6% vs. 93.3%; N\u0026thinsp;=\u0026thinsp;8: 98.6% vs. 94.2%) and the RT\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed method (N\u0026thinsp;=\u0026thinsp;6: 96.3% vs. 93.3%; N\u0026thinsp;=\u0026thinsp;8: 97.5% vs. 94.2%). The Average Individual Avg (grey dashed line) was consistently around 81.5%.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eReticle On Low Workload\u003c/b\u003e\u003c/p\u003e\u003cp\u003eUnder low workload conditions for the Reticle On epoch, individual SVM classifiers performed with an average accuracy of M\u0026thinsp;=\u0026thinsp;0.710 (SD\u0026thinsp;=\u0026thinsp;0.116), ranging from 0.497 to 0.891 (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e). While generally above chance, one participant's SVM performed below chance. The features most influential for SVM decisions in this low workload condition differed somewhat from the high workload scenario. Features such as Theta_Power_FC2, Theta_Power_F7, and Alpha_Power_Fp2 were among the most frequently selected by the SVMs. In terms of impact, Beta_Power_CP6 and Mean_CP1 exhibited high average importance scores when selected. Raw agreement between simple human majority and BCI majority decisions was higher than under high workload, increasing from approximately 74.1% for N\u0026thinsp;=\u0026thinsp;2 to 90.3% for N\u0026thinsp;=\u0026thinsp;8\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eUnder Low Workload conditions with EEG data epoched to the Reticle On event (Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e), overall team performance was considerably higher than under high workload, with many methods approaching ceiling levels, particularly for larger group sizes. The standard Majority Human method achieved accuracies ranging from 92.4% (N\u0026thinsp;=\u0026thinsp;2) to 99.6% (N\u0026thinsp;=\u0026thinsp;8). Human-only behavioral weighting methods generally performed very well: RT Weighted Human accuracies ranged from 94.7% (N\u0026thinsp;=\u0026thinsp;2) to 99.8% (N\u0026thinsp;=\u0026thinsp;8); Subjective Confidence Weighted Human from 94.5% (N\u0026thinsp;=\u0026thinsp;2) to 99.6% (N\u0026thinsp;=\u0026thinsp;8); and RT\u0026thinsp;+\u0026thinsp;Subjective Confidence Human from 95.0% (N\u0026thinsp;=\u0026thinsp;2) to 99.7% (N\u0026thinsp;=\u0026thinsp;8). The SVM Confidence Weighted BCI method performed lower, with accuracies from 76.5% (N\u0026thinsp;=\u0026thinsp;2) to 90.9% (N\u0026thinsp;=\u0026thinsp;8). The mixed methods integrating human and BCI information also achieved very high accuracies. Specifically, RT\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed yielded accuracies from 95.1% (N\u0026thinsp;=\u0026thinsp;2) to 99.8% (N\u0026thinsp;=\u0026thinsp;8), and RT\u0026thinsp;+\u0026thinsp;Subjective Confidence\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed produced similar results from 95.3% (N\u0026thinsp;=\u0026thinsp;2) to 99.8% (N\u0026thinsp;=\u0026thinsp;8). The Subjective Confidence\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed method also performed strongly, with accuracies from 95.1% (N\u0026thinsp;=\u0026thinsp;2) to 99.8% (N\u0026thinsp;=\u0026thinsp;8). While these advanced aggregation methods, such as RT\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed and RT\u0026thinsp;+\u0026thinsp;Subjective Confidence\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed, slightly outperformed the Average Individual Avg (approximately 94.4%), their advantage over simpler, high-performing methods like RT Weighted Human or even the standard Majority Human vote was minimal in this low workload context. Furthermore, these cBCI methods generally did not surpass the average Best Individual Avg, which itself was extremely high (ranging from 98.0% for N\u0026thinsp;=\u0026thinsp;2 to 99.9% for N\u0026thinsp;=\u0026thinsp;8). The relative gains from the more complex cBCI aggregation strategies were thus substantially smaller and less distinct than those observed under high workload conditions.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eButton Press High Workload\u003c/b\u003e\u003c/p\u003e\u003cp\u003eFor the high workload button press epoch, the individual participant-specific SVM classifiers performed above chance on average (M\u0026thinsp;=\u0026thinsp;0.692, SD\u0026thinsp;=\u0026thinsp;0.101), with accuracies ranging from 0.587 to 0.863 across participants (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e). This average SVM performance was slightly higher than that observed for the ReticleOn epochs under high workload. Exploratory analysis of SVM feature importance for Button Press epochs revealed a distinct set of influential features compared to ReticleOn; the top selected features for this condition were Var_F10, Var_F7, and Theta_Power_Oz were the most frequently selected features (each 17.6%), while Var_F4 showed the highest average importance score (3.660). Analysis of raw agreement between simple human majority and BCI majority decisions showed agreement increasing from approximately 65.8% for N\u0026thinsp;=\u0026thinsp;2 to 86.5% for N\u0026thinsp;=\u0026thinsp;8, rates slightly higher than in the ReticleOn high workload condition.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eUnder High Workload conditions with EEG data epoched to the Button Press event (Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e), a similar pattern of team performance to the ReticleOn high workload analysis emerged. The standard Majority Human method achieved accuracies ranging from 81.8% (N\u0026thinsp;=\u0026thinsp;2) to 94.7% (N\u0026thinsp;=\u0026thinsp;8). Behavioral weighting methods, such as RT Weighted Human and Subjective Confidence Weighted Human, yielded accuracies from 83.4\u0026ndash;94.4% and 86.9\u0026ndash;97.0%, respectively. The combined human-only method, RT\u0026thinsp;+\u0026thinsp;Subjective Confidence Human, resulted in accuracies from 86.7% (N\u0026thinsp;=\u0026thinsp;2) to 96.4% (N\u0026thinsp;=\u0026thinsp;8). The BCI-only aggregation, SVM Confidence Weighted BCI, produced accuracies in the range of 75.3% (N\u0026thinsp;=\u0026thinsp;2) to 91.1% (N\u0026thinsp;=\u0026thinsp;8). The mixed methods integrating human and BCI information again demonstrated strong performance. The Subjective Confidence\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed method showed the highest accuracies, increasing from 89.9% (N\u0026thinsp;=\u0026thinsp;2) to 95.6% (N\u0026thinsp;=\u0026thinsp;4), 97.8% (N\u0026thinsp;=\u0026thinsp;6), and reaching 98.6% for eight-person teams. The RT\u0026thinsp;+\u0026thinsp;Subjective Confidence\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed method also performed robustly, with accuracies from 89.8% (N\u0026thinsp;=\u0026thinsp;2) to 95.4% (N\u0026thinsp;=\u0026thinsp;4), 97.5% (N\u0026thinsp;=\u0026thinsp;6), and 98.4% (N\u0026thinsp;=\u0026thinsp;8). Similarly, the RT\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed method yielded accuracies from 87.8% (N\u0026thinsp;=\u0026thinsp;2) to 97.7% (N\u0026thinsp;=\u0026thinsp;8). These mixed methods consistently outperformed simpler aggregation strategies and, for larger team sizes, surpassed the average accuracy of the Best Individual Avg; for instance, Subjective Confidence\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed (N\u0026thinsp;=\u0026thinsp;6: 97.8% vs. 93.1%; N\u0026thinsp;=\u0026thinsp;8: 98.6% vs. 94.0%), RT\u0026thinsp;+\u0026thinsp;Subjective Confidence\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed (N\u0026thinsp;=\u0026thinsp;6: 97.5% vs. 93.1%; N\u0026thinsp;=\u0026thinsp;8: 98.4% vs. 94.0%), and RT\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed (N\u0026thinsp;=\u0026thinsp;6: 96.5% vs. 93.1%; N\u0026thinsp;=\u0026thinsp;8: 97.7% vs. 94.0%). The Average Individual Avg was consistently around 81.4%.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eButton Press Low Workload\u003c/b\u003e\u003c/p\u003e\u003cp\u003eFor the low workload button press epoch, the individual participant-specific SVM classifiers performed with an average accuracy of M\u0026thinsp;=\u0026thinsp;0.700 (SD\u0026thinsp;=\u0026thinsp;0.078), with accuracies ranging from 0.541 to 0.861 (Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e). The features most influential for SVM decisions under these conditions were, Beta_Power_FC5 and Theta_Power_C4 were among the most frequently selected features (each 17.6%), while Var_O2 (average importance: 0.586) and Var_F10 (average importance: 0.368) showed high average importance scores. Raw agreement between simple human majority and BCI majority decisions was high, increasing from approximately 73.0% for N\u0026thinsp;=\u0026thinsp;2 to 87.5% for N\u0026thinsp;=\u0026thinsp;8.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTeam performance under Low Workload for ButtonPress epochs showed very high accuracies across most methods, consistent with the low task demand. The standard Majority Human method yielded accuracies from 92.3% (N\u0026thinsp;=\u0026thinsp;2) to 99.6% (N\u0026thinsp;=\u0026thinsp;8). Behavioral weighting methods generally performed strongly: RT Weighted Human accuracies ranged from 94.8% (N\u0026thinsp;=\u0026thinsp;2) to 99.7% (N\u0026thinsp;=\u0026thinsp;8); Subjective Confidence Weighted Human from 94.6% (N\u0026thinsp;=\u0026thinsp;2) to 99.6% (N\u0026thinsp;=\u0026thinsp;8); and RT\u0026thinsp;+\u0026thinsp;Subjective Confidence Human from 95.0% (N\u0026thinsp;=\u0026thinsp;2) to 99.7% (N\u0026thinsp;=\u0026thinsp;8). The SVM Confidence Weighted BCI method was less accurate, with values from 75.3% (N\u0026thinsp;=\u0026thinsp;2) to 88.6% (N\u0026thinsp;=\u0026thinsp;8). The mixed methods integrating human and BCI information also achieved very high accuracies. The RT\u0026thinsp;+\u0026thinsp;Subjective Confidence\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed method produced accuracies from 95.4% (N\u0026thinsp;=\u0026thinsp;2) to 99.8% (N\u0026thinsp;=\u0026thinsp;8). Similarly, RT\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed yielded accuracies from 95.2% (N\u0026thinsp;=\u0026thinsp;2) to 99.8% (N\u0026thinsp;=\u0026thinsp;8), and Subjective Confidence\u0026thinsp;+\u0026thinsp;SVM Confidence Mixed achieved accuracies from 95.0% (N\u0026thinsp;=\u0026thinsp;2) to 99.7% (N\u0026thinsp;=\u0026thinsp;8). However, similar to other low workload conditions, the incremental benefit of these complex combined methods, while achieving high absolute accuracies, was marginal compared to simpler high-performing aggregations like RT Weighted Human, or the Best Individual Avg (which ranged from 98.0% for N\u0026thinsp;=\u0026thinsp;2 to 99.8% for N\u0026thinsp;=\u0026thinsp;8). The Average Individual Avg was consistently around 94.5% (Fig.\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e14\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThe current study investigated the potential for a collaborative Brain-Computer Interface (cBCI), integrating EEG-derived decision confidence (SVM_Confidence) with behavioural metrics (RT, subjective confidence), to enhance team perceptual decision-making accuracy in a VR drone task under varying cognitive workload. Our primary finding demonstrated that under High Workload, specific mixed cBCI aggregation methods enabled simulated teams (N=2-8) to significantly surpass the average performance of their best individual member. This synergistic outcome, where the collective effectively transcended individual capabilities through efficient information fusion\u003csup\u003e29,32\u003c/sup\u003e, was critically dependent on high task demand; under Low Workload, where individual performance neared ceiling, the benefits of these advanced integration methods were minimal. This pronounced workload dependency highlights the neuroergonomic utility of cBCIs in challenging operational contexts where human performance variability is high\u003csup\u003e38,39\u003c/sup\u003e.\u003c/p\u003e\n\n\u003cp\u003eIntriguingly, these cBCI benefits under high cognitive load were robust across EEG data epoched to both anticipatory pre-response (\u0026apos;ReticleOn\u0026apos;) and peri-/post-response (\u0026apos;ButtonPress\u0026apos;) events. This suggests that exploitable neural correlates of decision quality and relevant behavioral markers are present and integrable from both time windows, with the ReticleOn findings notably indicating a predictive capacity for team-level decision weighting\u003csup\u003e40\u0026ndash;42\u003c/sup\u003e. Comparing aggregation strategies, mixed methods integrating human-derived information with BCI-derived SVM confidence consistently yielded the highest accuracies and demonstrated synergy. The \u0026apos;Subjective Confidence + SVM Confidence Mixed\u0026apos; method frequently emerged as the top performer, suggesting a potent fusion of self-reported certainty and neurally-derived decision quality. While the comprehensive \u0026apos;RT + Subjective Confidence + SVM Confidence Mixed\u0026apos; method was competitive, it did not consistently outperform the two-component SC+SVM mix in high-demand scenarios. Simpler aggregation methods (e.g., Majority vote, BCI-only) were generally less effective under high load, aligning with broader findings on the benefits of integrating multiple, diverse information streams in cBCI and team decision-making\u003csup\u003e43,44\u003c/sup\u003e.\u003c/p\u003e\n\n\u003cp\u003eExploratory analyses of trait impulsivity and risk-taking (BIS-11, BART) revealed no significant associations with individual performance metrics in this task, suggesting that workload and aggregation method were more dominant factors, or that the task demands were not optimally suited to elicit strong trait-based effects\u003csup\u003e45,46\u003c/sup\u003e.\u003c/p\u003e\n\n\u003cp\u003eThe current study has several limitations. The sample size (N=17) may limit the generalizability of individual difference analyses. The VR drone task, while designed for ecological relevance, is specific, and findings require replication across other domains and team compositions. Our BCI approach used SVM-derived confidence for weighting; future work could explore incorporating overall classifier reliability\u003csup\u003e47\u0026ndash;49\u003c/sup\u003e. Furthermore, team performance was assessed via offline simulations, which, while enabling rigorous control, do not capture real-time team dynamics.\u003c/p\u003e\n\n\u003cp\u003eIn conclusion, this research provides compelling evidence that cBCIs integrating diverse human-derived and EEG-based information can markedly enhance team perceptual decision-making beyond best individual performance, particularly under high cognitive workload. This workload-dependent synergy, achievable using both anticipatory and post-response neural data, underscores the potential of neuroergonomically-informed cBCIs for supporting teams in demanding operational contexts. Future research should focus on replicating these findings with larger samples, investigating adaptive, AI-driven aggregation strategies\u003csup\u003e50,51\u003c/sup\u003e, and crucially, validating these cBCI frameworks in real-time, interactive team settings to further the development of systems that genuinely augment collective human intelligence\u003csup\u003e52,53\u003c/sup\u003e.\u003c/p\u003e"},{"header":"Methods","content":"\u003ch2\u003eParticipants\u003c/h2\u003e\n\u003cp\u003eSeventeen healthy adults (N=17; [10 Female]; Mean age \u0026plusmn; SD = [24.8 \u0026plusmn; 6.0]) participated in the study and were included in the final analysis dataset. Participants were recruited from the university undergraduate and postgraduate community. An initial cohort of 27 individuals was recruited; however, data from ten participants were excluded prior to the main analysis due to criteria established in preliminary quality control checks, including insufficient data quality after EEG preprocessing, or significant deviations in trial sequence alignment during task execution due to technical issues and the need for alignment of trials across all combinations of team group size. All included participants reported normal or corrected-to-normal vision and no history of neurological disorders or particular susceptibility to VR-induced motion sickness. Prior to the experiment, participants provided written informed consent and completed the Barratt Impulsiveness Scale (BIS-11)\u003ca href=\"https://www.zotero.org/google-docs/?HUfHX5\"\u003e\u003csup\u003e54\u003c/sup\u003e\u003c/a\u003e and the Balloon Analogue Risk Task (BART) prior to commencing the study. Participants received monetary compensation for their participation. The experimental protocol was approved by UK MoDREC, App No: 2309/MODREC/24 Ref: RQ0000037929 and all procedures were conducted in accordance with the ethical standards outlined in the Declaration of Helsinki.\u003c/p\u003e\n\u003ch2\u003eStudy Design\u003c/h2\u003e\n\u003cp\u003eThe study employed a within-subject repeated measures design to evaluate the effectiveness of a collaborative Brain-Computer Interface (cBCI) system designed to enhance group decision-making in a dynamic virtual reality (VR) environment. The primary within-subject factor manipulated was cognitive workload, presented at two levels (Low vs. High) across distinct experimental blocks. Electroencephalographic (EEG) data were analyzed time-locked to two different events: the onset of a targeting reticle (\u0026apos;ReticleOn\u0026apos;), representing a pre-stimulus anticipatory period, and the participant\u0026apos;s response (\u0026apos;ButtonPress\u0026apos;), capturing peri- and post-decisional neural activity. Team performance was assessed through offline simulations, evaluating decision accuracy for simulated groups of varying sizes (2, 4, 6, and 8 members) under different decision aggregation algorithms. Individual behavioral metrics, including accuracy, response time (RT), and subjective confidence ratings, served as primary dependent variables at the individual level. Additionally, personality traits related to impulsivity and risk-taking were assessed for exploratory analysis.\u003c/p\u003e\n\u003ch2\u003eVR Drone Task and Procedure\u003c/h2\u003e\n\u003cp\u003eParticipants were seated in the laboratory while they viewed the virtual environment via a Varjo Aero HMD. The simulation, developed in Unreal Engine 5 and rendered on a high-performance PC, depicted the viewpoint of a quadcopter type drone flying over a simulated landscape designed to be semi realistic i.e.textured realistically for a temperate climate and populated with sparse foliage but tailored to not interfere with stimuli acquisition (See Fig. 15). Task stimuli consisted of 3D models appearing on the ground: Non-Targets (See Fig. 16) and Targets (See Fig. 17). Participants completed six 8-minute blocks, with each block containing 50 trials (30 Non-Targets, 20 Targets). Stimuli were evenly spaced 300m apart, but varied in lateral placement (-30m, -15m, 0m, +15m, +30m)] and rotation (0\u0026deg;, 90\u0026deg;, 180\u0026deg;, 270\u0026deg;) .\u003c/p\u003e\n\u003cp\u003eEach trial followed a consistent sequence. Initially, the drone flew to a set virtual height (25m) and advanced across the landscape. As stimuli came into view a targeting reticle appeared over the stimuli (\u0026apos;ReticleOn\u0026apos;) for 2500ms (See Fig. 18). Participants were tasked with discriminating the stimulus type and responding as quickly and accurately as possible by pressing the designated joystick button for Target or Non-Target. Participants were instructed to respond whilst the reticle was on-screen. Any responses made after the reticle disappeared were not counted, and those trials were classed as misses. The reticle initially appeared blue, then changed to red after 2000ms, to warn the participants that they only had 500ms to make a decision (See Fig. 19). Following this primary response (\u0026apos;ButtonPress\u0026apos;), a prompt appeared, requiring participants to rate their confidence in the preceding decision via tilting the joystick to select a value on a 0-100 scale (0 = Not Confident, 100 = Very Confident) (See Fig. 20).\u003c/p\u003e\n\u003cp\u003eCognitive workload was manipulated across different blocks of trials. The Low Workload condition had the virtual light level fixed as approximate for midday average daylight. Conversely, the High Workload condition reduced the light level by 50%\u003ca id=\"_anchor_2\" href=\"#_msocom_2\" language=\"JavaScript\" name=\"_msoanchor_2\"\u003e\u003c/a\u003e\u003ca id=\"_anchor_3\" href=\"#_msocom_3\" language=\"JavaScript\" name=\"_msoanchor_3\"\u003e\u003c/a\u003e, and the solar angle by 9\u0026deg; simulating a night condition (See Fig. 21). Participants first completed a practice block for each workload condition, followed by three blocks under each workload condition.\u003c/p\u003e\n\u003ch2\u003eData Acquisition\u003c/h2\u003e\n\u003cp\u003eContinuous EEG data were recorded using a 32-channel LiveAmp system (Brain Products GmbH). Electrodes were arranged according to the international 10-20 configuration using an electrode cap (actiCAP snap electrode cap). Average referencing was utilised, the actiCAP utilised a dedicated ground electrode point at the front and centre of the head, between Fp1 and Fp2. Electrode impedances were kept below 30K\u0026Omega; throughout the recording. The EEG signal was recorded at a sampling rate of 500Hz. Behavioral data, including response button presses (for RT calculation) and subsequent confidence ratings, were logged via joystick actions which generated LabStreamingLayer\u003ca href=\"https://www.zotero.org/google-docs/?m57Sgj\"\u003e\u003csup\u003e55\u003c/sup\u003e\u003c/a\u003e (LSL) markers.. Event markers corresponding to critical task events (e.g., ReticleOn, StimulusOn, ButtonPress) were generated by the Unreal Engine environment using the LSL UE5 plugin and transmitted via LSL. Both EEG and event marker streams were simultaneously recorded and synchronized using LabRecorder and OpenSignals software, resulting in .xdf files for each session.\u003c/p\u003e\n\u003ch2\u003eProcedure\u003c/h2\u003e\n\u003cp\u003eUpon arrival at the laboratory, participants were briefed on the study\u0026apos;s objectives and procedures. After providing written informed consent, they completed the BART and BIS-11 questionnaires. Participants were then fitted with the EEG cap and the VR HMD. Electrode impedances were checked and adjusted to be below.\u003c/p\u003e\n\u003cp\u003eAfter baselining, participants undertook a training task of two blocks (one for each workload condition) then the VR drone task. They received instructions on the target/non-target discrimination, the response mapping (button presses), and the confidence rating scale. The main experiment consisted of six \u0026nbsp;blocks, divided equally between the Low Workload and High Workload conditions (three blocks per condition). The entire experimental session, including setup, task execution, and debriefing, lasted approximately 150 minutes per participant. Upon completion, participants were debriefed and received their monetary compensation.\u003c/p\u003e\n\u003ch2\u003eEEG Signal Processing\u003c/h2\u003e\n\u003cp\u003eEEG data were processed offline using MNE-Python\u003csup\u003e56\u003c/sup\u003e and custom Python scripts. The processing pipeline included:\u0026nbsp;\u003c/p\u003e\n\u003col start=\"1\" type=\"1\"\u003e\n \u003cli\u003eLoading and Filtering: Data from each XDF file were loaded. A band-pass filter (0.1\u0026ndash;30 Hz, FIR) and a 50 Hz notch filter were applied to the data.\u003c/li\u003e\n \u003cli\u003eTrimming: The continuous recording was trimmed to the duration of the experimental task using the first and last LSL markers.\u003c/li\u003e\n \u003cli\u003eArtifact Rejection (ICA): Independent Component Analysis (ICA) was employed to identify and remove stereotypical artifacts such as eye blinks and lateral eye movements. To enhance ICA quality, data from all sessions for a given participant were concatenated. FastICA was run on this concatenated data. Components reflecting ocular or other non-neural artifacts were manually identified by visual inspection of their topography and time course and were flagged for removal. The corresponding ICA demixing matrix was then applied to the individual session files to project out the artifactual components.\u003c/li\u003e\n \u003cli\u003eEpoching: The cleaned continuous data were segmented into epochs relative to two primary events: (i) \u0026apos;ReticleOn\u0026apos; epochs (-200 ms to +800 ms relative to marker onset) and (ii) \u0026apos;ButtonPress\u0026apos; epochs (-500 ms to +300 ms relative to marker onset).\u003c/li\u003e\n \u003cli\u003eBaseline Correction: Epochs were baseline-corrected using the pre-event interval: -200 ms to 0 ms for ReticleOn epochs, and -500 ms to -200 ms for ButtonPress epochs.\u003c/li\u003e\n \u003cli\u003eTrial Validation: Specific trials identified as problematic during preliminary checks (e.g., Trial 18) were excluded during the creation of the final metadata associated with the epochs.\u003c/li\u003e\n\u003c/ol\u003e\n\u003ch3\u003eBCI Feature Extraction and Classification\u003c/h3\u003e\n\u003cp\u003eA participant-specific BCI was developed using an SVM classifier to predict the likelihood of decision correctness based on EEG features.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFeature Extraction: For each epoch (ReticleOn or ButtonPress), features were extracted per channel (32 channels). Time-domain features included the mean amplitude, maximum amplitude, and variance across the epoch. Frequency-domain features included the average Power Spectral Density (PSD) within the Theta (4\u0026ndash;8 Hz), Alpha (8\u0026ndash;13 Hz), and Beta (13\u0026ndash;30 Hz) bands, estimated using Welch\u0026apos;s method (1-40 Hz range).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eSVM Training: For each participant, an SVM classifier (sklearn.svm.SVC\u003csup\u003e57\u003c/sup\u003e) was trained. Feature vectors were first standardized (StandardScaler). SelectKBest with mutual_info_classif scoring identified the top 5 features. To handle potential class imbalance (more correct than incorrect trials), the feature-selected training data was balanced using RandomUnderSampler. Hyperparameters (kernel type: \u0026apos;rbf\u0026apos;/\u0026apos;linear\u0026apos;; C: 0.1-100; gamma: \u0026apos;scale\u0026apos;/\u0026apos;auto\u0026apos;) were optimized via 5-fold stratified cross-validation using GridSearchCV on this balanced training set, maximizing accuracy.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003ePrediction and Confidence Score: The optimized SVM for each participant was then applied to predict the label (1=Correct Prediction, 0=Incorrect Prediction) for all of their valid, feature-selected (but unbalanced) trials. The classifier\u0026apos;s decision function output for each trial, representing the signed distance from the separating hyperplane, was recorded as the SVM Confidence score. This score served as a BCI-derived measure of decision certainty.\u003c/p\u003e\n\u003ch2\u003eTeam Simulation Procedure and Aggregation Methods\u003c/h2\u003e\n\u003cp\u003ePerformance of simulated teams was evaluated offline. To ensure a comprehensive and robust analysis, an exhaustive combinatorial approach was employed. For each of the approximately 150 valid experimental trials within each workload condition (Low and High), team performance was simulated independently for every possible unique combination of participants drawn from the final N=17 cohort for each specified team size (m = 2, 4, 6, and 8).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThis meant that for any single trial, decisions were simulated for:\u0026nbsp;\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003eall 136 unique two-person teams (the number of distinct combinations of 2 participants from 17),\u003c/li\u003e\n \u003cli\u003eall 2,380 unique four-person teams (the number of distinct combinations of 4 participants from 17),\u003c/li\u003e\n \u003cli\u003eall 12,376 unique six-person teams (the number of distinct combinations of 6 participants from 17),\u003c/li\u003e\n \u003cli\u003eall 24,310 unique eight-person teams (the number of distinct combinations of 8 participants from 17).\u0026nbsp;\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eGiven approximately 150 trials per workload condition, and two workload conditions (Low and High, totaling ~300 trials available for analysis after individual trial validation), the total number of simulated team decisions generated was substantial:\u0026nbsp;\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003eFor two-person teams, this involved 136 unique team combinations, each assessed over approximately 300 trials, resulting in approximately 40,800 simulated decisions.\u003c/li\u003e\n \u003cli\u003eFor four-person teams, this involved 2,380 unique team combinations, each assessed over approximately 300 trials, resulting in approximately 714,000 simulated decisions.\u003c/li\u003e\n \u003cli\u003eFor six-person teams, this involved 12,376 unique team combinations, each assessed over approximately 300 trials, resulting in approximately 3,712,800 simulated decisions.\u003c/li\u003e\n \u003cli\u003eFor eight-person teams, this involved 24,310 unique team combinations, each assessed over approximately 300 trials, resulting in approximately 7,293,000 simulated decisions.\u0026nbsp;\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eIn total, this comprehensive simulation strategy yielded over 11.7 million individual team decision points for analysis across all team sizes and trials. For these per-trial, per-team-composition simulations, data were grouped by a unique trial identifier (Trial_Number) ensuring that for each simulated team decision, only data (e.g., individual response, RT, SVM confidence) from participants who had experienced that identical trial were included when constituting that specific team instance.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTeam decisions were aggregated using several methods (summarized in Table 1 and detailed below), which include various human-only, BCI-only, and mixed human-BCI approaches:\u0026nbsp;\u003c/p\u003e\n\u003col start=\"1\" type=\"1\"\u003e\n \u003cli\u003eMajority Human: A simple count of human responses (Target/Non-Target). Ties were typically resolved randomly or by a predefined rule (e.g., favouring Target classification).\u0026nbsp;\u003c/li\u003e\n\u003c/ol\u003e\n\u003col start=\"2\" type=\"1\"\u003e\n \u003cli\u003eRT Weighted Human: Human responses were weighted by their corresponding normalized response time (a 0-1 scale where higher values indicate faster RT). The team decision favoured the response type (Target or Non-Target) with the higher sum of weights. Ties were resolved by favouring Target classification.\u0026nbsp;\u003c/li\u003e\n\u003c/ol\u003e\n\u003col start=\"3\" type=\"1\"\u003e\n \u003cli\u003eSubjective Confidence Weighted Human: Human responses were weighted by their normalized subjective confidence (0-1 scale). The team decision favoured the response type (Target or Non-Target) with the higher sum of weights. Ties were resolved by favouring Target classification.\u0026nbsp;\u003c/li\u003e\n\u003c/ol\u003e\n\u003col start=\"4\" type=\"1\"\u003e\n \u003cli\u003eRT + Subjective Confidence Human: For each team member, their normalized RT and normalized subjective confidence were averaged to form a single decision weight. This weight was assigned to their actual human response. The team decision favoured the response type with the higher total summed weight across members. Ties were resolved by favouring Target classification.\u0026nbsp;\u003c/li\u003e\n\u003c/ol\u003e\n\u003col start=\"5\" type=\"1\"\u003e\n \u003cli\u003eSVM Confidence Weighted BCI: SVM-predicted labels (Target/Non-Target) from each team member were weighted by their corresponding normalized SVM confidence (a 0-1 scale derived from the absolute SVM decision function output). The team decision favoured the predicted label type with the higher sum of confidence weights. Ties were resolved by favouring Target classification.\u0026nbsp;\u003c/li\u003e\n\u003c/ol\u003e\n\u003col start=\"6\" type=\"1\"\u003e\n \u003cli\u003eRT + SVM Confidence Mixed: This method integrated human response time and BCI-SVM confidence. For each individual, the evidence supporting a Target or Non-Target decision was calculated by giving equal 0.5 weighting to their RT-weighted human response and their SVM confidence-weighted BCI prediction. For example, the evidence for a particular choice was determined as half the human-derived score for that choice plus half the BCI-derived score for that choice. The team decision favoured the response type with the higher total summed evidence across members. Ties were resolved by favouring Target classification.\u0026nbsp;\u003c/li\u003e\n\u003c/ol\u003e\n\u003col start=\"7\" type=\"1\"\u003e\n \u003cli\u003eSubjective Confidence + SVM Confidence Mixed: This method integrated human subjective confidence and BCI-SVM confidence. For each individual, the evidence supporting a Target or Non-Target decision was calculated by giving equal 0.5 weighting to their subjective confidence-weighted human response and their SVM confidence-weighted BCI prediction. For example, the evidence for a particular choice was determined as half the human-derived score for that choice plus half the BCI-derived score for that choice. The team decision favoured the response type with the higher total summed evidence across members. Ties were resolved by favouring Target classification.\u0026nbsp;\u003c/li\u003e\n\u003c/ol\u003e\n\u003col start=\"8\" type=\"1\"\u003e\n \u003cli\u003eRT + Subjective Confidence + SVM Confidence Mixed: This comprehensive method integrated human RT, human subjective confidence, and BCI-SVM confidence. For each individual, a \u0026quot;human component\u0026quot; was derived from the average of their normalized RT and normalized subjective confidence. This human component and the individual\u0026apos;s \u0026quot;BCI component\u0026quot; (normalized SVM confidence) were then each given a 0.5 weighting. These two weighted components (one reflecting the human response and one reflecting the BCI prediction) were summed to determine the individual\u0026apos;s contribution to the team\u0026apos;s evidence for a Target or Non-Target decision. The team decision favoured the response type with the higher total summed evidence. Ties were resolved by favouring Target classification.\u0026nbsp;\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eIndividual Performance Baselines: Team performance for these aggregation methods was also compared against baseline accuracies derived from the average accuracy of the best, worst, and average human performer within each specific simulated group combination on each trial.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe main outcome measure reported is the mean team accuracy, averaged across all unique team combinations and valid trials for each group size and decision method. Unless otherwise specified for a particular method (e.g., Majority Human), ties in accumulated evidence for weighted methods were resolved by favouring Target classification.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTable 1: Calculation Summary of Key Team Decision Methods\u0026nbsp;\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"755\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 196px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMethod Label in Plot\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 286px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCore Logic Description\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 273px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eKey Information Used per Team Member (Trial-Level)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 196px;\"\u003e\n \u003cp\u003eWorst Individual Avg\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 286px;\"\u003e\n \u003cp\u003eAverage of the lowest individual accuracy observed within each unique simulated team composition.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 273px;\"\u003e\n \u003cp\u003eOverall Human Accuracy (of the worst performer in each group iteration)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 196px;\"\u003e\n \u003cp\u003eAverage Individual Avg\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 286px;\"\u003e\n \u003cp\u003eAverage of the mean individual accuracy observed within each unique simulated team composition. \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 273px;\"\u003e\n \u003cp\u003eOverall Human Accuracy (mean of individuals in each group iteration)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 196px;\"\u003e\n \u003cp\u003eBest Individual Avg\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 286px;\"\u003e\n \u003cp\u003eAverage of the highest individual accuracy observed within each unique simulated team composition. \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 273px;\"\u003e\n \u003cp\u003eOverall Human Accuracy (of the best performer in each group iteration)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 196px;\"\u003e\n \u003cp\u003eMajority Human \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 286px;\"\u003e\n \u003cp\u003eEach member\u0026apos;s binary human response (Target/Non-Target) contributes one vote. Team decision is the most frequent response. Ties typically resolved randomly or by pre-defined rule. \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 273px;\"\u003e\n \u003cp\u003eHuman Response (Target/Non-Target)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 196px;\"\u003e\n \u003cp\u003eRT Weighted Human \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 286px;\"\u003e\n \u003cp\u003eHuman responses are weighted by normalized RT (faster RT = higher weight). Team decision by sum of weights. Ties favour Target. \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 273px;\"\u003e\n \u003cp\u003eHuman Response, Normalized RT\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 196px;\"\u003e\n \u003cp\u003eSubjective Confidence Weighted Human \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 286px;\"\u003e\n \u003cp\u003eHuman responses are weighted by normalized subjective confidence. Team decision by sum of weights. Ties favour Target. \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 273px;\"\u003e\n \u003cp\u003eHuman Response, Normalized Subjective Confidence\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 196px;\"\u003e\n \u003cp\u003eRT + Subjective Confidence Human \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 286px;\"\u003e\n \u003cp\u003eHuman responses weighted by the average of normalized RT and normalized subjective confidence. Team decision by sum of weights. Ties favour Target. \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 273px;\"\u003e\n \u003cp\u003eHuman Response, Normalized RT, Normalized Subjective Confidence\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 196px;\"\u003e\n \u003cp\u003eSVM Confidence Weighted BCI \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 286px;\"\u003e\n \u003cp\u003eSVM-predicted labels are weighted by normalized SVM confidence. Team decision by sum of weights. Ties favour Target. \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 273px;\"\u003e\n \u003cp\u003eSVM Predicted Label, Normalized SVM Confidence\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 196px;\"\u003e\n \u003cp\u003eRT + SVM Confidence Mixed \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 286px;\"\u003e\n \u003cp\u003eFor each member, evidence for a decision (Target/Non-Target) is 0.5 * (Human Score from RT) + 0.5 * (BCI Score from SVM Confidence). Team decision by sum of total evidence. Ties favour Target. \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 273px;\"\u003e\n \u003cp\u003eHuman Response, Normalized RT, SVM Predicted Label, Normalized SVM Confidence\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 196px;\"\u003e\n \u003cp\u003eSubjective Confidence + SVM Confidence Mixed \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 286px;\"\u003e\n \u003cp\u003eFor each member, evidence for a decision (Target/Non-Target) is 0.5 * (Human Score from Subjective Confidence) + 0.5 * (BCI Score from SVM Confidence). Team decision by sum of total evidence. Ties favour Target. \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 273px;\"\u003e\n \u003cp\u003eHuman Response, Normalized Subjective Confidence, SVM Predicted Label, Normalized SVM Confidence\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 196px;\"\u003e\n \u003cp\u003eRT + Subjective Confidence + SVM Confidence Mixed \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 286px;\"\u003e\n \u003cp\u003eFor each member, a \u0026quot;human component\u0026quot; score is the average of Normalized RT and Normalized Subjective Confidence. Evidence for a decision (Target/Non-Target) is 0.5 * (Human Component Score) + 0.5 * (BCI Score from SVM Confidence). Team decision by sum of total evidence. Ties favour Target. \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 273px;\"\u003e\n \u003cp\u003eHuman Response, Normalized RT, Normalized Subjective Confidence, SVM Predicted Label, Normalized SVM Confidence\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003ch2\u003eStatistical Analysis\u003c/h2\u003e\n\u003cp\u003eIndividual behavioral data (accuracy, RT, confidence) were analyzed to assess the impact of the Workload manipulation (Low vs. High). Depending on data distributions, paired t-tests or Wilcoxon signed-rank tests were used for continuous variables (RT, confidence), while Chi-square tests were used for accuracy (comparing counts of correct/incorrect decisions). For simulated team performance, mean accuracies for the proposed cBCI weighting method(s) were compared against baseline methods (Majority, Best/Average Individual) for each group size using paired t-tests or Wilcoxon tests. Corrections for multiple comparisons (e.g., Bonferroni) were applied where appropriate. Statistical significance was defined at an alpha level of p \u0026lt; 0.05. All statistical analyses were performed using Python\u003csup\u003e58\u003c/sup\u003e and its scientific computing libraries, primarily SciPy\u003csup\u003e59\u003c/sup\u003e, for significance testing. Data processing and manipulation were conducted using Pandas\u003csup\u003e60\u003c/sup\u003e and NumPy\u003csup\u003e61\u003c/sup\u003e. Visualizations were generated with Matplotlib\u003csup\u003e62\u003c/sup\u003e and Seaborn\u003csup\u003e63\u003c/sup\u003e.\u003c/p\u003e\n\u003ch2\u003eHypotheses\u003c/h2\u003e\n\u003col start=\"1\" type=\"1\"\u003e\n \u003cli\u003eIndividual Performance (Manipulation Check): We hypothesised that the High Workload condition would significantly impair individual performance compared to the Low Workload condition, manifesting as:\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003ea. Lower decision accuracy.\u003c/p\u003e\n\u003cp\u003eb. Slower response times (RT).\u003c/p\u003e\n\u003cp\u003ec. Lower subjective confidence ratings.\u0026nbsp;\u003c/p\u003e\n\u003col start=\"2\" type=\"1\"\u003e\n \u003cli\u003eTeam accuracy using the cBCI method would be significantly higher than the average accuracy of the individual members comprising the team.\u0026nbsp;\u003c/li\u003e\n\u003c/ol\u003e\n\u003col start=\"3\" type=\"1\"\u003e\n \u003cli\u003eWorkload Interaction: We hypothesised that the performance benefit conferred by the cBCI method would be dependent on cognitive workload. Specifically:\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003ea. The accuracy gain provided by the cBCI method (relative to Majority vote or Average Individual performance) would be significantly greater under the High Workload condition compared to the Low Workload condition.\u003c/p\u003e\n\u003cp\u003eb. (Stronger/More Specific version of 2b related to synergy): Under High Workload conditions, team accuracy using the cBCI method might surpass the average accuracy of the best-performing individual member within the team, particularly for larger group sizes.\u0026nbsp;\u003c/p\u003e\n\u003col start=\"4\" type=\"1\"\u003e\n \u003cli\u003eEpoch Timing: We hypothesised that while the underlying neural signals and individual SVM performance might differ between the anticipatory (\u0026apos;ReticleOn\u0026apos;) and peri-/post-response (\u0026apos;ButtonPress\u0026apos;) epochs, the overall pattern of cBCI-driven team performance enhancement (relative to baselines and across workload conditions) would be comparable between the two epoch types.\u0026nbsp;\u003c/li\u003e\n\u003c/ol\u003e\n\u003col start=\"5\" type=\"1\"\u003e\n \u003cli\u003eExploratory - Individual Differences: We exploratorily hypothesised that individual differences in trait impulsivity (measured by BIS-11) and risk-taking propensity (measured by BART) might be associated with individual performance metrics (accuracy, RT, confidence) within the VR drone task.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Declarations","content":"\u003ch2\u003eEthics approval and consent to participate\u003c/h2\u003e\n\u003cp\u003eThe experimental protocol was approved by the UK Ministry of Defence Research Ethics Committee (MoDREC), Application Number: 2309/MODREC/24, Reference: RQ0000037929. All procedures were conducted in accordance with the ethical standards outlined in the Declaration of Helsinki. Written informed consent was obtained from all individual participants included in the study.\u003c/p\u003e\n\u003ch2\u003eData Availability Statement\u003c/h2\u003e\n\u003cp\u003eThe datasets generated and/or analysed during the current study are not publicly available due to restrictions imposed by the funding body (Defence Science and Technology Laboratory - Dstl). However, data are available from the corresponding author (CB) on reasonable request and subject to a data sharing agreement, if appropriate and in accordance with Dstl policy.\u003c/p\u003e\n\u003ch2\u003eCompeting interests\u003c/h2\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003ch2\u003eFunding\u003c/h2\u003e\n\u003cp\u003eThis research was funded by the Defence Science and Technology Laboratory (DSTL) via RQ0000037929.The funders contributed to the conceptualisation of the broader project aims. The funders did not have a direct role in the specific design of this study, data collection, detailed analysis, interpretation of data from this specific study, or in the writing of this manuscript beyond the contributions of the DSTL-affiliated co-author (T.R.) as described in the Author Contributions section.\u003c/p\u003e\n\u003ch2\u003eAuthors\u0026apos; contributions\u003c/h2\u003e\n\u003cp\u003eC.B.: Conceptualisation, Methodology, Software, Validation, Formal Analysis, Investigation, Data Curation, Writing \u0026ndash; Original Draft, Writing \u0026ndash; Review \u0026amp; Editing.\u003c/p\u003e\n\u003cp\u003eS.H.: Conceptualisation, Methodology, Formal Analysis, Investigation, Writing \u0026ndash; Review \u0026amp; Editing.\u003c/p\u003e\n\u003cp\u003eS.F.: Conceptualisation, Methodology, Software, Formal Analysis, Investigation, Writing \u0026ndash; Review \u0026amp; Editing.\u003c/p\u003e\n\u003cp\u003eA.N.: \u0026nbsp; Conceptualisation, Methodology, Writing \u0026ndash; Review \u0026amp; Editing.\u003c/p\u003e\n\u003cp\u003eR.P.: \u0026nbsp; Conceptualisation, Methodology, Writing \u0026ndash; Review \u0026amp; Editing.\u003c/p\u003e\n\u003cp\u003eC.C.: \u0026nbsp; Conceptualisation, Methodology, Writing \u0026ndash; Review \u0026amp; Editing.\u003c/p\u003e\n\u003cp\u003eT.R.: Writing \u0026ndash; Review \u0026amp; Editing.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eSalas, E., Reyes, D. L. \u0026amp; McDaniel, S. H. The science of teamwork: Progress, reflections, and the road ahead. \u003cem\u003eAm. Psychol. \u003c/em\u003e\u003cstrong\u003e73\u003c/strong\u003e, 593\u0026ndash;600 (2018).\u003c/li\u003e\n\u003cli\u003eHeinrich, M. K., Wahby, M., Dorigo, M. \u0026amp; Hamann, H. Swarm Robotics. (2022) doi:10.7551/mitpress/13780.003.0009.\u003c/li\u003e\n\u003cli\u003eSiean, A.-I., Gradinaru, B.-C., Gherman, O.-I., Danubianu, M. \u0026amp; Milici, L.-D. Opportunities and Challenges in Human-Swarm Interaction: Systematic Review and Research Implications. \u003cem\u003eInt. J. Adv. Comput. Sci. Appl. \u003c/em\u003e\u003cstrong\u003e14\u003c/strong\u003e, (2023).\u003c/li\u003e\n\u003cli\u003eHung, W. Team-based complex problem solving: a collective cognition perspective. \u003cem\u003eEduc. Technol. Res. Dev. \u003c/em\u003e\u003cstrong\u003e61\u003c/strong\u003e, 365\u0026ndash;384 (2013).\u003c/li\u003e\n\u003cli\u003eCanonico, L. B., Flathmann, C. \u0026amp; McNeese, N. Collectively Intelligent Teams: Integrating Team Cognition, Collective Intelligence, and AI for Future Teaming. \u003cem\u003eProc. Hum. Factors Ergon. Soc. Annu. Meet. \u003c/em\u003e\u003cstrong\u003e63\u003c/strong\u003e, 1466\u0026ndash;1470 (2019).\u003c/li\u003e\n\u003cli\u003ePoli, R. Super-Human and Super-AI Cognitive Augmentation of Human and Human-AI Teams Assisted by Brain Computer Interfaces. in \u003cem\u003eProceedings of the Genetic and Evolutionary Computation Conference\u003c/em\u003e 3 (Association for Computing Machinery, New York, NY, USA, 2023). doi:10.1145/3583131.3603554.\u003c/li\u003e\n\u003cli\u003evan Ments, L. \u0026amp; Treur, J. Reflections on dynamics, adaptation and control: A cognitive architecture for mental models. \u003cem\u003eCogn. Syst. Res. \u003c/em\u003e\u003cstrong\u003e70\u003c/strong\u003e, 1\u0026ndash;9 (2021).\u003c/li\u003e\n\u003cli\u003eXu, W., Chen, X., Dong, Y. \u0026amp; Chiclana, F. Impact of Decision Rules and Non-cooperative Behaviors on Minimum Consensus Cost in Group Decision Making. \u003cem\u003eGroup Decis. Negot. \u003c/em\u003e\u003cstrong\u003e30\u003c/strong\u003e, 1239\u0026ndash;1260 (2021).\u003c/li\u003e\n\u003cli\u003eBaltes, B. B., Dickson, M. W., Sherman, M. P., Bauer, C. C. \u0026amp; LaGanke, J. S. Computer-Mediated Communication and Group Decision Making: A Meta-Analysis. \u003cem\u003eOrgan. Behav. Hum. Decis. Process. \u003c/em\u003e\u003cstrong\u003e87\u003c/strong\u003e, 156\u0026ndash;179 (2002).\u003c/li\u003e\n\u003cli\u003eSheridan, T. B. HCI in supervisory control: Twelve dilemmas. in \u003cem\u003eHuman error and system design and management\u003c/em\u003e (eds. Elzer, P. F., Kluwe, R. H. \u0026amp; Boussoffara, B.) 1\u0026ndash;12 (Springer, London, 2000). doi:10.1007/BFb0110450.\u003c/li\u003e\n\u003cli\u003eParasuraman, R., Sheridan, T. B. \u0026amp; Wickens, C. D. A model for types and levels of human interaction with automation. \u003cem\u003eIEEE Trans. Syst. Man Cybern. - Part Syst. Hum. \u003c/em\u003e\u003cstrong\u003e30\u003c/strong\u003e, 286\u0026ndash;297 (2000).\u003c/li\u003e\n\u003cli\u003eWang, D. \u003cem\u003eet al.\u003c/em\u003e From Human-Human Collaboration to Human-AI Collaboration: Designing AI Systems That Can Work Together with People. in \u003cem\u003eExtended Abstracts of the 2020 CHI Conference on Human Factors in Computing Systems\u003c/em\u003e 1\u0026ndash;6 (Association for Computing Machinery, New York, NY, USA, 2020). doi:10.1145/3334480.3381069.\u003c/li\u003e\n\u003cli\u003eKaklauskas, A. Intelligent Decision Support Systems. in \u003cem\u003eBiometric and Intelligent Decision Making Support\u003c/em\u003e (ed. Kaklauskas, A.) 31\u0026ndash;85 (Springer International Publishing, Cham, 2015). doi:10.1007/978-3-319-13659-2_2.\u003c/li\u003e\n\u003cli\u003ePhillips-Wren, G. Intelligent Decision Support Systems. in \u003cem\u003eMulticriteria Decision Aid and Artificial Intelligence\u003c/em\u003e 25\u0026ndash;44 (John Wiley \u0026amp; Sons, Ltd, 2013). doi:10.1002/9781118522516.ch2.\u003c/li\u003e\n\u003cli\u003eSmith, P. J., Geddes, N. D. \u0026amp; Beatty, R. HUMAN-CENTERED DESIGN OF DECISION-SUPPORT SYSTEMS. in \u003cem\u003eHuman-Computer Interaction\u003c/em\u003e (CRC Press, 2009).\u003c/li\u003e\n\u003cli\u003eHerse, S., Vitale ,Jonathan \u0026amp; and Williams, M.-A. Using Agent Features to Influence User Trust, Decision Making and Task Outcome during Human-Agent Collaboration. \u003cem\u003eInt. J. Human\u0026ndash;Computer Interact. \u003c/em\u003e\u003cstrong\u003e39\u003c/strong\u003e, 1740\u0026ndash;1761 (2023).\u003c/li\u003e\n\u003cli\u003eMa, S. \u003cem\u003eet al.\u003c/em\u003e \u0026ldquo;Are You Really Sure?\u0026rdquo; Understanding the Effects of Human Self-Confidence Calibration in AI-Assisted Decision Making. in \u003cem\u003eProceedings of the 2024 CHI Conference on Human Factors in Computing Systems\u003c/em\u003e 1\u0026ndash;20 (Association for Computing Machinery, New York, NY, USA, 2024). doi:10.1145/3613904.3642671.\u003c/li\u003e\n\u003cli\u003eArshad, S. Z., Zhou, J., Bridon, C., Chen, F. \u0026amp; Wang, Y. Investigating User Confidence for Uncertainty Presentation in Predictive Decision Making. in \u003cem\u003eProceedings of the Annual Meeting of the Australian Special Interest Group for Computer Human Interaction\u003c/em\u003e 352\u0026ndash;360 (Association for Computing Machinery, New York, NY, USA, 2015). doi:10.1145/2838739.2838753.\u003c/li\u003e\n\u003cli\u003eProaps, A. \u0026amp; Bliss, J. Team Performance as a Function of Task Difficulty in a Computer Game. \u003cem\u003eProc. Hum. Factors Ergon. Soc. Annu. Meet. \u003c/em\u003e\u003cstrong\u003e54\u003c/strong\u003e, 2413\u0026ndash;2416 (2010).\u003c/li\u003e\n\u003cli\u003eCaldwell, B. S. \u0026amp; Wang, E. Delays and User Performance in Human-Computer-Network Interaction Tasks. \u003cem\u003eHum. Factors \u003c/em\u003e\u003cstrong\u003e51\u003c/strong\u003e, 813\u0026ndash;830 (2009).\u003c/li\u003e\n\u003cli\u003eLongo, L. Experienced mental workload, perception of usability, their interaction and impact on task performance. \u003cem\u003ePLOS ONE \u003c/em\u003e\u003cstrong\u003e13\u003c/strong\u003e, e0199661 (2018).\u003c/li\u003e\n\u003cli\u003eMITTA, D. A. \u0026amp; and PACKEBUSH, S. J. Improving interface quality: an investigation of human-computer interaction task learning. \u003cem\u003eErgonomics \u003c/em\u003e\u003cstrong\u003e38\u003c/strong\u003e, 1307\u0026ndash;1325 (1995).\u003c/li\u003e\n\u003cli\u003eMcFarlane, D. C. Coordinating the Interruption of People in Human- Computer Interaction.\u003c/li\u003e\n\u003cli\u003eFritz, T., Begel, A., M\u0026uuml;ller, S. C., Yigit-Elliott, S. \u0026amp; Z\u0026uuml;ger, M. Using psycho-physiological measures to assess task difficulty in software development. in \u003cem\u003eProceedings of the 36th International Conference on Software Engineering\u003c/em\u003e 402\u0026ndash;413 (Association for Computing Machinery, New York, NY, USA, 2014). doi:10.1145/2568225.2568266.\u003c/li\u003e\n\u003cli\u003eNeerincx, M. A., van der Waa, J., Kaptein, F. \u0026amp; van Diggelen, J. Using Perceptual and Cognitive Explanations for Enhanced Human-Agent Team Performance. in \u003cem\u003eEngineering Psychology and Cognitive Ergonomics\u003c/em\u003e (ed. Harris, D.) 204\u0026ndash;214 (Springer International Publishing, Cham, 2018). doi:10.1007/978-3-319-91122-9_18.\u003c/li\u003e\n\u003cli\u003eDirican, A. C. \u0026amp; G\u0026ouml;kt\u0026uuml;rk, M. Psychophysiological measures of human cognitive states applied in human computer interaction. \u003cem\u003eProcedia Comput. Sci. \u003c/em\u003e\u003cstrong\u003e3\u003c/strong\u003e, 1361\u0026ndash;1367 (2011).\u003c/li\u003e\n\u003cli\u003eZhou, Y. \u003cem\u003eet al.\u003c/em\u003e Cognitive Workload Recognition Using EEG Signals and Machine Learning: A Review. \u003cem\u003eIEEE Trans. Cogn. Dev. Syst. \u003c/em\u003e\u003cstrong\u003e14\u003c/strong\u003e, 799\u0026ndash;818 (2022).\u003c/li\u003e\n\u003cli\u003eHassan, J., Reza, Md. S., Ahmed, S. U., Anik, N. H. \u0026amp; Khan, Md. O. EEG workload estimation and classification: a systematic review. \u003cem\u003eJ. Neural Eng.\u003c/em\u003e (2024) doi:10.1088/1741-2552/ad705e.\u003c/li\u003e\n\u003cli\u003ePoli, R., Valeriani, D. \u0026amp; Cinel, C. Collaborative Brain-Computer Interface for Aiding Decision-Making. \u003cem\u003ePLoS ONE \u003c/em\u003e\u003cstrong\u003e9\u003c/strong\u003e, e102693 (2014).\u003c/li\u003e\n\u003cli\u003eValeriani, D., Poli, R. \u0026amp; Cinel, C. A collaborative Brain-Computer Interface for improving group detection of visual targets in complex natural environments. in \u003cem\u003e2015 7th International IEEE/EMBS Conference on Neural Engineering (NER)\u003c/em\u003e vols 2015-J 25\u0026ndash;28 (IEEE, 2015).\u003c/li\u003e\n\u003cli\u003eValeriani, D., Poli, R. \u0026amp; Cinel, C. Enhancement of Group Perception via a Collaborative Brain\u0026ndash;Computer Interface. \u003cem\u003eIEEE Trans. Biomed. Eng. \u003c/em\u003e\u003cstrong\u003e64\u003c/strong\u003e, 1238\u0026ndash;1248 (2017).\u003c/li\u003e\n\u003cli\u003eBhattacharyya, S., Valeriani, D., Cinel, C., Citi, L. \u0026amp; Poli, R. Anytime collaborative brain\u0026ndash;computer interfaces for enhancing perceptual group decision-making. \u003cem\u003eSci. Rep. \u003c/em\u003e\u003cstrong\u003e11\u003c/strong\u003e, 17008 (2021).\u003c/li\u003e\n\u003cli\u003eValeriani, D., O\u0026rsquo;Flynn, L. C., Worthley, A., Sichani, A. H. \u0026amp; Simonyan, K. Multimodal collaborative brain-computer interfaces aid human-machine team decision-making in a pandemic scenario. \u003cem\u003eJ. Neural Eng. \u003c/em\u003e\u003cstrong\u003e19\u003c/strong\u003e, 056036 (2022).\u003c/li\u003e\n\u003cli\u003eSalvatore, C., Valeriani, D., Piccialli, V. \u0026amp; Bianchi, L. Optimized Collaborative Brain-Computer Interfaces for Enhancing Face Recognition. \u003cem\u003eIEEE Trans. Neural Syst. Rehabil. Eng. \u003c/em\u003e\u003cstrong\u003e30\u003c/strong\u003e, 1223\u0026ndash;1232 (2022).\u003c/li\u003e\n\u003cli\u003eZuo, Z., Liu, C., Han, Q.-L. \u0026amp; Song, J. Unmanned Aerial Vehicles: Control Methods and Future Challenges. \u003cem\u003eIEEECAA J. Autom. Sin. \u003c/em\u003e\u003cstrong\u003e9\u003c/strong\u003e, 601\u0026ndash;614 (2022).\u003c/li\u003e\n\u003cli\u003eHamada, D., Nakayama, M. \u0026amp; Saiki, J. Wisdom of crowds and collective decision-making in a survival situation with complex information integration. \u003cem\u003eCogn. Res. Princ. Implic. \u003c/em\u003e\u003cstrong\u003e5\u003c/strong\u003e, 48 (2020).\u003c/li\u003e\n\u003cli\u003eThomas, E., Dyson, M. \u0026amp; Clerc, M. An analysis of performance evaluation for motor-imagery based BCI. \u003cem\u003eJ. Neural Eng. \u003c/em\u003e\u003cstrong\u003e10\u003c/strong\u003e, 031001 (2013).\u003c/li\u003e\n\u003cli\u003eDorneich, M. C., Mathan, S., Ververs, P. M. \u0026amp; Whitlow, S. D. An Evaluation of Real-Time Cognitive State Classification in a Harsh Operational Environment. \u003cem\u003eProc. Hum. Factors Ergon. Soc. Annu. Meet. \u003c/em\u003e\u003cstrong\u003e51\u003c/strong\u003e, 146\u0026ndash;150 (2007).\u003c/li\u003e\n\u003cli\u003eBagheri, M. \u0026amp; Power, S. D. Simultaneous Classification of Both Mental Workload and Stress Level Suitable for an Online Passive Brain\u0026ndash;Computer Interface. \u003cem\u003eSensors \u003c/em\u003e\u003cstrong\u003e22\u003c/strong\u003e, 535 (2022).\u003c/li\u003e\n\u003cli\u003eCinel, C., Fernandez-Vargas, J., Tremmel, C., Citi, L. \u0026amp; Poli, R. Enhancing performance with multisensory cues in a realistic target discrimination task. \u003cem\u003ePLOS ONE \u003c/em\u003e\u003cstrong\u003e17\u003c/strong\u003e, e0272320 (2022).\u003c/li\u003e\n\u003cli\u003eKelley, B. Do You See It? Aspects of Effective Visual Cueing. in \u003cem\u003e2024 IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct)\u003c/em\u003e 658\u0026ndash;659 (2024). doi:10.1109/ISMAR-Adjunct64951.2024.00198.\u003c/li\u003e\n\u003cli\u003eKelley, B. \u003cem\u003eet al.\u003c/em\u003e Addressing Human Factors Related to Artificial Intelligence Integrated Visual Cueing. in \u003cem\u003e2024 IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct)\u003c/em\u003e 209\u0026ndash;213 (2024). doi:10.1109/ISMAR-Adjunct64951.2024.00051.\u003c/li\u003e\n\u003cli\u003eKosch, T. \u003cem\u003eet al.\u003c/em\u003e A Survey on Measuring Cognitive Workload in Human-Computer Interaction. \u003cem\u003eACM Comput Surv \u003c/em\u003e\u003cstrong\u003e55\u003c/strong\u003e, 283:1-283:39 (2023).\u003c/li\u003e\n\u003cli\u003eWoolley, A. W., Chabris, C. F., Pentland, A., Hashmi, N. \u0026amp; Malone, T. W. Evidence for a Collective Intelligence Factor in the Performance of Human Groups. \u003cem\u003eScience \u003c/em\u003e\u003cstrong\u003e330\u003c/strong\u003e, 686\u0026ndash;688 (2010).\u003c/li\u003e\n\u003cli\u003eTrumbo, M. \u003cem\u003eet al.\u003c/em\u003e Individual Differences and the Science of Human Performance. in \u003cem\u003eFoundations of Augmented Cognition. Directing the Future of Adaptive Systems\u003c/em\u003e (eds. Schmorrow, D. D. \u0026amp; Fidopiastis, C. M.) 46\u0026ndash;54 (Springer, Berlin, Heidelberg, 2011). doi:10.1007/978-3-642-21852-1_6.\u003c/li\u003e\n\u003cli\u003eAgrigoroaie, R. \u0026amp; Tapus, A. Cognitive Performance and Physiological Response Analysis. \u003cem\u003eInt. J. Soc. Robot. \u003c/em\u003e\u003cstrong\u003e12\u003c/strong\u003e, 47\u0026ndash;64 (2020).\u003c/li\u003e\n\u003cli\u003eChowdhury, M. R. Development of Multiple Class Human-Computer Interaction System using Machine Learning Algorithm for Eyeball Movement. (Khulna University of Engineering \u0026amp; Technology (KUET), Khulna, Bangladesht, 2019).\u003c/li\u003e\n\u003cli\u003eQuitadamo, L. R. \u003cem\u003eet al.\u003c/em\u003e Support vector machines to detect physiological patterns for EEG and EMG-based human\u0026ndash;computer interaction: a review. \u003cem\u003eJ. Neural Eng. \u003c/em\u003e\u003cstrong\u003e14\u003c/strong\u003e, 011001 (2017).\u003c/li\u003e\n\u003cli\u003eVidaurre, C., Sannelli, C., M\u0026uuml;ller, K.-R. \u0026amp; Blankertz, B. Machine-Learning-Based Coadaptive Calibration for Brain-Computer Interfaces. \u003cem\u003eNeural Comput. \u003c/em\u003e\u003cstrong\u003e23\u003c/strong\u003e, 791\u0026ndash;816 (2011).\u003c/li\u003e\n\u003cli\u003eHauptman, A. I., Schelble, B. G., McNeese, N. J. \u0026amp; Madathil, K. C. Adapt and overcome: Perceptions of adaptive autonomous agents for human-AI teaming. \u003cem\u003eComput. Hum. Behav. \u003c/em\u003e\u003cstrong\u003e138\u003c/strong\u003e, 107451 (2023).\u003c/li\u003e\n\u003cli\u003eZhao, M., Simmons, R. \u0026amp; Admoni, H. The Role of Adaptation in Collective Human\u0026ndash;AI Teaming. \u003cem\u003eTop. Cogn. Sci. \u003c/em\u003e\u003cstrong\u003e17\u003c/strong\u003e, 291\u0026ndash;323 (2025).\u003c/li\u003e\n\u003cli\u003eZheng, N. \u003cem\u003eet al.\u003c/em\u003e Hybrid-augmented intelligence: collaboration and cognition. \u003cem\u003eFront. Inf. Technol. Electron. Eng. \u003c/em\u003e\u003cstrong\u003e18\u003c/strong\u003e, 153\u0026ndash;179 (2017).\u003c/li\u003e\n\u003cli\u003eCui, H. \u0026amp; Yasseri, T. AI-enhanced collective intelligence. \u003cem\u003ePatterns \u003c/em\u003e\u003cstrong\u003e5\u003c/strong\u003e, (2024).\u003c/li\u003e\n\u003cli\u003eBarratt, E. S. Impulsiveness Defined Within a Systems Model of Personality. in \u003cem\u003eAdvances in Personality Assessment\u003c/em\u003e (Routledge, 1985).\u003c/li\u003e\n\u003cli\u003eKothe, C. \u003cem\u003eet al.\u003c/em\u003e The Lab Streaming Layer for Synchronized Multimodal Recording. \u003cem\u003ebioRxiv\u003c/em\u003e 2024.02.13.580071 (2024) doi:10.1101/2024.02.13.580071.\u003c/li\u003e\n\u003cli\u003eGramfort, A. \u003cem\u003eet al.\u003c/em\u003e MEG and EEG data analysis with MNE-python. \u003cem\u003eFront. Neurosci. \u003c/em\u003e\u003cstrong\u003e7\u003c/strong\u003e, 267 (2013).\u003c/li\u003e\n\u003cli\u003ePedregosa, F. \u003cem\u003eet al.\u003c/em\u003e Scikit-learn: Machine Learning in Python. \u003cem\u003eJ. Mach. Learn. Res. \u003c/em\u003e\u003cstrong\u003e12\u003c/strong\u003e, 2825\u0026ndash;2830 (2011).\u003c/li\u003e\n\u003cli\u003evan Rossum, G. \u0026amp; de Boer, J. Interactively testing remote servers using the Python programming language. \u003cem\u003eCWI Q. \u003c/em\u003e\u003cstrong\u003e4\u003c/strong\u003e, 283\u0026ndash;304 (1991).\u003c/li\u003e\n\u003cli\u003eVirtanen, P. \u003cem\u003eet al.\u003c/em\u003e SciPy 1.0: fundamental algorithms for scientific computing in Python. \u003cem\u003eNat. Methods \u003c/em\u003e\u003cstrong\u003e17\u003c/strong\u003e, 261\u0026ndash;272 (2020).\u003c/li\u003e\n\u003cli\u003eMcKinney, W. Data Structures for Statistical Computing in Python. in 56\u0026ndash;61 (Austin, Texas, 2010). doi:10.25080/Majora-92bf1922-00a.\u003c/li\u003e\n\u003cli\u003eHarris, C. R. \u003cem\u003eet al.\u003c/em\u003e Array programming with NumPy. \u003cem\u003eNature \u003c/em\u003e\u003cstrong\u003e585\u003c/strong\u003e, 357\u0026ndash;362 (2020).\u003c/li\u003e\n\u003cli\u003eHunter, J. D. Matplotlib: A 2D Graphics Environment. \u003cem\u003eComput. Sci. Eng. \u003c/em\u003e\u003cstrong\u003e9\u003c/strong\u003e, 90\u0026ndash;95 (2007).\u003c/li\u003e\n\u003cli\u003eWaskom, M. L. seaborn: statistical data visualization. \u003cem\u003eJ. Open Source Softw. \u003c/em\u003e\u003cstrong\u003e6\u003c/strong\u003e, 3021 (2021).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":true,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-6985673/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6985673/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eOptimizing team decision-making by appropriately weighting individual contributions is a significant challenge. Collaborative Brain-Computer Interfaces (cBCIs) offer a novel approach by integrating neurophysiological and behavioral data. This study evaluated a cBCI system incorporating EEG-derived SVM decision confidence, response times (RT), and subjective confidence ratings to enhance team accuracy in a VR drone target-detection task, particularly under varying cognitive workload (Low vs. High). Seventeen participants performed the task; individual SVMs were trained, and team performance (N\u0026thinsp;=\u0026thinsp;2\u0026ndash;8 members) was simulated using diverse aggregation methods. Under High Workload, mixed cBCI methods (e.g., combining subjective and SVM confidence) significantly improved team accuracy, surpassing even the best individual's average performance (e.g., N\u0026thinsp;=\u0026thinsp;8: 98.8% vs. 94.2%). This synergistic benefit was minimal under Low Workload due to ceiling effects in individual performance. These cBCI enhancements were evident for EEG data from both pre- and post-decision epochs. The findings demonstrate that cBCIs can markedly improve team decision-making in demanding contexts, facilitating a \"superorganism\" effect where team capabilities exceed those of the best individual.\u003c/p\u003e","manuscriptTitle":"Integrating EEG-SVM Confidence and RT via cBCI Enhances Team Decisions in a VR Drone Task","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-23 08:38:30","doi":"10.21203/rs.3.rs-6985673/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"675ddb92-49f2-4335-b103-e9a985ab701d","owner":[],"postedDate":"July 23rd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":51804545,"name":"Physical sciences/Engineering"},{"id":51804546,"name":"Physical sciences/Mathematics and computing"},{"id":51804547,"name":"Biological sciences/Neuroscience"}],"tags":[],"updatedAt":"2025-12-15T14:24:39+00:00","versionOfRecord":[],"versionCreatedAt":"2025-07-23 08:38:30","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6985673","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6985673","identity":"rs-6985673","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}